vm_page.c revision 79248
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 79248 2001-07-04 20:15:18Z dillon $ 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 * Resident memory management module. 69 */ 70 71#include <sys/param.h> 72#include <sys/systm.h> 73#include <sys/lock.h> 74#include <sys/malloc.h> 75#include <sys/mutex.h> 76#include <sys/proc.h> 77#include <sys/vmmeter.h> 78#include <sys/vnode.h> 79 80#include <vm/vm.h> 81#include <vm/vm_param.h> 82#include <vm/vm_kern.h> 83#include <vm/vm_object.h> 84#include <vm/vm_page.h> 85#include <vm/vm_pageout.h> 86#include <vm/vm_pager.h> 87#include <vm/vm_extern.h> 88 89static void vm_page_queue_init __P((void)); 90static vm_page_t vm_page_select_cache __P((vm_object_t, vm_pindex_t)); 91 92/* 93 * Associated with page of user-allocatable memory is a 94 * page structure. 95 */ 96 97static struct vm_page **vm_page_buckets; /* Array of buckets */ 98static int vm_page_bucket_count; /* How big is array? */ 99static int vm_page_hash_mask; /* Mask for hash function */ 100static volatile int vm_page_bucket_generation; 101 102struct vpgqueues vm_page_queues[PQ_COUNT]; 103 104static void 105vm_page_queue_init(void) 106{ 107 int i; 108 109 for (i = 0; i < PQ_L2_SIZE; i++) { 110 vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count; 111 } 112 for (i = 0; i < PQ_L2_SIZE; i++) { 113 vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count; 114 } 115 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 116 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 117 118 for (i = 0; i < PQ_COUNT; i++) { 119 TAILQ_INIT(&vm_page_queues[i].pl); 120 } 121} 122 123vm_page_t vm_page_array = 0; 124int vm_page_array_size = 0; 125long first_page = 0; 126int vm_page_zero_count = 0; 127 128static vm_page_t _vm_page_list_find(int basequeue, int index); 129 130/* 131 * vm_set_page_size: 132 * 133 * Sets the page size, perhaps based upon the memory 134 * size. Must be called before any use of page-size 135 * dependent functions. 136 */ 137void 138vm_set_page_size(void) 139{ 140 if (cnt.v_page_size == 0) 141 cnt.v_page_size = PAGE_SIZE; 142 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 143 panic("vm_set_page_size: page size not a power of two"); 144} 145 146/* 147 * vm_add_new_page: 148 * 149 * Add a new page to the freelist for use by the system. 150 * Must be called at splhigh(). 151 */ 152vm_page_t 153vm_add_new_page(vm_offset_t pa) 154{ 155 vm_page_t m; 156 157 GIANT_REQUIRED; 158 159 ++cnt.v_page_count; 160 ++cnt.v_free_count; 161 m = PHYS_TO_VM_PAGE(pa); 162 m->phys_addr = pa; 163 m->flags = 0; 164 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 165 m->queue = m->pc + PQ_FREE; 166 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 167 vm_page_queues[m->queue].lcnt++; 168 return (m); 169} 170 171/* 172 * vm_page_startup: 173 * 174 * Initializes the resident memory module. 175 * 176 * Allocates memory for the page cells, and 177 * for the object/offset-to-page hash table headers. 178 * Each page cell is initialized and placed on the free list. 179 */ 180 181vm_offset_t 182vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 183{ 184 vm_offset_t mapped; 185 struct vm_page **bucket; 186 vm_size_t npages, page_range; 187 vm_offset_t new_end; 188 int i; 189 vm_offset_t pa; 190 int nblocks; 191 vm_offset_t last_pa; 192 193 /* the biggest memory array is the second group of pages */ 194 vm_offset_t end; 195 vm_offset_t biggestone, biggestsize; 196 197 vm_offset_t total; 198 199 total = 0; 200 biggestsize = 0; 201 biggestone = 0; 202 nblocks = 0; 203 vaddr = round_page(vaddr); 204 205 for (i = 0; phys_avail[i + 1]; i += 2) { 206 phys_avail[i] = round_page(phys_avail[i]); 207 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 208 } 209 210 for (i = 0; phys_avail[i + 1]; i += 2) { 211 int size = phys_avail[i + 1] - phys_avail[i]; 212 213 if (size > biggestsize) { 214 biggestone = i; 215 biggestsize = size; 216 } 217 ++nblocks; 218 total += size; 219 } 220 221 end = phys_avail[biggestone+1]; 222 223 /* 224 * Initialize the queue headers for the free queue, the active queue 225 * and the inactive queue. 226 */ 227 228 vm_page_queue_init(); 229 230 /* 231 * Allocate (and initialize) the hash table buckets. 232 * 233 * The number of buckets MUST BE a power of 2, and the actual value is 234 * the next power of 2 greater than the number of physical pages in 235 * the system. 236 * 237 * We make the hash table approximately 2x the number of pages to 238 * reduce the chain length. This is about the same size using the 239 * singly-linked list as the 1x hash table we were using before 240 * using TAILQ but the chain length will be smaller. 241 * 242 * Note: This computation can be tweaked if desired. 243 */ 244 if (vm_page_bucket_count == 0) { 245 vm_page_bucket_count = 1; 246 while (vm_page_bucket_count < atop(total)) 247 vm_page_bucket_count <<= 1; 248 } 249 vm_page_bucket_count <<= 1; 250 vm_page_hash_mask = vm_page_bucket_count - 1; 251 252 /* 253 * Validate these addresses. 254 */ 255 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 256 new_end = trunc_page(new_end); 257 mapped = pmap_map(&vaddr, new_end, end, 258 VM_PROT_READ | VM_PROT_WRITE); 259 bzero((caddr_t) mapped, end - new_end); 260 261 vm_page_buckets = (struct vm_page **)mapped; 262 bucket = vm_page_buckets; 263 for (i = 0; i < vm_page_bucket_count; i++) { 264 *bucket = NULL; 265 bucket++; 266 } 267 268 /* 269 * Compute the number of pages of memory that will be available for 270 * use (taking into account the overhead of a page structure per 271 * page). 272 */ 273 274 first_page = phys_avail[0] / PAGE_SIZE; 275 276 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 277 npages = (total - (page_range * sizeof(struct vm_page)) - 278 (end - new_end)) / PAGE_SIZE; 279 280 end = new_end; 281 282 /* 283 * Initialize the mem entry structures now, and put them in the free 284 * queue. 285 */ 286 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 287 mapped = pmap_map(&vaddr, new_end, end, 288 VM_PROT_READ | VM_PROT_WRITE); 289 vm_page_array = (vm_page_t) mapped; 290 291 /* 292 * Clear all of the page structures 293 */ 294 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 295 vm_page_array_size = page_range; 296 297 /* 298 * Construct the free queue(s) in descending order (by physical 299 * address) so that the first 16MB of physical memory is allocated 300 * last rather than first. On large-memory machines, this avoids 301 * the exhaustion of low physical memory before isa_dmainit has run. 302 */ 303 cnt.v_page_count = 0; 304 cnt.v_free_count = 0; 305 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 306 pa = phys_avail[i]; 307 if (i == biggestone) 308 last_pa = new_end; 309 else 310 last_pa = phys_avail[i + 1]; 311 while (pa < last_pa && npages-- > 0) { 312 vm_add_new_page(pa); 313 pa += PAGE_SIZE; 314 } 315 } 316 return (vaddr); 317} 318 319/* 320 * vm_page_hash: 321 * 322 * Distributes the object/offset key pair among hash buckets. 