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 70 /* 71 * Page fault handling module. 72 */ 73 74 #include <sys/cdefs.h> 75 __FBSDID("$FreeBSD$"); 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/lock.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/resourcevar.h> 84 #include <sys/sysctl.h> 85 #include <sys/vmmeter.h> 86 #include <sys/vnode.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vnode_pager.h> 98 #include <vm/vm_extern.h> 99 100 #define PFBAK 4 101 #define PFFOR 4 102 #define PAGEORDER_SIZE (PFBAK+PFFOR) 103 104 static int prefault_pageorder[] = { 105 -1 * PAGE_SIZE, 1 * PAGE_SIZE, 106 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 107 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 108 -4 * PAGE_SIZE, 4 * PAGE_SIZE 109 }; 110 111 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 112 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t); 113 114 #define VM_FAULT_READ_AHEAD 8 115 #define VM_FAULT_READ_BEHIND 7 116 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 117 118 struct faultstate { 119 vm_page_t m; 120 vm_object_t object; 121 vm_pindex_t pindex; 122 vm_page_t first_m; 123 vm_object_t first_object; 124 vm_pindex_t first_pindex; 125 vm_map_t map; 126 vm_map_entry_t entry; 127 int lookup_still_valid; 128 struct vnode *vp; 129 }; 130 131 static __inline void 132 release_page(struct faultstate *fs) 133 { 134 vm_page_lock_queues(); 135 vm_page_wakeup(fs->m); 136 vm_page_deactivate(fs->m); 137 vm_page_unlock_queues(); 138 fs->m = NULL; 139 } 140 141 static __inline void 142 unlock_map(struct faultstate *fs) 143 { 144 if (fs->lookup_still_valid) { 145 vm_map_lookup_done(fs->map, fs->entry); 146 fs->lookup_still_valid = FALSE; 147 } 148 } 149 150 static void 151 _unlock_things(struct faultstate *fs, int dealloc) 152 { 153 154 vm_object_pip_wakeup(fs->object); 155 VM_OBJECT_UNLOCK(fs->object); 156 if (fs->object != fs->first_object) { 157 VM_OBJECT_LOCK(fs->first_object); 158 vm_page_lock_queues(); 159 vm_page_free(fs->first_m); 160 vm_page_unlock_queues(); 161 vm_object_pip_wakeup(fs->first_object); 162 VM_OBJECT_UNLOCK(fs->first_object); 163 fs->first_m = NULL; 164 } 165 if (dealloc) { 166 vm_object_deallocate(fs->first_object); 167 } 168 unlock_map(fs); 169 if (fs->vp != NULL) { 170 vput(fs->vp); 171 fs->vp = NULL; 172 } 173 if (dealloc) 174 mtx_unlock(&Giant); 175 } 176 177 #define unlock_things(fs) _unlock_things(fs, 0) 178 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 179 180 /* 181 * TRYPAGER - used by vm_fault to calculate whether the pager for the 182 * current object *might* contain the page. 183 * 184 * default objects are zero-fill, there is no real pager. 185 */ 186 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 187 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 188 189 /* 190 * vm_fault: 191 * 192 * Handle a page fault occurring at the given address, 193 * requiring the given permissions, in the map specified. 194 * If successful, the page is inserted into the 195 * associated physical map. 196 * 197 * NOTE: the given address should be truncated to the 198 * proper page address. 199 * 200 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 201 * a standard error specifying why the fault is fatal is returned. 202 * 203 * 204 * The map in question must be referenced, and remains so. 205 * Caller may hold no locks. 206 */ 207 int 208 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 209 int fault_flags) 210 { 211 vm_prot_t prot; 212 int is_first_object_locked, result; 213 boolean_t growstack, wired; 214 int map_generation; 215 vm_object_t next_object; 216 vm_page_t marray[VM_FAULT_READ]; 217 int hardfault; 218 int faultcount; 219 struct faultstate fs; 220 221 hardfault = 0; 222 growstack = TRUE; 223 atomic_add_int(&cnt.v_vm_faults, 1); 224 225 RetryFault:; 226 227 /* 228 * Find the backing store object and offset into it to begin the 229 * search. 230 */ 231 fs.map = map; 232 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 233 &fs.first_object, &fs.first_pindex, &prot, &wired); 234 if (result != KERN_SUCCESS) { 235 if (result != KERN_PROTECTION_FAILURE || 236 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) { 237 if (growstack && result == KERN_INVALID_ADDRESS && 238 map != kernel_map && curproc != NULL) { 239 result = vm_map_growstack(curproc, vaddr); 240 if (result != KERN_SUCCESS) 241 return (KERN_FAILURE); 242 growstack = FALSE; 243 goto RetryFault; 244 } 245 return (result); 246 } 247 248 /* 249 * If we are user-wiring a r/w segment, and it is COW, then 250 * we need to do the COW operation. Note that we don't COW 251 * currently RO sections now, because it is NOT desirable 252 * to COW .text. We simply keep .text from ever being COW'ed 253 * and take the heat that one cannot debug wired .text sections. 