vm_page.c revision 73936
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 73936 2001-03-07 05:29:21Z 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 * Resident memory management module. 69 */ 70 71#include <sys/param.h> 72#include <sys/systm.h> 73#include <sys/malloc.h> 74#include <sys/proc.h> 75#include <sys/vmmeter.h> 76#include <sys/vnode.h> 77 78#include <vm/vm.h> 79#include <vm/vm_param.h> 80#include <sys/lock.h> 81#include <vm/vm_kern.h> 82#include <vm/vm_object.h> 83#include <vm/vm_page.h> 84#include <vm/vm_pageout.h> 85#include <vm/vm_pager.h> 86#include <vm/vm_extern.h> 87 88static void vm_page_queue_init __P((void)); 89static vm_page_t vm_page_select_cache __P((vm_object_t, vm_pindex_t)); 90 91/* 92 * Associated with page of user-allocatable memory is a 93 * page structure. 94 */ 95 96static struct vm_page **vm_page_buckets; /* Array of buckets */ 97static int vm_page_bucket_count; /* How big is array? */ 98static int vm_page_hash_mask; /* Mask for hash function */ 99static volatile int vm_page_bucket_generation; 100 101struct vpgqueues vm_page_queues[PQ_COUNT]; 102 103static void 104vm_page_queue_init(void) { 105 int i; 106 107 for(i=0;i<PQ_L2_SIZE;i++) { 108 vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count; 109 } 110 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 111 112 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 113 for(i=0;i<PQ_L2_SIZE;i++) { 114 vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count; 115 } 116 for(i=0;i<PQ_COUNT;i++) { 117 TAILQ_INIT(&vm_page_queues[i].pl); 118 } 119} 120 121vm_page_t vm_page_array = 0; 122int vm_page_array_size = 0; 123long first_page = 0; 124int vm_page_zero_count = 0; 125 126static __inline int vm_page_hash __P((vm_object_t object, vm_pindex_t pindex)); 127static void vm_page_free_wakeup __P((void)); 128 129/* 130 * vm_set_page_size: 131 * 132 * Sets the page size, perhaps based upon the memory 133 * size. Must be called before any use of page-size 134 * dependent functions. 135 */ 136void 137vm_set_page_size() 138{ 139 if (cnt.v_page_size == 0) 140 cnt.v_page_size = PAGE_SIZE; 141 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 142 panic("vm_set_page_size: page size not a power of two"); 143} 144 145/* 146 * vm_add_new_page: 147 * 148 * Add a new page to the freelist for use by the system. 149 * Must be called at splhigh(). 150 */ 151vm_page_t 152vm_add_new_page(pa) 153 vm_offset_t pa; 154{ 155 vm_page_t m; 156 157 ++cnt.v_page_count; 158 ++cnt.v_free_count; 159 m = PHYS_TO_VM_PAGE(pa); 160 m->phys_addr = pa; 161 m->flags = 0; 162 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 163 m->queue = m->pc + PQ_FREE; 164 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 165 vm_page_queues[m->queue].lcnt++; 166 return (m); 167} 168 169/* 170 * vm_page_startup: 171 * 172 * Initializes the resident memory module. 173 * 174 * Allocates memory for the page cells, and 175 * for the object/offset-to-page hash table headers. 176 * Each page cell is initialized and placed on the free list. 177 */ 178 179vm_offset_t 180vm_page_startup(starta, enda, vaddr) 181 register vm_offset_t starta; 182 vm_offset_t enda; 183 vm_offset_t vaddr; 184{ 185 register vm_offset_t mapped; 186 register struct vm_page **bucket; 187 vm_size_t npages, page_range; 188 register vm_offset_t new_end; 189 int i; 190 vm_offset_t pa; 191 int nblocks; 192 vm_offset_t last_pa; 193 194 /* the biggest memory array is the second group of pages */ 195 vm_offset_t end; 196 vm_offset_t biggestone, biggestsize; 197 198 vm_offset_t total; 199 200 total = 0; 201 biggestsize = 0; 202 biggestone = 0; 203 nblocks = 0; 204 vaddr = round_page(vaddr); 205 206 for (i = 0; phys_avail[i + 1]; i += 2) { 207 phys_avail[i] = round_page(phys_avail[i]); 208 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 209 } 210 211 for (i = 0; phys_avail[i + 1]; i += 2) { 212 int size = phys_avail[i + 1] - phys_avail[i]; 213 214 if (size > biggestsize) { 215 biggestone = i; 216 biggestsize = size; 217 } 218 ++nblocks; 219 total += size; 220 } 221 222 end = phys_avail[biggestone+1]; 223 224 /* 225 * Initialize the queue headers for the free queue, the active queue 226 * and the inactive queue. 227 */ 228 229 vm_page_queue_init(); 230 231 /* 232 * Allocate (and initialize) the hash table buckets. 233 * 234 * The number of buckets MUST BE a power of 2, and the actual value is 235 * the next power of 2 greater than the number of physical pages in 236 * the system. 237 * 238 * We make the hash table approximately 2x the number of pages to 239 * reduce the chain length. This is about the same size using the 240 * singly-linked list as the 1x hash table we were using before 241 * using TAILQ but the chain length will be smaller. 242 * 243 * Note: This computation can be tweaked if desired. 244 */ 245 if (vm_page_bucket_count == 0) { 246 vm_page_bucket_count = 1; 247 while (vm_page_bucket_count < atop(total)) 248 vm_page_bucket_count <<= 1; 249 } 250 vm_page_bucket_count <<= 1; 251 vm_page_hash_mask = vm_page_bucket_count - 1; 252 253 /* 254 * Validate these addresses. 255 */ 256 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 257 new_end = trunc_page(new_end); 258 mapped = pmap_map(&vaddr, new_end, end, 259 VM_PROT_READ | VM_PROT_WRITE); 260 bzero((caddr_t) mapped, end - new_end); 261 262 vm_page_buckets = (struct vm_page **)mapped; 263 bucket = vm_page_buckets; 264 for (i = 0; i < vm_page_bucket_count; i++) { 265 *bucket = NULL; 266 bucket++; 267 } 268 269 /* 270 * Compute the number of pages of memory that will be available for 271 * use (taking into account the overhead of a page structure per 272 * page). 273 */ 274 275 first_page = phys_avail[0] / PAGE_SIZE; 276 277 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 278 npages = (total - (page_range * sizeof(struct vm_page)) - 279 (end - new_end)) / PAGE_SIZE; 280 281 end = new_end; 282 283 /* 284 * Initialize the mem entry structures now, and put them in the free 285 * queue. 