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: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $ 70 * $DragonFly: src/sys/vm/vm_fault.c,v 1.18 2004/10/12 19:29:34 dillon Exp $ 71 */ 72 73 /* 74 * Page fault handling module. 75 */ 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/proc.h> 80 #include <sys/vnode.h> 81 #include <sys/resourcevar.h> 82 #include <sys/vmmeter.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_param.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 #include <sys/thread2.h> 98 #include <vm/vm_page2.h> 99 100 static int vm_fault_additional_pages (vm_page_t, int, 101 int, vm_page_t *, int *); 102 103 #define VM_FAULT_READ_AHEAD 8 104 #define VM_FAULT_READ_BEHIND 7 105 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 106 107 struct faultstate { 108 vm_page_t m; 109 vm_object_t object; 110 vm_pindex_t pindex; 111 vm_page_t first_m; 112 vm_object_t first_object; 113 vm_pindex_t first_pindex; 114 vm_map_t map; 115 vm_map_entry_t entry; 116 int lookup_still_valid; 117 struct vnode *vp; 118 }; 119 120 static __inline void 121 release_page(struct faultstate *fs) 122 { 123 vm_page_wakeup(fs->m); 124 vm_page_deactivate(fs->m); 125 fs->m = NULL; 126 } 127 128 static __inline void 129 unlock_map(struct faultstate *fs) 130 { 131 if (fs->lookup_still_valid) { 132 vm_map_lookup_done(fs->map, fs->entry, 0); 133 fs->lookup_still_valid = FALSE; 134 } 135 } 136 137 static void 138 _unlock_things(struct faultstate *fs, int dealloc) 139 { 140 vm_object_pip_wakeup(fs->object); 141 if (fs->object != fs->first_object) { 142 vm_page_free(fs->first_m); 143 vm_object_pip_wakeup(fs->first_object); 144 fs->first_m = NULL; 145 } 146 if (dealloc) { 147 vm_object_deallocate(fs->first_object); 148 } 149 unlock_map(fs); 150 if (fs->vp != NULL) { 151 vput(fs->vp); 152 fs->vp = NULL; 153 } 154 } 155 156 #define unlock_things(fs) _unlock_things(fs, 0) 157 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 158 159 /* 160 * TRYPAGER - used by vm_fault to calculate whether the pager for the 161 * current object *might* contain the page. 162 * 163 * default objects are zero-fill, there is no real pager. 164 */ 165 166 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 167 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 168 169 /* 170 * vm_fault: 171 * 172 * Handle a page fault occurring at the given address, 173 * requiring the given permissions, in the map specified. 174 * If successful, the page is inserted into the 175 * associated physical map. 176 * 177 * NOTE: the given address should be truncated to the 178 * proper page address. 179 * 180 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 181 * a standard error specifying why the fault is fatal is returned. 182 * 183 * 184 * The map in question must be referenced, and remains so. 185 * Caller may hold no locks. 186 */ 187 int 188 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 189 { 190 vm_prot_t prot; 191 int result; 192 boolean_t wired; 193 int map_generation; 194 vm_object_t next_object; 195 vm_page_t marray[VM_FAULT_READ]; 196 int hardfault; 197 int faultcount; 198 struct faultstate fs; 199 200 mycpu->gd_cnt.v_vm_faults++; 201 hardfault = 0; 202 203 RetryFault: 204 /* 205 * Find the backing store object and offset into it to begin the 206 * search. 207 */ 208 fs.map = map; 209 if ((result = vm_map_lookup(&fs.map, vaddr, 210 fault_type, &fs.entry, &fs.first_object, 211 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) { 212 if ((result != KERN_PROTECTION_FAILURE) || 213 ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) { 214 return result; 215 } 216 217 /* 218 * If we are user-wiring a r/w segment, and it is COW, then 219 * we need to do the COW operation. Note that we don't COW 220 * currently RO sections now, because it is NOT desirable 221 * to COW .text. We simply keep .text from ever being COW'ed 222 * and take the heat that one cannot debug wired .text sections. 