Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_fault.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 42 * 43 * 44 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 45 * All rights reserved. 46 * 47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 * 69 * $FreeBSD$ 70 */ 71 72 /* 73 * Page fault handling module. 74 */ 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/proc.h> 79 #include <sys/vnode.h> 80 #include <sys/resourcevar.h> 81 #include <sys/vmmeter.h> 82 83 #include <vm/vm.h> 84 #include <vm/vm_param.h> 85 #include <vm/vm_prot.h> 86 #include <sys/lock.h> 87 #include <vm/pmap.h> 88 #include <vm/vm_map.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_page.h> 91 #include <vm/vm_pageout.h> 92 #include <vm/vm_kern.h> 93 #include <vm/vm_pager.h> 94 #include <vm/vnode_pager.h> 95 #include <vm/vm_extern.h> 96 97 static int vm_fault_additional_pages __P((vm_page_t, int, 98 int, vm_page_t *, int *)); 99 100 #define VM_FAULT_READ_AHEAD 8 101 #define VM_FAULT_READ_BEHIND 7 102 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 103 104 struct faultstate { 105 vm_page_t m; 106 vm_object_t object; 107 vm_pindex_t pindex; 108 vm_page_t first_m; 109 vm_object_t first_object; 110 vm_pindex_t first_pindex; 111 vm_map_t map; 112 vm_map_entry_t entry; 113 int lookup_still_valid; 114 struct vnode *vp; 115 }; 116 117 static void 118 release_page(struct faultstate *fs) 119 { 120 vm_page_wakeup(fs->m); 121 vm_page_deactivate(fs->m); 122 fs->m = NULL; 123 } 124 125 static void 126 unlock_map(struct faultstate *fs) 127 { 128 if (fs->lookup_still_valid) { 129 vm_map_lookup_done(fs->map, fs->entry); 130 fs->lookup_still_valid = FALSE; 131 } 132 } 133 134 static void 135 _unlock_things(struct faultstate *fs, int dealloc) 136 { 137 vm_object_pip_wakeup(fs->object); 138 if (fs->object != fs->first_object) { 139 vm_page_free(fs->first_m); 140 vm_object_pip_wakeup(fs->first_object); 141 fs->first_m = NULL; 142 } 143 if (dealloc) { 144 vm_object_deallocate(fs->first_object); 145 } 146 unlock_map(fs); 147 if (fs->vp != NULL) { 148 vput(fs->vp); 149 fs->vp = NULL; 150 } 151 } 152 153 #define unlock_things(fs) _unlock_things(fs, 0) 154 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 155 156 /* 157 * vm_fault: 158 * 159 * Handle a page fault occuring at the given address, 160 * requiring the given permissions, in the map specified. 161 * If successful, the page is inserted into the 162 * associated physical map. 163 * 164 * NOTE: the given address should be truncated to the 165 * proper page address. 166 * 167 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 168 * a standard error specifying why the fault is fatal is returned. 169 * 170 * 171 * The map in question must be referenced, and remains so. 172 * Caller may hold no locks. 173 */ 174 int 175 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 176 { 177 vm_prot_t prot; 178 int result; 179 boolean_t wired; 180 int map_generation; 181 vm_page_t old_m; 182 vm_object_t next_object; 183 vm_page_t marray[VM_FAULT_READ]; 184 int hardfault; 185 int faultcount; 186 struct faultstate fs; 187 188 cnt.v_vm_faults++; /* needs lock XXX */ 189 hardfault = 0; 190 191 RetryFault:; 192 fs.map = map; 193 194 /* 195 * Find the backing store object and offset into it to begin the 196 * search. 