vm_page.c revision 81397
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 81397 2001-08-10 06:37:05Z jhb $ 38 */ 39 40/* 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 */ 66 67/* 68 * GENERAL RULES ON VM_PAGE MANIPULATION 69 * 70 * - a pageq mutex is required when adding or removing a page from a 71 * page queue (vm_page_queue[]), regardless of other mutexes or the 72 * busy state of a page. 73 * 74 * - a hash chain mutex is required when associating or disassociating 75 * a page from the VM PAGE CACHE hash table (vm_page_buckets), 76 * regardless of other mutexes or the busy state of a page. 77 * 78 * - either a hash chain mutex OR a busied page is required in order 79 * to modify the page flags. A hash chain mutex must be obtained in 80 * order to busy a page. A page's flags cannot be modified by a 81 * hash chain mutex if the page is marked busy. 82 * 83 * - The object memq mutex is held when inserting or removing 84 * pages from an object (vm_page_insert() or vm_page_remove()). This 85 * is different from the object's main mutex. 86 * 87 * Generally speaking, you have to be aware of side effects when running 88 * vm_page ops. A vm_page_lookup() will return with the hash chain 89 * locked, whether it was able to lookup the page or not. vm_page_free(), 90 * vm_page_cache(), vm_page_activate(), and a number of other routines 91 * will release the hash chain mutex for you. Intermediate manipulation 92 * routines such as vm_page_flag_set() expect the hash chain to be held 93 * on entry and the hash chain will remain held on return. 94 * 95 * pageq scanning can only occur with the pageq in question locked. 96 * We have a known bottleneck with the active queue, but the cache 97 * and free queues are actually arrays already. 98 */ 99 100/* 101 * Resident memory management module. 102 */ 103 104#include <sys/param.h> 105#include <sys/systm.h> 106#include <sys/lock.h> 107#include <sys/malloc.h> 108#include <sys/mutex.h> 109#include <sys/proc.h> 110#include <sys/vmmeter.h> 111#include <sys/vnode.h> 112 113#include <vm/vm.h> 114#include <vm/vm_param.h> 115#include <vm/vm_kern.h> 116#include <vm/vm_object.h> 117#include <vm/vm_page.h> 118#include <vm/vm_pageout.h> 119#include <vm/vm_pager.h> 120#include <vm/vm_extern.h> 121 122/* 123 * Associated with page of user-allocatable memory is a 124 * page structure. 125 */ 126 127static struct vm_page **vm_page_buckets; /* Array of buckets */ 128static int vm_page_bucket_count; /* How big is array? */ 129static int vm_page_hash_mask; /* Mask for hash function */ 130static volatile int vm_page_bucket_generation; 131static struct mtx vm_buckets_mtx[BUCKET_HASH_SIZE]; 132 133vm_page_t vm_page_array = 0; 134int vm_page_array_size = 0; 135long first_page = 0; 136int vm_page_zero_count = 0; 137 138/* 139 * vm_set_page_size: 140 * 141 * Sets the page size, perhaps based upon the memory 142 * size. Must be called before any use of page-size 143 * dependent functions. 144 */ 145void 146vm_set_page_size(void) 147{ 148 if (cnt.v_page_size == 0) 149 cnt.v_page_size = PAGE_SIZE; 150 if (((cnt.v_page_size - 1) & cnt.v_page_size) != 0) 151 panic("vm_set_page_size: page size not a power of two"); 152} 153 154/* 155 * vm_page_startup: 156 * 157 * Initializes the resident memory module. 158 * 159 * Allocates memory for the page cells, and 160 * for the object/offset-to-page hash table headers. 161 * Each page cell is initialized and placed on the free list. 162 */ 163 164vm_offset_t 165vm_page_startup(vm_offset_t starta, vm_offset_t enda, vm_offset_t vaddr) 166{ 167 vm_offset_t mapped; 168 struct vm_page **bucket; 169 vm_size_t npages, page_range; 170 vm_offset_t new_end; 171 int i; 172 vm_offset_t pa; 173 int nblocks; 174 vm_offset_t last_pa; 175 176 /* the biggest memory array is the second group of pages */ 177 vm_offset_t end; 178 vm_offset_t biggestone, biggestsize; 179 180 vm_offset_t total; 181 182 total = 0; 183 biggestsize = 0; 184 biggestone = 0; 185 nblocks = 0; 186 vaddr = round_page(vaddr); 187 188 for (i = 0; phys_avail[i + 1]; i += 2) { 189 phys_avail[i] = round_page(phys_avail[i]); 190 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]); 191 } 192 193 for (i = 0; phys_avail[i + 1]; i += 2) { 194 int size = phys_avail[i + 1] - phys_avail[i]; 195 196 if (size > biggestsize) { 197 biggestone = i; 198 biggestsize = size; 199 } 200 ++nblocks; 201 total += size; 202 } 203 204 end = phys_avail[biggestone+1]; 205 206 /* 207 * Initialize the queue headers for the free queue, the active queue 208 * and the inactive queue. 209 */ 210 211 vm_pageq_init(); 212 213 /* 214 * Allocate (and initialize) the hash table buckets. 215 * 216 * The number of buckets MUST BE a power of 2, and the actual value is 217 * the next power of 2 greater than the number of physical pages in 218 * the system. 219 * 220 * We make the hash table approximately 2x the number of pages to 221 * reduce the chain length. This is about the same size using the 222 * singly-linked list as the 1x hash table we were using before 223 * using TAILQ but the chain length will be smaller. 224 * 225 * Note: This computation can be tweaked if desired. 226 */ 227 if (vm_page_bucket_count == 0) { 228 vm_page_bucket_count = 1; 229 while (vm_page_bucket_count < atop(total)) 230 vm_page_bucket_count <<= 1; 231 } 232 vm_page_bucket_count <<= 1; 233 vm_page_hash_mask = vm_page_bucket_count - 1; 234 235 /* 236 * Validate these addresses. 