vm_page.c revision 93818
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 93818 2002-04-04 21:03:38Z jhb $ 38 */ 39 40/* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 67/* 68 * GENERAL RULES ON VM_PAGE MANIPULATION 69 * 70 * - a pageq mutex is required when adding or removing a page from a 71 * page queue (vm_page_queue[]), regardless of other mutexes or the 72 * busy state of a page. 73 * 74 * - a hash chain mutex is required when associating or disassociating 75 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 76 * regardless of other mutexes or the busy state of a page. 77 * 78 * - either a hash chain mutex OR a busied page is required in order 79 * to modify the page flags. A hash chain mutex must be obtained in 80 * order to busy a page. A page's flags cannot be modified by a 81 * hash chain mutex if the page is marked busy. 82 * 83 * - The object memq mutex is held when inserting or removing 84 * pages from an object (vm_page_insert() or vm_page_remove()). This 85 * is different from the object's main mutex. 86 * 87 * Generally speaking, you have to be aware of side effects when running 88 * vm_page ops. A vm_page_lookup() will return with the hash chain 89 * locked, whether it was able to lookup the page or not. vm_page_free(), 90 * vm_page_cache(), vm_page_activate(), and a number of other routines 91 * will release the hash chain mutex for you. Intermediate manipulation 92 * routines such as vm_page_flag_set() expect the hash chain to be held 93 * on entry and the hash chain will remain held on return. 94 * 95 * pageq scanning can only occur with the pageq in question locked. 96 * We have a known bottleneck with the active queue, but the cache 97 * and free queues are actually arrays already. 98 */ 99 100/* 101 * Resident memory management module. 102 */ 103 104#include <sys/param.h> 105#include <sys/systm.h> 106#include <sys/lock.h> 107#include <sys/malloc.h> 108#include <sys/mutex.h> 109#include <sys/proc.h> 110#include <sys/vmmeter.h> 111#include <sys/vnode.h> 112 113#include <vm/vm.h> 114#include <vm/vm_param.h> 115#include <vm/vm_kern.h> 116#include <vm/vm_object.h> 117#include <vm/vm_page.h> 118#include <vm/vm_pageout.h> 119#include <vm/vm_pager.h> 120#include <vm/vm_extern.h> 121#include <vm/uma.h> 122#include <vm/uma_int.h> 123 124/* 125 * Associated with page of user-allocatable memory is a 126 * page structure. 127 */ 128static struct vm_page **vm_page_buckets; /* Array of buckets */ 129static int vm_page_bucket_count; /* How big is array? */ 130static int vm_page_hash_mask; /* Mask for hash function */ 131static volatile int vm_page_bucket_generation; 132static struct mtx vm_buckets_mtx[BUCKET_HASH_SIZE]; 133 134vm_page_t vm_page_array = 0; 135int vm_page_array_size = 0; 136long first_page = 0; 137int vm_page_zero_count = 0; 138 139/* 140 * vm_set_page_size: 141 * 142 * Sets the page size, perhaps based upon the memory 143 * size. Must be called before any use of page-size 144 * dependent functions. 145 */ 146void 147vm_set_page_size(void) 148{ 149 if (cnt.v_page_size == 0) 150 cnt.v_page_size = PAGE_SIZE; 151 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 152 panic("vm_set_page_size: page size not a power of two"); 153} 154 155/* 156 * vm_page_startup: 157 * 158 * Initializes the resident memory module. 159 * 160 * Allocates memory for the page cells, and 161 * for the object/offset-to-page hash table headers. 162 * Each page cell is initialized and placed on the free list. 163 */ 164vm_offset_t 165vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 166{ 167 vm_offset_t mapped; 168 struct vm_page **bucket; 169 vm_size_t npages, page_range; 170 vm_offset_t new_end; 171 int i; 172 vm_offset_t pa; 173 int nblocks; 174 vm_offset_t last_pa; 175 176 /* the biggest memory array is the second group of pages */ 177 vm_offset_t end; 178 vm_offset_t biggestone, biggestsize; 179 180 vm_offset_t total; 181 vm_size_t bootpages; 182 183 total = 0; 184 biggestsize = 0; 185 biggestone = 0; 186 nblocks = 0; 187 vaddr = round_page(vaddr); 188 189 for (i = 0; phys_avail[i + 1]; i += 2) { 190 phys_avail[i] = round_page(phys_avail[i]); 191 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 192 } 193 194 for (i = 0; phys_avail[i + 1]; i += 2) { 195 vm_size_t size = phys_avail[i + 1] - phys_avail[i]; 196 197 if (size > biggestsize) { 198 biggestone = i; 199 biggestsize = size; 200 } 201 ++nblocks; 202 total += size; 203 } 204 205 end = phys_avail[biggestone+1]; 206 207 /* 208 * Initialize the queue headers for the free queue, the active queue 209 * and the inactive queue. 210 */ 211 vm_pageq_init(); 212 213 /* 214 * Allocate memory for use when boot strapping the kernel memory allocator 215 */ 216 bootpages = UMA_BOOT_PAGES * UMA_SLAB_SIZE; 217 new_end = end - bootpages; 218 new_end = trunc_page(new_end); 219 mapped = pmap_map(&vaddr, new_end, end, 220 VM_PROT_READ | VM_PROT_WRITE); 221 bzero((caddr_t) mapped, end - new_end); 222 uma_startup((caddr_t)mapped); 223 224 end = new_end; 225 226 /* 227 * Allocate (and initialize) the hash table buckets. 228 * 229 * The number of buckets MUST BE a power of 2, and the actual value is 230 * the next power of 2 greater than the number of physical pages in 231 * the system. 232 * 233 * We make the hash table approximately 2x the number of pages to 234 * reduce the chain length. This is about the same size using the 235 * singly-linked list as the 1x hash table we were using before 236 * using TAILQ but the chain length will be smaller. 237 * 238 * Note: This computation can be tweaked if desired. 239 */ 240 if (vm_page_bucket_count == 0) { 241 vm_page_bucket_count = 1; 242 while (vm_page_bucket_count < atop(total)) 243 vm_page_bucket_count <<= 1; 244 } 245 vm_page_bucket_count <<= 1; 246 vm_page_hash_mask = vm_page_bucket_count - 1; 247 248 /* 249 * Validate these addresses. 