286 */ 287 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 288 mapped = pmap_map(&vaddr, new_end, end, 289 VM_PROT_READ | VM_PROT_WRITE); 290 vm_page_array = (vm_page_t) mapped; 291 292 /* 293 * Clear all of the page structures 294 */ 295 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 296 vm_page_array_size = page_range; 297 298 /* 299 * Construct the free queue(s) in descending order (by physical 300 * address) so that the first 16MB of physical memory is allocated 301 * last rather than first. On large-memory machines, this avoids 302 * the exhaustion of low physical memory before isa_dmainit has run. 303 */ 304 cnt.v_page_count = 0; 305 cnt.v_free_count = 0; 306 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 307 pa = phys_avail[i]; 308 if (i == biggestone) 309 last_pa = new_end; 310 else 311 last_pa = phys_avail[i + 1]; 312 while (pa < last_pa && npages-- > 0) { 313 vm_add_new_page(pa); 314 pa += PAGE_SIZE; 315 } 316 } 317 return (vaddr); 318} 319 320/* 321 * vm_page_hash: 322 * 323 * Distributes the object/offset key pair among hash buckets. 324 * 325 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 326 * This routine may not block. 327 * 328 * We try to randomize the hash based on the object to spread the pages 329 * out in the hash table without it costing us too much. 330 */ 331static __inline int 332vm_page_hash(object, pindex) 333 vm_object_t object; 334 vm_pindex_t pindex; 335{ 336 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 337 338 return(i & vm_page_hash_mask); 339} 340 341/* 342 * vm_page_insert: [ internal use only ] 343 * 344 * Inserts the given mem entry into the object and object list. 345 * 346 * The pagetables are not updated but will presumably fault the page 347 * in if necessary, or if a kernel page the caller will at some point 348 * enter the page into the kernel's pmap. We are not allowed to block 349 * here so we *can't* do this anyway. 350 * 351 * The object and page must be locked, and must be splhigh. 352 * This routine may not block. 353 */ 354 355void 356vm_page_insert(m, object, pindex) 357 register vm_page_t m; 358 register vm_object_t object; 359 register vm_pindex_t pindex; 360{ 361 register struct vm_page **bucket; 362 363 if (m->object != NULL) 364 panic("vm_page_insert: already inserted"); 365 366 /* 367 * Record the object/offset pair in this page 368 */ 369 370 m->object = object; 371 m->pindex = pindex; 372 373 /* 374 * Insert it into the object_object/offset hash table 375 */ 376 377 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 378 m->hnext = *bucket; 379 *bucket = m; 380 vm_page_bucket_generation++; 381 382 /* 383 * Now link into the object's list of backed pages. 384 */ 385 386 TAILQ_INSERT_TAIL(&object->memq, m, listq); 387 object->generation++; 388 389 /* 390 * show that the object has one more resident page. 391 */ 392 393 object->resident_page_count++; 394 395 /* 396 * Since we are inserting a new and possibly dirty page, 397 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 398 */ 399 if (m->flags & PG_WRITEABLE) 400 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 401} 402 403/* 404 * vm_page_remove: 405 * NOTE: used by device pager as well -wfj 406 * 407 * Removes the given mem entry from the object/offset-page 408 * table and the object page list, but do not invalidate/terminate 409 * the backing store. 410 * 411 * The object and page must be locked, and at splhigh. 412 * The underlying pmap entry (if any) is NOT removed here. 413 * This routine may not block. 414 */ 415 416void 417vm_page_remove(m) 418 vm_page_t m; 419{ 420 vm_object_t object; 421 422 if (m->object == NULL) 423 return; 424 425 if ((m->flags & PG_BUSY) == 0) { 426 panic("vm_page_remove: page not busy"); 427 } 428 429 /* 430 * Basically destroy the page. 431 */ 432 433 vm_page_wakeup(m); 434 435 object = m->object; 436 437 /* 438 * Remove from the object_object/offset hash table. The object 439 * must be on the hash queue, we will panic if it isn't 440 * 441 * Note: we must NULL-out m->hnext to prevent loops in detached 442 * buffers with vm_page_lookup(). 443 */ 444 445 { 446 struct vm_page **bucket; 447 448 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 449 while (*bucket != m) { 450 if (*bucket == NULL) 451 panic("vm_page_remove(): page not found in hash"); 452 bucket = &(*bucket)->hnext; 453 } 454 *bucket = m->hnext; 455 m->hnext = NULL; 456 vm_page_bucket_generation++; 457 } 458 459 /* 460 * Now remove from the object's list of backed pages. 461 */ 462 463 TAILQ_REMOVE(&object->memq, m, listq); 464 465 /* 466 * And show that the object has one fewer resident page. 467 */ 468 469 object->resident_page_count--; 470 object->generation++; 471 472 m->object = NULL; 473} 474 475/* 476 * vm_page_lookup: 477 * 478 * Returns the page associated with the object/offset 479 * pair specified; if none is found, NULL is returned. 480 * 481 * NOTE: the code below does not lock. It will operate properly if 482 * an interrupt makes a change, but the generation algorithm will not 483 * operate properly in an SMP environment where both cpu's are able to run 484 * kernel code simultaneously. 485 * 486 * The object must be locked. No side effects. 487 * This routine may not block. 488 * This is a critical path routine 489 */ 490 491vm_page_t 492vm_page_lookup(object, pindex) 493 register vm_object_t object; 494 register vm_pindex_t pindex; 495{ 496 register vm_page_t m; 497 register struct vm_page **bucket; 498 int generation; 499 500 /* 501 * Search the hash table for this object/offset pair 502 */ 503 504retry: 505 generation = vm_page_bucket_generation; 506 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 507 for (m = *bucket; m != NULL; m = m->hnext) { 508 if ((m->object == object) && (m->pindex == pindex)) { 509 if (vm_page_bucket_generation != generation) 510 goto retry; 511 return (m); 512 } 513 } 514 if (vm_page_bucket_generation != generation) 515 goto retry; 516 return (NULL); 517} 518 519/* 520 * vm_page_rename: 521 * 522 * Move the given memory entry from its 523 * current object to the specified target object/offset. 524 * 525 * The object must be locked. 526 * This routine may not block. 527 * 528 * Note: this routine will raise itself to splvm(), the caller need not. 529 * 530 * Note: swap associated with the page must be invalidated by the move. We 531 * have to do this for several reasons: (1) we aren't freeing the 532 * page, (2) we are dirtying the page, (3) the VM system is probably 533 * moving the page from object A to B, and will then later move 534 * the backing store from A to B and we can't have a conflict. 