237 */ 238 new_end = end - vm_page_bucket_count * sizeof(struct vm_page *); 239 new_end = trunc_page(new_end); 240 mapped = pmap_map(&vaddr, new_end, end, 241 VM_PROT_READ | VM_PROT_WRITE); 242 bzero((caddr_t) mapped, end - new_end); 243 244 vm_page_buckets = (struct vm_page **)mapped; 245 bucket = vm_page_buckets; 246 for (i = 0; i < vm_page_bucket_count; i++) { 247 *bucket = NULL; 248 bucket++; 249 } 250 for (i = 0; i < BUCKET_HASH_SIZE; ++i) 251 mtx_init(&vm_buckets_mtx[i], "vm buckets hash mutexes", MTX_DEF); 252 253 /* 254 * Compute the number of pages of memory that will be available for 255 * use (taking into account the overhead of a page structure per 256 * page). 257 */ 258 259 first_page = phys_avail[0] / PAGE_SIZE; 260 261 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page; 262 npages = (total - (page_range * sizeof(struct vm_page)) - 263 (end - new_end)) / PAGE_SIZE; 264 265 end = new_end; 266 267 /* 268 * Initialize the mem entry structures now, and put them in the free 269 * queue. 270 */ 271 new_end = trunc_page(end - page_range * sizeof(struct vm_page)); 272 mapped = pmap_map(&vaddr, new_end, end, 273 VM_PROT_READ | VM_PROT_WRITE); 274 vm_page_array = (vm_page_t) mapped; 275 276 /* 277 * Clear all of the page structures 278 */ 279 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page)); 280 vm_page_array_size = page_range; 281 282 /* 283 * Construct the free queue(s) in descending order (by physical 284 * address) so that the first 16MB of physical memory is allocated 285 * last rather than first. On large-memory machines, this avoids 286 * the exhaustion of low physical memory before isa_dmainit has run. 287 */ 288 cnt.v_page_count = 0; 289 cnt.v_free_count = 0; 290 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) { 291 pa = phys_avail[i]; 292 if (i == biggestone) 293 last_pa = new_end; 294 else 295 last_pa = phys_avail[i + 1]; 296 while (pa < last_pa && npages-- > 0) { 297 vm_pageq_add_new_page(pa); 298 pa += PAGE_SIZE; 299 } 300 } 301 return (vaddr); 302} 303 304/* 305 * vm_page_hash: 306 * 307 * Distributes the object/offset key pair among hash buckets. 308 * 309 * NOTE: This macro depends on vm_page_bucket_count being a power of 2. 310 * This routine may not block. 311 * 312 * We try to randomize the hash based on the object to spread the pages 313 * out in the hash table without it costing us too much. 314 */ 315static __inline int 316vm_page_hash(vm_object_t object, vm_pindex_t pindex) 317{ 318 int i = ((uintptr_t)object + pindex) ^ object->hash_rand; 319 320 return(i & vm_page_hash_mask); 321} 322 323void 324vm_page_flag_set(vm_page_t m, unsigned short bits) 325{ 326 GIANT_REQUIRED; 327 m->flags |= bits; 328} 329 330void 331vm_page_flag_clear(vm_page_t m, unsigned short bits) 332{ 333 GIANT_REQUIRED; 334 m->flags &= ~bits; 335} 336 337void 338vm_page_busy(vm_page_t m) 339{ 340 KASSERT((m->flags & PG_BUSY) == 0, 341 ("vm_page_busy: page already busy!!!")); 342 vm_page_flag_set(m, PG_BUSY); 343} 344 345/* 346 * vm_page_flash: 347 * 348 * wakeup anyone waiting for the page. 349 */ 350 351void 352vm_page_flash(vm_page_t m) 353{ 354 if (m->flags & PG_WANTED) { 355 vm_page_flag_clear(m, PG_WANTED); 356 wakeup(m); 357 } 358} 359 360/* 361 * vm_page_wakeup: 362 * 363 * clear the PG_BUSY flag and wakeup anyone waiting for the 364 * page. 365 * 366 */ 367 368void 369vm_page_wakeup(vm_page_t m) 370{ 371 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!")); 372 vm_page_flag_clear(m, PG_BUSY); 373 vm_page_flash(m); 374} 375 376/* 377 * 378 * 379 */ 380 381void 382vm_page_io_start(vm_page_t m) 383{ 384 GIANT_REQUIRED; 385 m->busy++; 386} 387 388void 389vm_page_io_finish(vm_page_t m) 390{ 391 GIANT_REQUIRED; 392 m->busy--; 393 if (m->busy == 0) 394 vm_page_flash(m); 395} 396 397/* 398 * Keep page from being freed by the page daemon 399 * much of the same effect as wiring, except much lower 400 * overhead and should be used only for *very* temporary 401 * holding ("wiring"). 402 */ 403void 404vm_page_hold(vm_page_t mem) 405{ 406 GIANT_REQUIRED; 407 mem->hold_count++; 408} 409 410void 411vm_page_unhold(vm_page_t mem) 412{ 413 GIANT_REQUIRED; 414 --mem->hold_count; 415 KASSERT(mem->hold_count >= 0, ("vm_page_unhold: hold count < 0!!!")); 416} 417 418/* 419 * vm_page_protect: 420 * 421 * Reduce the protection of a page. This routine never raises the 422 * protection and therefore can be safely called if the page is already 423 * at VM_PROT_NONE (it will be a NOP effectively ). 424 */ 425 426void 427vm_page_protect(vm_page_t mem, int prot) 428{ 429 if (prot == VM_PROT_NONE) { 430 if (mem->flags & (PG_WRITEABLE|PG_MAPPED)) { 431 pmap_page_protect(mem, VM_PROT_NONE); 432 vm_page_flag_clear(mem, PG_WRITEABLE|PG_MAPPED); 433 } 434 } else if ((prot == VM_PROT_READ) && (mem->flags & PG_WRITEABLE)) { 435 pmap_page_protect(mem, VM_PROT_READ); 436 vm_page_flag_clear(mem, PG_WRITEABLE); 437 } 438} 439/* 440 * vm_page_zero_fill: 441 * 442 * Zero-fill the specified page. 443 * Written as a standard pagein routine, to 444 * be used by the zero-fill object. 445 */ 446boolean_t 447vm_page_zero_fill(vm_page_t m) 448{ 449 pmap_zero_page(VM_PAGE_TO_PHYS(m)); 450 return (TRUE); 451} 452 453/* 454 * vm_page_copy: 455 * 456 * Copy one page to another 457 */ 458void 459vm_page_copy(vm_page_t src_m, vm_page_t dest_m) 460{ 461 pmap_copy_page(VM_PAGE_TO_PHYS(src_m), VM_PAGE_TO_PHYS(dest_m)); 462 dest_m->valid = VM_PAGE_BITS_ALL; 463} 464 465/* 466 * vm_page_free: 467 * 468 * Free a page 469 * 470 * The clearing of PG_ZERO is a temporary safety until the code can be 471 * reviewed to determine that PG_ZERO is being properly cleared on 472 * write faults or maps. PG_ZERO was previously cleared in 473 * vm_page_alloc(). 