323 * 324 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 325 * This routine may not block. 326 * 327 * We try to randomize the hash based on the object to spread the pages 328 * out in the hash table without it costing us too much. 329 */ 330static __inline int 331vm_page_hash(vm_object_t object, vm_pindex_t pindex) 332{ 333 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 334 335 return(i & vm_page_hash_mask); 336} 337 338void 339vm_page_flag_set(vm_page_t m, unsigned short bits) 340{ 341 GIANT_REQUIRED; 342 atomic_set_short(&(m)->flags, bits); 343 /* m->flags |= bits; */ 344} 345 346void 347vm_page_flag_clear(vm_page_t m, unsigned short bits) 348{ 349 GIANT_REQUIRED; 350 atomic_clear_short(&(m)->flags, bits); 351 /* m->flags &= ~bits; */ 352} 353 354void 355vm_page_busy(vm_page_t m) 356{ 357 KASSERT((m->flags & PG_BUSY) == 0, 358 ("vm_page_busy: page already busy!!!")); 359 vm_page_flag_set(m, PG_BUSY); 360} 361 362/* 363 * vm_page_flash: 364 * 365 * wakeup anyone waiting for the page. 366 */ 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 */ 384 385void 386vm_page_wakeup(vm_page_t m) 387{ 388 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 389 vm_page_flag_clear(m, PG_BUSY); 390 vm_page_flash(m); 391} 392 393/* 394 * 395 * 396 */ 397 398void 399vm_page_io_start(vm_page_t m) 400{ 401 GIANT_REQUIRED; 402 atomic_add_char(&(m)->busy, 1); 403} 404 405void 406vm_page_io_finish(vm_page_t m) 407{ 408 GIANT_REQUIRED; 409 atomic_subtract_char(&(m)->busy, 1); 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} 434 435/* 436 * vm_page_protect: 437 * 438 * Reduce the protection of a page. This routine never raises the 439 * protection and therefore can be safely called if the page is already 440 * at VM_PROT_NONE (it will be a NOP effectively ). 441 */ 442 443void 444vm_page_protect(vm_page_t mem, int prot) 445{ 446 if (prot == VM_PROT_NONE) { 447 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) { 448 pmap_page_protect(mem, VM_PROT_NONE); 449 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED); 450 } 451 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 452 pmap_page_protect(mem, VM_PROT_READ); 453 vm_page_flag_clear(mem, PG_WRITEABLE); 454 } 455} 456/* 457 * vm_page_zero_fill: 458 * 459 * Zero-fill the specified page. 460 * Written as a standard pagein routine, to 461 * be used by the zero-fill object. 462 */ 463boolean_t 464vm_page_zero_fill(vm_page_t m) 465{ 466 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 467 return (TRUE); 468} 469 470/* 471 * vm_page_copy: 472 * 473 * Copy one page to another 474 */ 475void 476vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 477{ 478 pmap_copy_page(VM_PAGE_TO_PHYS(src_m), VM_PAGE_TO_PHYS(dest_m)); 479 dest_m->valid = VM_PAGE_BITS_ALL; 480} 481 482/* 483 * vm_page_free: 484 * 485 * Free a page 486 * 487 * The clearing of PG_ZERO is a temporary safety until the code can be 488 * reviewed to determine that PG_ZERO is being properly cleared on 489 * write faults or maps. PG_ZERO was previously cleared in 490 * vm_page_alloc(). 491 */ 492void 493vm_page_free(vm_page_t m) 494{ 495 vm_page_flag_clear(m, PG_ZERO); 496 vm_page_free_toq(m); 497} 498 499/* 500 * vm_page_free_zero: 501 * 502 * Free a page to the zerod-pages queue 503 */ 504void 505vm_page_free_zero(vm_page_t m) 506{ 507 vm_page_flag_set(m, PG_ZERO); 508 vm_page_free_toq(m); 509} 510 511/* 512 * vm_page_sleep_busy: 513 * 514 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 515 * m->busy is zero. Returns TRUE if it had to sleep ( including if 516 * it almost had to sleep and made temporary spl*() mods), FALSE 517 * otherwise. 518 * 519 * This routine assumes that interrupts can only remove the busy 520 * status from a page, not set the busy status or change it from 521 * PG_BUSY to m->busy or vise versa (which would create a timing 522 * window). 523 */ 524 525int 526vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 527{ 528 GIANT_REQUIRED; 529 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 530 int s = splvm(); 531 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 532 /* 533 * Page is busy. Wait and retry. 534 */ 535 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 536 tsleep(m, PVM, msg, 0); 537 } 538 splx(s); 539 return(TRUE); 540 /* not reached */ 541 } 542 return(FALSE); 543} 544/* 545 * vm_page_dirty: 546 * 547 * make page all dirty 548 */ 549 550void 551vm_page_dirty(vm_page_t m) 552{ 553 KASSERT(m->queue - m->pc != PQ_CACHE, 554 ("vm_page_dirty: page in cache!")); 555 m->dirty = VM_PAGE_BITS_ALL; 556} 557 558/* 559 * vm_page_undirty: 560 * 561 * Set page to not be dirty. Note: does not clear pmap modify bits 562 */ 563 564void 565vm_page_undirty(vm_page_t m) 566{ 567 m->dirty = 0; 568} 569 570/* 571 * vm_page_insert: [ internal use only ] 572 * 573 * Inserts the given mem entry into the object and object list. 574 * 575 * The pagetables are not updated but will presumably fault the page 576 * in if necessary, or if a kernel page the caller will at some point 577 * enter the page into the kernel's pmap. We are not allowed to block 578 * here so we *can't* do this anyway. 579 * 580 * The object and page must be locked, and must be splhigh. 581 * This routine may not block. 582 */ 583 584void 585vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 586{ 587 struct vm_page **bucket; 588 589 GIANT_REQUIRED; 590 591 if (m->object != NULL) 592 panic("vm_page_insert: already inserted"); 593 594 /* 595 * Record the object/offset pair in this page 596 */ 597 598 m->object = object; 599 m->pindex = pindex; 600 601 /* 602 * Insert it into the object_object/offset hash table 603 */ 604 605 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 606 m->hnext = *bucket; 607 *bucket = m; 608 vm_page_bucket_generation++; 609 610 /* 611 * Now link into the object's list of backed pages. 612 */ 613 614 TAILQ_INSERT_TAIL(&object->memq, m, listq); 615 object->generation++; 616 617 /* 618 * show that the object has one more resident page. 619 */ 620 621 object->resident_page_count++; 622 623 /* 624 * Since we are inserting a new and possibly dirty page, 625 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 626 */ 627 if (m->flags & PG_WRITEABLE) 628 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 629} 630 631/* 632 * vm_page_remove: 633 * NOTE: used by device pager as well -wfj 634 * 635 * Removes the given mem entry from the object/offset-page 636 * table and the object page list, but do not invalidate/terminate 637 * the backing store. 