250 */ 251 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 252 new_end = trunc_page(new_end); 253 mapped = pmap_map(&vaddr, new_end, end, 254 VM_PROT_READ | VM_PROT_WRITE); 255 bzero((caddr_t) mapped, end - new_end); 256 257 vm_page_buckets = (struct vm_page **)mapped; 258 bucket = vm_page_buckets; 259 for (i = 0; i < vm_page_bucket_count; i++) { 260 *bucket = NULL; 261 bucket++; 262 } 263 for (i = 0; i < BUCKET_HASH_SIZE; ++i) 264 mtx_init(&vm_buckets_mtx[i], "vm buckets hash mutexes", NULL, 265 MTX_DEF); 266 267 /* 268 * Compute the number of pages of memory that will be available for 269 * use (taking into account the overhead of a page structure per 270 * page). 271 */ 272 first_page = phys_avail[0] / PAGE_SIZE; 273 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 274 npages = (total - (page_range * sizeof(struct vm_page)) - 275 (end - new_end)) / PAGE_SIZE; 276 end = new_end; 277 278 /* 279 * Initialize the mem entry structures now, and put them in the free 280 * queue. 281 */ 282 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 283 mapped = pmap_map(&vaddr, new_end, end, 284 VM_PROT_READ | VM_PROT_WRITE); 285 vm_page_array = (vm_page_t) mapped; 286 287 /* 288 * Clear all of the page structures 289 */ 290 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 291 vm_page_array_size = page_range; 292 293 /* 294 * Construct the free queue(s) in descending order (by physical 295 * address) so that the first 16MB of physical memory is allocated 296 * last rather than first. On large-memory machines, this avoids 297 * the exhaustion of low physical memory before isa_dmainit has run. 298 */ 299 cnt.v_page_count = 0; 300 cnt.v_free_count = 0; 301 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 302 pa = phys_avail[i]; 303 if (i == biggestone) 304 last_pa = new_end; 305 else 306 last_pa = phys_avail[i + 1]; 307 while (pa < last_pa && npages-- > 0) { 308 vm_pageq_add_new_page(pa); 309 pa += PAGE_SIZE; 310 } 311 } 312 return (vaddr); 313} 314 315/* 316 * vm_page_hash: 317 * 318 * Distributes the object/offset key pair among hash buckets. 319 * 320 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 321 * This routine may not block. 322 * 323 * We try to randomize the hash based on the object to spread the pages 324 * out in the hash table without it costing us too much. 325 */ 326static __inline int 327vm_page_hash(vm_object_t object, vm_pindex_t pindex) 328{ 329 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 330 331 return (i & vm_page_hash_mask); 332} 333 334void 335vm_page_flag_set(vm_page_t m, unsigned short bits) 336{ 337 GIANT_REQUIRED; 338 m->flags |= bits; 339} 340 341void 342vm_page_flag_clear(vm_page_t m, unsigned short bits) 343{ 344 GIANT_REQUIRED; 345 m->flags &= ~bits; 346} 347 348void 349vm_page_busy(vm_page_t m) 350{ 351 KASSERT((m->flags & PG_BUSY) == 0, 352 ("vm_page_busy: page already busy!!!")); 353 vm_page_flag_set(m, PG_BUSY); 354} 355 356/* 357 * vm_page_flash: 358 * 359 * wakeup anyone waiting for the page. 360 */ 361void 362vm_page_flash(vm_page_t m) 363{ 364 if (m->flags & PG_WANTED) { 365 vm_page_flag_clear(m, PG_WANTED); 366 wakeup(m); 367 } 368} 369 370/* 371 * vm_page_wakeup: 372 * 373 * clear the PG_BUSY flag and wakeup anyone waiting for the 374 * page. 375 * 376 */ 377void 378vm_page_wakeup(vm_page_t m) 379{ 380 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 381 vm_page_flag_clear(m, PG_BUSY); 382 vm_page_flash(m); 383} 384 385/* 386 * 387 * 388 */ 389void 390vm_page_io_start(vm_page_t m) 391{ 392 GIANT_REQUIRED; 393 m->busy++; 394} 395 396void 397vm_page_io_finish(vm_page_t m) 398{ 399 GIANT_REQUIRED; 400 m->busy--; 401 if (m->busy == 0) 402 vm_page_flash(m); 403} 404 405/* 406 * Keep page from being freed by the page daemon 407 * much of the same effect as wiring, except much lower 408 * overhead and should be used only for *very* temporary 409 * holding ("wiring"). 410 */ 411void 412vm_page_hold(vm_page_t mem) 413{ 414 GIANT_REQUIRED; 415 mem->hold_count++; 416} 417 418void 419vm_page_unhold(vm_page_t mem) 420{ 421 GIANT_REQUIRED; 422 --mem->hold_count; 423 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 424 if (mem->hold_count == 0 && mem->queue == PQ_HOLD) 425 vm_page_free_toq(mem); 426} 427 428/* 429 * vm_page_protect: 430 * 431 * Reduce the protection of a page. This routine never raises the 432 * protection and therefore can be safely called if the page is already 433 * at VM_PROT_NONE (it will be a NOP effectively ). 434 */ 435void 436vm_page_protect(vm_page_t mem, int prot) 437{ 438 if (prot == VM_PROT_NONE) { 439 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) { 440 pmap_page_protect(mem, VM_PROT_NONE); 441 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED); 442 } 443 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 444 pmap_page_protect(mem, VM_PROT_READ); 445 vm_page_flag_clear(mem, PG_WRITEABLE); 446 } 447} 448/* 449 * vm_page_zero_fill: 450 * 451 * Zero-fill the specified page. 452 * Written as a standard pagein routine, to 453 * be used by the zero-fill object. 454 */ 455boolean_t 456vm_page_zero_fill(vm_page_t m) 457{ 458 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 459 return (TRUE); 460} 461 462/* 463 * vm_page_copy: 464 * 465 * Copy one page to another 466 */ 467void 468vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 469{ 470 pmap_copy_page(VM_PAGE_TO_PHYS(src_m), VM_PAGE_TO_PHYS(dest_m)); 471 dest_m->valid = VM_PAGE_BITS_ALL; 472} 473 474/* 475 * vm_page_free: 476 * 477 * Free a page 478 * 479 * The clearing of PG_ZERO is a temporary safety until the code can be 480 * reviewed to determine that PG_ZERO is being properly cleared on 481 * write faults or maps. PG_ZERO was previously cleared in 482 * vm_page_alloc(). 