vm_page.c revision 76827
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 76827 2001-05-19 01:28:09Z alfred $ 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/mutex.h> 75#include <sys/malloc.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 int i; 107 108 for(i=0;i<PQ_L2_SIZE;i++) { 109 vm_page_queues[PQ_FREE+i].cnt = &cnt.v_free_count; 110 } 111 vm_page_queues[PQ_INACTIVE].cnt = &cnt.v_inactive_count; 112 113 vm_page_queues[PQ_ACTIVE].cnt = &cnt.v_active_count; 114 for(i=0;i<PQ_L2_SIZE;i++) { 115 vm_page_queues[PQ_CACHE+i].cnt = &cnt.v_cache_count; 116 } 117 for(i=0;i<PQ_COUNT;i++) { 118 TAILQ_INIT(&vm_page_queues[i].pl); 119 } 120} 121 122vm_page_t vm_page_array = 0; 123int vm_page_array_size = 0; 124long first_page = 0; 125int vm_page_zero_count = 0; 126 127static __inline int vm_page_hash __P((vm_object_t object, vm_pindex_t pindex)); 128static void vm_page_free_wakeup __P((void)); 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() 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 * Must be called with the vm_mtx held. 152 */ 153vm_page_t 154vm_add_new_page(pa) 155 vm_offset_t pa; 156{ 157 vm_page_t m; 158 159 mtx_assert(&vm_mtx, MA_OWNED); 160 ++cnt.v_page_count; 161 ++cnt.v_free_count; 162 m = PHYS_TO_VM_PAGE(pa); 163 m->phys_addr = pa; 164 m->flags = 0; 165 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK; 166 m->queue = m->pc + PQ_FREE; 167 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 168 vm_page_queues[m->queue].lcnt++; 169 return (m); 170} 171 172/* 173 * vm_page_startup: 174 * 175 * Initializes the resident memory module. 176 * 177 * Allocates memory for the page cells, and 178 * for the object/offset-to-page hash table headers. 179 * Each page cell is initialized and placed on the free list. 180 */ 181 182vm_offset_t 183vm_page_startup(starta, enda, vaddr) 184 register vm_offset_t starta; 185 vm_offset_t enda; 186 vm_offset_t vaddr; 187{ 188 register vm_offset_t mapped; 189 register struct vm_page **bucket; 190 vm_size_t npages, page_range; 191 register vm_offset_t new_end; 192 int i; 193 vm_offset_t pa; 194 int nblocks; 195 vm_offset_t last_pa; 196 197 /* the biggest memory array is the second group of pages */ 198 vm_offset_t end; 199 vm_offset_t biggestone, biggestsize; 200 201 vm_offset_t total; 202 203 total = 0; 204 biggestsize = 0; 205 biggestone = 0; 206 nblocks = 0; 207 vaddr = round_page(vaddr); 208 209 for (i = 0; phys_avail[i + 1]; i += 2) { 210 phys_avail[i] = round_page(phys_avail[i]); 211 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 212 } 213 214 for (i = 0; phys_avail[i + 1]; i += 2) { 215 int size = phys_avail[i + 1] - phys_avail[i]; 216 217 if (size > biggestsize) { 218 biggestone = i; 219 biggestsize = size; 220 } 221 ++nblocks; 222 total += size; 223 } 224 225 end = phys_avail[biggestone+1]; 226 227 /* 228 * Initialize the queue headers for the free queue, the active queue 229 * and the inactive queue. 230 */ 231 232 vm_page_queue_init(); 233 234 /* 235 * Allocate (and initialize) the hash table buckets. 236 * 237 * The number of buckets MUST BE a power of 2, and the actual value is 238 * the next power of 2 greater than the number of physical pages in 239 * the system. 240 * 241 * We make the hash table approximately 2x the number of pages to 242 * reduce the chain length. This is about the same size using the 243 * singly-linked list as the 1x hash table we were using before 244 * using TAILQ but the chain length will be smaller. 245 * 246 * Note: This computation can be tweaked if desired. 247 */ 248 if (vm_page_bucket_count == 0) { 249 vm_page_bucket_count = 1; 250 while (vm_page_bucket_count < atop(total)) 251 vm_page_bucket_count <<= 1; 252 } 253 vm_page_bucket_count <<= 1; 254 vm_page_hash_mask = vm_page_bucket_count - 1; 255 256 /* 257 * Validate these addresses. 258 */ 259 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 260 new_end = trunc_page(new_end); 261 mapped = pmap_map(&vaddr, new_end, end, 262 VM_PROT_READ | VM_PROT_WRITE); 263 bzero((caddr_t) mapped, end - new_end); 264 265 vm_page_buckets = (struct vm_page **)mapped; 266 bucket = vm_page_buckets; 267 for (i = 0; i < vm_page_bucket_count; i++) { 268 *bucket = NULL; 269 bucket++; 270 } 271 272 /* 273 * Compute the number of pages of memory that will be available for 274 * use (taking into account the overhead of a page structure per 275 * page). 276 */ 277 278 first_page = phys_avail[0] / PAGE_SIZE; 279 280 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 281 npages = (total - (page_range * sizeof(struct vm_page)) - 282 (end - new_end)) / PAGE_SIZE; 283 284 end = new_end; 285 286 /* 287 * Initialize the mem entry structures now, and put them in the free 288 * queue. 289 */ 290 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 291 mapped = pmap_map(&vaddr, new_end, end, 292 VM_PROT_READ | VM_PROT_WRITE); 293 vm_page_array = (vm_page_t) mapped; 294 295 /* 296 * Clear all of the page structures 297 */ 298 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 299 vm_page_array_size = page_range; 300 301 /* 302 * Construct the free queue(s) in descending order (by physical 303 * address) so that the first 16MB of physical memory is allocated 304 * last rather than first. On large-memory machines, this avoids 305 * the exhaustion of low physical memory before isa_dmainit has run. 306 */ 307 cnt.v_page_count = 0; 308 cnt.v_free_count = 0; 309 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 310 pa = phys_avail[i]; 311 if (i == biggestone) 312 last_pa = new_end; 313 else 314 last_pa = phys_avail[i + 1]; 315 while (pa < last_pa && npages-- > 0) { 316 vm_add_new_page(pa); 317 pa += PAGE_SIZE; 318 } 319 } 320 return (vaddr); 321} 322 323/* 324 * vm_page_hash: 325 * 326 * Distributes the object/offset key pair among hash buckets. 327 * 328 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 329 * This routine may not block. 