638 * 639 * The object and page must be locked, and at splhigh. 640 * The underlying pmap entry (if any) is NOT removed here. 641 * This routine may not block. 642 */ 643 644void 645vm_page_remove(vm_page_t m) 646{ 647 vm_object_t object; 648 649 GIANT_REQUIRED; 650 651 if (m->object == NULL) 652 return; 653 654 if ((m->flags & PG_BUSY) == 0) { 655 panic("vm_page_remove: page not busy"); 656 } 657 658 /* 659 * Basically destroy the page. 660 */ 661 662 vm_page_wakeup(m); 663 664 object = m->object; 665 666 /* 667 * Remove from the object_object/offset hash table. The object 668 * must be on the hash queue, we will panic if it isn't 669 * 670 * Note: we must NULL-out m->hnext to prevent loops in detached 671 * buffers with vm_page_lookup(). 672 */ 673 674 { 675 struct vm_page **bucket; 676 677 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 678 while (*bucket != m) { 679 if (*bucket == NULL) 680 panic("vm_page_remove(): page not found in hash"); 681 bucket = &(*bucket)->hnext; 682 } 683 *bucket = m->hnext; 684 m->hnext = NULL; 685 vm_page_bucket_generation++; 686 } 687 688 /* 689 * Now remove from the object's list of backed pages. 690 */ 691 692 TAILQ_REMOVE(&object->memq, m, listq); 693 694 /* 695 * And show that the object has one fewer resident page. 696 */ 697 698 object->resident_page_count--; 699 object->generation++; 700 701 m->object = NULL; 702} 703 704/* 705 * vm_page_lookup: 706 * 707 * Returns the page associated with the object/offset 708 * pair specified; if none is found, NULL is returned. 709 * 710 * NOTE: the code below does not lock. It will operate properly if 711 * an interrupt makes a change, but the generation algorithm will not 712 * operate properly in an SMP environment where both cpu's are able to run 713 * kernel code simultaneously. 714 * 715 * The object must be locked. No side effects. 716 * This routine may not block. 717 * This is a critical path routine 718 */ 719 720vm_page_t 721vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 722{ 723 vm_page_t m; 724 struct vm_page **bucket; 725 int generation; 726 727 /* 728 * Search the hash table for this object/offset pair 729 */ 730 731retry: 732 generation = vm_page_bucket_generation; 733 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 734 for (m = *bucket; m != NULL; m = m->hnext) { 735 if ((m->object == object) && (m->pindex == pindex)) { 736 if (vm_page_bucket_generation != generation) 737 goto retry; 738 return (m); 739 } 740 } 741 if (vm_page_bucket_generation != generation) 742 goto retry; 743 return (NULL); 744} 745 746/* 747 * vm_page_rename: 748 * 749 * Move the given memory entry from its 750 * current object to the specified target object/offset. 751 * 752 * The object must be locked. 753 * This routine may not block. 754 * 755 * Note: this routine will raise itself to splvm(), the caller need not. 756 * 757 * Note: swap associated with the page must be invalidated by the move. We 758 * have to do this for several reasons: (1) we aren't freeing the 759 * page, (2) we are dirtying the page, (3) the VM system is probably 760 * moving the page from object A to B, and will then later move 761 * the backing store from A to B and we can't have a conflict. 762 * 763 * Note: we *always* dirty the page. It is necessary both for the 764 * fact that we moved it, and because we may be invalidating 765 * swap. If the page is on the cache, we have to deactivate it 766 * or vm_page_dirty() will panic. Dirty pages are not allowed 767 * on the cache. 768 */ 769 770void 771vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 772{ 773 int s; 774 775 s = splvm(); 776 vm_page_remove(m); 777 vm_page_insert(m, new_object, new_pindex); 778 if (m->queue - m->pc == PQ_CACHE) 779 vm_page_deactivate(m); 780 vm_page_dirty(m); 781 splx(s); 782} 783 784/* 785 * vm_page_unqueue_nowakeup: 786 * 787 * vm_page_unqueue() without any wakeup 788 * 789 * This routine must be called at splhigh(). 790 * This routine may not block. 791 */ 792 793void 794vm_page_unqueue_nowakeup(vm_page_t m) 795{ 796 int queue = m->queue; 797 struct vpgqueues *pq; 798 if (queue != PQ_NONE) { 799 pq = &vm_page_queues[queue]; 800 m->queue = PQ_NONE; 801 TAILQ_REMOVE(&pq->pl, m, pageq); 802 (*pq->cnt)--; 803 pq->lcnt--; 804 } 805} 806 807/* 808 * vm_page_unqueue: 809 * 810 * Remove a page from its queue. 811 * 812 * This routine must be called at splhigh(). 813 * This routine may not block. 814 */ 815 816void 817vm_page_unqueue(vm_page_t m) 818{ 819 int queue = m->queue; 820 struct vpgqueues *pq; 821 822 GIANT_REQUIRED; 823 if (queue != PQ_NONE) { 824 m->queue = PQ_NONE; 825 pq = &vm_page_queues[queue]; 826 TAILQ_REMOVE(&pq->pl, m, pageq); 827 (*pq->cnt)--; 828 pq->lcnt--; 829 if ((queue - m->pc) == PQ_CACHE) { 830 if (vm_paging_needed()) 831 pagedaemon_wakeup(); 832 } 833 } 834} 835 836vm_page_t 837vm_page_list_find(int basequeue, int index, boolean_t prefer_zero) 838{ 839 vm_page_t m; 840 841 GIANT_REQUIRED; 842 843#if PQ_L2_SIZE > 1 844 if (prefer_zero) { 845 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl, pglist); 846 } else { 847 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl); 848 } 849 if (m == NULL) { 850 m = _vm_page_list_find(basequeue, index); 851 } 852#else 853 if (prefer_zero) { 854 m = TAILQ_LAST(&vm_page_queues[basequeue].pl, pglist); 855 } else { 856 m = TAILQ_FIRST(&vm_page_queues[basequeue].pl); 857 } 858#endif 859 return(m); 860} 861 862 863#if PQ_L2_SIZE > 1 864 865/* 866 * vm_page_list_find: 867 * 868 * Find a page on the specified queue with color optimization. 869 * 870 * The page coloring optimization attempts to locate a page 871 * that does not overload other nearby pages in the object in 872 * the cpu's L1 or L2 caches. We need this optimization because 873 * cpu caches tend to be physical caches, while object spaces tend 874 * to be virtual. 875 * 876 * This routine must be called at splvm(). 877 * This routine may not block. 878 * 879 * This routine may only be called from the vm_page_list_find() macro 880 * in vm_page.h 881 */ 882static vm_page_t 883_vm_page_list_find(int basequeue, int index) 884{ 885 int i; 886 vm_page_t m = NULL; 887 struct vpgqueues *pq; 888 889 GIANT_REQUIRED; 890 pq = &vm_page_queues[basequeue]; 891 892 /* 893 * Note that for the first loop, index+i and index-i wind up at the 894 * same place. Even though this is not totally optimal, we've already 895 * blown it by missing the cache case so we do not care. 896 */ 897 898 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 899 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 900 break; 901 902 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 903 break; 904 } 905 return(m); 906} 907 908#endif 909 910/* 911 * vm_page_select_cache: 912 * 913 * Find a page on the cache queue with color optimization. As pages 914 * might be found, but not applicable, they are deactivated. This 915 * keeps us from using potentially busy cached pages. 