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