223 */ 224 result = vm_map_lookup(&fs.map, vaddr, 225 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 226 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 227 if (result != KERN_SUCCESS) { 228 return result; 229 } 230 231 /* 232 * If we don't COW now, on a user wire, the user will never 233 * be able to write to the mapping. If we don't make this 234 * restriction, the bookkeeping would be nearly impossible. 235 */ 236 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 237 fs.entry->max_protection &= ~VM_PROT_WRITE; 238 } 239 240 map_generation = fs.map->timestamp; 241 242 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 243 panic("vm_fault: fault on nofault entry, addr: %lx", 244 (u_long)vaddr); 245 } 246 247 /* 248 * Make a reference to this object to prevent its disposal while we 249 * are messing with it. Once we have the reference, the map is free 250 * to be diddled. Since objects reference their shadows (and copies), 251 * they will stay around as well. 252 * 253 * Bump the paging-in-progress count to prevent size changes (e.g. 254 * truncation operations) during I/O. This must be done after 255 * obtaining the vnode lock in order to avoid possible deadlocks. 256 */ 257 vm_object_reference(fs.first_object); 258 fs.vp = vnode_pager_lock(fs.first_object); 259 vm_object_pip_add(fs.first_object, 1); 260 261 if ((fault_type & VM_PROT_WRITE) && 262 (fs.first_object->type == OBJT_VNODE)) { 263 vm_freeze_copyopts(fs.first_object, 264 fs.first_pindex, fs.first_pindex + 1); 265 } 266 267 fs.lookup_still_valid = TRUE; 268 269 if (wired) 270 fault_type = prot; 271 272 fs.first_m = NULL; 273 274 /* 275 * Search for the page at object/offset. 276 */ 277 278 fs.object = fs.first_object; 279 fs.pindex = fs.first_pindex; 280 281 while (TRUE) { 282 /* 283 * If the object is dead, we stop here 284 */ 285 286 if (fs.object->flags & OBJ_DEAD) { 287 unlock_and_deallocate(&fs); 288 return (KERN_PROTECTION_FAILURE); 289 } 290 291 /* 292 * See if page is resident. spl protection is required 293 * to avoid an interrupt unbusy/free race against our 294 * lookup. We must hold the protection through a page 295 * allocation or busy. 296 */ 297 crit_enter(); 298 fs.m = vm_page_lookup(fs.object, fs.pindex); 299 if (fs.m != NULL) { 300 int queue; 301 /* 302 * Wait/Retry if the page is busy. We have to do this 303 * if the page is busy via either PG_BUSY or 304 * vm_page_t->busy because the vm_pager may be using 305 * vm_page_t->busy for pageouts ( and even pageins if 306 * it is the vnode pager ), and we could end up trying 307 * to pagein and pageout the same page simultaneously. 308 * 309 * We can theoretically allow the busy case on a read 310 * fault if the page is marked valid, but since such 311 * pages are typically already pmap'd, putting that 312 * special case in might be more effort then it is 313 * worth. We cannot under any circumstances mess 314 * around with a vm_page_t->busy page except, perhaps, 315 * to pmap it. 316 */ 317 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 318 unlock_things(&fs); 319 vm_page_sleep_busy(fs.m, TRUE, "vmpfw"); 320 mycpu->gd_cnt.v_intrans++; 321 vm_object_deallocate(fs.first_object); 322 crit_exit(); 323 goto RetryFault; 324 } 325 326 queue = fs.m->queue; 327 vm_page_unqueue_nowakeup(fs.m); 328 329 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) { 330 vm_page_activate(fs.m); 331 unlock_and_deallocate(&fs); 332 vm_waitpfault(); 333 crit_exit(); 334 goto RetryFault; 335 } 336 337 /* 338 * Mark page busy for other processes, and the 339 * pagedaemon. If it still isn't completely valid 340 * (readable), jump to readrest, else break-out ( we 341 * found the page ). 342 * 343 * We can release the spl once we have marked the 344 * page busy. 345 */ 346 347 vm_page_busy(fs.m); 348 crit_exit(); 349 350 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 351 fs.m->object != kernel_object && fs.m->object != kmem_object) { 352 goto readrest; 353 } 354 355 break; 356 } 357 358 /* 359 * Page is not resident, If this is the search termination 360 * or the pager might contain the page, allocate a new page. 361 * 362 * note: we are still in splvm(). 