197 */ 198 if ((result = vm_map_lookup(&fs.map, vaddr, 199 fault_type, &fs.entry, &fs.first_object, 200 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) { 201 if ((result != KERN_PROTECTION_FAILURE) || 202 ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) { 203 return result; 204 } 205 206 /* 207 * If we are user-wiring a r/w segment, and it is COW, then 208 * we need to do the COW operation. Note that we don't COW 209 * currently RO sections now, because it is NOT desirable 210 * to COW .text. We simply keep .text from ever being COW'ed 211 * and take the heat that one cannot debug wired .text sections. 212 */ 213 result = vm_map_lookup(&fs.map, vaddr, 214 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 215 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 216 if (result != KERN_SUCCESS) { 217 return result; 218 } 219 220 /* 221 * If we don't COW now, on a user wire, the user will never 222 * be able to write to the mapping. If we don't make this 223 * restriction, the bookkeeping would be nearly impossible. 224 */ 225 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 226 fs.entry->max_protection &= ~VM_PROT_WRITE; 227 } 228 229 map_generation = fs.map->timestamp; 230 231 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 232 panic("vm_fault: fault on nofault entry, addr: %lx", 233 (u_long)vaddr); 234 } 235 236 /* 237 * Make a reference to this object to prevent its disposal while we 238 * are messing with it. Once we have the reference, the map is free 239 * to be diddled. Since objects reference their shadows (and copies), 240 * they will stay around as well. 241 */ 242 vm_object_reference(fs.first_object); 243 vm_object_pip_add(fs.first_object, 1); 244 245 fs.vp = vnode_pager_lock(fs.first_object); 246 if ((fault_type & VM_PROT_WRITE) && 247 (fs.first_object->type == OBJT_VNODE)) { 248 vm_freeze_copyopts(fs.first_object, 249 fs.first_pindex, fs.first_pindex + 1); 250 } 251 252 fs.lookup_still_valid = TRUE; 253 254 if (wired) 255 fault_type = prot; 256 257 fs.first_m = NULL; 258 259 /* 260 * Search for the page at object/offset. 261 */ 262 263 fs.object = fs.first_object; 264 fs.pindex = fs.first_pindex; 265 266 /* 267 * See whether this page is resident 268 */ 269 while (TRUE) { 270 /* 271 * If the object is dead, we stop here 272 */ 273 274 if (fs.object->flags & OBJ_DEAD) { 275 unlock_and_deallocate(&fs); 276 return (KERN_PROTECTION_FAILURE); 277 } 278 279 /* 280 * See if page is resident 281 */ 282 283 fs.m = vm_page_lookup(fs.object, fs.pindex); 284 if (fs.m != NULL) { 285 int queue, s; 286 /* 287 * Wait/Retry if the page is busy. We have to do this 288 * if the page is busy via either PG_BUSY or 289 * vm_page_t->busy because the vm_pager may be using 290 * vm_page_t->busy for pageouts ( and even pageins if 291 * it is the vnode pager ), and we could end up trying 292 * to pagein and pageout the same page simultaniously. 293 * 294 * We can theoretically allow the busy case on a read 295 * fault if the page is marked valid, but since such 296 * pages are typically already pmap'd, putting that 297 * special case in might be more effort then it is 298 * worth. We cannot under any circumstances mess 299 * around with a vm_page_t->busy page except, perhaps, 300 * to pmap it. 301 */ 302 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 303 unlock_things(&fs); 304 s = splvm(); 305 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 306 vm_page_flag_set(fs.m, PG_WANTED | PG_REFERENCED); 307 cnt.v_intrans++; 308 tsleep(fs.m, PSWP, "vmpfw", 0); 309 } 310 splx(s); 311 vm_object_deallocate(fs.first_object); 312 goto RetryFault; 313 } 314 315 queue = fs.m->queue; 316 s = splvm(); 317 vm_page_unqueue_nowakeup(fs.m); 318 splx(s); 319 320 /* 321 * Mark page busy for other processes, and the pagedaemon. 