330 * 331 * We try to randomize the hash based on the object to spread the pages 332 * out in the hash table without it costing us too much. 333 */ 334static __inline int 335vm_page_hash(object, pindex) 336 vm_object_t object; 337 vm_pindex_t pindex; 338{ 339 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 340 341 return(i & vm_page_hash_mask); 342} 343 344/* 345 * vm_page_insert: [ internal use only ] 346 * 347 * Inserts the given mem entry into the object and object list. 348 * 349 * The pagetables are not updated but will presumably fault the page 350 * in if necessary, or if a kernel page the caller will at some point 351 * enter the page into the kernel's pmap. We are not allowed to block 352 * here so we *can't* do this anyway. 353 * 354 * The object and page must be locked, and must be splhigh. 355 * This routine may not block. 356 */ 357 358void 359vm_page_insert(m, object, pindex) 360 register vm_page_t m; 361 register vm_object_t object; 362 register vm_pindex_t pindex; 363{ 364 register struct vm_page **bucket; 365 366 mtx_assert(&vm_mtx, MA_OWNED); 367 if (m->object != NULL) 368 panic("vm_page_insert: already inserted"); 369 370 /* 371 * Record the object/offset pair in this page 372 */ 373 374 m->object = object; 375 m->pindex = pindex; 376 377 /* 378 * Insert it into the object_object/offset hash table 379 */ 380 381 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 382 m->hnext = *bucket; 383 *bucket = m; 384 vm_page_bucket_generation++; 385 386 /* 387 * Now link into the object's list of backed pages. 388 */ 389 390 TAILQ_INSERT_TAIL(&object->memq, m, listq); 391 object->generation++; 392 393 /* 394 * show that the object has one more resident page. 395 */ 396 397 object->resident_page_count++; 398 399 /* 400 * Since we are inserting a new and possibly dirty page, 401 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 402 */ 403 if (m->flags & PG_WRITEABLE) 404 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 405} 406 407/* 408 * vm_page_remove: 409 * NOTE: used by device pager as well -wfj 410 * 411 * Removes the given mem entry from the object/offset-page 412 * table and the object page list, but do not invalidate/terminate 413 * the backing store. 414 * 415 * The object and page must be locked, and at splhigh. 416 * The underlying pmap entry (if any) is NOT removed here. 417 * This routine may not block. 418 */ 419 420void 421vm_page_remove(m) 422 vm_page_t m; 423{ 424 vm_object_t object; 425 426 mtx_assert(&vm_mtx, MA_OWNED); 427 if (m->object == NULL) 428 return; 429 430 if ((m->flags & PG_BUSY) == 0) { 431 panic("vm_page_remove: page not busy"); 432 } 433 434 /* 435 * Basically destroy the page. 436 */ 437 438 vm_page_wakeup(m); 439 440 object = m->object; 441 442 /* 443 * Remove from the object_object/offset hash table. The object 444 * must be on the hash queue, we will panic if it isn't 445 * 446 * Note: we must NULL-out m->hnext to prevent loops in detached 447 * buffers with vm_page_lookup(). 448 */ 449 450 { 451 struct vm_page **bucket; 452 453 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 454 while (*bucket != m) { 455 if (*bucket == NULL) 456 panic("vm_page_remove(): page not found in hash"); 457 bucket = &(*bucket)->hnext; 458 } 459 *bucket = m->hnext; 460 m->hnext = NULL; 461 vm_page_bucket_generation++; 462 } 463 464 /* 465 * Now remove from the object's list of backed pages. 466 */ 467 468 TAILQ_REMOVE(&object->memq, m, listq); 469 470 /* 471 * And show that the object has one fewer resident page. 472 */ 473 474 object->resident_page_count--; 475 object->generation++; 476 477 m->object = NULL; 478} 479 480/* 481 * vm_page_lookup: 482 * 483 * Returns the page associated with the object/offset 484 * pair specified; if none is found, NULL is returned. 485 * 486 * NOTE: the code below does not lock. It will operate properly if 487 * an interrupt makes a change, but the generation algorithm will not 488 * operate properly in an SMP environment where both cpu's are able to run 489 * kernel code simultaneously. 490 * NOTE: under the giant vm lock we should be ok, there should be 491 * no reason to check vm_page_bucket_generation 492 * 493 * The object must be locked. No side effects. 494 * This routine may not block. 495 * This is a critical path routine 496 */ 497 498vm_page_t 499vm_page_lookup(object, pindex) 500 register vm_object_t object; 501 register vm_pindex_t pindex; 502{ 503 register vm_page_t m; 504 register struct vm_page **bucket; 505 int generation; 506 507 /* 508 * Search the hash table for this object/offset pair 509 */ 510 511retry: 512 generation = vm_page_bucket_generation; 513 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 514 for (m = *bucket; m != NULL; m = m->hnext) { 515 if ((m->object == object) && (m->pindex == pindex)) { 516 if (vm_page_bucket_generation != generation) 517 goto retry; 518 return (m); 519 } 520 } 521 if (vm_page_bucket_generation != generation) 522 goto retry; 523 return (NULL); 524} 525 526/* 527 * vm_page_rename: 528 * 529 * Move the given memory entry from its 530 * current object to the specified target object/offset. 531 * 532 * The object must be locked. 533 * This routine may not block. 534 * 535 * Note: this routine will raise itself to splvm(), the caller need not. 536 * 537 * Note: swap associated with the page must be invalidated by the move. We 538 * have to do this for several reasons: (1) we aren't freeing the 539 * page, (2) we are dirtying the page, (3) the VM system is probably 540 * moving the page from object A to B, and will then later move 541 * the backing store from A to B and we can't have a conflict. 542 * 543 * Note: we *always* dirty the page. It is necessary both for the 544 * fact that we moved it, and because we may be invalidating 545 * swap. If the page is on the cache, we have to deactivate it 546 * or vm_page_dirty() will panic. Dirty pages are not allowed 547 * on the cache. 548 */ 549 550void 551vm_page_rename(m, new_object, new_pindex) 552 register vm_page_t m; 553 register vm_object_t new_object; 554 vm_pindex_t new_pindex; 555{ 556 int s; 557 558 s = splvm(); 559 vm_page_remove(m); 560 vm_page_insert(m, new_object, new_pindex); 561 if (m->queue - m->pc == PQ_CACHE) 562 vm_page_deactivate(m); 563 vm_page_dirty(m); 564 splx(s); 565} 566 567/* 568 * vm_page_unqueue_nowakeup: 569 * 570 * vm_page_unqueue() without any wakeup 571 * 572 * This routine must be called at splhigh(). 