363 */ 364 365 if (TRYPAGER || fs.object == fs.first_object) { 366 if (fs.pindex >= fs.object->size) { 367 crit_exit(); 368 unlock_and_deallocate(&fs); 369 return (KERN_PROTECTION_FAILURE); 370 } 371 372 /* 373 * Allocate a new page for this object/offset pair. 374 */ 375 fs.m = NULL; 376 if (!vm_page_count_severe()) { 377 fs.m = vm_page_alloc(fs.object, fs.pindex, 378 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 379 } 380 if (fs.m == NULL) { 381 crit_exit(); 382 unlock_and_deallocate(&fs); 383 vm_waitpfault(); 384 goto RetryFault; 385 } 386 } 387 crit_exit(); 388 389 readrest: 390 /* 391 * We have found a valid page or we have allocated a new page. 392 * The page thus may not be valid or may not be entirely 393 * valid. 394 * 395 * Attempt to fault-in the page if there is a chance that the 396 * pager has it, and potentially fault in additional pages 397 * at the same time. 398 * 399 * We are NOT in splvm here and if TRYPAGER is true then 400 * fs.m will be non-NULL and will be PG_BUSY for us. 401 */ 402 403 if (TRYPAGER) { 404 int rv; 405 int reqpage; 406 int ahead, behind; 407 u_char behavior = vm_map_entry_behavior(fs.entry); 408 409 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 410 ahead = 0; 411 behind = 0; 412 } else { 413 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 414 if (behind > VM_FAULT_READ_BEHIND) 415 behind = VM_FAULT_READ_BEHIND; 416 417 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 418 if (ahead > VM_FAULT_READ_AHEAD) 419 ahead = VM_FAULT_READ_AHEAD; 420 } 421 422 if ((fs.first_object->type != OBJT_DEVICE) && 423 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 424 (behavior != MAP_ENTRY_BEHAV_RANDOM && 425 fs.pindex >= fs.entry->lastr && 426 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) 427 ) { 428 vm_pindex_t firstpindex, tmppindex; 429 430 if (fs.first_pindex < 2 * VM_FAULT_READ) 431 firstpindex = 0; 432 else 433 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 434 435 /* 436 * note: partially valid pages cannot be 437 * included in the lookahead - NFS piecemeal 438 * writes will barf on it badly. 439 * 440 * spl protection is required to avoid races 441 * between the lookup and an interrupt 442 * unbusy/free sequence occuring prior to 443 * our busy check. 444 */ 445 crit_enter(); 446 for (tmppindex = fs.first_pindex - 1; 447 tmppindex >= firstpindex; 448 --tmppindex 449 ) { 450 vm_page_t mt; 451 mt = vm_page_lookup( fs.first_object, tmppindex); 452 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 453 break; 454 if (mt->busy || 455 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 456 mt->hold_count || 457 mt->wire_count) 458 continue; 459 if (mt->dirty == 0) 460 vm_page_test_dirty(mt); 461 if (mt->dirty) { 462 vm_page_protect(mt, VM_PROT_NONE); 463 vm_page_deactivate(mt); 464 } else { 465 vm_page_cache(mt); 466 } 467 } 468 crit_exit(); 469 470 ahead += behind; 471 behind = 0; 472 } 473 474 /* 475 * now we find out if any other pages should be paged 476 * in at this time this routine checks to see if the 477 * pages surrounding this fault reside in the same 478 * object as the page for this fault. If they do, 479 * then they are faulted in also into the object. The 480 * array "marray" returned contains an array of 481 * vm_page_t structs where one of them is the 482 * vm_page_t passed to the routine. The reqpage 483 * return value is the index into the marray for the 484 * vm_page_t passed to the routine. 485 * 486 * fs.m plus the additional pages are PG_BUSY'd. 487 */ 488 faultcount = vm_fault_additional_pages( 489 fs.m, behind, ahead, marray, &reqpage); 490 491 /* 492 * update lastr imperfectly (we do not know how much 493 * getpages will actually read), but good enough. 