916 * 917 * This routine must be called at splvm(). 918 * This routine may not block. 919 */ 920vm_page_t 921vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 922{ 923 vm_page_t m; 924 925 GIANT_REQUIRED; 926 while (TRUE) { 927 m = vm_page_list_find( 928 PQ_CACHE, 929 (pindex + object->pg_color) & PQ_L2_MASK, 930 FALSE 931 ); 932 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 933 m->hold_count || m->wire_count)) { 934 vm_page_deactivate(m); 935 continue; 936 } 937 return m; 938 } 939} 940 941/* 942 * vm_page_select_free: 943 * 944 * Find a free or zero page, with specified preference. 945 * 946 * This routine must be called at splvm(). 947 * This routine may not block. 948 */ 949 950static __inline vm_page_t 951vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 952{ 953 vm_page_t m; 954 955 m = vm_page_list_find( 956 PQ_FREE, 957 (pindex + object->pg_color) & PQ_L2_MASK, 958 prefer_zero 959 ); 960 return(m); 961} 962 963/* 964 * vm_page_alloc: 965 * 966 * Allocate and return a memory cell associated 967 * with this VM object/offset pair. 968 * 969 * page_req classes: 970 * VM_ALLOC_NORMAL normal process request 971 * VM_ALLOC_SYSTEM system *really* needs a page 972 * VM_ALLOC_INTERRUPT interrupt time request 973 * VM_ALLOC_ZERO zero page 974 * 975 * This routine may not block. 976 * 977 * Additional special handling is required when called from an 978 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 979 * the page cache in this case. 980 */ 981 982vm_page_t 983vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 984{ 985 vm_page_t m = NULL; 986 int s; 987 988 GIANT_REQUIRED; 989 990 KASSERT(!vm_page_lookup(object, pindex), 991 ("vm_page_alloc: page already allocated")); 992 993 /* 994 * The pager is allowed to eat deeper into the free page list. 995 */ 996 997 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 998 page_req = VM_ALLOC_SYSTEM; 999 }; 1000 1001 s = splvm(); 1002 1003loop: 1004 if (cnt.v_free_count > cnt.v_free_reserved) { 1005 /* 1006 * Allocate from the free queue if there are plenty of pages 1007 * in it. 1008 */ 1009 if (page_req == VM_ALLOC_ZERO) 1010 m = vm_page_select_free(object, pindex, TRUE); 1011 else 1012 m = vm_page_select_free(object, pindex, FALSE); 1013 } else if ( 1014 (page_req == VM_ALLOC_SYSTEM && 1015 cnt.v_cache_count == 0 && 1016 cnt.v_free_count > cnt.v_interrupt_free_min) || 1017 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 1018 ) { 1019 /* 1020 * Interrupt or system, dig deeper into the free list. 1021 */ 1022 m = vm_page_select_free(object, pindex, FALSE); 1023 } else if (page_req != VM_ALLOC_INTERRUPT) { 1024 /* 1025 * Allocatable from cache (non-interrupt only). On success, 1026 * we must free the page and try again, thus ensuring that 1027 * cnt.v_*_free_min counters are replenished. 1028 */ 1029 m = vm_page_select_cache(object, pindex); 1030 if (m == NULL) { 1031 splx(s); 1032#if defined(DIAGNOSTIC) 1033 if (cnt.v_cache_count > 0) 1034 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 1035#endif 1036 vm_pageout_deficit++; 1037 pagedaemon_wakeup(); 1038 return (NULL); 1039 } 1040 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 1041 vm_page_busy(m); 1042 vm_page_protect(m, VM_PROT_NONE); 1043 vm_page_free(m); 1044 goto loop; 1045 } else { 1046 /* 1047 * Not allocatable from cache from interrupt, give up. 1048 */ 1049 splx(s); 1050 vm_pageout_deficit++; 1051 pagedaemon_wakeup(); 1052 return (NULL); 1053 } 1054 1055 /* 1056 * At this point we had better have found a good page. 1057 */ 1058 1059 KASSERT( 1060 m != NULL, 1061 ("vm_page_alloc(): missing page on free queue\n") 1062 ); 1063 1064 /* 1065 * Remove from free queue 1066 */ 1067 1068 vm_page_unqueue_nowakeup(m); 1069 1070 /* 1071 * Initialize structure. Only the PG_ZERO flag is inherited. 1072 */ 1073 1074 if (m->flags & PG_ZERO) { 1075 vm_page_zero_count--; 1076 m->flags = PG_ZERO | PG_BUSY; 1077 } else { 1078 m->flags = PG_BUSY; 1079 } 1080 m->wire_count = 0; 1081 m->hold_count = 0; 1082 m->act_count = 0; 1083 m->busy = 0; 1084 m->valid = 0; 1085 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 1086 1087 /* 1088 * vm_page_insert() is safe prior to the splx(). Note also that 1089 * inserting a page here does not insert it into the pmap (which 1090 * could cause us to block allocating memory). We cannot block 1091 * anywhere. 1092 */ 1093 1094 vm_page_insert(m, object, pindex); 1095 1096 /* 1097 * Don't wakeup too often - wakeup the pageout daemon when 1098 * we would be nearly out of memory. 1099 */ 1100 if (vm_paging_needed()) 1101 pagedaemon_wakeup(); 1102 1103 splx(s); 1104 1105 return (m); 1106} 1107 1108/* 1109 * vm_wait: (also see VM_WAIT macro) 1110 * 1111 * Block until free pages are available for allocation 1112 */ 1113 1114void 1115vm_wait(void) 1116{ 1117 int s; 1118 1119 s = splvm(); 1120 if (curproc == pageproc) { 1121 vm_pageout_pages_needed = 1; 1122 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 1123 } else { 1124 if (!vm_pages_needed) { 1125 vm_pages_needed = 1; 1126 wakeup(&vm_pages_needed); 1127 } 1128 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 1129 } 1130 splx(s); 1131} 1132 1133/* 1134 * vm_await: (also see VM_AWAIT macro) 1135 * 1136 * asleep on an event that will signal when free pages are available 1137 * for allocation. 1138 */ 1139 1140void 1141vm_await(void) 1142{ 1143 int s; 1144 1145 s = splvm(); 1146 if (curproc == pageproc) { 1147 vm_pageout_pages_needed = 1; 1148 asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0); 1149 } else { 1150 if (!vm_pages_needed) { 1151 vm_pages_needed++; 1152 wakeup(&vm_pages_needed); 1153 } 1154 asleep(&cnt.v_free_count, PVM, "vmwait", 0); 1155 } 1156 splx(s); 1157} 1158 1159/* 1160 * vm_page_activate: 1161 * 1162 * Put the specified page on the active list (if appropriate). 1163 * Ensure that act_count is at least ACT_INIT but do not otherwise 1164 * mess with it. 1165 * 1166 * The page queues must be locked. 1167 * This routine may not block. 1168 */ 1169void 1170vm_page_activate(vm_page_t m) 1171{ 1172 int s; 1173 1174 GIANT_REQUIRED; 1175 s = splvm(); 1176 1177 if (m->queue != PQ_ACTIVE) { 1178 if ((m->queue - m->pc) == PQ_CACHE) 1179 cnt.v_reactivated++; 1180 1181 vm_page_unqueue(m); 1182 1183 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1184 m->queue = PQ_ACTIVE; 1185 vm_page_queues[PQ_ACTIVE].lcnt++; 1186 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1187 if (m->act_count < ACT_INIT) 1188 m->act_count = ACT_INIT; 1189 cnt.v_active_count++; 1190 } 1191 } else { 1192 if (m->act_count < ACT_INIT) 1193 m->act_count = ACT_INIT; 1194 } 1195 1196 splx(s); 1197} 1198 1199/* 1200 * vm_page_free_wakeup: 1201 * 1202 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1203 * routine is called when a page has been added to the cache or free 1204 * queues. 