573 * This routine may not block. 574 */ 575 576void 577vm_page_unqueue_nowakeup(m) 578 vm_page_t m; 579{ 580 int queue = m->queue; 581 struct vpgqueues *pq; 582 if (queue != PQ_NONE) { 583 pq = &vm_page_queues[queue]; 584 m->queue = PQ_NONE; 585 TAILQ_REMOVE(&pq->pl, m, pageq); 586 (*pq->cnt)--; 587 pq->lcnt--; 588 } 589} 590 591/* 592 * vm_page_unqueue: 593 * 594 * Remove a page from its queue. 595 * 596 * This routine must be called at splhigh(). 597 * This routine may not block. 598 */ 599 600void 601vm_page_unqueue(m) 602 vm_page_t m; 603{ 604 int queue = m->queue; 605 struct vpgqueues *pq; 606 607 mtx_assert(&vm_mtx, MA_OWNED); 608 if (queue != PQ_NONE) { 609 m->queue = PQ_NONE; 610 pq = &vm_page_queues[queue]; 611 TAILQ_REMOVE(&pq->pl, m, pageq); 612 (*pq->cnt)--; 613 pq->lcnt--; 614 if ((queue - m->pc) == PQ_CACHE) { 615 if (vm_paging_needed()) 616 pagedaemon_wakeup(); 617 } 618 } 619} 620 621#if PQ_L2_SIZE > 1 622 623/* 624 * vm_page_list_find: 625 * 626 * Find a page on the specified queue with color optimization. 627 * 628 * The page coloring optimization attempts to locate a page 629 * that does not overload other nearby pages in the object in 630 * the cpu's L1 or L2 caches. We need this optimization because 631 * cpu caches tend to be physical caches, while object spaces tend 632 * to be virtual. 633 * 634 * This routine must be called at splvm(). 635 * This routine may not block. 636 * 637 * This routine may only be called from the vm_page_list_find() macro 638 * in vm_page.h 639 */ 640vm_page_t 641_vm_page_list_find(basequeue, index) 642 int basequeue, index; 643{ 644 int i; 645 vm_page_t m = NULL; 646 struct vpgqueues *pq; 647 648 mtx_assert(&vm_mtx, MA_OWNED); 649 pq = &vm_page_queues[basequeue]; 650 651 /* 652 * Note that for the first loop, index+i and index-i wind up at the 653 * same place. Even though this is not totally optimal, we've already 654 * blown it by missing the cache case so we do not care. 655 */ 656 657 for(i = PQ_L2_SIZE / 2; i > 0; --i) { 658 if ((m = TAILQ_FIRST(&pq[(index + i) & PQ_L2_MASK].pl)) != NULL) 659 break; 660 661 if ((m = TAILQ_FIRST(&pq[(index - i) & PQ_L2_MASK].pl)) != NULL) 662 break; 663 } 664 return(m); 665} 666 667#endif 668 669/* 670 * vm_page_select_cache: 671 * 672 * Find a page on the cache queue with color optimization. As pages 673 * might be found, but not applicable, they are deactivated. This 674 * keeps us from using potentially busy cached pages. 675 * 676 * This routine must be called at splvm(). 677 * This routine may not block. 678 */ 679vm_page_t 680vm_page_select_cache(object, pindex) 681 vm_object_t object; 682 vm_pindex_t pindex; 683{ 684 vm_page_t m; 685 686 mtx_assert(&vm_mtx, MA_OWNED); 687 while (TRUE) { 688 m = vm_page_list_find( 689 PQ_CACHE, 690 (pindex + object->pg_color) & PQ_L2_MASK, 691 FALSE 692 ); 693 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 694 m->hold_count || m->wire_count)) { 695 vm_page_deactivate(m); 696 continue; 697 } 698 return m; 699 } 700} 701 702/* 703 * vm_page_select_free: 704 * 705 * Find a free or zero page, with specified preference. We attempt to 706 * inline the nominal case and fall back to _vm_page_select_free() 707 * otherwise. 708 * 709 * This routine must be called at splvm(). 710 * This routine may not block. 711 */ 712 713static __inline vm_page_t 714vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 715{ 716 vm_page_t m; 717 718 m = vm_page_list_find( 719 PQ_FREE, 720 (pindex + object->pg_color) & PQ_L2_MASK, 721 prefer_zero 722 ); 723 return(m); 724} 725 726/* 727 * vm_page_alloc: 728 * 729 * Allocate and return a memory cell associated 730 * with this VM object/offset pair. 731 * 732 * page_req classes: 733 * VM_ALLOC_NORMAL normal process request 734 * VM_ALLOC_SYSTEM system *really* needs a page 735 * VM_ALLOC_INTERRUPT interrupt time request 736 * VM_ALLOC_ZERO zero page 737 * 738 * vm_mtx must be locked. 739 * This routine may not block. 740 * 741 * Additional special handling is required when called from an 742 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 743 * the page cache in this case. 744 */ 745 746vm_page_t 747vm_page_alloc(object, pindex, page_req) 748 vm_object_t object; 749 vm_pindex_t pindex; 750 int page_req; 751{ 752 register vm_page_t m = NULL; 753 int s; 754 755 mtx_assert(&vm_mtx, MA_OWNED); 756 KASSERT(!vm_page_lookup(object, pindex), 757 ("vm_page_alloc: page already allocated")); 758 759 /* 760 * The pager is allowed to eat deeper into the free page list. 761 */ 762 763 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 764 page_req = VM_ALLOC_SYSTEM; 765 }; 766 767 s = splvm(); 768 769loop: 770 if (cnt.v_free_count > cnt.v_free_reserved) { 771 /* 772 * Allocate from the free queue if there are plenty of pages 773 * in it. 774 */ 775 if (page_req == VM_ALLOC_ZERO) 776 m = vm_page_select_free(object, pindex, TRUE); 777 else 778 m = vm_page_select_free(object, pindex, FALSE); 779 } else if ( 780 (page_req == VM_ALLOC_SYSTEM && 781 cnt.v_cache_count == 0 && 782 cnt.v_free_count > cnt.v_interrupt_free_min) || 783 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 784 ) { 785 /* 786 * Interrupt or system, dig deeper into the free list. 787 */ 788 m = vm_page_select_free(object, pindex, FALSE); 789 } else if (page_req != VM_ALLOC_INTERRUPT) { 790 /* 791 * Allocatable from cache (non-interrupt only). On success, 792 * we must free the page and try again, thus ensuring that 793 * cnt.v_*_free_min counters are replenished. 