322 */ 323 if (((queue - fs.m->pc) == PQ_CACHE) && 324 (cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min) { 325 vm_page_activate(fs.m); 326 unlock_and_deallocate(&fs); 327 VM_WAIT; 328 goto RetryFault; 329 } 330 331 vm_page_busy(fs.m); 332 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 333 fs.m->object != kernel_object && fs.m->object != kmem_object) { 334 goto readrest; 335 } 336 337 break; 338 } 339 if (((fs.object->type != OBJT_DEFAULT) && 340 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 341 || (fs.object == fs.first_object)) { 342 343 if (fs.pindex >= fs.object->size) { 344 unlock_and_deallocate(&fs); 345 return (KERN_PROTECTION_FAILURE); 346 } 347 348 /* 349 * Allocate a new page for this object/offset pair. 350 */ 351 fs.m = vm_page_alloc(fs.object, fs.pindex, 352 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); 353 354 if (fs.m == NULL) { 355 unlock_and_deallocate(&fs); 356 VM_WAIT; 357 goto RetryFault; 358 } 359 } 360 361 readrest: 362 if (fs.object->type != OBJT_DEFAULT && 363 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) { 364 int rv; 365 int reqpage; 366 int ahead, behind; 367 368 if (fs.first_object->behavior == OBJ_RANDOM) { 369 ahead = 0; 370 behind = 0; 371 } else { 372 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 373 if (behind > VM_FAULT_READ_BEHIND) 374 behind = VM_FAULT_READ_BEHIND; 375 376 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 377 if (ahead > VM_FAULT_READ_AHEAD) 378 ahead = VM_FAULT_READ_AHEAD; 379 } 380 381 if ((fs.first_object->type != OBJT_DEVICE) && 382 (fs.first_object->behavior == OBJ_SEQUENTIAL)) { 383 vm_pindex_t firstpindex, tmppindex; 384 if (fs.first_pindex < 385 2*(VM_FAULT_READ_BEHIND + VM_FAULT_READ_AHEAD + 1)) 386 firstpindex = 0; 387 else 388 firstpindex = fs.first_pindex - 389 2*(VM_FAULT_READ_BEHIND + VM_FAULT_READ_AHEAD + 1); 390 391 for(tmppindex = fs.first_pindex - 1; 392 tmppindex >= firstpindex; 393 --tmppindex) { 394 vm_page_t mt; 395 mt = vm_page_lookup( fs.first_object, tmppindex); 396 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 397 break; 398 if (mt->busy || 399 (mt->flags & (PG_BUSY | PG_FICTITIOUS)) || 400 mt->hold_count || 401 mt->wire_count) 402 continue; 403 if (mt->dirty == 0) 404 vm_page_test_dirty(mt); 405 if (mt->dirty) { 406 vm_page_protect(mt, VM_PROT_NONE); 407 vm_page_deactivate(mt); 408 } else { 409 vm_page_cache(mt); 410 } 411 } 412 413 ahead += behind; 414 behind = 0; 415 } 416 417 /* 418 * now we find out if any other pages should be paged 419 * in at this time this routine checks to see if the 420 * pages surrounding this fault reside in the same 421 * object as the page for this fault. If they do, 422 * then they are faulted in also into the object. The 423 * array "marray" returned contains an array of 424 * vm_page_t structs where one of them is the 425 * vm_page_t passed to the routine. The reqpage 426 * return value is the index into the marray for the 427 * vm_page_t passed to the routine. 428 */ 429 faultcount = vm_fault_additional_pages( 430 fs.m, behind, ahead, marray, &reqpage); 431 432 /* 433 * Call the pager to retrieve the data, if any, after 434 * releasing the lock on the map. 435 */ 436 unlock_map(&fs); 437 438 rv = faultcount ? 439 vm_pager_get_pages(fs.object, marray, faultcount, 440 reqpage) : VM_PAGER_FAIL; 441 442 if (rv == VM_PAGER_OK) { 443 /* 444 * Found the page. Leave it busy while we play 445 * with it. 446 */ 447 448 /* 449 * Relookup in case pager changed page. Pager 450 * is responsible for disposition of old page 451 * if moved. 