535 * 536 * Note: we *always* dirty the page. It is necessary both for the 537 * fact that we moved it, and because we may be invalidating 538 * swap. If the page is on the cache, we have to deactivate it 539 * or vm_page_dirty() will panic. Dirty pages are not allowed 540 * on the cache. 541 */ 542 543void 544vm_page_rename(m, new_object, new_pindex) 545 register vm_page_t m; 546 register vm_object_t new_object; 547 vm_pindex_t new_pindex; 548{ 549 int s; 550 551 s = splvm(); 552 vm_page_remove(m); 553 vm_page_insert(m, new_object, new_pindex); 554 if (m->queue - m->pc == PQ_CACHE) 555 vm_page_deactivate(m); 556 vm_page_dirty(m); 557 splx(s); 558} 559 560/* 561 * vm_page_unqueue_nowakeup: 562 * 563 * vm_page_unqueue() without any wakeup 564 * 565 * This routine must be called at splhigh(). 566 * This routine may not block. 567 */ 568 569void 570vm_page_unqueue_nowakeup(m) 571 vm_page_t m; 572{ 573 int queue = m->queue; 574 struct vpgqueues *pq; 575 if (queue != PQ_NONE) { 576 pq = &vm_page_queues[queue]; 577 m->queue = PQ_NONE; 578 TAILQ_REMOVE(&pq->pl, m, pageq); 579 (*pq->cnt)--; 580 pq->lcnt--; 581 } 582} 583 584/* 585 * vm_page_unqueue: 586 * 587 * Remove a page from its queue. 588 * 589 * This routine must be called at splhigh(). 590 * This routine may not block. 591 */ 592 593void 594vm_page_unqueue(m) 595 vm_page_t m; 596{ 597 int queue = m->queue; 598 struct vpgqueues *pq; 599 if (queue != PQ_NONE) { 600 m->queue = PQ_NONE; 601 pq = &vm_page_queues[queue]; 602 TAILQ_REMOVE(&pq->pl, m, pageq); 603 (*pq->cnt)--; 604 pq->lcnt--; 605 if ((queue - m->pc) == PQ_CACHE) { 606 if (vm_paging_needed()) 607 pagedaemon_wakeup(); 608 } 609 } 610} 611 612#if PQ_L2_SIZE > 1 613 614/* 615 * vm_page_list_find: 616 * 617 * Find a page on the specified queue with color optimization. 618 * 619 * The page coloring optimization attempts to locate a page 620 * that does not overload other nearby pages in the object in 621 * the cpu's L1 or L2 caches. We need this optimization because 622 * cpu caches tend to be physical caches, while object spaces tend 623 * to be virtual. 624 * 625 * This routine must be called at splvm(). 626 * This routine may not block. 627 * 628 * This routine may only be called from the vm_page_list_find() macro 629 * in vm_page.h 630 */ 631vm_page_t 632_vm_page_list_find(basequeue, index) 633 int basequeue, index; 634{ 635 int i; 636 vm_page_t m = NULL; 637 struct vpgqueues *pq; 638 639 pq = &vm_page_queues[basequeue]; 640 641 /* 642 * Note that for the first loop, index+i and index-i wind up at the 643 * same place. Even though this is not totally optimal, we've already 644 * blown it by missing the cache case so we do not care. 645 */ 646 647 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 648 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 649 break; 650 651 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 652 break; 653 } 654 return(m); 655} 656 657#endif 658 659/* 660 * vm_page_select_cache: 661 * 662 * Find a page on the cache queue with color optimization. As pages 663 * might be found, but not applicable, they are deactivated. This 664 * keeps us from using potentially busy cached pages. 665 * 666 * This routine must be called at splvm(). 667 * This routine may not block. 668 */ 669vm_page_t 670vm_page_select_cache(object, pindex) 671 vm_object_t object; 672 vm_pindex_t pindex; 673{ 674 vm_page_t m; 675 676 while (TRUE) { 677 m = vm_page_list_find( 678 PQ_CACHE, 679 (pindex + object->pg_color) & PQ_L2_MASK, 680 FALSE 681 ); 682 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 683 m->hold_count || m->wire_count)) { 684 vm_page_deactivate(m); 685 continue; 686 } 687 return m; 688 } 689} 690 691/* 692 * vm_page_select_free: 693 * 694 * Find a free or zero page, with specified preference. We attempt to 695 * inline the nominal case and fall back to _vm_page_select_free() 696 * otherwise. 697 * 698 * This routine must be called at splvm(). 699 * This routine may not block. 700 */ 701 702static __inline vm_page_t 703vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 704{ 705 vm_page_t m; 706 707 m = vm_page_list_find( 708 PQ_FREE, 709 (pindex + object->pg_color) & PQ_L2_MASK, 710 prefer_zero 711 ); 712 return(m); 713} 714 715/* 716 * vm_page_alloc: 717 * 718 * Allocate and return a memory cell associated 719 * with this VM object/offset pair. 720 * 721 * page_req classes: 722 * VM_ALLOC_NORMAL normal process request 723 * VM_ALLOC_SYSTEM system *really* needs a page 724 * VM_ALLOC_INTERRUPT interrupt time request 725 * VM_ALLOC_ZERO zero page 726 * 727 * Object must be locked. 728 * This routine may not block. 729 * 730 * Additional special handling is required when called from an 731 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 732 * the page cache in this case. 733 */ 734 735vm_page_t 736vm_page_alloc(object, pindex, page_req) 737 vm_object_t object; 738 vm_pindex_t pindex; 739 int page_req; 740{ 741 register vm_page_t m = NULL; 742 int s; 743 744 KASSERT(!vm_page_lookup(object, pindex), 745 ("vm_page_alloc: page already allocated")); 746 747 /* 748 * The pager is allowed to eat deeper into the free page list. 749 */ 750 751 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 752 page_req = VM_ALLOC_SYSTEM; 753 }; 754 755 s = splvm(); 756 757loop: 758 if (cnt.v_free_count > cnt.v_free_reserved) { 759 /* 760 * Allocate from the free queue if there are plenty of pages 761 * in it. 762 */ 763 if (page_req == VM_ALLOC_ZERO) 764 m = vm_page_select_free(object, pindex, TRUE); 765 else 766 m = vm_page_select_free(object, pindex, FALSE); 767 } else if ( 768 (page_req == VM_ALLOC_SYSTEM && 769 cnt.v_cache_count == 0 && 770 cnt.v_free_count > cnt.v_interrupt_free_min) || 771 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 772 ) { 773 /* 774 * Interrupt or system, dig deeper into the free list. 775 */ 776 m = vm_page_select_free(object, pindex, FALSE); 777 } else if (page_req != VM_ALLOC_INTERRUPT) { 778 /* 779 * Allocatable from cache (non-interrupt only). On success, 780 * we must free the page and try again, thus ensuring that 781 * cnt.v_*_free_min counters are replenished. 