794 */ 795 m = vm_page_select_cache(object, pindex); 796 if (m == NULL) { 797 splx(s); 798#if defined(DIAGNOSTIC) 799 if (cnt.v_cache_count > 0) 800 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 801#endif 802 vm_pageout_deficit++; 803 pagedaemon_wakeup(); 804 return (NULL); 805 } 806 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 807 vm_page_busy(m); 808 vm_page_protect(m, VM_PROT_NONE); 809 vm_page_free(m); 810 goto loop; 811 } else { 812 /* 813 * Not allocatable from cache from interrupt, give up. 814 */ 815 splx(s); 816 vm_pageout_deficit++; 817 pagedaemon_wakeup(); 818 return (NULL); 819 } 820 821 /* 822 * At this point we had better have found a good page. 823 */ 824 825 KASSERT( 826 m != NULL, 827 ("vm_page_alloc(): missing page on free queue\n") 828 ); 829 830 /* 831 * Remove from free queue 832 */ 833 834 vm_page_unqueue_nowakeup(m); 835 836 /* 837 * Initialize structure. Only the PG_ZERO flag is inherited. 838 */ 839 840 if (m->flags & PG_ZERO) { 841 vm_page_zero_count--; 842 m->flags = PG_ZERO | PG_BUSY; 843 } else { 844 m->flags = PG_BUSY; 845 } 846 m->wire_count = 0; 847 m->hold_count = 0; 848 m->act_count = 0; 849 m->busy = 0; 850 m->valid = 0; 851 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 852 853 /* 854 * vm_page_insert() is safe prior to the splx(). Note also that 855 * inserting a page here does not insert it into the pmap (which 856 * could cause us to block allocating memory). We cannot block 857 * anywhere. 858 */ 859 860 vm_page_insert(m, object, pindex); 861 862 /* 863 * Don't wakeup too often - wakeup the pageout daemon when 864 * we would be nearly out of memory. 865 */ 866 if (vm_paging_needed()) 867 pagedaemon_wakeup(); 868 869 splx(s); 870 871 return (m); 872} 873 874/* 875 * vm_wait: (also see VM_WAIT macro) 876 * 877 * Block until free pages are available for allocation 878 */ 879 880void 881vm_wait() 882{ 883 int s; 884 885 s = splvm(); 886 if (curproc == pageproc) { 887 vm_pageout_pages_needed = 1; 888 msleep(&vm_pageout_pages_needed, &vm_mtx, PSWP, "VMWait", 0); 889 } else { 890 if (!vm_pages_needed) { 891 vm_pages_needed = 1; 892 wakeup(&vm_pages_needed); 893 } 894 msleep(&cnt.v_free_count, &vm_mtx, PVM, "vmwait", 0); 895 } 896 splx(s); 897} 898 899/* 900 * vm_await: (also see VM_AWAIT macro) 901 * 902 * asleep on an event that will signal when free pages are available 903 * for allocation. 904 */ 905 906void 907vm_await() 908{ 909 int s; 910 911 s = splvm(); 912 if (curproc == pageproc) { 913 vm_pageout_pages_needed = 1; 914 asleep(&vm_pageout_pages_needed, PSWP, "vmwait", 0); 915 } else { 916 if (!vm_pages_needed) { 917 vm_pages_needed++; 918 wakeup(&vm_pages_needed); 919 } 920 asleep(&cnt.v_free_count, PVM, "vmwait", 0); 921 } 922 splx(s); 923} 924 925/* 926 * vm_page_activate: 927 * 928 * Put the specified page on the active list (if appropriate). 929 * Ensure that act_count is at least ACT_INIT but do not otherwise 930 * mess with it. 931 * 932 * The page queues must be locked. 933 * This routine may not block. 934 */ 935void 936vm_page_activate(m) 937 register vm_page_t m; 938{ 939 int s; 940 941 s = splvm(); 942 mtx_assert(&vm_mtx, MA_OWNED); 943 if (m->queue != PQ_ACTIVE) { 944 if ((m->queue - m->pc) == PQ_CACHE) 945 cnt.v_reactivated++; 946 947 vm_page_unqueue(m); 948 949 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 950 m->queue = PQ_ACTIVE; 951 vm_page_queues[PQ_ACTIVE].lcnt++; 952 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 953 if (m->act_count < ACT_INIT) 954 m->act_count = ACT_INIT; 955 cnt.v_active_count++; 956 } 957 } else { 958 if (m->act_count < ACT_INIT) 959 m->act_count = ACT_INIT; 960 } 961 962 splx(s); 963} 964 965/* 966 * vm_page_free_wakeup: 967 * 968 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 969 * routine is called when a page has been added to the cache or free 970 * queues. 971 * 972 * This routine may not block. 973 * This routine must be called at splvm() 974 */ 975static __inline void 976vm_page_free_wakeup() 977{ 978 /* 979 * if pageout daemon needs pages, then tell it that there are 980 * some free. 981 */ 982 if (vm_pageout_pages_needed && 983 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 984 wakeup(&vm_pageout_pages_needed); 985 vm_pageout_pages_needed = 0; 986 } 987 /* 988 * wakeup processes that are waiting on memory if we hit a 989 * high water mark. And wakeup scheduler process if we have 990 * lots of memory. this process will swapin processes. 991 */ 992 if (vm_pages_needed && !vm_page_count_min()) { 993 vm_pages_needed = 0; 994 wakeup(&cnt.v_free_count); 995 } 996} 997 998/* 999 * vm_page_free_toq: 1000 * 1001 * Returns the given page to the PQ_FREE list, 1002 * disassociating it with any VM object. 1003 * 1004 * Object and page must be locked prior to entry. 1005 * This routine may not block. 1006 */ 1007 1008void 1009vm_page_free_toq(vm_page_t m) 1010{ 1011 int s; 1012 struct vpgqueues *pq; 1013 vm_object_t object = m->object; 1014 1015 s = splvm(); 1016 1017 mtx_assert(&vm_mtx, MA_OWNED); 1018 cnt.v_tfree++; 1019 1020 if (m->busy || ((m->queue - m->pc) == PQ_FREE) || 1021 (m->hold_count != 0)) { 1022 printf( 1023 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1024 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1025 m->hold_count); 1026 if ((m->queue - m->pc) == PQ_FREE) 1027 panic("vm_page_free: freeing free page"); 1028 else 1029 panic("vm_page_free: freeing busy page"); 1030 } 1031 1032 /* 1033 * unqueue, then remove page. Note that we cannot destroy 1034 * the page here because we do not want to call the pager's 1035 * callback routine until after we've put the page on the 1036 * appropriate free queue. 1037 */ 1038 1039 vm_page_unqueue_nowakeup(m); 1040 vm_page_remove(m); 1041 1042 /* 1043 * If fictitious remove object association and 1044 * return, otherwise delay object association removal. 