452 */ 453 fs.m = vm_page_lookup(fs.object, fs.pindex); 454 if(!fs.m) { 455 unlock_and_deallocate(&fs); 456 goto RetryFault; 457 } 458 459 hardfault++; 460 break; 461 } 462 /* 463 * Remove the bogus page (which does not exist at this 464 * object/offset); before doing so, we must get back 465 * our object lock to preserve our invariant. 466 * 467 * Also wake up any other process that may want to bring 468 * in this page. 469 * 470 * If this is the top-level object, we must leave the 471 * busy page to prevent another process from rushing 472 * past us, and inserting the page in that object at 473 * the same time that we are. 474 */ 475 476 if (rv == VM_PAGER_ERROR) 477 printf("vm_fault: pager read error, pid %d (%s)\n", 478 curproc->p_pid, curproc->p_comm); 479 /* 480 * Data outside the range of the pager or an I/O error 481 */ 482 /* 483 * XXX - the check for kernel_map is a kludge to work 484 * around having the machine panic on a kernel space 485 * fault w/ I/O error. 486 */ 487 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 488 (rv == VM_PAGER_BAD)) { 489 vm_page_free(fs.m); 490 fs.m = NULL; 491 unlock_and_deallocate(&fs); 492 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 493 } 494 if (fs.object != fs.first_object) { 495 vm_page_free(fs.m); 496 fs.m = NULL; 497 /* 498 * XXX - we cannot just fall out at this 499 * point, m has been freed and is invalid! 500 */ 501 } 502 } 503 /* 504 * We get here if the object has default pager (or unwiring) or the 505 * pager doesn't have the page. 506 */ 507 if (fs.object == fs.first_object) 508 fs.first_m = fs.m; 509 510 /* 511 * Move on to the next object. Lock the next object before 512 * unlocking the current one. 513 */ 514 515 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 516 next_object = fs.object->backing_object; 517 if (next_object == NULL) { 518 /* 519 * If there's no object left, fill the page in the top 520 * object with zeros. 521 */ 522 if (fs.object != fs.first_object) { 523 vm_object_pip_wakeup(fs.object); 524 525 fs.object = fs.first_object; 526 fs.pindex = fs.first_pindex; 527 fs.m = fs.first_m; 528 } 529 fs.first_m = NULL; 530 531 if ((fs.m->flags & PG_ZERO) == 0) { 532 vm_page_zero_fill(fs.m); 533 } 534 else 535 cnt.v_ozfod++; 536 cnt.v_zfod++; 537 break; 538 } else { 539 if (fs.object != fs.first_object) { 540 vm_object_pip_wakeup(fs.object); 541 } 542 fs.object = next_object; 543 vm_object_pip_add(fs.object, 1); 544 } 545 } 546 KASSERT((fs.m->flags & PG_BUSY) != 0, 547 ("vm_fault: not busy after main loop")); 548 549 /* 550 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 551 * is held.] 552 */ 553 554 old_m = fs.m; /* save page that would be copied */ 555 556 /* 557 * If the page is being written, but isn't already owned by the 558 * top-level object, we have to copy it into a new page owned by the 559 * top-level object. 560 */ 561 562 if (fs.object != fs.first_object) { 563 /* 564 * We only really need to copy if we want to write it. 565 */ 566 567 if (fault_type & VM_PROT_WRITE) { 568 569 /* 570 * This allows pages to be virtually copied from a backing_object 571 * into the first_object, where the backing object has no other 572 * refs to it, and cannot gain any more refs. Instead of a 573 * bcopy, we just move the page from the backing object to the 574 * first object. Note that we must mark the page dirty in the 575 * first object so that it will go out to swap when needed. 