494 */ 495 fs.entry->lastr = fs.pindex + faultcount - behind; 496 497 /* 498 * Call the pager to retrieve the data, if any, after 499 * releasing the lock on the map. We hold a ref on 500 * fs.object and the pages are PG_BUSY'd. 501 */ 502 unlock_map(&fs); 503 504 rv = faultcount ? 505 vm_pager_get_pages(fs.object, marray, faultcount, 506 reqpage) : VM_PAGER_FAIL; 507 508 if (rv == VM_PAGER_OK) { 509 /* 510 * Found the page. Leave it busy while we play 511 * with it. 512 */ 513 514 /* 515 * Relookup in case pager changed page. Pager 516 * is responsible for disposition of old page 517 * if moved. 518 * 519 * XXX other code segments do relookups too. 520 * It's a bad abstraction that needs to be 521 * fixed/removed. 522 */ 523 fs.m = vm_page_lookup(fs.object, fs.pindex); 524 if (fs.m == NULL) { 525 unlock_and_deallocate(&fs); 526 goto RetryFault; 527 } 528 529 hardfault++; 530 break; /* break to PAGE HAS BEEN FOUND */ 531 } 532 /* 533 * Remove the bogus page (which does not exist at this 534 * object/offset); before doing so, we must get back 535 * our object lock to preserve our invariant. 536 * 537 * Also wake up any other process that may want to bring 538 * in this page. 539 * 540 * If this is the top-level object, we must leave the 541 * busy page to prevent another process from rushing 542 * past us, and inserting the page in that object at 543 * the same time that we are. 544 */ 545 546 if (rv == VM_PAGER_ERROR) 547 printf("vm_fault: pager read error, pid %d (%s)\n", 548 curproc->p_pid, curproc->p_comm); 549 /* 550 * Data outside the range of the pager or an I/O error 551 */ 552 /* 553 * XXX - the check for kernel_map is a kludge to work 554 * around having the machine panic on a kernel space 555 * fault w/ I/O error. 556 */ 557 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 558 (rv == VM_PAGER_BAD)) { 559 vm_page_free(fs.m); 560 fs.m = NULL; 561 unlock_and_deallocate(&fs); 562 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 563 } 564 if (fs.object != fs.first_object) { 565 vm_page_free(fs.m); 566 fs.m = NULL; 567 /* 568 * XXX - we cannot just fall out at this 569 * point, m has been freed and is invalid! 570 */ 571 } 572 } 573 574 /* 575 * We get here if the object has default pager (or unwiring) 576 * or the pager doesn't have the page. 577 */ 578 if (fs.object == fs.first_object) 579 fs.first_m = fs.m; 580 581 /* 582 * Move on to the next object. Lock the next object before 583 * unlocking the current one. 584 */ 585 586 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 587 next_object = fs.object->backing_object; 588 if (next_object == NULL) { 589 /* 590 * If there's no object left, fill the page in the top 591 * object with zeros. 592 */ 593 if (fs.object != fs.first_object) { 594 vm_object_pip_wakeup(fs.object); 595 596 fs.object = fs.first_object; 597 fs.pindex = fs.first_pindex; 598 fs.m = fs.first_m; 599 } 600 fs.first_m = NULL; 601 602 /* 603 * Zero the page if necessary and mark it valid. 604 */ 605 if ((fs.m->flags & PG_ZERO) == 0) { 606 vm_page_zero_fill(fs.m); 607 } else { 608 mycpu->gd_cnt.v_ozfod++; 609 } 610 mycpu->gd_cnt.v_zfod++; 611 fs.m->valid = VM_PAGE_BITS_ALL; 612 break; /* break to PAGE HAS BEEN FOUND */ 613 } else { 614 if (fs.object != fs.first_object) { 615 vm_object_pip_wakeup(fs.object); 616 } 617 KASSERT(fs.object != next_object, ("object loop %p", next_object)); 618 fs.object = next_object; 619 vm_object_pip_add(fs.object, 1); 620 } 621 } 622 623 KASSERT((fs.m->flags & PG_BUSY) != 0, 624 ("vm_fault: not busy after main loop")); 625 626 /* 627 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 628 * is held.] 629 */ 630 631 /* 632 * If the page is being written, but isn't already owned by the 633 * top-level object, we have to copy it into a new page owned by the 634 * top-level object. 