1045 */ 1046 1047 if ((m->flags & PG_FICTITIOUS) != 0) { 1048 splx(s); 1049 return; 1050 } 1051 1052 m->valid = 0; 1053 vm_page_undirty(m); 1054 1055 if (m->wire_count != 0) { 1056 if (m->wire_count > 1) { 1057 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1058 m->wire_count, (long)m->pindex); 1059 } 1060 panic("vm_page_free: freeing wired page\n"); 1061 } 1062 1063 /* 1064 * If we've exhausted the object's resident pages we want to free 1065 * it up. 1066 */ 1067 1068 if (object && 1069 (object->type == OBJT_VNODE) && 1070 ((object->flags & OBJ_DEAD) == 0) 1071 ) { 1072 struct vnode *vp = (struct vnode *)object->handle; 1073 1074 if (vp && VSHOULDFREE(vp)) 1075 vfree(vp); 1076 } 1077 1078 /* 1079 * Clear the UNMANAGED flag when freeing an unmanaged page. 1080 */ 1081 1082 if (m->flags & PG_UNMANAGED) { 1083 m->flags &= ~PG_UNMANAGED; 1084 } else { 1085#ifdef __alpha__ 1086 pmap_page_is_free(m); 1087#endif 1088 } 1089 1090 m->queue = PQ_FREE + m->pc; 1091 pq = &vm_page_queues[m->queue]; 1092 pq->lcnt++; 1093 ++(*pq->cnt); 1094 1095 /* 1096 * Put zero'd pages on the end ( where we look for zero'd pages 1097 * first ) and non-zerod pages at the head. 1098 */ 1099 1100 if (m->flags & PG_ZERO) { 1101 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1102 ++vm_page_zero_count; 1103 } else { 1104 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1105 } 1106 1107 vm_page_free_wakeup(); 1108 1109 splx(s); 1110} 1111 1112/* 1113 * vm_page_unmanage: 1114 * 1115 * Prevent PV management from being done on the page. The page is 1116 * removed from the paging queues as if it were wired, and as a 1117 * consequence of no longer being managed the pageout daemon will not 1118 * touch it (since there is no way to locate the pte mappings for the 1119 * page). madvise() calls that mess with the pmap will also no longer 1120 * operate on the page. 1121 * 1122 * Beyond that the page is still reasonably 'normal'. Freeing the page 1123 * will clear the flag. 1124 * 1125 * This routine is used by OBJT_PHYS objects - objects using unswappable 1126 * physical memory as backing store rather then swap-backed memory and 1127 * will eventually be extended to support 4MB unmanaged physical 1128 * mappings. 1129 */ 1130 1131void 1132vm_page_unmanage(vm_page_t m) 1133{ 1134 int s; 1135 1136 s = splvm(); 1137 if ((m->flags & PG_UNMANAGED) == 0) { 1138 if (m->wire_count == 0) 1139 vm_page_unqueue(m); 1140 } 1141 vm_page_flag_set(m, PG_UNMANAGED); 1142 splx(s); 1143} 1144 1145/* 1146 * vm_page_wire: 1147 * 1148 * Mark this page as wired down by yet 1149 * another map, removing it from paging queues 1150 * as necessary. 1151 * 1152 * The page queues must be locked. 1153 * This routine may not block. 1154 */ 1155void 1156vm_page_wire(m) 1157 register vm_page_t m; 1158{ 1159 int s; 1160 1161 /* 1162 * Only bump the wire statistics if the page is not already wired, 1163 * and only unqueue the page if it is on some queue (if it is unmanaged 1164 * it is already off the queues). 1165 */ 1166 s = splvm(); 1167 if (m->wire_count == 0) { 1168 if ((m->flags & PG_UNMANAGED) == 0) 1169 vm_page_unqueue(m); 1170 cnt.v_wire_count++; 1171 } 1172 m->wire_count++; 1173 splx(s); 1174 vm_page_flag_set(m, PG_MAPPED); 1175} 1176 1177/* 1178 * vm_page_unwire: 1179 * 1180 * Release one wiring of this page, potentially 1181 * enabling it to be paged again. 1182 * 1183 * Many pages placed on the inactive queue should actually go 1184 * into the cache, but it is difficult to figure out which. What 1185 * we do instead, if the inactive target is well met, is to put 1186 * clean pages at the head of the inactive queue instead of the tail. 1187 * This will cause them to be moved to the cache more quickly and 1188 * if not actively re-referenced, freed more quickly. If we just 1189 * stick these pages at the end of the inactive queue, heavy filesystem 1190 * meta-data accesses can cause an unnecessary paging load on memory bound 1191 * processes. This optimization causes one-time-use metadata to be 1192 * reused more quickly. 1193 * 1194 * BUT, if we are in a low-memory situation we have no choice but to 1195 * put clean pages on the cache queue. 1196 * 1197 * A number of routines use vm_page_unwire() to guarantee that the page 1198 * will go into either the inactive or active queues, and will NEVER 1199 * be placed in the cache - for example, just after dirtying a page. 1200 * dirty pages in the cache are not allowed. 1201 * 1202 * The page queues must be locked. 1203 * This routine may not block. 1204 */ 1205void 1206vm_page_unwire(m, activate) 1207 register vm_page_t m; 1208 int activate; 1209{ 1210 int s; 1211 1212 s = splvm(); 1213 1214 if (m->wire_count > 0) { 1215 m->wire_count--; 1216 if (m->wire_count == 0) { 1217 cnt.v_wire_count--; 1218 if (m->flags & PG_UNMANAGED) { 1219 ; 1220 } else if (activate) { 1221 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1222 m->queue = PQ_ACTIVE; 1223 vm_page_queues[PQ_ACTIVE].lcnt++; 1224 cnt.v_active_count++; 1225 } else { 1226 vm_page_flag_clear(m, PG_WINATCFLS); 1227 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1228 m->queue = PQ_INACTIVE; 1229 vm_page_queues[PQ_INACTIVE].lcnt++; 1230 cnt.v_inactive_count++; 1231 } 1232 } 1233 } else { 1234 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1235 } 1236 splx(s); 1237} 1238 1239 1240/* 1241 * Move the specified page to the inactive queue. If the page has 1242 * any associated swap, the swap is deallocated. 1243 * 1244 * Normally athead is 0 resulting in LRU operation. athead is set 1245 * to 1 if we want this page to be 'as if it were placed in the cache', 1246 * except without unmapping it from the process address space. 1247 * 1248 * This routine may not block. 1249 */ 1250static __inline void 1251_vm_page_deactivate(vm_page_t m, int athead) 1252{ 1253 int s; 1254 1255 mtx_assert(&vm_mtx, MA_OWNED); 1256 /* 1257 * Ignore if already inactive. 1258 */ 1259 if (m->queue == PQ_INACTIVE) 1260 return; 1261 1262 s = splvm(); 1263 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1264 if ((m->queue - m->pc) == PQ_CACHE) 1265 cnt.v_reactivated++; 1266 vm_page_flag_clear(m, PG_WINATCFLS); 1267 vm_page_unqueue(m); 1268 if (athead) 1269 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1270 else 1271 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1272 m->queue = PQ_INACTIVE; 1273 vm_page_queues[PQ_INACTIVE].lcnt++; 1274 cnt.v_inactive_count++; 1275 } 1276 splx(s); 1277} 1278 1279void 1280vm_page_deactivate(vm_page_t m) 1281{ 1282 _vm_page_deactivate(m, 0); 1283} 1284 1285/* 1286 * vm_page_try_to_cache: 1287 * 1288 * Returns 0 on failure, 1 on success 1289 */ 1290int 1291vm_page_try_to_cache(vm_page_t m) 1292{ 1293 1294 mtx_assert(VM_PAGE_MTX(m), MA_OWNED); 1295 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1296 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1297 return(0); 1298 } 1299 vm_page_test_dirty(m); 1300 if (m->dirty) 1301 return(0); 1302 vm_page_cache(m); 1303 return(1); 1304} 1305 1306/* 1307 * vm_page_cache 1308 * 1309 * Put the specified page onto the page cache queue (if appropriate). 