474 */ 475void 476vm_page_free(vm_page_t m) 477{ 478 vm_page_flag_clear(m, PG_ZERO); 479 vm_page_free_toq(m); 480} 481 482/* 483 * vm_page_free_zero: 484 * 485 * Free a page to the zerod-pages queue 486 */ 487void 488vm_page_free_zero(vm_page_t m) 489{ 490 vm_page_flag_set(m, PG_ZERO); 491 vm_page_free_toq(m); 492} 493 494/* 495 * vm_page_sleep_busy: 496 * 497 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE) 498 * m->busy is zero. Returns TRUE if it had to sleep ( including if 499 * it almost had to sleep and made temporary spl*() mods), FALSE 500 * otherwise. 501 * 502 * This routine assumes that interrupts can only remove the busy 503 * status from a page, not set the busy status or change it from 504 * PG_BUSY to m->busy or vise versa (which would create a timing 505 * window). 506 */ 507 508int 509vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg) 510{ 511 GIANT_REQUIRED; 512 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 513 int s = splvm(); 514 if ((m->flags & PG_BUSY) || (also_m_busy && m->busy)) { 515 /* 516 * Page is busy. Wait and retry. 517 */ 518 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 519 tsleep(m, PVM, msg, 0); 520 } 521 splx(s); 522 return(TRUE); 523 /* not reached */ 524 } 525 return(FALSE); 526} 527/* 528 * vm_page_dirty: 529 * 530 * make page all dirty 531 */ 532 533void 534vm_page_dirty(vm_page_t m) 535{ 536 KASSERT(m->queue - m->pc != PQ_CACHE, 537 ("vm_page_dirty: page in cache!")); 538 m->dirty = VM_PAGE_BITS_ALL; 539} 540 541/* 542 * vm_page_undirty: 543 * 544 * Set page to not be dirty. Note: does not clear pmap modify bits 545 */ 546 547void 548vm_page_undirty(vm_page_t m) 549{ 550 m->dirty = 0; 551} 552 553/* 554 * vm_page_insert: [ internal use only ] 555 * 556 * Inserts the given mem entry into the object and object list. 557 * 558 * The pagetables are not updated but will presumably fault the page 559 * in if necessary, or if a kernel page the caller will at some point 560 * enter the page into the kernel's pmap. We are not allowed to block 561 * here so we *can't* do this anyway. 562 * 563 * The object and page must be locked, and must be splhigh. 564 * This routine may not block. 565 */ 566 567void 568vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex) 569{ 570 struct vm_page **bucket; 571 572 GIANT_REQUIRED; 573 574 if (m->object != NULL) 575 panic("vm_page_insert: already inserted"); 576 577 /* 578 * Record the object/offset pair in this page 579 */ 580 581 m->object = object; 582 m->pindex = pindex; 583 584 /* 585 * Insert it into the object_object/offset hash table 586 */ 587 588 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 589 m->hnext = *bucket; 590 *bucket = m; 591 vm_page_bucket_generation++; 592 593 /* 594 * Now link into the object's list of backed pages. 595 */ 596 597 TAILQ_INSERT_TAIL(&object->memq, m, listq); 598 object->generation++; 599 600 /* 601 * show that the object has one more resident page. 602 */ 603 604 object->resident_page_count++; 605 606 /* 607 * Since we are inserting a new and possibly dirty page, 608 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags. 609 */ 610 if (m->flags & PG_WRITEABLE) 611 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 612} 613 614/* 615 * vm_page_remove: 616 * NOTE: used by device pager as well -wfj 617 * 618 * Removes the given mem entry from the object/offset-page 619 * table and the object page list, but do not invalidate/terminate 620 * the backing store. 621 * 622 * The object and page must be locked, and at splhigh. 623 * The underlying pmap entry (if any) is NOT removed here. 624 * This routine may not block. 625 */ 626 627void 628vm_page_remove(vm_page_t m) 629{ 630 vm_object_t object; 631 632 GIANT_REQUIRED; 633 634 if (m->object == NULL) 635 return; 636 637 if ((m->flags & PG_BUSY) == 0) { 638 panic("vm_page_remove: page not busy"); 639 } 640 641 /* 642 * Basically destroy the page. 643 */ 644 645 vm_page_wakeup(m); 646 647 object = m->object; 648 649 /* 650 * Remove from the object_object/offset hash table. The object 651 * must be on the hash queue, we will panic if it isn't 652 * 653 * Note: we must NULL-out m->hnext to prevent loops in detached 654 * buffers with vm_page_lookup(). 655 */ 656 657 { 658 struct vm_page **bucket; 659 660 bucket = &vm_page_buckets[vm_page_hash(m->object, m->pindex)]; 661 while (*bucket != m) { 662 if (*bucket == NULL) 663 panic("vm_page_remove(): page not found in hash"); 664 bucket = &(*bucket)->hnext; 665 } 666 *bucket = m->hnext; 667 m->hnext = NULL; 668 vm_page_bucket_generation++; 669 } 670 671 /* 672 * Now remove from the object's list of backed pages. 673 */ 674 675 TAILQ_REMOVE(&object->memq, m, listq); 676 677 /* 678 * And show that the object has one fewer resident page. 679 */ 680 681 object->resident_page_count--; 682 object->generation++; 683 684 m->object = NULL; 685} 686 687/* 688 * vm_page_lookup: 689 * 690 * Returns the page associated with the object/offset 691 * pair specified; if none is found, NULL is returned. 692 * 693 * NOTE: the code below does not lock. It will operate properly if 694 * an interrupt makes a change, but the generation algorithm will not 695 * operate properly in an SMP environment where both cpu's are able to run 696 * kernel code simultaneously. 697 * 698 * The object must be locked. No side effects. 699 * This routine may not block. 700 * This is a critical path routine 701 */ 702 703vm_page_t 704vm_page_lookup(vm_object_t object, vm_pindex_t pindex) 705{ 706 vm_page_t m; 707 struct vm_page **bucket; 708 int generation; 709 710 /* 711 * Search the hash table for this object/offset pair 712 */ 713 714retry: 715 generation = vm_page_bucket_generation; 716 bucket = &vm_page_buckets[vm_page_hash(object, pindex)]; 717 for (m = *bucket; m != NULL; m = m->hnext) { 718 if ((m->object == object) && (m->pindex == pindex)) { 719 if (vm_page_bucket_generation != generation) 720 goto retry; 721 return (m); 722 } 723 } 724 if (vm_page_bucket_generation != generation) 725 goto retry; 726 return (NULL); 727} 728 729/* 730 * vm_page_rename: 731 * 732 * Move the given memory entry from its 733 * current object to the specified target object/offset. 