576 */ 577 if (map_generation == fs.map->timestamp && 578 /* 579 * Only one shadow object 580 */ 581 (fs.object->shadow_count == 1) && 582 /* 583 * No COW refs, except us 584 */ 585 (fs.object->ref_count == 1) && 586 /* 587 * Noone else can look this object up 588 */ 589 (fs.object->handle == NULL) && 590 /* 591 * No other ways to look the object up 592 */ 593 ((fs.object->type == OBJT_DEFAULT) || 594 (fs.object->type == OBJT_SWAP)) && 595 /* 596 * We don't chase down the shadow chain 597 */ 598 (fs.object == fs.first_object->backing_object) && 599 600 /* 601 * grab the lock if we need to 602 */ 603 (fs.lookup_still_valid || 604 (((fs.entry->eflags & MAP_ENTRY_IS_A_MAP) == 0) && 605 lockmgr(&fs.map->lock, 606 LK_EXCLUSIVE|LK_NOWAIT, (void *)0, curproc) == 0))) { 607 608 fs.lookup_still_valid = 1; 609 /* 610 * get rid of the unnecessary page 611 */ 612 vm_page_protect(fs.first_m, VM_PROT_NONE); 613 vm_page_free(fs.first_m); 614 fs.first_m = NULL; 615 616 /* 617 * grab the page and put it into the process'es object 618 */ 619 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 620 fs.first_m = fs.m; 621 fs.first_m->dirty = VM_PAGE_BITS_ALL; 622 vm_page_busy(fs.first_m); 623 fs.m = NULL; 624 cnt.v_cow_optim++; 625 } else { 626 /* 627 * Oh, well, lets copy it. 628 */ 629 vm_page_copy(fs.m, fs.first_m); 630 } 631 632 if (fs.m) { 633 /* 634 * We no longer need the old page or object. 635 */ 636 release_page(&fs); 637 } 638 639 vm_object_pip_wakeup(fs.object); 640 /* 641 * Only use the new page below... 642 */ 643 644 cnt.v_cow_faults++; 645 fs.m = fs.first_m; 646 fs.object = fs.first_object; 647 fs.pindex = fs.first_pindex; 648 649 } else { 650 prot &= ~VM_PROT_WRITE; 651 } 652 } 653 654 /* 655 * We must verify that the maps have not changed since our last 656 * lookup. 657 */ 658 659 if (!fs.lookup_still_valid && 660 (fs.map->timestamp != map_generation)) { 661 vm_object_t retry_object; 662 vm_pindex_t retry_pindex; 663 vm_prot_t retry_prot; 664 665 /* 666 * Since map entries may be pageable, make sure we can take a 667 * page fault on them. 668 */ 669 670 671 /* 672 * Unlock vnode before the lookup to avoid deadlock. E.G. 673 * avoid a deadlock between the inode and exec_map that can 674 * occur due to locks being obtained in different orders. 675 */ 676 677 if (fs.vp != NULL) { 678 vput(fs.vp); 679 fs.vp = NULL; 680 } 681 682 /* 683 * To avoid trying to write_lock the map while another process 684 * has it read_locked (in vm_map_pageable), we do not try for 685 * write permission. If the page is still writable, we will 686 * get write permission. If it is not, or has been marked 687 * needs_copy, we enter the mapping without write permission, 688 * and will merely take another fault. 689 */ 690 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE, 691 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 692 map_generation = fs.map->timestamp; 693 694 /* 695 * If we don't need the page any longer, put it on the active 696 * list (the easiest thing to do here). If no one needs it, 697 * pageout will grab it eventually. 