635 */ 636 637 if (fs.object != fs.first_object) { 638 /* 639 * We only really need to copy if we want to write it. 640 */ 641 642 if (fault_type & VM_PROT_WRITE) { 643 /* 644 * This allows pages to be virtually copied from a 645 * backing_object into the first_object, where the 646 * backing object has no other refs to it, and cannot 647 * gain any more refs. Instead of a bcopy, we just 648 * move the page from the backing object to the 649 * first object. Note that we must mark the page 650 * dirty in the first object so that it will go out 651 * to swap when needed. 652 */ 653 if (map_generation == fs.map->timestamp && 654 /* 655 * Only one shadow object 656 */ 657 (fs.object->shadow_count == 1) && 658 /* 659 * No COW refs, except us 660 */ 661 (fs.object->ref_count == 1) && 662 /* 663 * No one else can look this object up 664 */ 665 (fs.object->handle == NULL) && 666 /* 667 * No other ways to look the object up 668 */ 669 ((fs.object->type == OBJT_DEFAULT) || 670 (fs.object->type == OBJT_SWAP)) && 671 /* 672 * We don't chase down the shadow chain 673 */ 674 (fs.object == fs.first_object->backing_object) && 675 676 /* 677 * grab the lock if we need to 678 */ 679 (fs.lookup_still_valid || 680 lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT, NULL, curthread) == 0) 681 ) { 682 683 fs.lookup_still_valid = 1; 684 /* 685 * get rid of the unnecessary page 686 */ 687 vm_page_protect(fs.first_m, VM_PROT_NONE); 688 vm_page_free(fs.first_m); 689 fs.first_m = NULL; 690 691 /* 692 * grab the page and put it into the 693 * process'es object. The page is 694 * automatically made dirty. 695 */ 696 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 697 fs.first_m = fs.m; 698 vm_page_busy(fs.first_m); 699 fs.m = NULL; 700 mycpu->gd_cnt.v_cow_optim++; 701 } else { 702 /* 703 * Oh, well, lets copy it. 704 */ 705 vm_page_copy(fs.m, fs.first_m); 706 } 707 708 if (fs.m) { 709 /* 710 * We no longer need the old page or object. 711 */ 712 release_page(&fs); 713 } 714 715 /* 716 * fs.object != fs.first_object due to above 717 * conditional 718 */ 719 720 vm_object_pip_wakeup(fs.object); 721 722 /* 723 * Only use the new page below... 724 */ 725 726 mycpu->gd_cnt.v_cow_faults++; 727 fs.m = fs.first_m; 728 fs.object = fs.first_object; 729 fs.pindex = fs.first_pindex; 730 731 } else { 732 prot &= ~VM_PROT_WRITE; 733 } 734 } 735 736 /* 737 * We must verify that the maps have not changed since our last 738 * lookup. 739 */ 740 741 if (!fs.lookup_still_valid && 742 (fs.map->timestamp != map_generation)) { 743 vm_object_t retry_object; 744 vm_pindex_t retry_pindex; 745 vm_prot_t retry_prot; 746 747 /* 748 * Since map entries may be pageable, make sure we can take a 749 * page fault on them. 750 */ 751 752 /* 753 * Unlock vnode before the lookup to avoid deadlock. E.G. 754 * avoid a deadlock between the inode and exec_map that can 755 * occur due to locks being obtained in different orders. 756 */ 757 758 if (fs.vp != NULL) { 759 vput(fs.vp); 760 fs.vp = NULL; 761 } 762 763 if (fs.map->infork) { 764 release_page(&fs); 765 unlock_and_deallocate(&fs); 766 goto RetryFault; 767 } 768 769 /* 770 * To avoid trying to write_lock the map while another process 771 * has it read_locked (in vm_map_wire), we do not try for 772 * write permission. If the page is still writable, we will 773 * get write permission. If it is not, or has been marked 774 * needs_copy, we enter the mapping without write permission, 775 * and will merely take another fault. 776 */ 777 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE, 778 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 779 map_generation = fs.map->timestamp; 780 781 /* 782 * If we don't need the page any longer, put it on the active 783 * list (the easiest thing to do here). If no one needs it, 784 * pageout will grab it eventually. 