734 * 735 * The object must be locked. 736 * This routine may not block. 737 * 738 * Note: this routine will raise itself to splvm(), the caller need not. 739 * 740 * Note: swap associated with the page must be invalidated by the move. We 741 * have to do this for several reasons: (1) we aren't freeing the 742 * page, (2) we are dirtying the page, (3) the VM system is probably 743 * moving the page from object A to B, and will then later move 744 * the backing store from A to B and we can't have a conflict. 745 * 746 * Note: we *always* dirty the page. It is necessary both for the 747 * fact that we moved it, and because we may be invalidating 748 * swap. If the page is on the cache, we have to deactivate it 749 * or vm_page_dirty() will panic. Dirty pages are not allowed 750 * on the cache. 751 */ 752 753void 754vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex) 755{ 756 int s; 757 758 s = splvm(); 759 vm_page_remove(m); 760 vm_page_insert(m, new_object, new_pindex); 761 if (m->queue - m->pc == PQ_CACHE) 762 vm_page_deactivate(m); 763 vm_page_dirty(m); 764 splx(s); 765} 766 767/* 768 * vm_page_select_cache: 769 * 770 * Find a page on the cache queue with color optimization. As pages 771 * might be found, but not applicable, they are deactivated. This 772 * keeps us from using potentially busy cached pages. 773 * 774 * This routine must be called at splvm(). 775 * This routine may not block. 776 */ 777static vm_page_t 778vm_page_select_cache(vm_object_t object, vm_pindex_t pindex) 779{ 780 vm_page_t m; 781 782 GIANT_REQUIRED; 783 while (TRUE) { 784 m = vm_pageq_find( 785 PQ_CACHE, 786 (pindex + object->pg_color) & PQ_L2_MASK, 787 FALSE 788 ); 789 if (m && ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || 790 m->hold_count || m->wire_count)) { 791 vm_page_deactivate(m); 792 continue; 793 } 794 return m; 795 } 796} 797 798/* 799 * vm_page_select_free: 800 * 801 * Find a free or zero page, with specified preference. 802 * 803 * This routine must be called at splvm(). 804 * This routine may not block. 805 */ 806 807static __inline vm_page_t 808vm_page_select_free(vm_object_t object, vm_pindex_t pindex, boolean_t prefer_zero) 809{ 810 vm_page_t m; 811 812 m = vm_pageq_find( 813 PQ_FREE, 814 (pindex + object->pg_color) & PQ_L2_MASK, 815 prefer_zero 816 ); 817 return(m); 818} 819 820/* 821 * vm_page_alloc: 822 * 823 * Allocate and return a memory cell associated 824 * with this VM object/offset pair. 825 * 826 * page_req classes: 827 * VM_ALLOC_NORMAL normal process request 828 * VM_ALLOC_SYSTEM system *really* needs a page 829 * VM_ALLOC_INTERRUPT interrupt time request 830 * VM_ALLOC_ZERO zero page 831 * 832 * This routine may not block. 833 * 834 * Additional special handling is required when called from an 835 * interrupt (VM_ALLOC_INTERRUPT). We are not allowed to mess with 836 * the page cache in this case. 837 */ 838 839vm_page_t 840vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req) 841{ 842 vm_page_t m = NULL; 843 int s; 844 845 GIANT_REQUIRED; 846 847 KASSERT(!vm_page_lookup(object, pindex), 848 ("vm_page_alloc: page already allocated")); 849 850 /* 851 * The pager is allowed to eat deeper into the free page list. 852 */ 853 854 if ((curproc == pageproc) && (page_req != VM_ALLOC_INTERRUPT)) { 855 page_req = VM_ALLOC_SYSTEM; 856 }; 857 858 s = splvm(); 859 860loop: 861 if (cnt.v_free_count > cnt.v_free_reserved) { 862 /* 863 * Allocate from the free queue if there are plenty of pages 864 * in it. 865 */ 866 if (page_req == VM_ALLOC_ZERO) 867 m = vm_page_select_free(object, pindex, TRUE); 868 else 869 m = vm_page_select_free(object, pindex, FALSE); 870 } else if ( 871 (page_req == VM_ALLOC_SYSTEM && 872 cnt.v_cache_count == 0 && 873 cnt.v_free_count > cnt.v_interrupt_free_min) || 874 (page_req == VM_ALLOC_INTERRUPT && cnt.v_free_count > 0) 875 ) { 876 /* 877 * Interrupt or system, dig deeper into the free list. 878 */ 879 m = vm_page_select_free(object, pindex, FALSE); 880 } else if (page_req != VM_ALLOC_INTERRUPT) { 881 /* 882 * Allocatable from cache (non-interrupt only). On success, 883 * we must free the page and try again, thus ensuring that 884 * cnt.v_*_free_min counters are replenished. 885 */ 886 m = vm_page_select_cache(object, pindex); 887 if (m == NULL) { 888 splx(s); 889#if defined(DIAGNOSTIC) 890 if (cnt.v_cache_count > 0) 891 printf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", cnt.v_cache_count); 892#endif 893 vm_pageout_deficit++; 894 pagedaemon_wakeup(); 895 return (NULL); 896 } 897 KASSERT(m->dirty == 0, ("Found dirty cache page %p", m)); 898 vm_page_busy(m); 899 vm_page_protect(m, VM_PROT_NONE); 900 vm_page_free(m); 901 goto loop; 902 } else { 903 /* 904 * Not allocatable from cache from interrupt, give up. 905 */ 906 splx(s); 907 vm_pageout_deficit++; 908 pagedaemon_wakeup(); 909 return (NULL); 910 } 911 912 /* 913 * At this point we had better have found a good page. 914 */ 915 916 KASSERT( 917 m != NULL, 918 ("vm_page_alloc(): missing page on free queue\n") 919 ); 920 921 /* 922 * Remove from free queue 923 */ 924 925 vm_pageq_remove_nowakeup(m); 926 927 /* 928 * Initialize structure. Only the PG_ZERO flag is inherited. 