698 */ 699 700 if (result != KERN_SUCCESS) { 701 release_page(&fs); 702 unlock_and_deallocate(&fs); 703 return (result); 704 } 705 fs.lookup_still_valid = TRUE; 706 707 if ((retry_object != fs.first_object) || 708 (retry_pindex != fs.first_pindex)) { 709 release_page(&fs); 710 unlock_and_deallocate(&fs); 711 goto RetryFault; 712 } 713 /* 714 * Check whether the protection has changed or the object has 715 * been copied while we left the map unlocked. Changing from 716 * read to write permission is OK - we leave the page 717 * write-protected, and catch the write fault. Changing from 718 * write to read permission means that we can't mark the page 719 * write-enabled after all. 720 */ 721 prot &= retry_prot; 722 } 723 724 /* 725 * Put this page into the physical map. We had to do the unlock above 726 * because pmap_enter may cause other faults. We don't put the page 727 * back on the active queue until later so that the page-out daemon 728 * won't find us (yet). 729 */ 730 731 if (prot & VM_PROT_WRITE) { 732 vm_page_flag_set(fs.m, PG_WRITEABLE); 733 vm_object_set_flag(fs.m->object, 734 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY); 735 /* 736 * If the fault is a write, we know that this page is being 737 * written NOW. This will save on the pmap_is_modified() calls 738 * later. 739 */ 740 if (fault_flags & VM_FAULT_DIRTY) { 741 fs.m->dirty = VM_PAGE_BITS_ALL; 742 } 743 } 744 745 unlock_things(&fs); 746 fs.m->valid = VM_PAGE_BITS_ALL; 747 vm_page_flag_clear(fs.m, PG_ZERO); 748 749 pmap_enter(fs.map->pmap, vaddr, VM_PAGE_TO_PHYS(fs.m), prot, wired); 750 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 751 pmap_prefault(fs.map->pmap, vaddr, fs.entry); 752 } 753 754 vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED); 755 if (fault_flags & VM_FAULT_HOLD) 756 vm_page_hold(fs.m); 757 758 /* 759 * If the page is not wired down, then put it where the pageout daemon 760 * can find it. 761 */ 762 if (fault_flags & VM_FAULT_WIRE_MASK) { 763 if (wired) 764 vm_page_wire(fs.m); 765 else 766 vm_page_unwire(fs.m, 1); 767 } else { 768 vm_page_activate(fs.m); 769 } 770 771 if (curproc && (curproc->p_flag & P_INMEM) && curproc->p_stats) { 772 if (hardfault) { 773 curproc->p_stats->p_ru.ru_majflt++; 774 } else { 775 curproc->p_stats->p_ru.ru_minflt++; 776 } 777 } 778 779 /* 780 * Unlock everything, and return 781 */ 782 783 vm_page_wakeup(fs.m); 784 vm_object_deallocate(fs.first_object); 785 786 return (KERN_SUCCESS); 787 788 } 789 790 /* 791 * vm_fault_wire: 792 * 793 * Wire down a range of virtual addresses in a map. 794 */ 795 int 796 vm_fault_wire(map, start, end) 797 vm_map_t map; 798 vm_offset_t start, end; 799 { 800 801 register vm_offset_t va; 802 register pmap_t pmap; 803 int rv; 804 805 pmap = vm_map_pmap(map); 806 807 /* 808 * Inform the physical mapping system that the range of addresses may 809 * not fault, so that page tables and such can be locked down as well. 810 */ 811 812 pmap_pageable(pmap, start, end, FALSE); 813 814 /* 815 * We simulate a fault to get the page and enter it in the physical 816 * map. 817 */ 818 819 for (va = start; va < end; va += PAGE_SIZE) { 820 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 821 VM_FAULT_CHANGE_WIRING); 822 if (rv) { 823 if (va != start) 824 vm_fault_unwire(map, start, va); 825 return (rv); 826 } 827 } 828 return (KERN_SUCCESS); 829 } 830 831 /* 832 * vm_fault_user_wire: 833 * 834 * Wire down a range of virtual addresses in a map. This 835 * is for user mode though, so we only ask for read access 836 * on currently read only sections. 