483 */ 484void 485vm_page_free(vm_page_t m) 486{ 487 vm_page_flag_clear(m, PG_ZERO); 488 vm_page_free_toq(m); 489 vm_page_zero_idle_wakeup(); 490} 491 492/* 493 * vm_page_free_zero: 494 * 495 * Free a page to the zerod-pages queue 496 */ 497void 498vm_page_free_zero(vm_page_t m) 499{ 500 vm_page_flag_set(m, PG_ZERO); 501 vm_page_free_toq(m); 502} 503 504/* 505 * vm_page_sleep_busy: 506 * 507 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 508 * m->busy is zero. Returns TRUE if it had to sleep ( including if 509 * it almost had to sleep and made temporary spl*() mods), FALSE 510 * otherwise. 511 * 512 * This routine assumes that interrupts can only remove the busy 513 * status from a page, not set the busy status or change it from 514 * PG_BUSY to m->busy or vise versa (which would create a timing 515 * window). 516 */ 517int 518vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 519{ 520 GIANT_REQUIRED; 521 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 522 int s = splvm(); 523 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 524 /* 525 * Page is busy. Wait and retry. 526 */ 527 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 528 tsleep(m, PVM, msg, 0); 529 } 530 splx(s); 531 return (TRUE); 532 /* not reached */ 533 } 534 return (FALSE); 535} 536/* 537 * vm_page_dirty: 538 * 539 * make page all dirty 540 */ 541void 542vm_page_dirty(vm_page_t m) 543{ 544 KASSERT(m->queue - m->pc != PQ_CACHE, 545 ("vm_page_dirty: page in cache!")); 546 m->dirty = VM_PAGE_BITS_ALL; 547} 548 549/* 550 * vm_page_undirty: 551 * 552 * Set page to not be dirty. Note: does not clear pmap modify bits 553 */ 554void 555vm_page_undirty(vm_page_t m) 556{ 557 m->dirty = 0; 558} 559 560/* 561 * vm_page_insert: [ internal use only ] 562 * 563 * Inserts the given mem entry into the object and object list. 564 * 565 * The pagetables are not updated but will presumably fault the page 566 * in if necessary, or if a kernel page the caller will at some point 567 * enter the page into the kernel's pmap. We are not allowed to block 568 * here so we *can't* do this anyway. 569 * 570 * The object and page must be locked, and must be splhigh. 571 * This routine may not block. 572 */ 573void 574vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 575{ 576 struct vm_page **bucket; 577 578 GIANT_REQUIRED; 579 580 if (m->object != NULL) 581 panic("vm_page_insert: already inserted"); 582 583 /* 584 * Record the object/offset pair in this page 585 */ 586 m->object = object; 587 m->pindex = pindex; 588 589 /* 590 * Insert it into the object_object/offset hash table 591 */ 592 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 593 m->hnext = *bucket; 594 *bucket = m; 595 vm_page_bucket_generation++; 596 597 /* 598 * Now link into the object's list of backed pages. 599 */ 600 TAILQ_INSERT_TAIL(&object->memq, m, listq); 601 object->generation++; 602 603 /* 604 * show that the object has one more resident page. 605 */ 606 object->resident_page_count++; 607 608 /* 609 * Since we are inserting a new and possibly dirty page, 610 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 611 */ 612 if (m->flags & PG_WRITEABLE) 613 vm_object_set_writeable_dirty(object); 614} 615 616/* 617 * vm_page_remove: 618 * NOTE: used by device pager as well -wfj 619 * 620 * Removes the given mem entry from the object/offset-page 621 * table and the object page list, but do not invalidate/terminate 622 * the backing store. 623 * 624 * The object and page must be locked, and at splhigh. 625 * The underlying pmap entry (if any) is NOT removed here. 626 * This routine may not block. 627 */ 628void 629vm_page_remove(vm_page_t m) 630{ 631 vm_object_t object; 632 633 GIANT_REQUIRED; 634 635 if (m->object == NULL) 636 return; 637 638 if ((m->flags & PG_BUSY) == 0) { 639 panic("vm_page_remove: page not busy"); 640 } 641 642 /* 643 * Basically destroy the page. 644 */ 645 vm_page_wakeup(m); 646 647 object = m->object; 648 649 /* 650 * Remove from the object_object/offset hash table. The object 651 * must be on the hash queue, we will panic if it isn't 652 * 653 * Note: we must NULL-out m->hnext to prevent loops in detached 654 * buffers with vm_page_lookup(). 655 */ 656 { 657 struct vm_page **bucket; 658 659 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 660 while (*bucket != m) { 661 if (*bucket == NULL) 662 panic("vm_page_remove(): page not found in hash"); 663 bucket = &(*bucket)->hnext; 664 } 665 *bucket = m->hnext; 666 m->hnext = NULL; 667 vm_page_bucket_generation++; 668 } 669 670 /* 671 * Now remove from the object's list of backed pages. 672 */ 673 TAILQ_REMOVE(&object->memq, m, listq); 674 675 /* 676 * And show that the object has one fewer resident page. 677 */ 678 object->resident_page_count--; 679 object->generation++; 680 681 m->object = NULL; 682} 683 684/* 685 * vm_page_lookup: 686 * 687 * Returns the page associated with the object/offset 688 * pair specified; if none is found, NULL is returned. 689 * 690 * NOTE: the code below does not lock. It will operate properly if 691 * an interrupt makes a change, but the generation algorithm will not 692 * operate properly in an SMP environment where both cpu's are able to run 693 * kernel code simultaneously. 694 * 695 * The object must be locked. No side effects. 696 * This routine may not block. 