929 */ 930 931 if (m->flags & PG_ZERO) { 932 vm_page_zero_count--; 933 m->flags = PG_ZERO | PG_BUSY; 934 } else { 935 m->flags = PG_BUSY; 936 } 937 m->wire_count = 0; 938 m->hold_count = 0; 939 m->act_count = 0; 940 m->busy = 0; 941 m->valid = 0; 942 KASSERT(m->dirty == 0, ("vm_page_alloc: free/cache page %p was dirty", m)); 943 944 /* 945 * vm_page_insert() is safe prior to the splx(). Note also that 946 * inserting a page here does not insert it into the pmap (which 947 * could cause us to block allocating memory). We cannot block 948 * anywhere. 949 */ 950 951 vm_page_insert(m, object, pindex); 952 953 /* 954 * Don't wakeup too often - wakeup the pageout daemon when 955 * we would be nearly out of memory. 956 */ 957 if (vm_paging_needed()) 958 pagedaemon_wakeup(); 959 960 splx(s); 961 962 return (m); 963} 964 965/* 966 * vm_wait: (also see VM_WAIT macro) 967 * 968 * Block until free pages are available for allocation 969 */ 970 971void 972vm_wait(void) 973{ 974 int s; 975 976 s = splvm(); 977 if (curproc == pageproc) { 978 vm_pageout_pages_needed = 1; 979 tsleep(&vm_pageout_pages_needed, PSWP, "VMWait", 0); 980 } else { 981 if (!vm_pages_needed) { 982 vm_pages_needed = 1; 983 wakeup(&vm_pages_needed); 984 } 985 tsleep(&cnt.v_free_count, PVM, "vmwait", 0); 986 } 987 splx(s); 988} 989 990/* 991 * vm_page_activate: 992 * 993 * Put the specified page on the active list (if appropriate). 994 * Ensure that act_count is at least ACT_INIT but do not otherwise 995 * mess with it. 996 * 997 * The page queues must be locked. 998 * This routine may not block. 999 */ 1000void 1001vm_page_activate(vm_page_t m) 1002{ 1003 int s; 1004 1005 GIANT_REQUIRED; 1006 s = splvm(); 1007 1008 if (m->queue != PQ_ACTIVE) { 1009 if ((m->queue - m->pc) == PQ_CACHE) 1010 cnt.v_reactivated++; 1011 1012 vm_pageq_remove(m); 1013 1014 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1015 m->queue = PQ_ACTIVE; 1016 vm_page_queues[PQ_ACTIVE].lcnt++; 1017 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1018 if (m->act_count < ACT_INIT) 1019 m->act_count = ACT_INIT; 1020 cnt.v_active_count++; 1021 } 1022 } else { 1023 if (m->act_count < ACT_INIT) 1024 m->act_count = ACT_INIT; 1025 } 1026 1027 splx(s); 1028} 1029 1030/* 1031 * vm_page_free_wakeup: 1032 * 1033 * Helper routine for vm_page_free_toq() and vm_page_cache(). This 1034 * routine is called when a page has been added to the cache or free 1035 * queues. 1036 * 1037 * This routine may not block. 1038 * This routine must be called at splvm() 1039 */ 1040static __inline void 1041vm_page_free_wakeup(void) 1042{ 1043 /* 1044 * if pageout daemon needs pages, then tell it that there are 1045 * some free. 1046 */ 1047 if (vm_pageout_pages_needed && 1048 cnt.v_cache_count + cnt.v_free_count >= cnt.v_pageout_free_min) { 1049 wakeup(&vm_pageout_pages_needed); 1050 vm_pageout_pages_needed = 0; 1051 } 1052 /* 1053 * wakeup processes that are waiting on memory if we hit a 1054 * high water mark. And wakeup scheduler process if we have 1055 * lots of memory. this process will swapin processes. 1056 */ 1057 if (vm_pages_needed && !vm_page_count_min()) { 1058 vm_pages_needed = 0; 1059 wakeup(&cnt.v_free_count); 1060 } 1061} 1062 1063/* 1064 * vm_page_free_toq: 1065 * 1066 * Returns the given page to the PQ_FREE list, 1067 * disassociating it with any VM object. 1068 * 1069 * Object and page must be locked prior to entry. 1070 * This routine may not block. 1071 */ 1072 1073void 1074vm_page_free_toq(vm_page_t m) 1075{ 1076 int s; 1077 struct vpgqueues *pq; 1078 vm_object_t object = m->object; 1079 1080 GIANT_REQUIRED; 1081 s = splvm(); 1082 cnt.v_tfree++; 1083 1084 if (m->busy || ((m->queue - m->pc) == PQ_FREE) || 1085 (m->hold_count != 0)) { 1086 printf( 1087 "vm_page_free: pindex(%lu), busy(%d), PG_BUSY(%d), hold(%d)\n", 1088 (u_long)m->pindex, m->busy, (m->flags & PG_BUSY) ? 1 : 0, 1089 m->hold_count); 1090 if ((m->queue - m->pc) == PQ_FREE) 1091 panic("vm_page_free: freeing free page"); 1092 else 1093 panic("vm_page_free: freeing busy page"); 1094 } 1095 1096 /* 1097 * unqueue, then remove page. Note that we cannot destroy 1098 * the page here because we do not want to call the pager's 1099 * callback routine until after we've put the page on the 1100 * appropriate free queue. 1101 */ 1102 1103 vm_pageq_remove_nowakeup(m); 1104 vm_page_remove(m); 1105 1106 /* 1107 * If fictitious remove object association and 1108 * return, otherwise delay object association removal. 1109 */ 1110 1111 if ((m->flags & PG_FICTITIOUS) != 0) { 1112 splx(s); 1113 return; 1114 } 1115 1116 m->valid = 0; 1117 vm_page_undirty(m); 1118 1119 if (m->wire_count != 0) { 1120 if (m->wire_count > 1) { 1121 panic("vm_page_free: invalid wire count (%d), pindex: 0x%lx", 1122 m->wire_count, (long)m->pindex); 1123 } 1124 panic("vm_page_free: freeing wired page\n"); 1125 } 1126 1127 /* 1128 * If we've exhausted the object's resident pages we want to free 1129 * it up. 1130 */ 1131 1132 if (object && 1133 (object->type == OBJT_VNODE) && 1134 ((object->flags & OBJ_DEAD) == 0) 1135 ) { 1136 struct vnode *vp = (struct vnode *)object->handle; 1137 1138 if (vp && VSHOULDFREE(vp)) 1139 vfree(vp); 1140 } 1141 1142 /* 1143 * Clear the UNMANAGED flag when freeing an unmanaged page. 1144 */ 1145 1146 if (m->flags & PG_UNMANAGED) { 1147 m->flags &= ~PG_UNMANAGED; 1148 } else { 1149#ifdef __alpha__ 1150 pmap_page_is_free(m); 1151#endif 1152 } 1153 1154 m->queue = PQ_FREE + m->pc; 1155 pq = &vm_page_queues[m->queue]; 1156 pq->lcnt++; 1157 ++(*pq->cnt); 1158 1159 /* 1160 * Put zero'd pages on the end ( where we look for zero'd pages 1161 * first ) and non-zerod pages at the head. 1162 */ 1163 1164 if (m->flags & PG_ZERO) { 1165 TAILQ_INSERT_TAIL(&pq->pl, m, pageq); 1166 ++vm_page_zero_count; 1167 } else { 1168 TAILQ_INSERT_HEAD(&pq->pl, m, pageq); 1169 } 1170 1171 vm_page_free_wakeup(); 1172 1173 splx(s); 1174} 1175 1176/* 1177 * vm_page_unmanage: 1178 * 1179 * Prevent PV management from being done on the page. The page is 1180 * removed from the paging queues as if it were wired, and as a 1181 * consequence of no longer being managed the pageout daemon will not 1182 * touch it (since there is no way to locate the pte mappings for the 1183 * page). madvise() calls that mess with the pmap will also no longer 1184 * operate on the page. 1185 * 1186 * Beyond that the page is still reasonably 'normal'. Freeing the page 1187 * will clear the flag. 1188 * 1189 * This routine is used by OBJT_PHYS objects - objects using unswappable 1190 * physical memory as backing store rather then swap-backed memory and 1191 * will eventually be extended to support 4MB unmanaged physical 1192 * mappings. 