782 */ 783 m = vm_page_select_cache(object, pindex); 784 if (m == NULL) { 785 splx(s); 786#if defined(DIAGNOSTIC) 787 if (cnt.v_cache_count > 0) 788 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 789#endif 790 vm_pageout_deficit++; 791 pagedaemon_wakeup(); 792 return (NULL); 793 } 794 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 795 vm_page_busy(m); 796 vm_page_protect(m, VM_PROT_NONE); 797 vm_page_free(m); 798 goto loop; 799 } else { 800 /* 801 * Not allocatable from cache from interrupt, give up. 802 */ 803 splx(s); 804 vm_pageout_deficit++; 805 pagedaemon_wakeup(); 806 return (NULL); 807 } 808 809 /* 810 * At this point we had better have found a good page. 811 */ 812 813 KASSERT( 814 m != NULL, 815 ("vm_page_alloc(): missing page on free queue\n") 816 ); 817 818 /* 819 * Remove from free queue 820 */ 821 822 vm_page_unqueue_nowakeup(m); 823 824 /* 825 * Initialize structure. Only the PG_ZERO flag is inherited. 826 */ 827 828 if (m->flags & PG_ZERO) { 829 vm_page_zero_count--; 830 m->flags = PG_ZERO | PG_BUSY; 831 } else { 832 m->flags = PG_BUSY; 833 } 834 m->wire_count = 0; 835 m->hold_count = 0; 836 m->act_count = 0; 837 m->busy = 0; 838 m->valid = 0; 839 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 840 841 /* 842 * vm_page_insert() is safe prior to the splx(). Note also that 843 * inserting a page here does not insert it into the pmap (which 844 * could cause us to block allocating memory). We cannot block 845 * anywhere. 846 */ 847 848 vm_page_insert(m, object, pindex); 849 850 /* 851 * Don't wakeup too often - wakeup the pageout daemon when 852 * we would be nearly out of memory. 853 */ 854 if (vm_paging_needed()) 855 pagedaemon_wakeup(); 856 857 splx(s); 858 859 return (m); 860} 861 862/* 863 * vm_wait: (also see VM_WAIT macro) 864 * 865 * Block until free pages are available for allocation 866 */ 867 868void 869vm_wait() 870{ 871 int s; 872 873 s = splvm(); 874 if (curproc == pageproc) { 875 vm_pageout_pages_needed = 1; 876 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 877 } else { 878 if (!vm_pages_needed) { 879 vm_pages_needed = 1; 880 wakeup(&vm_pages_needed); 881 } 882 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 883 } 884 splx(s); 885} 886 887/* 888 * vm_await: (also see VM_AWAIT macro) 889 * 890 * asleep on an event that will signal when free pages are available 891 * for allocation. 892 */ 893 894void 895vm_await() 896{ 897 int s; 898 899 s = splvm(); 900 if (curproc == pageproc) { 901 vm_pageout_pages_needed = 1; 902 asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0); 903 } else { 904 if (!vm_pages_needed) { 905 vm_pages_needed++; 906 wakeup(&vm_pages_needed); 907 } 908 asleep(&cnt.v_free_count, PVM, "vmwait", 0); 909 } 910 splx(s); 911} 912 913#if 0 914/* 915 * vm_page_sleep: 916 * 917 * Block until page is no longer busy. 918 */ 919 920int 921vm_page_sleep(vm_page_t m, char *msg, char *busy) { 922 int slept = 0; 923 if ((busy && *busy) || (m->flags & PG_BUSY)) { 924 int s; 925 s = splvm(); 926 if ((busy && *busy) || (m->flags & PG_BUSY)) { 927 vm_page_flag_set(m, PG_WANTED); 928 tsleep(m, PVM, msg, 0); 929 slept = 1; 930 } 931 splx(s); 932 } 933 return slept; 934} 935 936#endif 937 938#if 0 939 940/* 941 * vm_page_asleep: 942 * 943 * Similar to vm_page_sleep(), but does not block. Returns 0 if 944 * the page is not busy, or 1 if the page is busy. 945 * 946 * This routine has the side effect of calling asleep() if the page 947 * was busy (1 returned). 948 */ 949 950int 951vm_page_asleep(vm_page_t m, char *msg, char *busy) { 952 int slept = 0; 953 if ((busy && *busy) || (m->flags & PG_BUSY)) { 954 int s; 955 s = splvm(); 956 if ((busy && *busy) || (m->flags & PG_BUSY)) { 957 vm_page_flag_set(m, PG_WANTED); 958 asleep(m, PVM, msg, 0); 959 slept = 1; 960 } 961 splx(s); 962 } 963 return slept; 964} 965 966#endif 967 968/* 969 * vm_page_activate: 970 * 971 * Put the specified page on the active list (if appropriate). 972 * Ensure that act_count is at least ACT_INIT but do not otherwise 973 * mess with it. 974 * 975 * The page queues must be locked. 976 * This routine may not block. 977 */ 978void 979vm_page_activate(m) 980 register vm_page_t m; 981{ 982 int s; 983 984 s = splvm(); 985 if (m->queue != PQ_ACTIVE) { 986 if ((m->queue - m->pc) == PQ_CACHE) 987 cnt.v_reactivated++; 988 989 vm_page_unqueue(m); 990 991 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 992 m->queue = PQ_ACTIVE; 993 vm_page_queues[PQ_ACTIVE].lcnt++; 994 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 995 if (m->act_count < ACT_INIT) 996 m->act_count = ACT_INIT; 997 cnt.v_active_count++; 998 } 999 } else { 1000 if (m->act_count < ACT_INIT) 1001 m->act_count = ACT_INIT; 1002 } 1003 1004 splx(s); 1005} 1006 1007/* 1008 * vm_page_free_wakeup: 1009 * 1010 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1011 * routine is called when a page has been added to the cache or free 1012 * queues. 1013 * 1014 * This routine may not block. 1015 * This routine must be called at splvm() 1016 */ 1017static __inline void 1018vm_page_free_wakeup() 1019{ 1020 /* 1021 * if pageout daemon needs pages, then tell it that there are 1022 * some free. 1023 */ 1024 if (vm_pageout_pages_needed && 1025 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1026 wakeup(&vm_pageout_pages_needed); 1027 vm_pageout_pages_needed = 0; 1028 } 1029 /* 1030 * wakeup processes that are waiting on memory if we hit a 1031 * high water mark. And wakeup scheduler process if we have 1032 * lots of memory. this process will swapin processes. 1033 */ 1034 if (vm_pages_needed && !vm_page_count_min()) { 1035 vm_pages_needed = 0; 1036 wakeup(&cnt.v_free_count); 1037 } 1038} 1039 1040/* 1041 * vm_page_free_toq: 1042 * 1043 * Returns the given page to the PQ_FREE list, 1044 * disassociating it with any VM object. 1045 * 1046 * Object and page must be locked prior to entry. 1047 * This routine may not block. 