1310 * 1311 * This routine may not block. 1312 */ 1313void 1314vm_page_cache(m) 1315 register vm_page_t m; 1316{ 1317 int s; 1318 1319 mtx_assert(&vm_mtx, MA_OWNED); 1320 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1321 printf("vm_page_cache: attempting to cache busy page\n"); 1322 return; 1323 } 1324 if ((m->queue - m->pc) == PQ_CACHE) 1325 return; 1326 1327 /* 1328 * Remove all pmaps and indicate that the page is not 1329 * writeable or mapped. 1330 */ 1331 1332 vm_page_protect(m, VM_PROT_NONE); 1333 if (m->dirty != 0) { 1334 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1335 (long)m->pindex); 1336 } 1337 s = splvm(); 1338 vm_page_unqueue_nowakeup(m); 1339 m->queue = PQ_CACHE + m->pc; 1340 vm_page_queues[m->queue].lcnt++; 1341 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1342 cnt.v_cache_count++; 1343 vm_page_free_wakeup(); 1344 splx(s); 1345} 1346 1347/* 1348 * vm_page_dontneed 1349 * 1350 * Cache, deactivate, or do nothing as appropriate. This routine 1351 * is typically used by madvise() MADV_DONTNEED. 1352 * 1353 * Generally speaking we want to move the page into the cache so 1354 * it gets reused quickly. However, this can result in a silly syndrome 1355 * due to the page recycling too quickly. Small objects will not be 1356 * fully cached. On the otherhand, if we move the page to the inactive 1357 * queue we wind up with a problem whereby very large objects 1358 * unnecessarily blow away our inactive and cache queues. 1359 * 1360 * The solution is to move the pages based on a fixed weighting. We 1361 * either leave them alone, deactivate them, or move them to the cache, 1362 * where moving them to the cache has the highest weighting. 1363 * By forcing some pages into other queues we eventually force the 1364 * system to balance the queues, potentially recovering other unrelated 1365 * space from active. The idea is to not force this to happen too 1366 * often. 1367 */ 1368 1369void 1370vm_page_dontneed(m) 1371 vm_page_t m; 1372{ 1373 static int dnweight; 1374 int dnw; 1375 int head; 1376 1377 mtx_assert(&vm_mtx, MA_OWNED); 1378 dnw = ++dnweight; 1379 1380 /* 1381 * occassionally leave the page alone 1382 */ 1383 1384 if ((dnw & 0x01F0) == 0 || 1385 m->queue == PQ_INACTIVE || 1386 m->queue - m->pc == PQ_CACHE 1387 ) { 1388 if (m->act_count >= ACT_INIT) 1389 --m->act_count; 1390 return; 1391 } 1392 1393 if (m->dirty == 0) 1394 vm_page_test_dirty(m); 1395 1396 if (m->dirty || (dnw & 0x0070) == 0) { 1397 /* 1398 * Deactivate the page 3 times out of 32. 1399 */ 1400 head = 0; 1401 } else { 1402 /* 1403 * Cache the page 28 times out of every 32. Note that 1404 * the page is deactivated instead of cached, but placed 1405 * at the head of the queue instead of the tail. 1406 */ 1407 head = 1; 1408 } 1409 _vm_page_deactivate(m, head); 1410} 1411 1412/* 1413 * Grab a page, waiting until we are waken up due to the page 1414 * changing state. We keep on waiting, if the page continues 1415 * to be in the object. If the page doesn't exist, allocate it. 1416 * 1417 * This routine may block. 1418 * Requires vm_mtx. 1419 */ 1420vm_page_t 1421vm_page_grab(object, pindex, allocflags) 1422 vm_object_t object; 1423 vm_pindex_t pindex; 1424 int allocflags; 1425{ 1426 vm_page_t m; 1427 int s, generation; 1428 1429 mtx_assert(&vm_mtx, MA_OWNED); 1430retrylookup: 1431 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1432 if (m->busy || (m->flags & PG_BUSY)) { 1433 generation = object->generation; 1434 1435 s = splvm(); 1436 while ((object->generation == generation) && 1437 (m->busy || (m->flags & PG_BUSY))) { 1438 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1439 msleep(m, &vm_mtx, PVM, "pgrbwt", 0); 1440 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1441 splx(s); 1442 return NULL; 1443 } 1444 } 1445 splx(s); 1446 goto retrylookup; 1447 } else { 1448 vm_page_busy(m); 1449 return m; 1450 } 1451 } 1452 1453 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1454 if (m == NULL) { 1455 VM_WAIT; 1456 if ((allocflags & VM_ALLOC_RETRY) == 0) 1457 return NULL; 1458 goto retrylookup; 1459 } 1460 1461 return m; 1462} 1463 1464/* 1465 * Mapping function for valid bits or for dirty bits in 1466 * a page. May not block. 1467 * 1468 * Inputs are required to range within a page. 1469 */ 1470 1471__inline int 1472vm_page_bits(int base, int size) 1473{ 1474 int first_bit; 1475 int last_bit; 1476 1477 KASSERT( 1478 base + size <= PAGE_SIZE, 1479 ("vm_page_bits: illegal base/size %d/%d", base, size) 1480 ); 1481 1482 if (size == 0) /* handle degenerate case */ 1483 return(0); 1484 1485 first_bit = base >> DEV_BSHIFT; 1486 last_bit = (base + size - 1) >> DEV_BSHIFT; 1487 1488 return ((2 << last_bit) - (1 << first_bit)); 1489} 1490 1491/* 1492 * vm_page_set_validclean: 1493 * 1494 * Sets portions of a page valid and clean. The arguments are expected 1495 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1496 * of any partial chunks touched by the range. The invalid portion of 1497 * such chunks will be zero'd. 1498 * 1499 * This routine may not block. 1500 * 1501 * (base + size) must be less then or equal to PAGE_SIZE. 1502 * 1503 * vm_mtx needs to be held 1504 */ 1505void 1506vm_page_set_validclean(m, base, size) 1507 vm_page_t m; 1508 int base; 1509 int size; 1510{ 1511 int pagebits; 1512 int frag; 1513 int endoff; 1514 1515 mtx_assert(&vm_mtx, MA_OWNED); 1516 if (size == 0) /* handle degenerate case */ 1517 return; 1518 1519 /* 1520 * If the base is not DEV_BSIZE aligned and the valid 1521 * bit is clear, we have to zero out a portion of the 1522 * first block. 