1193 */ 1194 1195void 1196vm_page_unmanage(vm_page_t m) 1197{ 1198 int s; 1199 1200 s = splvm(); 1201 if ((m->flags & PG_UNMANAGED) == 0) { 1202 if (m->wire_count == 0) 1203 vm_pageq_remove(m); 1204 } 1205 vm_page_flag_set(m, PG_UNMANAGED); 1206 splx(s); 1207} 1208 1209/* 1210 * vm_page_wire: 1211 * 1212 * Mark this page as wired down by yet 1213 * another map, removing it from paging queues 1214 * as necessary. 1215 * 1216 * The page queues must be locked. 1217 * This routine may not block. 1218 */ 1219void 1220vm_page_wire(vm_page_t m) 1221{ 1222 int s; 1223 1224 /* 1225 * Only bump the wire statistics if the page is not already wired, 1226 * and only unqueue the page if it is on some queue (if it is unmanaged 1227 * it is already off the queues). 1228 */ 1229 s = splvm(); 1230 if (m->wire_count == 0) { 1231 if ((m->flags & PG_UNMANAGED) == 0) 1232 vm_pageq_remove(m); 1233 cnt.v_wire_count++; 1234 } 1235 m->wire_count++; 1236 splx(s); 1237 vm_page_flag_set(m, PG_MAPPED); 1238} 1239 1240/* 1241 * vm_page_unwire: 1242 * 1243 * Release one wiring of this page, potentially 1244 * enabling it to be paged again. 1245 * 1246 * Many pages placed on the inactive queue should actually go 1247 * into the cache, but it is difficult to figure out which. What 1248 * we do instead, if the inactive target is well met, is to put 1249 * clean pages at the head of the inactive queue instead of the tail. 1250 * This will cause them to be moved to the cache more quickly and 1251 * if not actively re-referenced, freed more quickly. If we just 1252 * stick these pages at the end of the inactive queue, heavy filesystem 1253 * meta-data accesses can cause an unnecessary paging load on memory bound 1254 * processes. This optimization causes one-time-use metadata to be 1255 * reused more quickly. 1256 * 1257 * BUT, if we are in a low-memory situation we have no choice but to 1258 * put clean pages on the cache queue. 1259 * 1260 * A number of routines use vm_page_unwire() to guarantee that the page 1261 * will go into either the inactive or active queues, and will NEVER 1262 * be placed in the cache - for example, just after dirtying a page. 1263 * dirty pages in the cache are not allowed. 1264 * 1265 * The page queues must be locked. 1266 * This routine may not block. 1267 */ 1268void 1269vm_page_unwire(vm_page_t m, int activate) 1270{ 1271 int s; 1272 1273 s = splvm(); 1274 1275 if (m->wire_count > 0) { 1276 m->wire_count--; 1277 if (m->wire_count == 0) { 1278 cnt.v_wire_count--; 1279 if (m->flags & PG_UNMANAGED) { 1280 ; 1281 } else if (activate) { 1282 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq); 1283 m->queue = PQ_ACTIVE; 1284 vm_page_queues[PQ_ACTIVE].lcnt++; 1285 cnt.v_active_count++; 1286 } else { 1287 vm_page_flag_clear(m, PG_WINATCFLS); 1288 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1289 m->queue = PQ_INACTIVE; 1290 vm_page_queues[PQ_INACTIVE].lcnt++; 1291 cnt.v_inactive_count++; 1292 } 1293 } 1294 } else { 1295 panic("vm_page_unwire: invalid wire count: %d\n", m->wire_count); 1296 } 1297 splx(s); 1298} 1299 1300 1301/* 1302 * Move the specified page to the inactive queue. If the page has 1303 * any associated swap, the swap is deallocated. 1304 * 1305 * Normally athead is 0 resulting in LRU operation. athead is set 1306 * to 1 if we want this page to be 'as if it were placed in the cache', 1307 * except without unmapping it from the process address space. 1308 * 1309 * This routine may not block. 1310 */ 1311static __inline void 1312_vm_page_deactivate(vm_page_t m, int athead) 1313{ 1314 int s; 1315 1316 GIANT_REQUIRED; 1317 /* 1318 * Ignore if already inactive. 1319 */ 1320 if (m->queue == PQ_INACTIVE) 1321 return; 1322 1323 s = splvm(); 1324 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) { 1325 if ((m->queue - m->pc) == PQ_CACHE) 1326 cnt.v_reactivated++; 1327 vm_page_flag_clear(m, PG_WINATCFLS); 1328 vm_pageq_remove(m); 1329 if (athead) 1330 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1331 else 1332 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq); 1333 m->queue = PQ_INACTIVE; 1334 vm_page_queues[PQ_INACTIVE].lcnt++; 1335 cnt.v_inactive_count++; 1336 } 1337 splx(s); 1338} 1339 1340void 1341vm_page_deactivate(vm_page_t m) 1342{ 1343 _vm_page_deactivate(m, 0); 1344} 1345 1346/* 1347 * vm_page_try_to_cache: 1348 * 1349 * Returns 0 on failure, 1 on success 1350 */ 1351int 1352vm_page_try_to_cache(vm_page_t m) 1353{ 1354 GIANT_REQUIRED; 1355 1356 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1357 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1358 return(0); 1359 } 1360 vm_page_test_dirty(m); 1361 if (m->dirty) 1362 return(0); 1363 vm_page_cache(m); 1364 return(1); 1365} 1366 1367/* 1368 * vm_page_try_to_free() 1369 * 1370 * Attempt to free the page. If we cannot free it, we do nothing. 