1205 * 1206 * This routine may not block. 1207 * This routine must be called at splvm() 1208 */ 1209static __inline void 1210vm_page_free_wakeup(void) 1211{ 1212 /* 1213 * if pageout daemon needs pages, then tell it that there are 1214 * some free. 1215 */ 1216 if (vm_pageout_pages_needed && 1217 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1218 wakeup(&vm_pageout_pages_needed); 1219 vm_pageout_pages_needed = 0; 1220 } 1221 /* 1222 * wakeup processes that are waiting on memory if we hit a 1223 * high water mark. And wakeup scheduler process if we have 1224 * lots of memory. this process will swapin processes. 1225 */ 1226 if (vm_pages_needed && !vm_page_count_min()) { 1227 vm_pages_needed = 0; 1228 wakeup(&cnt.v_free_count); 1229 } 1230} 1231 1232/* 1233 * vm_page_free_toq: 1234 * 1235 * Returns the given page to the PQ_FREE list, 1236 * disassociating it with any VM object. 1237 * 1238 * Object and page must be locked prior to entry. 1239 * This routine may not block. 1240 */ 1241 1242void 1243vm_page_free_toq(vm_page_t m) 1244{ 1245 int s; 1246 struct vpgqueues *pq; 1247 vm_object_t object = m->object; 1248 1249 GIANT_REQUIRED; 1250 s = splvm(); 1251 cnt.v_tfree++; 1252 1253 if (m->busy || ((m->queue - m->pc) == PQ_FREE) || 1254 (m->hold_count != 0)) { 1255 printf( 1256 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1257 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1258 m->hold_count); 1259 if ((m->queue - m->pc) == PQ_FREE) 1260 panic("vm_page_free: freeing free page"); 1261 else 1262 panic("vm_page_free: freeing busy page"); 1263 } 1264 1265 /* 1266 * unqueue, then remove page. Note that we cannot destroy 1267 * the page here because we do not want to call the pager's 1268 * callback routine until after we've put the page on the 1269 * appropriate free queue. 1270 */ 1271 1272 vm_page_unqueue_nowakeup(m); 1273 vm_page_remove(m); 1274 1275 /* 1276 * If fictitious remove object association and 1277 * return, otherwise delay object association removal. 1278 */ 1279 1280 if ((m->flags & PG_FICTITIOUS) != 0) { 1281 splx(s); 1282 return; 1283 } 1284 1285 m->valid = 0; 1286 vm_page_undirty(m); 1287 1288 if (m->wire_count != 0) { 1289 if (m->wire_count > 1) { 1290 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1291 m->wire_count, (long)m->pindex); 1292 } 1293 panic("vm_page_free: freeing wired page\n"); 1294 } 1295 1296 /* 1297 * If we've exhausted the object's resident pages we want to free 1298 * it up. 1299 */ 1300 1301 if (object && 1302 (object->type == OBJT_VNODE) && 1303 ((object->flags & OBJ_DEAD) == 0) 1304 ) { 1305 struct vnode *vp = (struct vnode *)object->handle; 1306 1307 if (vp && VSHOULDFREE(vp)) 1308 vfree(vp); 1309 } 1310 1311 /* 1312 * Clear the UNMANAGED flag when freeing an unmanaged page. 1313 */ 1314 1315 if (m->flags & PG_UNMANAGED) { 1316 m->flags &= ~PG_UNMANAGED; 1317 } else { 1318#ifdef __alpha__ 1319 pmap_page_is_free(m); 1320#endif 1321 } 1322 1323 m->queue = PQ_FREE + m->pc; 1324 pq = &vm_page_queues[m->queue]; 1325 pq->lcnt++; 1326 ++(*pq->cnt); 1327 1328 /* 1329 * Put zero'd pages on the end ( where we look for zero'd pages 1330 * first ) and non-zerod pages at the head. 1331 */ 1332 1333 if (m->flags & PG_ZERO) { 1334 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1335 ++vm_page_zero_count; 1336 } else { 1337 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1338 } 1339 1340 vm_page_free_wakeup(); 1341 1342 splx(s); 1343} 1344 1345/* 1346 * vm_page_unmanage: 1347 * 1348 * Prevent PV management from being done on the page. The page is 1349 * removed from the paging queues as if it were wired, and as a 1350 * consequence of no longer being managed the pageout daemon will not 1351 * touch it (since there is no way to locate the pte mappings for the 1352 * page). madvise() calls that mess with the pmap will also no longer 1353 * operate on the page. 1354 * 1355 * Beyond that the page is still reasonably 'normal'. Freeing the page 1356 * will clear the flag. 1357 * 1358 * This routine is used by OBJT_PHYS objects - objects using unswappable 1359 * physical memory as backing store rather then swap-backed memory and 1360 * will eventually be extended to support 4MB unmanaged physical 1361 * mappings. 1362 */ 1363 1364void 1365vm_page_unmanage(vm_page_t m) 1366{ 1367 int s; 1368 1369 s = splvm(); 1370 if ((m->flags & PG_UNMANAGED) == 0) { 1371 if (m->wire_count == 0) 1372 vm_page_unqueue(m); 1373 } 1374 vm_page_flag_set(m, PG_UNMANAGED); 1375 splx(s); 1376} 1377 1378/* 1379 * vm_page_wire: 1380 * 1381 * Mark this page as wired down by yet 1382 * another map, removing it from paging queues 1383 * as necessary. 1384 * 1385 * The page queues must be locked. 1386 * This routine may not block. 1387 */ 1388void 1389vm_page_wire(vm_page_t m) 1390{ 1391 int s; 1392 1393 /* 1394 * Only bump the wire statistics if the page is not already wired, 1395 * and only unqueue the page if it is on some queue (if it is unmanaged 1396 * it is already off the queues). 1397 */ 1398 s = splvm(); 1399 if (m->wire_count == 0) { 1400 if ((m->flags & PG_UNMANAGED) == 0) 1401 vm_page_unqueue(m); 1402 cnt.v_wire_count++; 1403 } 1404 m->wire_count++; 1405 splx(s); 1406 vm_page_flag_set(m, PG_MAPPED); 1407} 1408 1409/* 1410 * vm_page_unwire: 1411 * 1412 * Release one wiring of this page, potentially 1413 * enabling it to be paged again. 1414 * 1415 * Many pages placed on the inactive queue should actually go 1416 * into the cache, but it is difficult to figure out which. What 1417 * we do instead, if the inactive target is well met, is to put 1418 * clean pages at the head of the inactive queue instead of the tail. 1419 * This will cause them to be moved to the cache more quickly and 1420 * if not actively re-referenced, freed more quickly. If we just 1421 * stick these pages at the end of the inactive queue, heavy filesystem 1422 * meta-data accesses can cause an unnecessary paging load on memory bound 1423 * processes. This optimization causes one-time-use metadata to be 1424 * reused more quickly. 1425 * 1426 * BUT, if we are in a low-memory situation we have no choice but to 1427 * put clean pages on the cache queue. 1428 * 1429 * A number of routines use vm_page_unwire() to guarantee that the page 1430 * will go into either the inactive or active queues, and will NEVER 1431 * be placed in the cache - for example, just after dirtying a page. 1432 * dirty pages in the cache are not allowed. 