1523 */ 1524 1525 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1526 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1527 ) { 1528 pmap_zero_page_area( 1529 VM_PAGE_TO_PHYS(m), 1530 frag, 1531 base - frag 1532 ); 1533 } 1534 1535 /* 1536 * If the ending offset is not DEV_BSIZE aligned and the 1537 * valid bit is clear, we have to zero out a portion of 1538 * the last block. 1539 */ 1540 1541 endoff = base + size; 1542 1543 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1544 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1545 ) { 1546 pmap_zero_page_area( 1547 VM_PAGE_TO_PHYS(m), 1548 endoff, 1549 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1550 ); 1551 } 1552 1553 /* 1554 * Set valid, clear dirty bits. If validating the entire 1555 * page we can safely clear the pmap modify bit. We also 1556 * use this opportunity to clear the PG_NOSYNC flag. If a process 1557 * takes a write fault on a MAP_NOSYNC memory area the flag will 1558 * be set again. 1559 */ 1560 1561 pagebits = vm_page_bits(base, size); 1562 m->valid |= pagebits; 1563 m->dirty &= ~pagebits; 1564 if (base == 0 && size == PAGE_SIZE) { 1565 pmap_clear_modify(m); 1566 vm_page_flag_clear(m, PG_NOSYNC); 1567 } 1568} 1569 1570#if 0 1571 1572void 1573vm_page_set_dirty(m, base, size) 1574 vm_page_t m; 1575 int base; 1576 int size; 1577{ 1578 m->dirty |= vm_page_bits(base, size); 1579} 1580 1581#endif 1582 1583void 1584vm_page_clear_dirty(m, base, size) 1585 vm_page_t m; 1586 int base; 1587 int size; 1588{ 1589 1590 mtx_assert(&vm_mtx, MA_OWNED); 1591 m->dirty &= ~vm_page_bits(base, size); 1592} 1593 1594/* 1595 * vm_page_set_invalid: 1596 * 1597 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1598 * valid and dirty bits for the effected areas are cleared. 1599 * 1600 * May not block. 1601 */ 1602void 1603vm_page_set_invalid(m, base, size) 1604 vm_page_t m; 1605 int base; 1606 int size; 1607{ 1608 int bits; 1609 1610 mtx_assert(&vm_mtx, MA_OWNED); 1611 bits = vm_page_bits(base, size); 1612 m->valid &= ~bits; 1613 m->dirty &= ~bits; 1614 m->object->generation++; 1615} 1616 1617/* 1618 * vm_page_zero_invalid() 1619 * 1620 * The kernel assumes that the invalid portions of a page contain 1621 * garbage, but such pages can be mapped into memory by user code. 1622 * When this occurs, we must zero out the non-valid portions of the 1623 * page so user code sees what it expects. 1624 * 1625 * Pages are most often semi-valid when the end of a file is mapped 1626 * into memory and the file's size is not page aligned. 1627 */ 1628 1629void 1630vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1631{ 1632 int b; 1633 int i; 1634 1635 /* 1636 * Scan the valid bits looking for invalid sections that 1637 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1638 * valid bit may be set ) have already been zerod by 1639 * vm_page_set_validclean(). 1640 */ 1641 1642 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1643 if (i == (PAGE_SIZE / DEV_BSIZE) || 1644 (m->valid & (1 << i)) 1645 ) { 1646 if (i > b) { 1647 pmap_zero_page_area( 1648 VM_PAGE_TO_PHYS(m), 1649 b << DEV_BSHIFT, 1650 (i - b) << DEV_BSHIFT 1651 ); 1652 } 1653 b = i + 1; 1654 } 1655 } 1656 1657 /* 1658 * setvalid is TRUE when we can safely set the zero'd areas 1659 * as being valid. We can do this if there are no cache consistancy 1660 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1661 */ 1662 1663 if (setvalid) 1664 m->valid = VM_PAGE_BITS_ALL; 1665} 1666 1667/* 1668 * vm_page_is_valid: 1669 * 1670 * Is (partial) page valid? Note that the case where size == 0 1671 * will return FALSE in the degenerate case where the page is 1672 * entirely invalid, and TRUE otherwise. 1673 * 1674 * May not block. 1675 */ 1676 1677int 1678vm_page_is_valid(m, base, size) 1679 vm_page_t m; 1680 int base; 1681 int size; 1682{ 1683 int bits = vm_page_bits(base, size); 1684 1685 if (m->valid && ((m->valid & bits) == bits)) 1686 return 1; 1687 else 1688 return 0; 1689} 1690 1691/* 1692 * update dirty bits from pmap/mmu. May not block. 1693 */ 1694 1695void 1696vm_page_test_dirty(m) 1697 vm_page_t m; 1698{ 1699 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1700 vm_page_dirty(m); 1701 } 1702} 1703 1704/* 1705 * This interface is for merging with malloc() someday. 1706 * Even if we never implement compaction so that contiguous allocation 1707 * works after initialization time, malloc()'s data structures are good 1708 * for statistics and for allocations of less than a page. 1709 */ 1710void * 1711contigmalloc1(size, type, flags, low, high, alignment, boundary, map) 1712 unsigned long size; /* should be size_t here and for malloc() */ 1713 struct malloc_type *type; 1714 int flags; 1715 unsigned long low; 1716 unsigned long high; 1717 unsigned long alignment; 1718 unsigned long boundary; 1719 vm_map_t map; 1720{ 1721 int i, s, start; 1722 vm_offset_t addr, phys, tmp_addr; 1723 int pass; 1724 vm_page_t pga = vm_page_array; 1725 1726 size = round_page(size); 1727 if (size == 0) 1728 panic("contigmalloc1: size must not be 0"); 1729 if ((alignment & (alignment - 1)) != 0) 1730 panic("contigmalloc1: alignment must be a power of 2"); 1731 if ((boundary & (boundary - 1)) != 0) 1732 panic("contigmalloc1: boundary must be a power of 2"); 1733 1734 start = 0; 1735 for (pass = 0; pass <= 1; pass++) { 1736 s = splvm(); 1737again: 1738 /* 1739 * Find first page in array that is free, within range, aligned, and 1740 * such that the boundary won't be crossed. 