697 * This is a critical path routine 698 */ 699vm_page_t 700vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 701{ 702 vm_page_t m; 703 struct vm_page **bucket; 704 int generation; 705 706 /* 707 * Search the hash table for this object/offset pair 708 */ 709retry: 710 generation = vm_page_bucket_generation; 711 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 712 for (m = *bucket; m != NULL; m = m->hnext) { 713 if ((m->object == object) && (m->pindex == pindex)) { 714 if (vm_page_bucket_generation != generation) 715 goto retry; 716 return (m); 717 } 718 } 719 if (vm_page_bucket_generation != generation) 720 goto retry; 721 return (NULL); 722} 723 724/* 725 * vm_page_rename: 726 * 727 * Move the given memory entry from its 728 * current object to the specified target object/offset. 729 * 730 * The object must be locked. 731 * This routine may not block. 732 * 733 * Note: this routine will raise itself to splvm(), the caller need not. 734 * 735 * Note: swap associated with the page must be invalidated by the move. We 736 * have to do this for several reasons: (1) we aren't freeing the 737 * page, (2) we are dirtying the page, (3) the VM system is probably 738 * moving the page from object A to B, and will then later move 739 * the backing store from A to B and we can't have a conflict. 740 * 741 * Note: we *always* dirty the page. It is necessary both for the 742 * fact that we moved it, and because we may be invalidating 743 * swap. If the page is on the cache, we have to deactivate it 744 * or vm_page_dirty() will panic. Dirty pages are not allowed 745 * on the cache. 746 */ 747void 748vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 749{ 750 int s; 751 752 s = splvm(); 753 vm_page_remove(m); 754 vm_page_insert(m, new_object, new_pindex); 755 if (m->queue - m->pc == PQ_CACHE) 756 vm_page_deactivate(m); 757 vm_page_dirty(m); 758 splx(s); 759} 760 761/* 762 * vm_page_select_cache: 763 * 764 * Find a page on the cache queue with color optimization. As pages 765 * might be found, but not applicable, they are deactivated. This 766 * keeps us from using potentially busy cached pages. 767 * 768 * This routine must be called at splvm(). 769 * This routine may not block. 770 */ 771static vm_page_t 772vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 773{ 774 vm_page_t m; 775 776 GIANT_REQUIRED; 777 while (TRUE) { 778 m = vm_pageq_find( 779 PQ_CACHE, 780 (pindex + object->pg_color) & PQ_L2_MASK, 781 FALSE 782 ); 783 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 784 m->hold_count || m->wire_count)) { 785 vm_page_deactivate(m); 786 continue; 787 } 788 return m; 789 } 790} 791 792/* 793 * vm_page_select_free: 794 * 795 * Find a free or zero page, with specified preference. 796 * 797 * This routine must be called at splvm(). 798 * This routine may not block. 799 */ 800static __inline vm_page_t 801vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 802{ 803 vm_page_t m; 804 805 m = vm_pageq_find( 806 PQ_FREE, 807 (pindex + object->pg_color) & PQ_L2_MASK, 808 prefer_zero 809 ); 810 return (m); 811} 812 813/* 814 * vm_page_alloc: 815 * 816 * Allocate and return a memory cell associated 817 * with this VM object/offset pair. 818 * 819 * page_req classes: 820 * VM_ALLOC_NORMAL normal process request 821 * VM_ALLOC_SYSTEM system *really* needs a page 822 * VM_ALLOC_INTERRUPT interrupt time request 823 * VM_ALLOC_ZERO zero page 824 * 825 * This routine may not block. 826 * 827 * Additional special handling is required when called from an 828 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 829 * the page cache in this case. 830 */ 831vm_page_t 832vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 833{ 834 vm_page_t m = NULL; 835 int s; 836 837 GIANT_REQUIRED; 838 839 KASSERT(!vm_page_lookup(object, pindex), 840 ("vm_page_alloc: page already allocated")); 841 842 /* 843 * The pager is allowed to eat deeper into the free page list. 844 */ 845 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 846 page_req = VM_ALLOC_SYSTEM; 847 }; 848 849 s = splvm(); 850 851loop: 852 if (cnt.v_free_count > cnt.v_free_reserved) { 853 /* 854 * Allocate from the free queue if there are plenty of pages 855 * in it. 856 */ 857 if (page_req == VM_ALLOC_ZERO) 858 m = vm_page_select_free(object, pindex, TRUE); 859 else 860 m = vm_page_select_free(object, pindex, FALSE); 861 } else if ( 862 (page_req == VM_ALLOC_SYSTEM && 863 cnt.v_cache_count == 0 && 864 cnt.v_free_count > cnt.v_interrupt_free_min) || 865 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 866 ) { 867 /* 868 * Interrupt or system, dig deeper into the free list. 869 */ 870 m = vm_page_select_free(object, pindex, FALSE); 871 } else if (page_req != VM_ALLOC_INTERRUPT) { 872 /* 873 * Allocatable from cache (non-interrupt only). On success, 874 * we must free the page and try again, thus ensuring that 875 * cnt.v_*_free_min counters are replenished. 876 */ 877 m = vm_page_select_cache(object, pindex); 878 if (m == NULL) { 879 splx(s); 880#if defined(DIAGNOSTIC) 881 if (cnt.v_cache_count > 0) 882 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 883#endif 884 vm_pageout_deficit++; 885 pagedaemon_wakeup(); 886 return (NULL); 887 } 888 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 889 vm_page_busy(m); 890 vm_page_protect(m, VM_PROT_NONE); 891 vm_page_free(m); 892 goto loop; 893 } else { 894 /* 895 * Not allocatable from cache from interrupt, give up. 896 */ 897 splx(s); 898 vm_pageout_deficit++; 899 pagedaemon_wakeup(); 900 return (NULL); 901 } 902 903 /* 904 * At this point we had better have found a good page. 905 */ 906 907 KASSERT( 908 m != NULL, 909 ("vm_page_alloc(): missing page on free queue\n") 910 ); 911 912 /* 913 * Remove from free queue 914 */ 915 916 vm_pageq_remove_nowakeup(m); 917 918 /* 919 * Initialize structure. Only the PG_ZERO flag is inherited. 