837 */ 838 int 839 vm_fault_user_wire(map, start, end) 840 vm_map_t map; 841 vm_offset_t start, end; 842 { 843 844 register vm_offset_t va; 845 register pmap_t pmap; 846 int rv; 847 848 pmap = vm_map_pmap(map); 849 850 /* 851 * Inform the physical mapping system that the range of addresses may 852 * not fault, so that page tables and such can be locked down as well. 853 */ 854 855 pmap_pageable(pmap, start, end, FALSE); 856 857 /* 858 * We simulate a fault to get the page and enter it in the physical 859 * map. 860 */ 861 for (va = start; va < end; va += PAGE_SIZE) { 862 rv = vm_fault(map, va, VM_PROT_READ, VM_FAULT_USER_WIRE); 863 if (rv) { 864 if (va != start) 865 vm_fault_unwire(map, start, va); 866 return (rv); 867 } 868 } 869 return (KERN_SUCCESS); 870 } 871 872 873 /* 874 * vm_fault_unwire: 875 * 876 * Unwire a range of virtual addresses in a map. 877 */ 878 void 879 vm_fault_unwire(map, start, end) 880 vm_map_t map; 881 vm_offset_t start, end; 882 { 883 884 register vm_offset_t va, pa; 885 register pmap_t pmap; 886 887 pmap = vm_map_pmap(map); 888 889 /* 890 * Since the pages are wired down, we must be able to get their 891 * mappings from the physical map system. 892 */ 893 894 for (va = start; va < end; va += PAGE_SIZE) { 895 pa = pmap_extract(pmap, va); 896 if (pa != (vm_offset_t) 0) { 897 pmap_change_wiring(pmap, va, FALSE); 898 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 899 } 900 } 901 902 /* 903 * Inform the physical mapping system that the range of addresses may 904 * fault, so that page tables and such may be unwired themselves. 905 */ 906 907 pmap_pageable(pmap, start, end, TRUE); 908 909 } 910 911 /* 912 * Routine: 913 * vm_fault_copy_entry 914 * Function: 915 * Copy all of the pages from a wired-down map entry to another. 916 * 917 * In/out conditions: 918 * The source and destination maps must be locked for write. 919 * The source map entry must be wired down (or be a sharing map 920 * entry corresponding to a main map entry that is wired down). 921 */ 922 923 void 924 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) 925 vm_map_t dst_map; 926 vm_map_t src_map; 927 vm_map_entry_t dst_entry; 928 vm_map_entry_t src_entry; 929 { 930 vm_object_t dst_object; 931 vm_object_t src_object; 932 vm_ooffset_t dst_offset; 933 vm_ooffset_t src_offset; 934 vm_prot_t prot; 935 vm_offset_t vaddr; 936 vm_page_t dst_m; 937 vm_page_t src_m; 938 939 #ifdef lint 940 src_map++; 941 #endif /* lint */ 942 943 src_object = src_entry->object.vm_object; 944 src_offset = src_entry->offset; 945 946 /* 947 * Create the top-level object for the destination entry. (Doesn't 948 * actually shadow anything - we copy the pages directly.) 949 */ 950 dst_object = vm_object_allocate(OBJT_DEFAULT, 951 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start)); 952 953 dst_entry->object.vm_object = dst_object; 954 dst_entry->offset = 0; 955 956 prot = dst_entry->max_protection; 957 958 /* 959 * Loop through all of the pages in the entry's range, copying each 960 * one from the source object (it should be there) to the destination 961 * object. 962 */ 963 for (vaddr = dst_entry->start, dst_offset = 0; 964 vaddr < dst_entry->end; 965 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 966 967 /* 968 * Allocate a page in the destination object 969 */ 970 do { 971 dst_m = vm_page_alloc(dst_object, 972 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 973 if (dst_m == NULL) { 974 VM_WAIT; 975 } 976 } while (dst_m == NULL); 977 978 /* 979 * Find the page in the source object, and copy it in. 980 * (Because the source is wired down, the page will be in 981 * memory.) 