785 */ 786 787 if (result != KERN_SUCCESS) { 788 release_page(&fs); 789 unlock_and_deallocate(&fs); 790 return (result); 791 } 792 fs.lookup_still_valid = TRUE; 793 794 if ((retry_object != fs.first_object) || 795 (retry_pindex != fs.first_pindex)) { 796 release_page(&fs); 797 unlock_and_deallocate(&fs); 798 goto RetryFault; 799 } 800 /* 801 * Check whether the protection has changed or the object has 802 * been copied while we left the map unlocked. Changing from 803 * read to write permission is OK - we leave the page 804 * write-protected, and catch the write fault. Changing from 805 * write to read permission means that we can't mark the page 806 * write-enabled after all. 807 */ 808 prot &= retry_prot; 809 } 810 811 /* 812 * Put this page into the physical map. We had to do the unlock above 813 * because pmap_enter may cause other faults. We don't put the page 814 * back on the active queue until later so that the page-out daemon 815 * won't find us (yet). 816 */ 817 818 if (prot & VM_PROT_WRITE) { 819 vm_page_flag_set(fs.m, PG_WRITEABLE); 820 vm_object_set_writeable_dirty(fs.m->object); 821 822 /* 823 * If the fault is a write, we know that this page is being 824 * written NOW so dirty it explicitly to save on 825 * pmap_is_modified() calls later. 826 * 827 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 828 * if the page is already dirty to prevent data written with 829 * the expectation of being synced from not being synced. 830 * Likewise if this entry does not request NOSYNC then make 831 * sure the page isn't marked NOSYNC. Applications sharing 832 * data should use the same flags to avoid ping ponging. 833 * 834 * Also tell the backing pager, if any, that it should remove 835 * any swap backing since the page is now dirty. 836 */ 837 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 838 if (fs.m->dirty == 0) 839 vm_page_flag_set(fs.m, PG_NOSYNC); 840 } else { 841 vm_page_flag_clear(fs.m, PG_NOSYNC); 842 } 843 if (fault_flags & VM_FAULT_DIRTY) { 844 crit_enter(); 845 vm_page_dirty(fs.m); 846 vm_pager_page_unswapped(fs.m); 847 crit_exit(); 848 } 849 } 850 851 /* 852 * Page had better still be busy 853 */ 854 855 KASSERT(fs.m->flags & PG_BUSY, 856 ("vm_fault: page %p not busy!", fs.m)); 857 858 unlock_things(&fs); 859 860 /* 861 * Sanity check: page must be completely valid or it is not fit to 862 * map into user space. vm_pager_get_pages() ensures this. 863 */ 864 865 if (fs.m->valid != VM_PAGE_BITS_ALL) { 866 vm_page_zero_invalid(fs.m, TRUE); 867 printf("Warning: page %p partially invalid on fault\n", fs.m); 868 } 869 870 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 871 872 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 873 pmap_prefault(fs.map->pmap, vaddr, fs.entry); 874 } 875 876 vm_page_flag_clear(fs.m, PG_ZERO); 877 vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED); 878 if (fault_flags & VM_FAULT_HOLD) 879 vm_page_hold(fs.m); 880 881 /* 882 * If the page is not wired down, then put it where the pageout daemon 883 * can find it. 884 */ 885 886 if (fault_flags & VM_FAULT_WIRE_MASK) { 887 if (wired) 888 vm_page_wire(fs.m); 889 else 890 vm_page_unwire(fs.m, 1); 891 } else { 892 vm_page_activate(fs.m); 893 } 894 895 if (curproc && (curproc->p_flag & P_INMEM) && curproc->p_stats) { 896 if (hardfault) { 897 curproc->p_stats->p_ru.ru_majflt++; 898 } else { 899 curproc->p_stats->p_ru.ru_minflt++; 900 } 901 } 902 903 /* 904 * Unlock everything, and return 905 */ 906 907 vm_page_wakeup(fs.m); 908 vm_object_deallocate(fs.first_object); 909 910 return (KERN_SUCCESS); 911 912 } 913 914 /* 915 * quick version of vm_fault 916 */ 917 int 918 vm_fault_quick(caddr_t v, int prot) 919 { 920 int r; 921 922 if (prot & VM_PROT_WRITE) 923 r = subyte(v, fubyte(v)); 924 else 925 r = fubyte(v); 926 return(r); 927 } 928 929 /* 930 * Wire down a range of virtual addresses in a map. The entry in question 931 * should be marked in-transition and the map must be locked. We must 932 * release the map temporarily while faulting-in the page to avoid a 933 * deadlock. Note that the entry may be clipped while we are blocked but 934 * will never be freed. 