1371 * 1 is returned on success, 0 on failure. 1372 */ 1373int 1374vm_page_try_to_free(vm_page_t m) 1375{ 1376 if (m->dirty || m->hold_count || m->busy || m->wire_count || 1377 (m->flags & (PG_BUSY|PG_UNMANAGED))) { 1378 return(0); 1379 } 1380 vm_page_test_dirty(m); 1381 if (m->dirty) 1382 return(0); 1383 vm_page_busy(m); 1384 vm_page_protect(m, VM_PROT_NONE); 1385 vm_page_free(m); 1386 return(1); 1387} 1388 1389/* 1390 * vm_page_cache 1391 * 1392 * Put the specified page onto the page cache queue (if appropriate). 1393 * 1394 * This routine may not block. 1395 */ 1396void 1397vm_page_cache(vm_page_t m) 1398{ 1399 int s; 1400 1401 GIANT_REQUIRED; 1402 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) || m->busy || m->wire_count) { 1403 printf("vm_page_cache: attempting to cache busy page\n"); 1404 return; 1405 } 1406 if ((m->queue - m->pc) == PQ_CACHE) 1407 return; 1408 1409 /* 1410 * Remove all pmaps and indicate that the page is not 1411 * writeable or mapped. 1412 */ 1413 1414 vm_page_protect(m, VM_PROT_NONE); 1415 if (m->dirty != 0) { 1416 panic("vm_page_cache: caching a dirty page, pindex: %ld", 1417 (long)m->pindex); 1418 } 1419 s = splvm(); 1420 vm_pageq_remove_nowakeup(m); 1421 m->queue = PQ_CACHE + m->pc; 1422 vm_page_queues[m->queue].lcnt++; 1423 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq); 1424 cnt.v_cache_count++; 1425 vm_page_free_wakeup(); 1426 splx(s); 1427} 1428 1429/* 1430 * vm_page_dontneed 1431 * 1432 * Cache, deactivate, or do nothing as appropriate. This routine 1433 * is typically used by madvise() MADV_DONTNEED. 1434 * 1435 * Generally speaking we want to move the page into the cache so 1436 * it gets reused quickly. However, this can result in a silly syndrome 1437 * due to the page recycling too quickly. Small objects will not be 1438 * fully cached. On the otherhand, if we move the page to the inactive 1439 * queue we wind up with a problem whereby very large objects 1440 * unnecessarily blow away our inactive and cache queues. 1441 * 1442 * The solution is to move the pages based on a fixed weighting. We 1443 * either leave them alone, deactivate them, or move them to the cache, 1444 * where moving them to the cache has the highest weighting. 1445 * By forcing some pages into other queues we eventually force the 1446 * system to balance the queues, potentially recovering other unrelated 1447 * space from active. The idea is to not force this to happen too 1448 * often. 1449 */ 1450 1451void 1452vm_page_dontneed(vm_page_t m) 1453{ 1454 static int dnweight; 1455 int dnw; 1456 int head; 1457 1458 GIANT_REQUIRED; 1459 dnw = ++dnweight; 1460 1461 /* 1462 * occassionally leave the page alone 1463 */ 1464 1465 if ((dnw & 0x01F0) == 0 || 1466 m->queue == PQ_INACTIVE || 1467 m->queue - m->pc == PQ_CACHE 1468 ) { 1469 if (m->act_count >= ACT_INIT) 1470 --m->act_count; 1471 return; 1472 } 1473 1474 if (m->dirty == 0) 1475 vm_page_test_dirty(m); 1476 1477 if (m->dirty || (dnw & 0x0070) == 0) { 1478 /* 1479 * Deactivate the page 3 times out of 32. 1480 */ 1481 head = 0; 1482 } else { 1483 /* 1484 * Cache the page 28 times out of every 32. Note that 1485 * the page is deactivated instead of cached, but placed 1486 * at the head of the queue instead of the tail. 1487 */ 1488 head = 1; 1489 } 1490 _vm_page_deactivate(m, head); 1491} 1492 1493/* 1494 * Grab a page, waiting until we are waken up due to the page 1495 * changing state. We keep on waiting, if the page continues 1496 * to be in the object. If the page doesn't exist, allocate it. 1497 * 1498 * This routine may block. 1499 */ 1500vm_page_t 1501vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags) 1502{ 1503 vm_page_t m; 1504 int s, generation; 1505 1506 GIANT_REQUIRED; 1507retrylookup: 1508 if ((m = vm_page_lookup(object, pindex)) != NULL) { 1509 if (m->busy || (m->flags & PG_BUSY)) { 1510 generation = object->generation; 1511 1512 s = splvm(); 1513 while ((object->generation == generation) && 1514 (m->busy || (m->flags & PG_BUSY))) { 1515 vm_page_flag_set(m, PG_WANTED | PG_REFERENCED); 1516 tsleep(m, PVM, "pgrbwt", 0); 1517 if ((allocflags & VM_ALLOC_RETRY) == 0) { 1518 splx(s); 1519 return NULL; 1520 } 1521 } 1522 splx(s); 1523 goto retrylookup; 1524 } else { 1525 vm_page_busy(m); 1526 return m; 1527 } 1528 } 1529 1530 m = vm_page_alloc(object, pindex, allocflags & ~VM_ALLOC_RETRY); 1531 if (m == NULL) { 1532 VM_WAIT; 1533 if ((allocflags & VM_ALLOC_RETRY) == 0) 1534 return NULL; 1535 goto retrylookup; 1536 } 1537 1538 return m; 1539} 1540 1541/* 1542 * Mapping function for valid bits or for dirty bits in 1543 * a page. May not block. 1544 * 1545 * Inputs are required to range within a page. 1546 */ 1547 1548__inline int 1549vm_page_bits(int base, int size) 1550{ 1551 int first_bit; 1552 int last_bit; 1553 1554 KASSERT( 1555 base + size <= PAGE_SIZE, 1556 ("vm_page_bits: illegal base/size %d/%d", base, size) 1557 ); 1558 1559 if (size == 0) /* handle degenerate case */ 1560 return(0); 1561 1562 first_bit = base >> DEV_BSHIFT; 1563 last_bit = (base + size - 1) >> DEV_BSHIFT; 1564 1565 return ((2 << last_bit) - (1 << first_bit)); 1566} 1567 1568/* 1569 * vm_page_set_validclean: 1570 * 1571 * Sets portions of a page valid and clean. The arguments are expected 1572 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive 1573 * of any partial chunks touched by the range. The invalid portion of 1574 * such chunks will be zero'd. 1575 * 1576 * This routine may not block. 