1048 */ 1049 1050void 1051vm_page_free_toq(vm_page_t m) 1052{ 1053 int s; 1054 struct vpgqueues *pq; 1055 vm_object_t object = m->object; 1056 1057 s = splvm(); 1058 1059 cnt.v_tfree++; 1060 1061 if (m->busy || ((m->queue - m->pc) == PQ_FREE) || 1062 (m->hold_count != 0)) { 1063 printf( 1064 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1065 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1066 m->hold_count); 1067 if ((m->queue - m->pc) == PQ_FREE) 1068 panic("vm_page_free: freeing free page"); 1069 else 1070 panic("vm_page_free: freeing busy page"); 1071 } 1072 1073 /* 1074 * unqueue, then remove page. Note that we cannot destroy 1075 * the page here because we do not want to call the pager's 1076 * callback routine until after we've put the page on the 1077 * appropriate free queue. 1078 */ 1079 1080 vm_page_unqueue_nowakeup(m); 1081 vm_page_remove(m); 1082 1083 /* 1084 * If fictitious remove object association and 1085 * return, otherwise delay object association removal. 1086 */ 1087 1088 if ((m->flags & PG_FICTITIOUS) != 0) { 1089 splx(s); 1090 return; 1091 } 1092 1093 m->valid = 0; 1094 vm_page_undirty(m); 1095 1096 if (m->wire_count != 0) { 1097 if (m->wire_count > 1) { 1098 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1099 m->wire_count, (long)m->pindex); 1100 } 1101 panic("vm_page_free: freeing wired page\n"); 1102 } 1103 1104 /* 1105 * If we've exhausted the object's resident pages we want to free 1106 * it up. 1107 */ 1108 1109 if (object && 1110 (object->type == OBJT_VNODE) && 1111 ((object->flags & OBJ_DEAD) == 0) 1112 ) { 1113 struct vnode *vp = (struct vnode *)object->handle; 1114 1115 if (vp && VSHOULDFREE(vp)) 1116 vfree(vp); 1117 } 1118 1119 /* 1120 * Clear the UNMANAGED flag when freeing an unmanaged page. 1121 */ 1122 1123 if (m->flags & PG_UNMANAGED) { 1124 m->flags &= ~PG_UNMANAGED; 1125 } else { 1126#ifdef __alpha__ 1127 pmap_page_is_free(m); 1128#endif 1129 } 1130 1131 m->queue = PQ_FREE + m->pc; 1132 pq = &vm_page_queues[m->queue]; 1133 pq->lcnt++; 1134 ++(*pq->cnt); 1135 1136 /* 1137 * Put zero'd pages on the end ( where we look for zero'd pages 1138 * first ) and non-zerod pages at the head. 1139 */ 1140 1141 if (m->flags & PG_ZERO) { 1142 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1143 ++vm_page_zero_count; 1144 } else { 1145 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1146 } 1147 1148 vm_page_free_wakeup(); 1149 1150 splx(s); 1151} 1152 1153/* 1154 * vm_page_unmanage: 1155 * 1156 * Prevent PV management from being done on the page. The page is 1157 * removed from the paging queues as if it were wired, and as a 1158 * consequence of no longer being managed the pageout daemon will not 1159 * touch it (since there is no way to locate the pte mappings for the 1160 * page). madvise() calls that mess with the pmap will also no longer 1161 * operate on the page. 1162 * 1163 * Beyond that the page is still reasonably 'normal'. Freeing the page 1164 * will clear the flag. 1165 * 1166 * This routine is used by OBJT_PHYS objects - objects using unswappable 1167 * physical memory as backing store rather then swap-backed memory and 1168 * will eventually be extended to support 4MB unmanaged physical 1169 * mappings. 1170 */ 1171 1172void 1173vm_page_unmanage(vm_page_t m) 1174{ 1175 int s; 1176 1177 s = splvm(); 1178 if ((m->flags & PG_UNMANAGED) == 0) { 1179 if (m->wire_count == 0) 1180 vm_page_unqueue(m); 1181 } 1182 vm_page_flag_set(m, PG_UNMANAGED); 1183 splx(s); 1184} 1185 1186/* 1187 * vm_page_wire: 1188 * 1189 * Mark this page as wired down by yet 1190 * another map, removing it from paging queues 1191 * as necessary. 1192 * 1193 * The page queues must be locked. 1194 * This routine may not block. 1195 */ 1196void 1197vm_page_wire(m) 1198 register vm_page_t m; 1199{ 1200 int s; 1201 1202 /* 1203 * Only bump the wire statistics if the page is not already wired, 1204 * and only unqueue the page if it is on some queue (if it is unmanaged 1205 * it is already off the queues). 1206 */ 1207 s = splvm(); 1208 if (m->wire_count == 0) { 1209 if ((m->flags & PG_UNMANAGED) == 0) 1210 vm_page_unqueue(m); 1211 cnt.v_wire_count++; 1212 } 1213 m->wire_count++; 1214 splx(s); 1215 vm_page_flag_set(m, PG_MAPPED); 1216} 1217 1218/* 1219 * vm_page_unwire: 1220 * 1221 * Release one wiring of this page, potentially 1222 * enabling it to be paged again. 1223 * 1224 * Many pages placed on the inactive queue should actually go 1225 * into the cache, but it is difficult to figure out which. What 1226 * we do instead, if the inactive target is well met, is to put 1227 * clean pages at the head of the inactive queue instead of the tail. 1228 * This will cause them to be moved to the cache more quickly and 1229 * if not actively re-referenced, freed more quickly. If we just 1230 * stick these pages at the end of the inactive queue, heavy filesystem 1231 * meta-data accesses can cause an unnecessary paging load on memory bound 1232 * processes. This optimization causes one-time-use metadata to be 1233 * reused more quickly. 1234 * 1235 * BUT, if we are in a low-memory situation we have no choice but to 1236 * put clean pages on the cache queue. 1237 * 1238 * A number of routines use vm_page_unwire() to guarantee that the page 1239 * will go into either the inactive or active queues, and will NEVER 1240 * be placed in the cache - for example, just after dirtying a page. 1241 * dirty pages in the cache are not allowed. 1242 * 1243 * The page queues must be locked. 1244 * This routine may not block. 1245 */ 1246void 1247vm_page_unwire(m, activate) 1248 register vm_page_t m; 1249 int activate; 1250{ 1251 int s; 1252 1253 s = splvm(); 1254 1255 if (m->wire_count > 0) { 1256 m->wire_count--; 1257 if (m->wire_count == 0) { 1258 cnt.v_wire_count--; 1259 if (m->flags & PG_UNMANAGED) { 1260 ; 1261 } else if (activate) { 1262 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1263 m->queue = PQ_ACTIVE; 1264 vm_page_queues[PQ_ACTIVE].lcnt++; 1265 cnt.v_active_count++; 1266 } else { 1267 vm_page_flag_clear(m, PG_WINATCFLS); 1268 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1269 m->queue = PQ_INACTIVE; 1270 vm_page_queues[PQ_INACTIVE].lcnt++; 1271 cnt.v_inactive_count++; 1272 } 1273 } 1274 } else { 1275 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1276 } 1277 splx(s); 1278} 1279 1280 1281/* 1282 * Move the specified page to the inactive queue. If the page has 1283 * any associated swap, the swap is deallocated. 