935 */ 936 int 937 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 938 { 939 boolean_t fictitious; 940 vm_offset_t start; 941 vm_offset_t end; 942 vm_offset_t va; 943 vm_paddr_t pa; 944 pmap_t pmap; 945 int rv; 946 947 pmap = vm_map_pmap(map); 948 start = entry->start; 949 end = entry->end; 950 fictitious = entry->object.vm_object && 951 (entry->object.vm_object->type == OBJT_DEVICE); 952 953 vm_map_unlock(map); 954 map->timestamp++; 955 956 /* 957 * We simulate a fault to get the page and enter it in the physical 958 * map. 959 */ 960 for (va = start; va < end; va += PAGE_SIZE) { 961 if (user_wire) { 962 rv = vm_fault(map, va, VM_PROT_READ, 963 VM_FAULT_USER_WIRE); 964 } else { 965 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 966 VM_FAULT_CHANGE_WIRING); 967 } 968 if (rv) { 969 while (va > start) { 970 va -= PAGE_SIZE; 971 if ((pa = pmap_extract(pmap, va)) == 0) 972 continue; 973 pmap_change_wiring(pmap, va, FALSE); 974 if (!fictitious) 975 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 976 } 977 vm_map_lock(map); 978 return (rv); 979 } 980 } 981 vm_map_lock(map); 982 return (KERN_SUCCESS); 983 } 984 985 /* 986 * Unwire a range of virtual addresses in a map. The map should be 987 * locked. 988 */ 989 void 990 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 991 { 992 boolean_t fictitious; 993 vm_offset_t start; 994 vm_offset_t end; 995 vm_offset_t va; 996 vm_paddr_t pa; 997 pmap_t pmap; 998 999 pmap = vm_map_pmap(map); 1000 start = entry->start; 1001 end = entry->end; 1002 fictitious = entry->object.vm_object && 1003 (entry->object.vm_object->type == OBJT_DEVICE); 1004 1005 /* 1006 * Since the pages are wired down, we must be able to get their 1007 * mappings from the physical map system. 1008 */ 1009 for (va = start; va < end; va += PAGE_SIZE) { 1010 pa = pmap_extract(pmap, va); 1011 if (pa != 0) { 1012 pmap_change_wiring(pmap, va, FALSE); 1013 if (!fictitious) 1014 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1015 } 1016 } 1017 } 1018 1019 /* 1020 * Routine: 1021 * vm_fault_copy_entry 1022 * Function: 1023 * Copy all of the pages from a wired-down map entry to another. 1024 * 1025 * In/out conditions: 1026 * The source and destination maps must be locked for write. 1027 * The source map entry must be wired down (or be a sharing map 1028 * entry corresponding to a main map entry that is wired down). 1029 */ 1030 1031 void 1032 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1033 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 1034 { 1035 vm_object_t dst_object; 1036 vm_object_t src_object; 1037 vm_ooffset_t dst_offset; 1038 vm_ooffset_t src_offset; 1039 vm_prot_t prot; 1040 vm_offset_t vaddr; 1041 vm_page_t dst_m; 1042 vm_page_t src_m; 1043 1044 #ifdef lint 1045 src_map++; 1046 #endif /* lint */ 1047 1048 src_object = src_entry->object.vm_object; 1049 src_offset = src_entry->offset; 1050 1051 /* 1052 * Create the top-level object for the destination entry. (Doesn't 1053 * actually shadow anything - we copy the pages directly.) 1054 */ 1055 dst_object = vm_object_allocate(OBJT_DEFAULT, 1056 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1057 1058 dst_entry->object.vm_object = dst_object; 1059 dst_entry->offset = 0; 1060 1061 prot = dst_entry->max_protection; 1062 1063 /* 1064 * Loop through all of the pages in the entry's range, copying each 1065 * one from the source object (it should be there) to the destination 1066 * object. 1067 */ 1068 for (vaddr = dst_entry->start, dst_offset = 0; 1069 vaddr < dst_entry->end; 1070 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1071 1072 /* 1073 * Allocate a page in the destination object 1074 */ 1075 do { 1076 dst_m = vm_page_alloc(dst_object, 1077 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1078 if (dst_m == NULL) { 1079 vm_wait(); 1080 } 1081 } while (dst_m == NULL); 1082 1083 /* 1084 * Find the page in the source object, and copy it in. 1085 * (Because the source is wired down, the page will be in 1086 * memory.) 