1577 * 1578 * (base + size) must be less then or equal to PAGE_SIZE. 1579 */ 1580void 1581vm_page_set_validclean(vm_page_t m, int base, int size) 1582{ 1583 int pagebits; 1584 int frag; 1585 int endoff; 1586 1587 GIANT_REQUIRED; 1588 if (size == 0) /* handle degenerate case */ 1589 return; 1590 1591 /* 1592 * If the base is not DEV_BSIZE aligned and the valid 1593 * bit is clear, we have to zero out a portion of the 1594 * first block. 1595 */ 1596 1597 if ((frag = base & ~(DEV_BSIZE - 1)) != base && 1598 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0 1599 ) { 1600 pmap_zero_page_area( 1601 VM_PAGE_TO_PHYS(m), 1602 frag, 1603 base - frag 1604 ); 1605 } 1606 1607 /* 1608 * If the ending offset is not DEV_BSIZE aligned and the 1609 * valid bit is clear, we have to zero out a portion of 1610 * the last block. 1611 */ 1612 1613 endoff = base + size; 1614 1615 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff && 1616 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0 1617 ) { 1618 pmap_zero_page_area( 1619 VM_PAGE_TO_PHYS(m), 1620 endoff, 1621 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)) 1622 ); 1623 } 1624 1625 /* 1626 * Set valid, clear dirty bits. If validating the entire 1627 * page we can safely clear the pmap modify bit. We also 1628 * use this opportunity to clear the PG_NOSYNC flag. If a process 1629 * takes a write fault on a MAP_NOSYNC memory area the flag will 1630 * be set again. 1631 */ 1632 1633 pagebits = vm_page_bits(base, size); 1634 m->valid |= pagebits; 1635 m->dirty &= ~pagebits; 1636 if (base == 0 && size == PAGE_SIZE) { 1637 pmap_clear_modify(m); 1638 vm_page_flag_clear(m, PG_NOSYNC); 1639 } 1640} 1641 1642#if 0 1643 1644void 1645vm_page_set_dirty(vm_page_t m, int base, int size) 1646{ 1647 m->dirty |= vm_page_bits(base, size); 1648} 1649 1650#endif 1651 1652void 1653vm_page_clear_dirty(vm_page_t m, int base, int size) 1654{ 1655 GIANT_REQUIRED; 1656 m->dirty &= ~vm_page_bits(base, size); 1657} 1658 1659/* 1660 * vm_page_set_invalid: 1661 * 1662 * Invalidates DEV_BSIZE'd chunks within a page. Both the 1663 * valid and dirty bits for the effected areas are cleared. 1664 * 1665 * May not block. 1666 */ 1667void 1668vm_page_set_invalid(vm_page_t m, int base, int size) 1669{ 1670 int bits; 1671 1672 GIANT_REQUIRED; 1673 bits = vm_page_bits(base, size); 1674 m->valid &= ~bits; 1675 m->dirty &= ~bits; 1676 m->object->generation++; 1677} 1678 1679/* 1680 * vm_page_zero_invalid() 1681 * 1682 * The kernel assumes that the invalid portions of a page contain 1683 * garbage, but such pages can be mapped into memory by user code. 1684 * When this occurs, we must zero out the non-valid portions of the 1685 * page so user code sees what it expects. 1686 * 1687 * Pages are most often semi-valid when the end of a file is mapped 1688 * into memory and the file's size is not page aligned. 1689 */ 1690 1691void 1692vm_page_zero_invalid(vm_page_t m, boolean_t setvalid) 1693{ 1694 int b; 1695 int i; 1696 1697 /* 1698 * Scan the valid bits looking for invalid sections that 1699 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the 1700 * valid bit may be set ) have already been zerod by 1701 * vm_page_set_validclean(). 1702 */ 1703 1704 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) { 1705 if (i == (PAGE_SIZE / DEV_BSIZE) || 1706 (m->valid & (1 << i)) 1707 ) { 1708 if (i > b) { 1709 pmap_zero_page_area( 1710 VM_PAGE_TO_PHYS(m), 1711 b << DEV_BSHIFT, 1712 (i - b) << DEV_BSHIFT 1713 ); 1714 } 1715 b = i + 1; 1716 } 1717 } 1718 1719 /* 1720 * setvalid is TRUE when we can safely set the zero'd areas 1721 * as being valid. We can do this if there are no cache consistancy 1722 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS. 1723 */ 1724 1725 if (setvalid) 1726 m->valid = VM_PAGE_BITS_ALL; 1727} 1728 1729/* 1730 * vm_page_is_valid: 1731 * 1732 * Is (partial) page valid? Note that the case where size == 0 1733 * will return FALSE in the degenerate case where the page is 1734 * entirely invalid, and TRUE otherwise. 1735 * 1736 * May not block. 1737 */ 1738 1739int 1740vm_page_is_valid(vm_page_t m, int base, int size) 1741{ 1742 int bits = vm_page_bits(base, size); 1743 1744 if (m->valid && ((m->valid & bits) == bits)) 1745 return 1; 1746 else 1747 return 0; 1748} 1749 1750/* 1751 * update dirty bits from pmap/mmu. May not block. 1752 */ 1753 1754void 1755vm_page_test_dirty(vm_page_t m) 1756{ 1757 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) { 1758 vm_page_dirty(m); 1759 } 1760} 1761 1762#include "opt_ddb.h" 1763#ifdef DDB 1764#include <sys/kernel.h> 1765 1766#include <ddb/ddb.h> 1767 1768DB_SHOW_COMMAND(page, vm_page_print_page_info) 1769{ 1770 db_printf("cnt.v_free_count: %d\n", cnt.v_free_count); 1771 db_printf("cnt.v_cache_count: %d\n", cnt.v_cache_count); 1772 db_printf("cnt.v_inactive_count: %d\n", cnt.v_inactive_count); 1773 db_printf("cnt.v_active_count: %d\n", cnt.v_active_count); 1774 db_printf("cnt.v_wire_count: %d\n", cnt.v_wire_count); 1775 db_printf("cnt.v_free_reserved: %d\n", cnt.v_free_reserved); 1776 db_printf("cnt.v_free_min: %d\n", cnt.v_free_min); 1777 db_printf("cnt.v_free_target: %d\n", cnt.v_free_target); 1778 db_printf("cnt.v_cache_min: %d\n", cnt.v_cache_min); 1779 db_printf("cnt.v_inactive_target: %d\n", cnt.v_inactive_target); 1780} 1781 1782DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info) 1783{ 1784 int i; 1785 db_printf("PQ_FREE:"); 1786 for (i = 0; i < PQ_L2_SIZE; i++) { 1787 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt); 1788 } 1789 db_printf("\n"); 1790 1791 db_printf("PQ_CACHE:"); 1792 for (i = 0; i < PQ_L2_SIZE; i++) { 1793 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt); 1794 } 1795 db_printf("\n"); 1796 1797 db_printf("PQ_ACTIVE: %d, PQ_INACTIVE: %d\n", 1798 vm_page_queues[PQ_ACTIVE].lcnt, 1799 vm_page_queues[PQ_INACTIVE].lcnt); 1800} 1801#endif /* DDB */ 1802