920 */ 921 if (m->flags & PG_ZERO) { 922 vm_page_zero_count--; 923 m->flags = PG_ZERO | PG_BUSY; 924 } else { 925 m->flags = PG_BUSY; 926 } 927 m->wire_count = 0; 928 m->hold_count = 0; 929 m->act_count = 0; 930 m->busy = 0; 931 m->valid = 0; 932 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 933 934 /* 935 * vm_page_insert() is safe prior to the splx(). Note also that 936 * inserting a page here does not insert it into the pmap (which 937 * could cause us to block allocating memory). We cannot block 938 * anywhere. 939 */ 940 vm_page_insert(m, object, pindex); 941 942 /* 943 * Don't wakeup too often - wakeup the pageout daemon when 944 * we would be nearly out of memory. 945 */ 946 if (vm_paging_needed()) 947 pagedaemon_wakeup(); 948 949 splx(s); 950 return (m); 951} 952 953/* 954 * vm_wait: (also see VM_WAIT macro) 955 * 956 * Block until free pages are available for allocation 957 * - Called in various places before memory allocations. 958 */ 959void 960vm_wait(void) 961{ 962 int s; 963 964 s = splvm(); 965 if (curproc == pageproc) { 966 vm_pageout_pages_needed = 1; 967 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 968 } else { 969 if (!vm_pages_needed) { 970 vm_pages_needed = 1; 971 wakeup(&vm_pages_needed); 972 } 973 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 974 } 975 splx(s); 976} 977 978/* 979 * vm_waitpfault: (also see VM_WAITPFAULT macro) 980 * 981 * Block until free pages are available for allocation 982 * - Called only in vm_fault so that processes page faulting 983 * can be easily tracked. 984 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing 985 * processes will be able to grab memory first. Do not change 986 * this balance without careful testing first. 987 */ 988void 989vm_waitpfault(void) 990{ 991 int s; 992 993 s = splvm(); 994 if (!vm_pages_needed) { 995 vm_pages_needed = 1; 996 wakeup(&vm_pages_needed); 997 } 998 tsleep(&cnt.v_free_count, PUSER, "pfault", 0); 999 splx(s); 1000} 1001 1002/* 1003 * vm_page_activate: 1004 * 1005 * Put the specified page on the active list (if appropriate). 1006 * Ensure that act_count is at least ACT_INIT but do not otherwise 1007 * mess with it. 1008 * 1009 * The page queues must be locked. 1010 * This routine may not block. 1011 */ 1012void 1013vm_page_activate(vm_page_t m) 1014{ 1015 int s; 1016 1017 GIANT_REQUIRED; 1018 s = splvm(); 1019 if (m->queue != PQ_ACTIVE) { 1020 if ((m->queue - m->pc) == PQ_CACHE) 1021 cnt.v_reactivated++; 1022 vm_pageq_remove(m); 1023 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1024 if (m->act_count < ACT_INIT) 1025 m->act_count = ACT_INIT; 1026 vm_pageq_enqueue(PQ_ACTIVE, m); 1027 } 1028 } else { 1029 if (m->act_count < ACT_INIT) 1030 m->act_count = ACT_INIT; 1031 } 1032 splx(s); 1033} 1034 1035/* 1036 * vm_page_free_wakeup: 1037 * 1038 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1039 * routine is called when a page has been added to the cache or free 1040 * queues. 1041 * 1042 * This routine may not block. 1043 * This routine must be called at splvm() 1044 */ 1045static __inline void 1046vm_page_free_wakeup(void) 1047{ 1048 /* 1049 * if pageout daemon needs pages, then tell it that there are 1050 * some free. 1051 */ 1052 if (vm_pageout_pages_needed && 1053 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1054 wakeup(&vm_pageout_pages_needed); 1055 vm_pageout_pages_needed = 0; 1056 } 1057 /* 1058 * wakeup processes that are waiting on memory if we hit a 1059 * high water mark. And wakeup scheduler process if we have 1060 * lots of memory. this process will swapin processes. 1061 */ 1062 if (vm_pages_needed && !vm_page_count_min()) { 1063 vm_pages_needed = 0; 1064 wakeup(&cnt.v_free_count); 1065 } 1066} 1067 1068/* 1069 * vm_page_free_toq: 1070 * 1071 * Returns the given page to the PQ_FREE list, 1072 * disassociating it with any VM object. 1073 * 1074 * Object and page must be locked prior to entry. 1075 * This routine may not block. 1076 */ 1077 1078void 1079vm_page_free_toq(vm_page_t m) 1080{ 1081 int s; 1082 struct vpgqueues *pq; 1083 vm_object_t object = m->object; 1084 1085 GIANT_REQUIRED; 1086 s = splvm(); 1087 cnt.v_tfree++; 1088 1089 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) { 1090 printf( 1091 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1092 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1093 m->hold_count); 1094 if ((m->queue - m->pc) == PQ_FREE) 1095 panic("vm_page_free: freeing free page"); 1096 else 1097 panic("vm_page_free: freeing busy page"); 1098 } 1099 1100 /* 1101 * unqueue, then remove page. Note that we cannot destroy 1102 * the page here because we do not want to call the pager's 1103 * callback routine until after we've put the page on the 1104 * appropriate free queue. 1105 */ 1106 vm_pageq_remove_nowakeup(m); 1107 vm_page_remove(m); 1108 1109 /* 1110 * If fictitious remove object association and 1111 * return, otherwise delay object association removal. 1112 */ 1113 if ((m->flags & PG_FICTITIOUS) != 0) { 1114 splx(s); 1115 return; 1116 } 1117 1118 m->valid = 0; 1119 vm_page_undirty(m); 1120 1121 if (m->wire_count != 0) { 1122 if (m->wire_count > 1) { 1123 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1124 m->wire_count, (long)m->pindex); 1125 } 1126 panic("vm_page_free: freeing wired page\n"); 1127 } 1128 1129 /* 1130 * If we've exhausted the object's resident pages we want to free 1131 * it up. 1132 */ 1133 if (object && 1134 (object->type == OBJT_VNODE) && 1135 ((object->flags & OBJ_DEAD) == 0) 1136 ) { 1137 struct vnode *vp = (struct vnode *)object->handle; 1138 1139 if (vp && VSHOULDFREE(vp)) 1140 vfree(vp); 1141 } 1142 1143 /* 1144 * Clear the UNMANAGED flag when freeing an unmanaged page. 1145 */ 1146 if (m->flags & PG_UNMANAGED) { 1147 m->flags &= ~PG_UNMANAGED; 1148 } else { 1149#ifdef __alpha__ 1150 pmap_page_is_free(m); 1151#endif 1152 } 1153 1154 if (m->hold_count != 0) { 1155 m->flags &= ~PG_ZERO; 1156 m->queue = PQ_HOLD; 1157 } else 1158 m->queue = PQ_FREE + m->pc; 1159 pq = &vm_page_queues[m->queue]; 1160 pq->lcnt++; 1161 ++(*pq->cnt); 1162 1163 /* 1164 * Put zero'd pages on the end ( where we look for zero'd pages 1165 * first ) and non-zerod pages at the head. 1166 */ 1167 if (m->flags & PG_ZERO) { 1168 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1169 ++vm_page_zero_count; 1170 } else { 1171 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1172 } 1173 vm_page_free_wakeup(); 1174 splx(s); 1175} 1176 1177/* 1178 * vm_page_unmanage: 1179 * 1180 * Prevent PV management from being done on the page. The page is 1181 * removed from the paging queues as if it were wired, and as a 1182 * consequence of no longer being managed the pageout daemon will not 1183 * touch it (since there is no way to locate the pte mappings for the 1184 * page). madvise() calls that mess with the pmap will also no longer 1185 * operate on the page. 1186 * 1187 * Beyond that the page is still reasonably 'normal'. Freeing the page 1188 * will clear the flag. 1189 * 1190 * This routine is used by OBJT_PHYS objects - objects using unswappable 1191 * physical memory as backing store rather then swap-backed memory and 1192 * will eventually be extended to support 4MB unmanaged physical 1193 * mappings. 1194 */ 1195void 1196vm_page_unmanage(vm_page_t m) 1197{ 1198 int s; 1199 1200 s = splvm(); 1201 if ((m->flags & PG_UNMANAGED) == 0) { 1202 if (m->wire_count == 0) 1203 vm_pageq_remove(m); 1204 } 1205 vm_page_flag_set(m, PG_UNMANAGED); 1206 splx(s); 1207} 1208 1209/* 1210 * vm_page_wire: 1211 * 1212 * Mark this page as wired down by yet 1213 * another map, removing it from paging queues 1214 * as necessary. 1215 * 1216 * The page queues must be locked. 1217 * This routine may not block. 1218 */ 1219void 1220vm_page_wire(vm_page_t m) 1221{ 1222 int s; 1223 1224 /* 1225 * Only bump the wire statistics if the page is not already wired, 1226 * and only unqueue the page if it is on some queue (if it is unmanaged 1227 * it is already off the queues). 1228 */ 1229 s = splvm(); 1230 if (m->wire_count == 0) { 1231 if ((m->flags & PG_UNMANAGED) == 0) 1232 vm_pageq_remove(m); 1233 cnt.v_wire_count++; 1234 } 1235 m->wire_count++; 1236 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m)); 1237 splx(s); 1238 vm_page_flag_set(m, PG_MAPPED); 1239} 1240 1241/* 1242 * vm_page_unwire: 1243 * 1244 * Release one wiring of this page, potentially 1245 * enabling it to be paged again. 1246 * 1247 * Many pages placed on the inactive queue should actually go 1248 * into the cache, but it is difficult to figure out which. What 1249 * we do instead, if the inactive target is well met, is to put 1250 * clean pages at the head of the inactive queue instead of the tail. 1251 * This will cause them to be moved to the cache more quickly and 1252 * if not actively re-referenced, freed more quickly. If we just 1253 * stick these pages at the end of the inactive queue, heavy filesystem 1254 * meta-data accesses can cause an unnecessary paging load on memory bound 1255 * processes. This optimization causes one-time-use metadata to be 1256 * reused more quickly. 1257 * 1258 * BUT, if we are in a low-memory situation we have no choice but to 1259 * put clean pages on the cache queue. 1260 * 1261 * A number of routines use vm_page_unwire() to guarantee that the page 1262 * will go into either the inactive or active queues, and will NEVER 1263 * be placed in the cache - for example, just after dirtying a page. 1264 * dirty pages in the cache are not allowed. 1265 * 1266 * The page queues must be locked. 1267 * This routine may not block. 1268 */ 1269void 1270vm_page_unwire(vm_page_t m, int activate) 1271{ 1272 int s; 1273 1274 s = splvm(); 1275 1276 if (m->wire_count > 0) { 1277 m->wire_count--; 1278 if (m->wire_count == 0) { 1279 cnt.v_wire_count--; 1280 if (m->flags & PG_UNMANAGED) { 1281 ; 1282 } else if (activate) 1283 vm_pageq_enqueue(PQ_ACTIVE, m); 1284 else { 1285 vm_page_flag_clear(m, PG_WINATCFLS); 1286 vm_pageq_enqueue(PQ_INACTIVE, m); 1287 } 1288 } 1289 } else { 1290 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1291 } 1292 splx(s); 1293} 1294 1295 1296/* 1297 * Move the specified page to the inactive queue. If the page has 1298 * any associated swap, the swap is deallocated. 1299 * 1300 * Normally athead is 0 resulting in LRU operation. athead is set 1301 * to 1 if we want this page to be 'as if it were placed in the cache', 1302 * except without unmapping it from the process address space. 1303 * 1304 * This routine may not block. 1305 */ 1306static __inline void 1307_vm_page_deactivate(vm_page_t m, int athead) 1308{ 1309 int s; 1310 1311 GIANT_REQUIRED; 1312 /* 1313 * Ignore if already inactive. 1314 */ 1315 if (m->queue == PQ_INACTIVE) 1316 return; 1317 1318 s = splvm(); 1319 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1320 if ((m->queue - m->pc) == PQ_CACHE) 1321 cnt.v_reactivated++; 1322 vm_page_flag_clear(m, PG_WINATCFLS); 1323 vm_pageq_remove(m); 1324 if (athead) 1325 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1326 else 1327 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1328 m->queue = PQ_INACTIVE; 1329 vm_page_queues[PQ_INACTIVE].lcnt++; 1330 cnt.v_inactive_count++; 1331 } 1332 splx(s); 1333} 1334 1335void 1336vm_page_deactivate(vm_page_t m) 1337{ 1338 _vm_page_deactivate(m, 0); 1339} 1340 1341/* 1342 * vm_page_try_to_cache: 1343 * 1344 * Returns 0 on failure, 1 on success 1345 */ 1346int 1347vm_page_try_to_cache(vm_page_t m) 1348{ 1349 GIANT_REQUIRED; 1350 1351 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1352 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1353 return (0); 1354 } 1355 vm_page_test_dirty(m); 1356 if (m->dirty) 1357 return (0); 1358 vm_page_cache(m); 1359 return (1); 1360} 1361 1362/* 1363 * vm_page_try_to_free() 1364 * 1365 * Attempt to free the page. If we cannot free it, we do nothing. 