1741 */ 1742 for (i = start; i < cnt.v_page_count; i++) { 1743 int pqtype; 1744 phys = VM_PAGE_TO_PHYS(&pga[i]); 1745 pqtype = pga[i].queue - pga[i].pc; 1746 if (((pqtype == PQ_FREE) || (pqtype == PQ_CACHE)) && 1747 (phys >= low) && (phys < high) && 1748 ((phys & (alignment - 1)) == 0) && 1749 (((phys ^ (phys + size - 1)) & ~(boundary - 1)) == 0)) 1750 break; 1751 } 1752 1753 /* 1754 * If the above failed or we will exceed the upper bound, fail. 1755 */ 1756 if ((i == cnt.v_page_count) || 1757 ((VM_PAGE_TO_PHYS(&pga[i]) + size) > high)) { 1758 vm_page_t m, next; 1759 1760again1: 1761 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 1762 m != NULL; 1763 m = next) { 1764 1765 KASSERT(m->queue == PQ_INACTIVE, 1766 ("contigmalloc1: page %p is not PQ_INACTIVE", m)); 1767 1768 next = TAILQ_NEXT(m, pageq); 1769 if (vm_page_sleep_busy(m, TRUE, "vpctw0")) 1770 goto again1; 1771 vm_page_test_dirty(m); 1772 if (m->dirty) { 1773 if (m->object->type == OBJT_VNODE) { 1774 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc); 1775 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC); 1776 VOP_UNLOCK(m->object->handle, 0, curproc); 1777 goto again1; 1778 } else if (m->object->type == OBJT_SWAP || 1779 m->object->type == OBJT_DEFAULT) { 1780 vm_pageout_flush(&m, 1, 0); 1781 goto again1; 1782 } 1783 } 1784 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0)) 1785 vm_page_cache(m); 1786 } 1787 1788 for (m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1789 m != NULL; 1790 m = next) { 1791 1792 KASSERT(m->queue == PQ_ACTIVE, 1793 ("contigmalloc1: page %p is not PQ_ACTIVE", m)); 1794 1795 next = TAILQ_NEXT(m, pageq); 1796 if (vm_page_sleep_busy(m, TRUE, "vpctw1")) 1797 goto again1; 1798 vm_page_test_dirty(m); 1799 if (m->dirty) { 1800 if (m->object->type == OBJT_VNODE) { 1801 vn_lock(m->object->handle, LK_EXCLUSIVE | LK_RETRY, curproc); 1802 vm_object_page_clean(m->object, 0, 0, OBJPC_SYNC); 1803 VOP_UNLOCK(m->object->handle, 0, curproc); 1804 goto again1; 1805 } else if (m->object->type == OBJT_SWAP || 1806 m->object->type == OBJT_DEFAULT) { 1807 vm_pageout_flush(&m, 1, 0); 1808 goto again1; 1809 } 1810 } 1811 if ((m->dirty == 0) && (m->busy == 0) && (m->hold_count == 0)) 1812 vm_page_cache(m); 1813 } 1814 1815 splx(s); 1816 continue; 1817 } 1818 start = i; 1819 1820 /* 1821 * Check successive pages for contiguous and free. 1822 */ 1823 for (i = start + 1; i < (start + size / PAGE_SIZE); i++) { 1824 int pqtype; 1825 pqtype = pga[i].queue - pga[i].pc; 1826 if ((VM_PAGE_TO_PHYS(&pga[i]) != 1827 (VM_PAGE_TO_PHYS(&pga[i - 1]) + PAGE_SIZE)) || 1828 ((pqtype != PQ_FREE) && (pqtype != PQ_CACHE))) { 1829 start++; 1830 goto again; 1831 } 1832 } 1833 1834 for (i = start; i < (start + size / PAGE_SIZE); i++) { 1835 int pqtype; 1836 vm_page_t m = &pga[i]; 1837 1838 pqtype = m->queue - m->pc; 1839 if (pqtype == PQ_CACHE) { 1840 vm_page_busy(m); 1841 vm_page_free(m); 1842 } 1843 1844 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq); 1845 vm_page_queues[m->queue].lcnt--; 1846 cnt.v_free_count--; 1847 m->valid = VM_PAGE_BITS_ALL; 1848 m->flags = 0; 1849 KASSERT(m->dirty == 0, ("contigmalloc1: page %p was dirty", m)); 1850 m->wire_count = 0; 1851 m->busy = 0; 1852 m->queue = PQ_NONE; 1853 m->object = NULL; 1854 vm_page_wire(m); 1855 } 1856 1857 /* 1858 * We've found a contiguous chunk that meets are requirements. 1859 * Allocate kernel VM, unfree and assign the physical pages to it and 1860 * return kernel VM pointer. 1861 */ 1862 tmp_addr = addr = kmem_alloc_pageable(map, size); 1863 if (addr == 0) { 1864 /* 1865 * XXX We almost never run out of kernel virtual 1866 * space, so we don't make the allocated memory 1867 * above available. 1868 */ 1869 splx(s); 1870 return (NULL); 1871 } 1872 1873 for (i = start; i < (start + size / PAGE_SIZE); i++) { 1874 vm_page_t m = &pga[i]; 1875 vm_page_insert(m, kernel_object, 1876 OFF_TO_IDX(tmp_addr - VM_MIN_KERNEL_ADDRESS)); 1877 pmap_kenter(tmp_addr, VM_PAGE_TO_PHYS(m)); 1878 tmp_addr += PAGE_SIZE; 1879 } 1880 1881 splx(s); 1882 return ((void *)addr); 1883 } 1884 return NULL; 1885} 1886 1887void * 1888contigmalloc(size, type, flags, low, high, alignment, boundary) 1889 unsigned long size; /* should be size_t here and for malloc() */ 1890 struct malloc_type *type; 1891 int flags; 1892 unsigned long low; 1893 unsigned long high; 1894 unsigned long alignment; 1895 unsigned long boundary; 1896{ 1897 void * ret; 1898 int hadvmlock; 1899 1900 hadvmlock = mtx_owned(&vm_mtx); 1901 if (!hadvmlock) 1902 mtx_lock(&vm_mtx); 1903 ret = contigmalloc1(size, type, flags, low, high, alignment, boundary, 1904 kernel_map); 1905 if (!hadvmlock) 1906 mtx_unlock(&vm_mtx); 1907 1908 return (ret); 1909 1910} 1911 1912void 1913contigfree(addr, size, type) 1914 void *addr; 1915 unsigned long size; 1916 struct malloc_type *type; 1917{ 1918 int hadvmlock; 1919 1920 hadvmlock = mtx_owned(&vm_mtx); 1921 if (!hadvmlock) 1922 mtx_lock(&vm_mtx); 1923 kmem_free(kernel_map, (vm_offset_t)addr, size); 1924 if (!hadvmlock) 1925 mtx_unlock(&vm_mtx); 1926} 1927 1928vm_offset_t 1929vm_page_alloc_contig(size, low, high, alignment) 1930 vm_offset_t size; 1931 vm_offset_t low; 1932 vm_offset_t high; 1933 vm_offset_t alignment; 1934{ 1935 vm_offset_t ret; 1936 int hadvmlock; 1937 1938 hadvmlock = mtx_owned(&vm_mtx); 1939 if (!hadvmlock) 1940 mtx_lock(&vm_mtx); 1941 ret = ((vm_offset_t)contigmalloc1(size, M_DEVBUF, M_NOWAIT, low, high, 1942 alignment, 0ul, kernel_map)); 1943 if (!hadvmlock) 1944 mtx_unlock(&vm_mtx); 1945 return (ret); 1946 1947} 1948 1949#include "opt_ddb.h" 1950#ifdef DDB 1951#include <sys/kernel.h> 1952 1953#include <ddb/ddb.h> 1954 1955DB_SHOW_COMMAND(page, vm_page_print_page_info) 1956{ 1957 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1958 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1959 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1960 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1961 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1962 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1963 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1964 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1965 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1966 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1967} 1968 1969DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1970{ 1971 int i; 1972 db_printf("PQ_FREE:"); 1973 for(i=0;i<PQ_L2_SIZE;i++) { 1974 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1975 } 1976 db_printf("\n"); 1977 1978 db_printf("PQ_CACHE:"); 1979 for(i=0;i<PQ_L2_SIZE;i++) { 1980 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1981 } 1982 db_printf("\n"); 1983 1984 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1985 vm_page_queues[PQ_ACTIVE].lcnt, 1986 vm_page_queues[PQ_INACTIVE].lcnt); 1987} 1988#endif /* DDB */ 1989