1284 * 1285 * Normally athead is 0 resulting in LRU operation. athead is set 1286 * to 1 if we want this page to be 'as if it were placed in the cache', 1287 * except without unmapping it from the process address space. 1288 * 1289 * This routine may not block. 1290 */ 1291static __inline void 1292_vm_page_deactivate(vm_page_t m, int athead) 1293{ 1294 int s; 1295 1296 /* 1297 * Ignore if already inactive. 1298 */ 1299 if (m->queue == PQ_INACTIVE) 1300 return; 1301 1302 s = splvm(); 1303 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1304 if ((m->queue - m->pc) == PQ_CACHE) 1305 cnt.v_reactivated++; 1306 vm_page_flag_clear(m, PG_WINATCFLS); 1307 vm_page_unqueue(m); 1308 if (athead) 1309 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1310 else 1311 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1312 m->queue = PQ_INACTIVE; 1313 vm_page_queues[PQ_INACTIVE].lcnt++; 1314 cnt.v_inactive_count++; 1315 } 1316 splx(s); 1317} 1318 1319void 1320vm_page_deactivate(vm_page_t m) 1321{ 1322 _vm_page_deactivate(m, 0); 1323} 1324 1325/* 1326 * vm_page_try_to_cache: 1327 * 1328 * Returns 0 on failure, 1 on success 1329 */ 1330int 1331vm_page_try_to_cache(vm_page_t m) 1332{ 1333 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1334 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1335 return(0); 1336 } 1337 vm_page_test_dirty(m); 1338 if (m->dirty) 1339 return(0); 1340 vm_page_cache(m); 1341 return(1); 1342} 1343 1344/* 1345 * vm_page_cache 1346 * 1347 * Put the specified page onto the page cache queue (if appropriate). 1348 * 1349 * This routine may not block. 1350 */ 1351void 1352vm_page_cache(m) 1353 register vm_page_t m; 1354{ 1355 int s; 1356 1357 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1358 printf("vm_page_cache: attempting to cache busy page\n"); 1359 return; 1360 } 1361 if ((m->queue - m->pc) == PQ_CACHE) 1362 return; 1363 1364 /* 1365 * Remove all pmaps and indicate that the page is not 1366 * writeable or mapped. 1367 */ 1368 1369 vm_page_protect(m, VM_PROT_NONE); 1370 if (m->dirty != 0) { 1371 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1372 (long)m->pindex); 1373 } 1374 s = splvm(); 1375 vm_page_unqueue_nowakeup(m); 1376 m->queue = PQ_CACHE + m->pc; 1377 vm_page_queues[m->queue].lcnt++; 1378 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1379 cnt.v_cache_count++; 1380 vm_page_free_wakeup(); 1381 splx(s); 1382} 1383 1384/* 1385 * vm_page_dontneed 1386 * 1387 * Cache, deactivate, or do nothing as appropriate. This routine 1388 * is typically used by madvise() MADV_DONTNEED. 1389 * 1390 * Generally speaking we want to move the page into the cache so 1391 * it gets reused quickly. However, this can result in a silly syndrome 1392 * due to the page recycling too quickly. Small objects will not be 1393 * fully cached. On the otherhand, if we move the page to the inactive 1394 * queue we wind up with a problem whereby very large objects 1395 * unnecessarily blow away our inactive and cache queues. 1396 * 1397 * The solution is to move the pages based on a fixed weighting. We 1398 * either leave them alone, deactivate them, or move them to the cache, 1399 * where moving them to the cache has the highest weighting. 1400 * By forcing some pages into other queues we eventually force the 1401 * system to balance the queues, potentially recovering other unrelated 1402 * space from active. The idea is to not force this to happen too 1403 * often. 1404 */ 1405 1406void 1407vm_page_dontneed(m) 1408 vm_page_t m; 1409{ 1410 static int dnweight; 1411 int dnw; 1412 int head; 1413 1414 dnw = ++dnweight; 1415 1416 /* 1417 * occassionally leave the page alone 1418 */ 1419 1420 if ((dnw & 0x01F0) == 0 || 1421 m->queue == PQ_INACTIVE || 1422 m->queue - m->pc == PQ_CACHE 1423 ) { 1424 if (m->act_count >= ACT_INIT) 1425 --m->act_count; 1426 return; 1427 } 1428 1429 if (m->dirty == 0) 1430 vm_page_test_dirty(m); 1431 1432 if (m->dirty || (dnw & 0x0070) == 0) { 1433 /* 1434 * Deactivate the page 3 times out of 32. 1435 */ 1436 head = 0; 1437 } else { 1438 /* 1439 * Cache the page 28 times out of every 32. Note that 1440 * the page is deactivated instead of cached, but placed 1441 * at the head of the queue instead of the tail. 1442 */ 1443 head = 1; 1444 } 1445 _vm_page_deactivate(m, head); 1446} 1447 1448/* 1449 * Grab a page, waiting until we are waken up due to the page 1450 * changing state. We keep on waiting, if the page continues 1451 * to be in the object. If the page doesn't exist, allocate it. 1452 * 1453 * This routine may block. 1454 */ 1455vm_page_t 1456vm_page_grab(object, pindex, allocflags) 1457 vm_object_t object; 1458 vm_pindex_t pindex; 1459 int allocflags; 1460{ 1461 1462 vm_page_t m; 1463 int s, generation; 1464 1465retrylookup: 1466 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1467 if (m->busy || (m->flags & PG_BUSY)) { 1468 generation = object->generation; 1469 1470 s = splvm(); 1471 while ((object->generation == generation) && 1472 (m->busy || (m->flags & PG_BUSY))) { 1473 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1474 tsleep(m, PVM, "pgrbwt", 0); 1475 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1476 splx(s); 1477 return NULL; 1478 } 1479 } 1480 splx(s); 1481 goto retrylookup; 1482 } else { 1483 vm_page_busy(m); 1484 return m; 1485 } 1486 } 1487 1488 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1489 if (m == NULL) { 1490 VM_WAIT; 1491 if ((allocflags & VM_ALLOC_RETRY) == 0) 1492 return NULL; 1493 goto retrylookup; 1494 } 1495 1496 return m; 1497} 1498 1499/* 1500 * Mapping function for valid bits or for dirty bits in 1501 * a page. May not block. 1502 * 1503 * Inputs are required to range within a page. 1504 */ 1505 1506__inline int 1507vm_page_bits(int base, int size) 1508{ 1509 int first_bit; 1510 int last_bit; 1511 1512 KASSERT( 1513 base + size <= PAGE_SIZE, 1514 ("vm_page_bits: illegal base/size %d/%d", base, size) 1515 ); 1516 1517 if (size == 0) /* handle degenerate case */ 1518 return(0); 1519 1520 first_bit = base >> DEV_BSHIFT; 1521 last_bit = (base + size - 1) >> DEV_BSHIFT; 1522 1523 return ((2 << last_bit) - (1 << first_bit)); 1524} 1525 1526/* 1527 * vm_page_set_validclean: 1528 * 1529 * Sets portions of a page valid and clean. The arguments are expected 1530 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1531 * of any partial chunks touched by the range. The invalid portion of 1532 * such chunks will be zero'd. 1533 * 1534 * This routine may not block. 