1366 * 1 is returned on success, 0 on failure. 1367 */ 1368int 1369vm_page_try_to_free(vm_page_t m) 1370{ 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_busy(m); 1379 vm_page_protect(m, VM_PROT_NONE); 1380 vm_page_free(m); 1381 return (1); 1382} 1383 1384/* 1385 * vm_page_cache 1386 * 1387 * Put the specified page onto the page cache queue (if appropriate). 1388 * 1389 * This routine may not block. 1390 */ 1391void 1392vm_page_cache(vm_page_t m) 1393{ 1394 int s; 1395 1396 GIANT_REQUIRED; 1397 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1398 printf("vm_page_cache: attempting to cache busy page\n"); 1399 return; 1400 } 1401 if ((m->queue - m->pc) == PQ_CACHE) 1402 return; 1403 1404 /* 1405 * Remove all pmaps and indicate that the page is not 1406 * writeable or mapped. 1407 */ 1408 vm_page_protect(m, VM_PROT_NONE); 1409 if (m->dirty != 0) { 1410 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1411 (long)m->pindex); 1412 } 1413 s = splvm(); 1414 vm_pageq_remove_nowakeup(m); 1415 vm_pageq_enqueue(PQ_CACHE + m->pc, m); 1416 vm_page_free_wakeup(); 1417 splx(s); 1418} 1419 1420/* 1421 * vm_page_dontneed 1422 * 1423 * Cache, deactivate, or do nothing as appropriate. This routine 1424 * is typically used by madvise() MADV_DONTNEED. 1425 * 1426 * Generally speaking we want to move the page into the cache so 1427 * it gets reused quickly. However, this can result in a silly syndrome 1428 * due to the page recycling too quickly. Small objects will not be 1429 * fully cached. On the otherhand, if we move the page to the inactive 1430 * queue we wind up with a problem whereby very large objects 1431 * unnecessarily blow away our inactive and cache queues. 1432 * 1433 * The solution is to move the pages based on a fixed weighting. We 1434 * either leave them alone, deactivate them, or move them to the cache, 1435 * where moving them to the cache has the highest weighting. 1436 * By forcing some pages into other queues we eventually force the 1437 * system to balance the queues, potentially recovering other unrelated 1438 * space from active. The idea is to not force this to happen too 1439 * often. 1440 */ 1441void 1442vm_page_dontneed(vm_page_t m) 1443{ 1444 static int dnweight; 1445 int dnw; 1446 int head; 1447 1448 GIANT_REQUIRED; 1449 dnw = ++dnweight; 1450 1451 /* 1452 * occassionally leave the page alone 1453 */ 1454 if ((dnw & 0x01F0) == 0 || 1455 m->queue == PQ_INACTIVE || 1456 m->queue - m->pc == PQ_CACHE 1457 ) { 1458 if (m->act_count >= ACT_INIT) 1459 --m->act_count; 1460 return; 1461 } 1462 1463 if (m->dirty == 0) 1464 vm_page_test_dirty(m); 1465 1466 if (m->dirty || (dnw & 0x0070) == 0) { 1467 /* 1468 * Deactivate the page 3 times out of 32. 1469 */ 1470 head = 0; 1471 } else { 1472 /* 1473 * Cache the page 28 times out of every 32. Note that 1474 * the page is deactivated instead of cached, but placed 1475 * at the head of the queue instead of the tail. 1476 */ 1477 head = 1; 1478 } 1479 _vm_page_deactivate(m, head); 1480} 1481 1482/* 1483 * Grab a page, waiting until we are waken up due to the page 1484 * changing state. We keep on waiting, if the page continues 1485 * to be in the object. If the page doesn't exist, allocate it. 1486 * 1487 * This routine may block. 1488 */ 1489vm_page_t 1490vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1491{ 1492 vm_page_t m; 1493 int s, generation; 1494 1495 GIANT_REQUIRED; 1496retrylookup: 1497 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1498 if (m->busy || (m->flags & PG_BUSY)) { 1499 generation = object->generation; 1500 1501 s = splvm(); 1502 while ((object->generation == generation) && 1503 (m->busy || (m->flags & PG_BUSY))) { 1504 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1505 tsleep(m, PVM, "pgrbwt", 0); 1506 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1507 splx(s); 1508 return NULL; 1509 } 1510 } 1511 splx(s); 1512 goto retrylookup; 1513 } else { 1514 vm_page_busy(m); 1515 return m; 1516 } 1517 } 1518 1519 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1520 if (m == NULL) { 1521 VM_WAIT; 1522 if ((allocflags & VM_ALLOC_RETRY) == 0) 1523 return NULL; 1524 goto retrylookup; 1525 } 1526 1527 return m; 1528} 1529 1530/* 1531 * Mapping function for valid bits or for dirty bits in 1532 * a page. May not block. 1533 * 1534 * Inputs are required to range within a page. 1535 */ 1536__inline int 1537vm_page_bits(int base, int size) 1538{ 1539 int first_bit; 1540 int last_bit; 1541 1542 KASSERT( 1543 base + size <= PAGE_SIZE, 1544 ("vm_page_bits: illegal base/size %d/%d", base, size) 1545 ); 1546 1547 if (size == 0) /* handle degenerate case */ 1548 return (0); 1549 1550 first_bit = base >> DEV_BSHIFT; 1551 last_bit = (base + size - 1) >> DEV_BSHIFT; 1552 1553 return ((2 << last_bit) - (1 << first_bit)); 1554} 1555 1556/* 1557 * vm_page_set_validclean: 1558 * 1559 * Sets portions of a page valid and clean. The arguments are expected 1560 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1561 * of any partial chunks touched by the range. The invalid portion of 1562 * such chunks will be zero'd. 1563 * 1564 * This routine may not block. 1565 * 1566 * (base + size) must be less then or equal to PAGE_SIZE. 1567 */ 1568void 1569vm_page_set_validclean(vm_page_t m, int base, int size) 1570{ 1571 int pagebits; 1572 int frag; 1573 int endoff; 1574 1575 GIANT_REQUIRED; 1576 if (size == 0) /* handle degenerate case */ 1577 return; 1578 1579 /* 1580 * If the base is not DEV_BSIZE aligned and the valid 1581 * bit is clear, we have to zero out a portion of the 1582 * first block. 