1433 * 1434 * The page queues must be locked. 1435 * This routine may not block. 1436 */ 1437void 1438vm_page_unwire(vm_page_t m, int activate) 1439{ 1440 int s; 1441 1442 s = splvm(); 1443 1444 if (m->wire_count > 0) { 1445 m->wire_count--; 1446 if (m->wire_count == 0) { 1447 cnt.v_wire_count--; 1448 if (m->flags & PG_UNMANAGED) { 1449 ; 1450 } else if (activate) { 1451 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1452 m->queue = PQ_ACTIVE; 1453 vm_page_queues[PQ_ACTIVE].lcnt++; 1454 cnt.v_active_count++; 1455 } else { 1456 vm_page_flag_clear(m, PG_WINATCFLS); 1457 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1458 m->queue = PQ_INACTIVE; 1459 vm_page_queues[PQ_INACTIVE].lcnt++; 1460 cnt.v_inactive_count++; 1461 } 1462 } 1463 } else { 1464 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1465 } 1466 splx(s); 1467} 1468 1469 1470/* 1471 * Move the specified page to the inactive queue. If the page has 1472 * any associated swap, the swap is deallocated. 1473 * 1474 * Normally athead is 0 resulting in LRU operation. athead is set 1475 * to 1 if we want this page to be 'as if it were placed in the cache', 1476 * except without unmapping it from the process address space. 1477 * 1478 * This routine may not block. 1479 */ 1480static __inline void 1481_vm_page_deactivate(vm_page_t m, int athead) 1482{ 1483 int s; 1484 1485 GIANT_REQUIRED; 1486 /* 1487 * Ignore if already inactive. 1488 */ 1489 if (m->queue == PQ_INACTIVE) 1490 return; 1491 1492 s = splvm(); 1493 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1494 if ((m->queue - m->pc) == PQ_CACHE) 1495 cnt.v_reactivated++; 1496 vm_page_flag_clear(m, PG_WINATCFLS); 1497 vm_page_unqueue(m); 1498 if (athead) 1499 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1500 else 1501 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1502 m->queue = PQ_INACTIVE; 1503 vm_page_queues[PQ_INACTIVE].lcnt++; 1504 cnt.v_inactive_count++; 1505 } 1506 splx(s); 1507} 1508 1509void 1510vm_page_deactivate(vm_page_t m) 1511{ 1512 _vm_page_deactivate(m, 0); 1513} 1514 1515/* 1516 * vm_page_try_to_cache: 1517 * 1518 * Returns 0 on failure, 1 on success 1519 */ 1520int 1521vm_page_try_to_cache(vm_page_t m) 1522{ 1523 GIANT_REQUIRED; 1524 1525 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1526 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1527 return(0); 1528 } 1529 vm_page_test_dirty(m); 1530 if (m->dirty) 1531 return(0); 1532 vm_page_cache(m); 1533 return(1); 1534} 1535 1536/* 1537 * vm_page_try_to_free() 1538 * 1539 * Attempt to free the page. If we cannot free it, we do nothing. 1540 * 1 is returned on success, 0 on failure. 1541 */ 1542int 1543vm_page_try_to_free(vm_page_t m) 1544{ 1545 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1546 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1547 return(0); 1548 } 1549 vm_page_test_dirty(m); 1550 if (m->dirty) 1551 return(0); 1552 vm_page_busy(m); 1553 vm_page_protect(m, VM_PROT_NONE); 1554 vm_page_free(m); 1555 return(1); 1556} 1557 1558/* 1559 * vm_page_cache 1560 * 1561 * Put the specified page onto the page cache queue (if appropriate). 1562 * 1563 * This routine may not block. 1564 */ 1565void 1566vm_page_cache(vm_page_t m) 1567{ 1568 int s; 1569 1570 GIANT_REQUIRED; 1571 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1572 printf("vm_page_cache: attempting to cache busy page\n"); 1573 return; 1574 } 1575 if ((m->queue - m->pc) == PQ_CACHE) 1576 return; 1577 1578 /* 1579 * Remove all pmaps and indicate that the page is not 1580 * writeable or mapped. 1581 */ 1582 1583 vm_page_protect(m, VM_PROT_NONE); 1584 if (m->dirty != 0) { 1585 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1586 (long)m->pindex); 1587 } 1588 s = splvm(); 1589 vm_page_unqueue_nowakeup(m); 1590 m->queue = PQ_CACHE + m->pc; 1591 vm_page_queues[m->queue].lcnt++; 1592 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1593 cnt.v_cache_count++; 1594 vm_page_free_wakeup(); 1595 splx(s); 1596} 1597 1598/* 1599 * vm_page_dontneed 1600 * 1601 * Cache, deactivate, or do nothing as appropriate. This routine 1602 * is typically used by madvise() MADV_DONTNEED. 1603 * 1604 * Generally speaking we want to move the page into the cache so 1605 * it gets reused quickly. However, this can result in a silly syndrome 1606 * due to the page recycling too quickly. Small objects will not be 1607 * fully cached. On the otherhand, if we move the page to the inactive 1608 * queue we wind up with a problem whereby very large objects 1609 * unnecessarily blow away our inactive and cache queues. 1610 * 1611 * The solution is to move the pages based on a fixed weighting. We 1612 * either leave them alone, deactivate them, or move them to the cache, 1613 * where moving them to the cache has the highest weighting. 1614 * By forcing some pages into other queues we eventually force the 1615 * system to balance the queues, potentially recovering other unrelated 1616 * space from active. The idea is to not force this to happen too 1617 * often. 1618 */ 1619 1620void 1621vm_page_dontneed(vm_page_t m) 1622{ 1623 static int dnweight; 1624 int dnw; 1625 int head; 1626 1627 GIANT_REQUIRED; 1628 dnw = ++dnweight; 1629 1630 /* 1631 * occassionally leave the page alone 1632 */ 1633 1634 if ((dnw & 0x01F0) == 0 || 1635 m->queue == PQ_INACTIVE || 1636 m->queue - m->pc == PQ_CACHE 1637 ) { 1638 if (m->act_count >= ACT_INIT) 1639 --m->act_count; 1640 return; 1641 } 1642 1643 if (m->dirty == 0) 1644 vm_page_test_dirty(m); 1645 1646 if (m->dirty || (dnw & 0x0070) == 0) { 1647 /* 1648 * Deactivate the page 3 times out of 32. 1649 */ 1650 head = 0; 1651 } else { 1652 /* 1653 * Cache the page 28 times out of every 32. Note that 1654 * the page is deactivated instead of cached, but placed 1655 * at the head of the queue instead of the tail. 1656 */ 1657 head = 1; 1658 } 1659 _vm_page_deactivate(m, head); 1660} 1661 1662/* 1663 * Grab a page, waiting until we are waken up due to the page 1664 * changing state. We keep on waiting, if the page continues 1665 * to be in the object. If the page doesn't exist, allocate it. 1666 * 1667 * This routine may block. 1668 */ 1669vm_page_t 1670vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1671{ 1672 vm_page_t m; 1673 int s, generation; 1674 1675 GIANT_REQUIRED; 1676retrylookup: 1677 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1678 if (m->busy || (m->flags & PG_BUSY)) { 1679 generation = object->generation; 1680 1681 s = splvm(); 1682 while ((object->generation == generation) && 1683 (m->busy || (m->flags & PG_BUSY))) { 1684 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1685 tsleep(m, PVM, "pgrbwt", 0); 1686 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1687 splx(s); 1688 return NULL; 1689 } 1690 } 1691 splx(s); 1692 goto retrylookup; 1693 } else { 1694 vm_page_busy(m); 1695 return m; 1696 } 1697 } 1698 1699 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1700 if (m == NULL) { 1701 VM_WAIT; 1702 if ((allocflags & VM_ALLOC_RETRY) == 0) 1703 return NULL; 1704 goto retrylookup; 1705 } 1706 1707 return m; 1708} 1709 1710/* 1711 * Mapping function for valid bits or for dirty bits in 1712 * a page. May not block. 