1535 * 1536 * (base + size) must be less then or equal to PAGE_SIZE. 1537 */ 1538void 1539vm_page_set_validclean(m, base, size) 1540 vm_page_t m; 1541 int base; 1542 int size; 1543{ 1544 int pagebits; 1545 int frag; 1546 int endoff; 1547 1548 if (size == 0) /* handle degenerate case */ 1549 return; 1550 1551 /* 1552 * If the base is not DEV_BSIZE aligned and the valid 1553 * bit is clear, we have to zero out a portion of the 1554 * first block. 1555 */ 1556 1557 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1558 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1559 ) { 1560 pmap_zero_page_area( 1561 VM_PAGE_TO_PHYS(m), 1562 frag, 1563 base - frag 1564 ); 1565 } 1566 1567 /* 1568 * If the ending offset is not DEV_BSIZE aligned and the 1569 * valid bit is clear, we have to zero out a portion of 1570 * the last block. 1571 */ 1572 1573 endoff = base + size; 1574 1575 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1576 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1577 ) { 1578 pmap_zero_page_area( 1579 VM_PAGE_TO_PHYS(m), 1580 endoff, 1581 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1582 ); 1583 } 1584 1585 /* 1586 * Set valid, clear dirty bits. If validating the entire 1587 * page we can safely clear the pmap modify bit. We also 1588 * use this opportunity to clear the PG_NOSYNC flag. If a process 1589 * takes a write fault on a MAP_NOSYNC memory area the flag will 1590 * be set again. 1591 */ 1592 1593 pagebits = vm_page_bits(base, size); 1594 m->valid |= pagebits; 1595 m->dirty &= ~pagebits; 1596 if (base == 0 && size == PAGE_SIZE) { 1597 pmap_clear_modify(m); 1598 vm_page_flag_clear(m, PG_NOSYNC); 1599 } 1600} 1601 1602#if 0 1603 1604void 1605vm_page_set_dirty(m, base, size) 1606 vm_page_t m; 1607 int base; 1608 int size; 1609{ 1610 m->dirty |= vm_page_bits(base, size); 1611} 1612 1613#endif 1614 1615void 1616vm_page_clear_dirty(m, base, size) 1617 vm_page_t m; 1618 int base; 1619 int size; 1620{ 1621 m->dirty &= ~vm_page_bits(base, size); 1622} 1623 1624/* 1625 * vm_page_set_invalid: 1626 * 1627 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1628 * valid and dirty bits for the effected areas are cleared. 1629 * 1630 * May not block. 1631 */ 1632void 1633vm_page_set_invalid(m, base, size) 1634 vm_page_t m; 1635 int base; 1636 int size; 1637{ 1638 int bits; 1639 1640 bits = vm_page_bits(base, size); 1641 m->valid &= ~bits; 1642 m->dirty &= ~bits; 1643 m->object->generation++; 1644} 1645 1646/* 1647 * vm_page_zero_invalid() 1648 * 1649 * The kernel assumes that the invalid portions of a page contain 1650 * garbage, but such pages can be mapped into memory by user code. 1651 * When this occurs, we must zero out the non-valid portions of the 1652 * page so user code sees what it expects. 1653 * 1654 * Pages are most often semi-valid when the end of a file is mapped 1655 * into memory and the file's size is not page aligned. 1656 */ 1657 1658void 1659vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1660{ 1661 int b; 1662 int i; 1663 1664 /* 1665 * Scan the valid bits looking for invalid sections that 1666 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1667 * valid bit may be set ) have already been zerod by 1668 * vm_page_set_validclean(). 1669 */ 1670 1671 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1672 if (i == (PAGE_SIZE / DEV_BSIZE) || 1673 (m->valid & (1 << i)) 1674 ) { 1675 if (i > b) { 1676 pmap_zero_page_area( 1677 VM_PAGE_TO_PHYS(m), 1678 b << DEV_BSHIFT, 1679 (i - b) << DEV_BSHIFT 1680 ); 1681 } 1682 b = i + 1; 1683 } 1684 } 1685 1686 /* 1687 * setvalid is TRUE when we can safely set the zero'd areas 1688 * as being valid. We can do this if there are no cache consistancy 1689 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1690 */ 1691 1692 if (setvalid) 1693 m->valid = VM_PAGE_BITS_ALL; 1694} 1695 1696/* 1697 * vm_page_is_valid: 1698 * 1699 * Is (partial) page valid? Note that the case where size == 0 1700 * will return FALSE in the degenerate case where the page is 1701 * entirely invalid, and TRUE otherwise. 1702 * 1703 * May not block. 1704 */ 1705 1706int 1707vm_page_is_valid(m, base, size) 1708 vm_page_t m; 1709 int base; 1710 int size; 1711{ 1712 int bits = vm_page_bits(base, size); 1713 1714 if (m->valid && ((m->valid & bits) == bits)) 1715 return 1; 1716 else 1717 return 0; 1718} 1719 1720/* 1721 * update dirty bits from pmap/mmu. May not block. 1722 */ 1723 1724void 1725vm_page_test_dirty(m) 1726 vm_page_t m; 1727{ 1728 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1729 vm_page_dirty(m); 1730 } 1731} 1732 1733/* 1734 * This interface is for merging with malloc() someday. 1735 * Even if we never implement compaction so that contiguous allocation 1736 * works after initialization time, malloc()'s data structures are good 1737 * for statistics and for allocations of less than a page. 1738 */ 1739void * 1740contigmalloc1(size, type, flags, low, high, alignment, boundary, map) 1741 unsigned long size; /* should be size_t here and for malloc() */ 1742 struct malloc_type *type; 1743 int flags; 1744 unsigned long low; 1745 unsigned long high; 1746 unsigned long alignment; 1747 unsigned long boundary; 1748 vm_map_t map; 1749{ 1750 int i, s, start; 1751 vm_offset_t addr, phys, tmp_addr; 1752 int pass; 1753 vm_page_t pga = vm_page_array; 1754 1755 size = round_page(size); 1756 if (size == 0) 1757 panic("contigmalloc1: size must not be 0"); 1758 if ((alignment & (alignment - 1)) != 0) 1759 panic("contigmalloc1: alignment must be a power of 2"); 1760 if ((boundary & (boundary - 1)) != 0) 1761 panic("contigmalloc1: boundary must be a power of 2"); 1762 1763 start = 0; 1764 for (pass = 0; pass <= 1; pass++) { 1765 s = splvm(); 1766again: 1767 /* 1768 * Find first page in array that is free, within range, aligned, and 1769 * such that the boundary won't be crossed. 