982 */ 983 src_m = vm_page_lookup(src_object, 984 OFF_TO_IDX(dst_offset + src_offset)); 985 if (src_m == NULL) 986 panic("vm_fault_copy_wired: page missing"); 987 988 vm_page_copy(src_m, dst_m); 989 990 /* 991 * Enter it in the pmap... 992 */ 993 994 vm_page_flag_clear(dst_m, PG_ZERO); 995 pmap_enter(dst_map->pmap, vaddr, VM_PAGE_TO_PHYS(dst_m), 996 prot, FALSE); 997 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED); 998 999 /* 1000 * Mark it no longer busy, and put it on the active list. 1001 */ 1002 vm_page_activate(dst_m); 1003 vm_page_wakeup(dst_m); 1004 } 1005 } 1006 1007 1008 /* 1009 * This routine checks around the requested page for other pages that 1010 * might be able to be faulted in. This routine brackets the viable 1011 * pages for the pages to be paged in. 1012 * 1013 * Inputs: 1014 * m, rbehind, rahead 1015 * 1016 * Outputs: 1017 * marray (array of vm_page_t), reqpage (index of requested page) 1018 * 1019 * Return value: 1020 * number of pages in marray 1021 */ 1022 static int 1023 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1024 vm_page_t m; 1025 int rbehind; 1026 int rahead; 1027 vm_page_t *marray; 1028 int *reqpage; 1029 { 1030 int i,j; 1031 vm_object_t object; 1032 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1033 vm_page_t rtm; 1034 int cbehind, cahead; 1035 1036 object = m->object; 1037 pindex = m->pindex; 1038 1039 /* 1040 * we don't fault-ahead for device pager 1041 */ 1042 if (object->type == OBJT_DEVICE) { 1043 *reqpage = 0; 1044 marray[0] = m; 1045 return 1; 1046 } 1047 1048 /* 1049 * if the requested page is not available, then give up now 1050 */ 1051 1052 if (!vm_pager_has_page(object, 1053 OFF_TO_IDX(object->paging_offset) + pindex, &cbehind, &cahead)) { 1054 return 0; 1055 } 1056 1057 if ((cbehind == 0) && (cahead == 0)) { 1058 *reqpage = 0; 1059 marray[0] = m; 1060 return 1; 1061 } 1062 1063 if (rahead > cahead) { 1064 rahead = cahead; 1065 } 1066 1067 if (rbehind > cbehind) { 1068 rbehind = cbehind; 1069 } 1070 1071 /* 1072 * try to do any readahead that we might have free pages for. 1073 */ 1074 if ((rahead + rbehind) > 1075 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) { 1076 pagedaemon_wakeup(); 1077 marray[0] = m; 1078 *reqpage = 0; 1079 return 1; 1080 } 1081 1082 /* 1083 * scan backward for the read behind pages -- in memory 1084 */ 1085 if (pindex > 0) { 1086 if (rbehind > pindex) { 1087 rbehind = pindex; 1088 startpindex = 0; 1089 } else { 1090 startpindex = pindex - rbehind; 1091 } 1092 1093 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1094 if (vm_page_lookup( object, tpindex)) { 1095 startpindex = tpindex + 1; 1096 break; 1097 } 1098 if (tpindex == 0) 1099 break; 1100 } 1101 1102 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1103 1104 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1105 if (rtm == NULL) { 1106 for (j = 0; j < i; j++) { 1107 vm_page_free(marray[j]); 1108 } 1109 marray[0] = m; 1110 *reqpage = 0; 1111 return 1; 1112 } 1113 1114 marray[i] = rtm; 1115 } 1116 } else { 1117 startpindex = 0; 1118 i = 0; 1119 } 1120 1121 marray[i] = m; 1122 /* page offset of the required page */ 1123 *reqpage = i; 1124 1125 tpindex = pindex + 1; 1126 i++; 1127 1128 /* 1129 * scan forward for the read ahead pages 1130 */ 1131 endpindex = tpindex + rahead; 1132 if (endpindex > object->size) 1133 endpindex = object->size; 1134 1135 for( ; tpindex < endpindex; i++, tpindex++) { 1136 1137 if (vm_page_lookup(object, tpindex)) { 1138 break; 1139 } 1140 1141 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1142 if (rtm == NULL) { 1143 break; 1144 } 1145 1146 marray[i] = rtm; 1147 } 1148 1149 /* return number of bytes of pages */ 1150 return i; 1151 }
Cache object: 1018199b2b892c946cc988a68c3726d6
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]