1713 * 1714 * Inputs are required to range within a page. 1715 */ 1716 1717__inline int 1718vm_page_bits(int base, int size) 1719{ 1720 int first_bit; 1721 int last_bit; 1722 1723 KASSERT( 1724 base + size <= PAGE_SIZE, 1725 ("vm_page_bits: illegal base/size %d/%d", base, size) 1726 ); 1727 1728 if (size == 0) /* handle degenerate case */ 1729 return(0); 1730 1731 first_bit = base >> DEV_BSHIFT; 1732 last_bit = (base + size - 1) >> DEV_BSHIFT; 1733 1734 return ((2 << last_bit) - (1 << first_bit)); 1735} 1736 1737/* 1738 * vm_page_set_validclean: 1739 * 1740 * Sets portions of a page valid and clean. The arguments are expected 1741 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1742 * of any partial chunks touched by the range. The invalid portion of 1743 * such chunks will be zero'd. 1744 * 1745 * This routine may not block. 1746 * 1747 * (base + size) must be less then or equal to PAGE_SIZE. 1748 */ 1749void 1750vm_page_set_validclean(vm_page_t m, int base, int size) 1751{ 1752 int pagebits; 1753 int frag; 1754 int endoff; 1755 1756 GIANT_REQUIRED; 1757 if (size == 0) /* handle degenerate case */ 1758 return; 1759 1760 /* 1761 * If the base is not DEV_BSIZE aligned and the valid 1762 * bit is clear, we have to zero out a portion of the 1763 * first block. 1764 */ 1765 1766 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1767 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1768 ) { 1769 pmap_zero_page_area( 1770 VM_PAGE_TO_PHYS(m), 1771 frag, 1772 base - frag 1773 ); 1774 } 1775 1776 /* 1777 * If the ending offset is not DEV_BSIZE aligned and the 1778 * valid bit is clear, we have to zero out a portion of 1779 * the last block. 1780 */ 1781 1782 endoff = base + size; 1783 1784 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1785 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1786 ) { 1787 pmap_zero_page_area( 1788 VM_PAGE_TO_PHYS(m), 1789 endoff, 1790 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1791 ); 1792 } 1793 1794 /* 1795 * Set valid, clear dirty bits. If validating the entire 1796 * page we can safely clear the pmap modify bit. We also 1797 * use this opportunity to clear the PG_NOSYNC flag. If a process 1798 * takes a write fault on a MAP_NOSYNC memory area the flag will 1799 * be set again. 1800 */ 1801 1802 pagebits = vm_page_bits(base, size); 1803 m->valid |= pagebits; 1804 m->dirty &= ~pagebits; 1805 if (base == 0 && size == PAGE_SIZE) { 1806 pmap_clear_modify(m); 1807 vm_page_flag_clear(m, PG_NOSYNC); 1808 } 1809} 1810 1811#if 0 1812 1813void 1814vm_page_set_dirty(vm_page_t m, int base, int size) 1815{ 1816 m->dirty |= vm_page_bits(base, size); 1817} 1818 1819#endif 1820 1821void 1822vm_page_clear_dirty(vm_page_t m, int base, int size) 1823{ 1824 GIANT_REQUIRED; 1825 m->dirty &= ~vm_page_bits(base, size); 1826} 1827 1828/* 1829 * vm_page_set_invalid: 1830 * 1831 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1832 * valid and dirty bits for the effected areas are cleared. 1833 * 1834 * May not block. 1835 */ 1836void 1837vm_page_set_invalid(vm_page_t m, int base, int size) 1838{ 1839 int bits; 1840 1841 GIANT_REQUIRED; 1842 bits = vm_page_bits(base, size); 1843 m->valid &= ~bits; 1844 m->dirty &= ~bits; 1845 m->object->generation++; 1846} 1847 1848/* 1849 * vm_page_zero_invalid() 1850 * 1851 * The kernel assumes that the invalid portions of a page contain 1852 * garbage, but such pages can be mapped into memory by user code. 1853 * When this occurs, we must zero out the non-valid portions of the 1854 * page so user code sees what it expects. 1855 * 1856 * Pages are most often semi-valid when the end of a file is mapped 1857 * into memory and the file's size is not page aligned. 1858 */ 1859 1860void 1861vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1862{ 1863 int b; 1864 int i; 1865 1866 /* 1867 * Scan the valid bits looking for invalid sections that 1868 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1869 * valid bit may be set ) have already been zerod by 1870 * vm_page_set_validclean(). 1871 */ 1872 1873 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1874 if (i == (PAGE_SIZE / DEV_BSIZE) || 1875 (m->valid & (1 << i)) 1876 ) { 1877 if (i > b) { 1878 pmap_zero_page_area( 1879 VM_PAGE_TO_PHYS(m), 1880 b << DEV_BSHIFT, 1881 (i - b) << DEV_BSHIFT 1882 ); 1883 } 1884 b = i + 1; 1885 } 1886 } 1887 1888 /* 1889 * setvalid is TRUE when we can safely set the zero'd areas 1890 * as being valid. We can do this if there are no cache consistancy 1891 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1892 */ 1893 1894 if (setvalid) 1895 m->valid = VM_PAGE_BITS_ALL; 1896} 1897 1898/* 1899 * vm_page_is_valid: 1900 * 1901 * Is (partial) page valid? Note that the case where size == 0 1902 * will return FALSE in the degenerate case where the page is 1903 * entirely invalid, and TRUE otherwise. 1904 * 1905 * May not block. 1906 */ 1907 1908int 1909vm_page_is_valid(vm_page_t m, int base, int size) 1910{ 1911 int bits = vm_page_bits(base, size); 1912 1913 if (m->valid && ((m->valid & bits) == bits)) 1914 return 1; 1915 else 1916 return 0; 1917} 1918 1919/* 1920 * update dirty bits from pmap/mmu. May not block. 1921 */ 1922 1923void 1924vm_page_test_dirty(vm_page_t m) 1925{ 1926 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1927 vm_page_dirty(m); 1928 } 1929} 1930 1931/* 1932 * This interface is for merging with malloc() someday. 1933 * Even if we never implement compaction so that contiguous allocation 1934 * works after initialization time, malloc()'s data structures are good 1935 * for statistics and for allocations of less than a page. 1936 */ 1937void * 1938contigmalloc1( 1939 unsigned long size, /* should be size_t here and for malloc() */ 1940 struct malloc_type *type, 1941 int flags, 1942 unsigned long low, 1943 unsigned long high, 1944 unsigned long alignment, 1945 unsigned long boundary, 1946 vm_map_t map) 1947{ 1948 int i, s, start; 1949 vm_offset_t addr, phys, tmp_addr; 1950 int pass; 1951 vm_page_t pga = vm_page_array; 1952 1953 size = round_page(size); 1954 if (size == 0) 1955 panic("contigmalloc1: size must not be 0"); 1956 if ((alignment & (alignment - 1)) != 0) 1957 panic("contigmalloc1: alignment must be a power of 2"); 1958 if ((boundary & (boundary - 1)) != 0) 1959 panic("contigmalloc1: boundary must be a power of 2"); 1960 1961 start = 0; 1962 for (pass = 0; pass <= 1; pass++) { 1963 s = splvm(); 1964again: 1965 /* 1966 * Find first page in array that is free, within range, aligned, and 1967 * such that the boundary won't be crossed. 