1770 */ 1771 for (i = start; i < cnt.v_page_count; i++) { 1772 int pqtype; 1773 phys = VM_PAGE_TO_PHYS(&pga[i]); 1774 pqtype = pga[i].queue - pga[i].pc; 1775 if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) && 1776 (phys >= low) && (phys < high) && 1777 ((phys & (alignment - 1)) == 0) && 1778 (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0)) 1779 break; 1780 } 1781 1782 /* 1783 * If the above failed or we will exceed the upper bound, fail. 1784 */ 1785 if ((i == cnt.v_page_count) || 1786 ((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) { 1787 vm_page_t m, next; 1788 1789again1: 1790 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 1791 m != NULL; 1792 m = next) { 1793 1794 KASSERT(m->queue == PQ_INACTIVE, 1795 ("contigmalloc1: page %p is not PQ_INACTIVE", m)); 1796 1797 next = TAILQ_NEXT(m, pageq); 1798 if (vm_page_sleep_busy(m, TRUE, "vpctw0")) 1799 goto again1; 1800 vm_page_test_dirty(m); 1801 if (m->dirty) { 1802 if (m->object->type == OBJT_VNODE) { 1803 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc); 1804 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC); 1805 VOP_UNLOCK(m->object->handle, 0, curproc); 1806 goto again1; 1807 } else if (m->object->type == OBJT_SWAP || 1808 m->object->type == OBJT_DEFAULT) { 1809 vm_pageout_flush(&m, 1, 0); 1810 goto again1; 1811 } 1812 } 1813 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0)) 1814 vm_page_cache(m); 1815 } 1816 1817 for (m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1818 m != NULL; 1819 m = next) { 1820 1821 KASSERT(m->queue == PQ_ACTIVE, 1822 ("contigmalloc1: page %p is not PQ_ACTIVE", m)); 1823 1824 next = TAILQ_NEXT(m, pageq); 1825 if (vm_page_sleep_busy(m, TRUE, "vpctw1")) 1826 goto again1; 1827 vm_page_test_dirty(m); 1828 if (m->dirty) { 1829 if (m->object->type == OBJT_VNODE) { 1830 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc); 1831 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC); 1832 VOP_UNLOCK(m->object->handle, 0, curproc); 1833 goto again1; 1834 } else if (m->object->type == OBJT_SWAP || 1835 m->object->type == OBJT_DEFAULT) { 1836 vm_pageout_flush(&m, 1, 0); 1837 goto again1; 1838 } 1839 } 1840 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0)) 1841 vm_page_cache(m); 1842 } 1843 1844 splx(s); 1845 continue; 1846 } 1847 start = i; 1848 1849 /* 1850 * Check successive pages for contiguous and free. 1851 */ 1852 for (i = start + 1; i < (start + size / PAGE_SIZE); i++) { 1853 int pqtype; 1854 pqtype = pga[i].queue - pga[i].pc; 1855 if ((VM_PAGE_TO_PHYS(&pga[i]) != 1856 (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) || 1857 ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) { 1858 start++; 1859 goto again; 1860 } 1861 } 1862 1863 for (i = start; i < (start + size / PAGE_SIZE); i++) { 1864 int pqtype; 1865 vm_page_t m = &pga[i]; 1866 1867 pqtype = m->queue - m->pc; 1868 if (pqtype == PQ_CACHE) { 1869 vm_page_busy(m); 1870 vm_page_free(m); 1871 } 1872 1873 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq); 1874 vm_page_queues[m->queue].lcnt--; 1875 cnt.v_free_count--; 1876 m->valid = VM_PAGE_BITS_ALL; 1877 m->flags = 0; 1878 KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m)); 1879 m->wire_count = 0; 1880 m->busy = 0; 1881 m->queue = PQ_NONE; 1882 m->object = NULL; 1883 vm_page_wire(m); 1884 } 1885 1886 /* 1887 * We've found a contiguous chunk that meets are requirements. 1888 * Allocate kernel VM, unfree and assign the physical pages to it and 1889 * return kernel VM pointer. 1890 */ 1891 tmp_addr = addr = kmem_alloc_pageable(map, size); 1892 if (addr == 0) { 1893 /* 1894 * XXX We almost never run out of kernel virtual 1895 * space, so we don't make the allocated memory 1896 * above available. 1897 */ 1898 splx(s); 1899 return (NULL); 1900 } 1901 1902 for (i = start; i < (start + size / PAGE_SIZE); i++) { 1903 vm_page_t m = &pga[i]; 1904 vm_page_insert(m, kernel_object, 1905 OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS)); 1906 pmap_kenter(tmp_addr, VM_PAGE_TO_PHYS(m)); 1907 tmp_addr += PAGE_SIZE; 1908 } 1909 1910 splx(s); 1911 return ((void *)addr); 1912 } 1913 return NULL; 1914} 1915 1916void * 1917contigmalloc(size, type, flags, low, high, alignment, boundary) 1918 unsigned long size; /* should be size_t here and for malloc() */ 1919 struct malloc_type *type; 1920 int flags; 1921 unsigned long low; 1922 unsigned long high; 1923 unsigned long alignment; 1924 unsigned long boundary; 1925{ 1926 return contigmalloc1(size, type, flags, low, high, alignment, boundary, 1927 kernel_map); 1928} 1929 1930void 1931contigfree(addr, size, type) 1932 void *addr; 1933 unsigned long size; 1934 struct malloc_type *type; 1935{ 1936 kmem_free(kernel_map, (vm_offset_t)addr, size); 1937} 1938 1939vm_offset_t 1940vm_page_alloc_contig(size, low, high, alignment) 1941 vm_offset_t size; 1942 vm_offset_t low; 1943 vm_offset_t high; 1944 vm_offset_t alignment; 1945{ 1946 return ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high, 1947 alignment, 0ul, kernel_map)); 1948} 1949 1950#include "opt_ddb.h" 1951#ifdef DDB 1952#include <sys/kernel.h> 1953 1954#include <ddb/ddb.h> 1955 1956DB_SHOW_COMMAND(page, vm_page_print_page_info) 1957{ 1958 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1959 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1960 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1961 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1962 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1963 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1964 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1965 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1966 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1967 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1968} 1969 1970DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1971{ 1972 int i; 1973 db_printf("PQ_FREE:"); 1974 for(i=0;i<PQ_L2_SIZE;i++) { 1975 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1976 } 1977 db_printf("\n"); 1978 1979 db_printf("PQ_CACHE:"); 1980 for(i=0;i<PQ_L2_SIZE;i++) { 1981 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1982 } 1983 db_printf("\n"); 1984 1985 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1986 vm_page_queues[PQ_ACTIVE].lcnt, 1987 vm_page_queues[PQ_INACTIVE].lcnt); 1988} 1989#endif /* DDB */ 1990