1583 */ 1584 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1585 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1586 ) { 1587 pmap_zero_page_area( 1588 VM_PAGE_TO_PHYS(m), 1589 frag, 1590 base - frag 1591 ); 1592 } 1593 1594 /* 1595 * If the ending offset is not DEV_BSIZE aligned and the 1596 * valid bit is clear, we have to zero out a portion of 1597 * the last block. 1598 */ 1599 endoff = base + size; 1600 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1601 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1602 ) { 1603 pmap_zero_page_area( 1604 VM_PAGE_TO_PHYS(m), 1605 endoff, 1606 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1607 ); 1608 } 1609 1610 /* 1611 * Set valid, clear dirty bits. If validating the entire 1612 * page we can safely clear the pmap modify bit. We also 1613 * use this opportunity to clear the PG_NOSYNC flag. If a process 1614 * takes a write fault on a MAP_NOSYNC memory area the flag will 1615 * be set again. 1616 * 1617 * We set valid bits inclusive of any overlap, but we can only 1618 * clear dirty bits for DEV_BSIZE chunks that are fully within 1619 * the range. 1620 */ 1621 pagebits = vm_page_bits(base, size); 1622 m->valid |= pagebits; 1623#if 0 /* NOT YET */ 1624 if ((frag = base & (DEV_BSIZE - 1)) != 0) { 1625 frag = DEV_BSIZE - frag; 1626 base += frag; 1627 size -= frag; 1628 if (size < 0) 1629 size = 0; 1630 } 1631 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1)); 1632#endif 1633 m->dirty &= ~pagebits; 1634 if (base == 0 && size == PAGE_SIZE) { 1635 pmap_clear_modify(m); 1636 vm_page_flag_clear(m, PG_NOSYNC); 1637 } 1638} 1639 1640#if 0 1641 1642void 1643vm_page_set_dirty(vm_page_t m, int base, int size) 1644{ 1645 m->dirty |= vm_page_bits(base, size); 1646} 1647 1648#endif 1649 1650void 1651vm_page_clear_dirty(vm_page_t m, int base, int size) 1652{ 1653 GIANT_REQUIRED; 1654 m->dirty &= ~vm_page_bits(base, size); 1655} 1656 1657/* 1658 * vm_page_set_invalid: 1659 * 1660 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1661 * valid and dirty bits for the effected areas are cleared. 1662 * 1663 * May not block. 1664 */ 1665void 1666vm_page_set_invalid(vm_page_t m, int base, int size) 1667{ 1668 int bits; 1669 1670 GIANT_REQUIRED; 1671 bits = vm_page_bits(base, size); 1672 m->valid &= ~bits; 1673 m->dirty &= ~bits; 1674 m->object->generation++; 1675} 1676 1677/* 1678 * vm_page_zero_invalid() 1679 * 1680 * The kernel assumes that the invalid portions of a page contain 1681 * garbage, but such pages can be mapped into memory by user code. 1682 * When this occurs, we must zero out the non-valid portions of the 1683 * page so user code sees what it expects. 1684 * 1685 * Pages are most often semi-valid when the end of a file is mapped 1686 * into memory and the file's size is not page aligned. 1687 */ 1688void 1689vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1690{ 1691 int b; 1692 int i; 1693 1694 /* 1695 * Scan the valid bits looking for invalid sections that 1696 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1697 * valid bit may be set ) have already been zerod by 1698 * vm_page_set_validclean(). 1699 */ 1700 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1701 if (i == (PAGE_SIZE / DEV_BSIZE) || 1702 (m->valid & (1 << i)) 1703 ) { 1704 if (i > b) { 1705 pmap_zero_page_area( 1706 VM_PAGE_TO_PHYS(m), 1707 b << DEV_BSHIFT, 1708 (i - b) << DEV_BSHIFT 1709 ); 1710 } 1711 b = i + 1; 1712 } 1713 } 1714 1715 /* 1716 * setvalid is TRUE when we can safely set the zero'd areas 1717 * as being valid. We can do this if there are no cache consistancy 1718 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1719 */ 1720 if (setvalid) 1721 m->valid = VM_PAGE_BITS_ALL; 1722} 1723 1724/* 1725 * vm_page_is_valid: 1726 * 1727 * Is (partial) page valid? Note that the case where size == 0 1728 * will return FALSE in the degenerate case where the page is 1729 * entirely invalid, and TRUE otherwise. 1730 * 1731 * May not block. 1732 */ 1733int 1734vm_page_is_valid(vm_page_t m, int base, int size) 1735{ 1736 int bits = vm_page_bits(base, size); 1737 1738 if (m->valid && ((m->valid & bits) == bits)) 1739 return 1; 1740 else 1741 return 0; 1742} 1743 1744/* 1745 * update dirty bits from pmap/mmu. May not block. 1746 */ 1747void 1748vm_page_test_dirty(vm_page_t m) 1749{ 1750 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1751 vm_page_dirty(m); 1752 } 1753} 1754 1755#include "opt_ddb.h" 1756#ifdef DDB 1757#include <sys/kernel.h> 1758 1759#include <ddb/ddb.h> 1760 1761DB_SHOW_COMMAND(page, vm_page_print_page_info) 1762{ 1763 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1764 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1765 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1766 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1767 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1768 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1769 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1770 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1771 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1772 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1773} 1774 1775DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1776{ 1777 int i; 1778 db_printf("PQ_FREE:"); 1779 for (i = 0; i < PQ_L2_SIZE; i++) { 1780 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1781 } 1782 db_printf("\n"); 1783 1784 db_printf("PQ_CACHE:"); 1785 for (i = 0; i < PQ_L2_SIZE; i++) { 1786 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1787 } 1788 db_printf("\n"); 1789 1790 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1791 vm_page_queues[PQ_ACTIVE].lcnt, 1792 vm_page_queues[PQ_INACTIVE].lcnt); 1793} 1794#endif /* DDB */ 1795