1968 */ 1969 for (i = start; i < cnt.v_page_count; i++) { 1970 int pqtype; 1971 phys = VM_PAGE_TO_PHYS(&pga[i]); 1972 pqtype = pga[i].queue - pga[i].pc; 1973 if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) && 1974 (phys >= low) && (phys < high) && 1975 ((phys & (alignment - 1)) == 0) && 1976 (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0)) 1977 break; 1978 } 1979 1980 /* 1981 * If the above failed or we will exceed the upper bound, fail. 1982 */ 1983 if ((i == cnt.v_page_count) || 1984 ((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) { 1985 vm_page_t m, next; 1986 1987again1: 1988 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 1989 m != NULL; 1990 m = next) { 1991 1992 KASSERT(m->queue == PQ_INACTIVE, 1993 ("contigmalloc1: page %p is not PQ_INACTIVE", m)); 1994 1995 next = TAILQ_NEXT(m, pageq); 1996 if (vm_page_sleep_busy(m, TRUE, "vpctw0")) 1997 goto again1; 1998 vm_page_test_dirty(m); 1999 if (m->dirty) { 2000 if (m->object->type == OBJT_VNODE) { 2001 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc); 2002 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC); 2003 VOP_UNLOCK(m->object->handle, 0, curproc); 2004 goto again1; 2005 } else if (m->object->type == OBJT_SWAP || 2006 m->object->type == OBJT_DEFAULT) { 2007 vm_pageout_flush(&m, 1, 0); 2008 goto again1; 2009 } 2010 } 2011 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0)) 2012 vm_page_cache(m); 2013 } 2014 2015 for (m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 2016 m != NULL; 2017 m = next) { 2018 2019 KASSERT(m->queue == PQ_ACTIVE, 2020 ("contigmalloc1: page %p is not PQ_ACTIVE", m)); 2021 2022 next = TAILQ_NEXT(m, pageq); 2023 if (vm_page_sleep_busy(m, TRUE, "vpctw1")) 2024 goto again1; 2025 vm_page_test_dirty(m); 2026 if (m->dirty) { 2027 if (m->object->type == OBJT_VNODE) { 2028 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc); 2029 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC); 2030 VOP_UNLOCK(m->object->handle, 0, curproc); 2031 goto again1; 2032 } else if (m->object->type == OBJT_SWAP || 2033 m->object->type == OBJT_DEFAULT) { 2034 vm_pageout_flush(&m, 1, 0); 2035 goto again1; 2036 } 2037 } 2038 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0)) 2039 vm_page_cache(m); 2040 } 2041 2042 splx(s); 2043 continue; 2044 } 2045 start = i; 2046 2047 /* 2048 * Check successive pages for contiguous and free. 2049 */ 2050 for (i = start + 1; i < (start + size / PAGE_SIZE); i++) { 2051 int pqtype; 2052 pqtype = pga[i].queue - pga[i].pc; 2053 if ((VM_PAGE_TO_PHYS(&pga[i]) != 2054 (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) || 2055 ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) { 2056 start++; 2057 goto again; 2058 } 2059 } 2060 2061 for (i = start; i < (start + size / PAGE_SIZE); i++) { 2062 int pqtype; 2063 vm_page_t m = &pga[i]; 2064 2065 pqtype = m->queue - m->pc; 2066 if (pqtype == PQ_CACHE) { 2067 vm_page_busy(m); 2068 vm_page_free(m); 2069 } 2070 2071 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq); 2072 vm_page_queues[m->queue].lcnt--; 2073 cnt.v_free_count--; 2074 m->valid = VM_PAGE_BITS_ALL; 2075 m->flags = 0; 2076 KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m)); 2077 m->wire_count = 0; 2078 m->busy = 0; 2079 m->queue = PQ_NONE; 2080 m->object = NULL; 2081 vm_page_wire(m); 2082 } 2083 2084 /* 2085 * We've found a contiguous chunk that meets are requirements. 2086 * Allocate kernel VM, unfree and assign the physical pages to it and 2087 * return kernel VM pointer. 2088 */ 2089 tmp_addr = addr = kmem_alloc_pageable(map, size); 2090 if (addr == 0) { 2091 /* 2092 * XXX We almost never run out of kernel virtual 2093 * space, so we don't make the allocated memory 2094 * above available. 2095 */ 2096 splx(s); 2097 return (NULL); 2098 } 2099 2100 for (i = start; i < (start + size / PAGE_SIZE); i++) { 2101 vm_page_t m = &pga[i]; 2102 vm_page_insert(m, kernel_object, 2103 OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS)); 2104 pmap_kenter(tmp_addr, VM_PAGE_TO_PHYS(m)); 2105 tmp_addr += PAGE_SIZE; 2106 } 2107 2108 splx(s); 2109 return ((void *)addr); 2110 } 2111 return NULL; 2112} 2113 2114void * 2115contigmalloc( 2116 unsigned long size, /* should be size_t here and for malloc() */ 2117 struct malloc_type *type, 2118 int flags, 2119 unsigned long low, 2120 unsigned long high, 2121 unsigned long alignment, 2122 unsigned long boundary) 2123{ 2124 void * ret; 2125 2126 GIANT_REQUIRED; 2127 ret = contigmalloc1(size, type, flags, low, high, alignment, boundary, 2128 kernel_map); 2129 return (ret); 2130 2131} 2132 2133void 2134contigfree(void *addr, unsigned long size, struct malloc_type *type) 2135{ 2136 GIANT_REQUIRED; 2137 kmem_free(kernel_map, (vm_offset_t)addr, size); 2138} 2139 2140vm_offset_t 2141vm_page_alloc_contig( 2142 vm_offset_t size, 2143 vm_offset_t low, 2144 vm_offset_t high, 2145 vm_offset_t alignment) 2146{ 2147 vm_offset_t ret; 2148 2149 GIANT_REQUIRED; 2150 ret = ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high, 2151 alignment, 0ul, kernel_map)); 2152 return (ret); 2153 2154} 2155 2156#include "opt_ddb.h" 2157#ifdef DDB 2158#include <sys/kernel.h> 2159 2160#include <ddb/ddb.h> 2161 2162DB_SHOW_COMMAND(page, vm_page_print_page_info) 2163{ 2164 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 2165 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 2166 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 2167 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 2168 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 2169 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 2170 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 2171 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 2172 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 2173 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 2174} 2175 2176DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 2177{ 2178 int i; 2179 db_printf("PQ_FREE:"); 2180 for (i = 0; i < PQ_L2_SIZE; i++) { 2181 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 2182 } 2183 db_printf("\n"); 2184 2185 db_printf("PQ_CACHE:"); 2186 for (i = 0; i < PQ_L2_SIZE; i++) { 2187 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 2188 } 2189 db_printf("\n"); 2190 2191 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 2192 vm_page_queues[PQ_ACTIVE].lcnt, 2193 vm_page_queues[PQ_INACTIVE].lcnt); 2194} 2195#endif /* DDB */ 2196