1087 */ 1088 src_m = vm_page_lookup(src_object, 1089 OFF_TO_IDX(dst_offset + src_offset)); 1090 if (src_m == NULL) 1091 panic("vm_fault_copy_wired: page missing"); 1092 1093 vm_page_copy(src_m, dst_m); 1094 1095 /* 1096 * Enter it in the pmap... 1097 */ 1098 1099 vm_page_flag_clear(dst_m, PG_ZERO); 1100 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1101 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED); 1102 1103 /* 1104 * Mark it no longer busy, and put it on the active list. 1105 */ 1106 vm_page_activate(dst_m); 1107 vm_page_wakeup(dst_m); 1108 } 1109 } 1110 1111 1112 /* 1113 * This routine checks around the requested page for other pages that 1114 * might be able to be faulted in. This routine brackets the viable 1115 * pages for the pages to be paged in. 1116 * 1117 * Inputs: 1118 * m, rbehind, rahead 1119 * 1120 * Outputs: 1121 * marray (array of vm_page_t), reqpage (index of requested page) 1122 * 1123 * Return value: 1124 * number of pages in marray 1125 */ 1126 static int 1127 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1128 vm_page_t *marray, int *reqpage) 1129 { 1130 int i,j; 1131 vm_object_t object; 1132 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1133 vm_page_t rtm; 1134 int cbehind, cahead; 1135 1136 object = m->object; 1137 pindex = m->pindex; 1138 1139 /* 1140 * we don't fault-ahead for device pager 1141 */ 1142 if (object->type == OBJT_DEVICE) { 1143 *reqpage = 0; 1144 marray[0] = m; 1145 return 1; 1146 } 1147 1148 /* 1149 * if the requested page is not available, then give up now 1150 */ 1151 1152 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1153 return 0; 1154 } 1155 1156 if ((cbehind == 0) && (cahead == 0)) { 1157 *reqpage = 0; 1158 marray[0] = m; 1159 return 1; 1160 } 1161 1162 if (rahead > cahead) { 1163 rahead = cahead; 1164 } 1165 1166 if (rbehind > cbehind) { 1167 rbehind = cbehind; 1168 } 1169 1170 /* 1171 * try to do any readahead that we might have free pages for. 1172 */ 1173 if ((rahead + rbehind) > 1174 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) { 1175 pagedaemon_wakeup(); 1176 marray[0] = m; 1177 *reqpage = 0; 1178 return 1; 1179 } 1180 1181 /* 1182 * scan backward for the read behind pages -- in memory 1183 * 1184 * Assume that if the page is not found an interrupt will not 1185 * create it. Theoretically interrupts can only remove (busy) 1186 * pages, not create new associations. 1187 */ 1188 if (pindex > 0) { 1189 if (rbehind > pindex) { 1190 rbehind = pindex; 1191 startpindex = 0; 1192 } else { 1193 startpindex = pindex - rbehind; 1194 } 1195 1196 crit_enter(); 1197 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1198 if (vm_page_lookup( object, tpindex)) { 1199 startpindex = tpindex + 1; 1200 break; 1201 } 1202 if (tpindex == 0) 1203 break; 1204 } 1205 1206 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1207 1208 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1209 if (rtm == NULL) { 1210 crit_exit(); 1211 for (j = 0; j < i; j++) { 1212 vm_page_free(marray[j]); 1213 } 1214 marray[0] = m; 1215 *reqpage = 0; 1216 return 1; 1217 } 1218 1219 marray[i] = rtm; 1220 } 1221 crit_exit(); 1222 } else { 1223 startpindex = 0; 1224 i = 0; 1225 } 1226 1227 marray[i] = m; 1228 /* page offset of the required page */ 1229 *reqpage = i; 1230 1231 tpindex = pindex + 1; 1232 i++; 1233 1234 /* 1235 * scan forward for the read ahead pages 1236 */ 1237 endpindex = tpindex + rahead; 1238 if (endpindex > object->size) 1239 endpindex = object->size; 1240 1241 crit_enter(); 1242 for( ; tpindex < endpindex; i++, tpindex++) { 1243 1244 if (vm_page_lookup(object, tpindex)) { 1245 break; 1246 } 1247 1248 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1249 if (rtm == NULL) { 1250 break; 1251 } 1252 1253 marray[i] = rtm; 1254 } 1255 crit_exit(); 1256 1257 /* return number of bytes of pages */ 1258 return i; 1259 } 1260