254 */ 255 result = vm_map_lookup(&fs.map, vaddr, 256 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 257 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 258 if (result != KERN_SUCCESS) 259 return (result); 260 261 /* 262 * If we don't COW now, on a user wire, the user will never 263 * be able to write to the mapping. If we don't make this 264 * restriction, the bookkeeping would be nearly impossible. 265 * 266 * XXX The following assignment modifies the map without 267 * holding a write lock on it. 268 */ 269 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 270 fs.entry->max_protection &= ~VM_PROT_WRITE; 271 } 272 273 map_generation = fs.map->timestamp; 274 275 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 276 panic("vm_fault: fault on nofault entry, addr: %lx", 277 (u_long)vaddr); 278 } 279 280 /* 281 * Make a reference to this object to prevent its disposal while we 282 * are messing with it. Once we have the reference, the map is free 283 * to be diddled. Since objects reference their shadows (and copies), 284 * they will stay around as well. 285 * 286 * Bump the paging-in-progress count to prevent size changes (e.g. 287 * truncation operations) during I/O. This must be done after 288 * obtaining the vnode lock in order to avoid possible deadlocks. 289 * 290 * XXX vnode_pager_lock() can block without releasing the map lock. 291 */ 292 mtx_lock(&Giant); 293 VM_OBJECT_LOCK(fs.first_object); 294 vm_object_reference_locked(fs.first_object); 295 fs.vp = vnode_pager_lock(fs.first_object); 296 vm_object_pip_add(fs.first_object, 1); 297 298 fs.lookup_still_valid = TRUE; 299 300 if (wired) 301 fault_type = prot; 302 303 fs.first_m = NULL; 304 305 /* 306 * Search for the page at object/offset. 307 */ 308 fs.object = fs.first_object; 309 fs.pindex = fs.first_pindex; 310 while (TRUE) { 311 /* 312 * If the object is dead, we stop here 313 */ 314 if (fs.object->flags & OBJ_DEAD) { 315 unlock_and_deallocate(&fs); 316 return (KERN_PROTECTION_FAILURE); 317 } 318 319 /* 320 * See if page is resident 321 */ 322 fs.m = vm_page_lookup(fs.object, fs.pindex); 323 if (fs.m != NULL) { 324 int queue, s; 325 326 /* 327 * check for page-based copy on write. 328 * We check fs.object == fs.first_object so 329 * as to ensure the legacy COW mechanism is 330 * used when the page in question is part of 331 * a shadow object. Otherwise, vm_page_cowfault() 332 * removes the page from the backing object, 333 * which is not what we want. 334 */ 335 vm_page_lock_queues(); 336 if ((fs.m->cow) && 337 (fault_type & VM_PROT_WRITE) && 338 (fs.object == fs.first_object)) { 339 s = splvm(); 340 vm_page_cowfault(fs.m); 341 splx(s); 342 vm_page_unlock_queues(); 343 unlock_and_deallocate(&fs); 344 goto RetryFault; 345 } 346 347 /* 348 * Wait/Retry if the page is busy. We have to do this 349 * if the page is busy via either PG_BUSY or 350 * vm_page_t->busy because the vm_pager may be using 351 * vm_page_t->busy for pageouts ( and even pageins if 352 * it is the vnode pager ), and we could end up trying 353 * to pagein and pageout the same page simultaneously. 354 * 355 * We can theoretically allow the busy case on a read 356 * fault if the page is marked valid, but since such 357 * pages are typically already pmap'd, putting that 358 * special case in might be more effort then it is 359 * worth. We cannot under any circumstances mess 360 * around with a vm_page_t->busy page except, perhaps, 361 * to pmap it. 362 */ 363 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 364 vm_page_unlock_queues(); 365 unlock_things(&fs); 366 vm_page_lock_queues(); 367 if (!vm_page_sleep_if_busy(fs.m, TRUE, "vmpfw")) 368 vm_page_unlock_queues(); 369 cnt.v_intrans++; 370 mtx_unlock(&Giant); 371 vm_object_deallocate(fs.first_object); 372 goto RetryFault; 373 } 374 queue = fs.m->queue; 375 376 s = splvm(); 377 vm_pageq_remove_nowakeup(fs.m); 378 splx(s); 379 380 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) { 381 vm_page_activate(fs.m); 382 vm_page_unlock_queues(); 383 unlock_and_deallocate(&fs); 384 VM_WAITPFAULT; 385 goto RetryFault; 386 } 387 388 /* 389 * Mark page busy for other processes, and the 390 * pagedaemon. If it still isn't completely valid 391 * (readable), jump to readrest, else break-out ( we 392 * found the page ). 393 */ 394 vm_page_busy(fs.m); 395 vm_page_unlock_queues(); 396 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 397 fs.m->object != kernel_object && fs.m->object != kmem_object) { 398 goto readrest; 399 } 400 401 break; 402 } 403 404 /* 405 * Page is not resident, If this is the search termination 406 * or the pager might contain the page, allocate a new page. 407 */ 408 if (TRYPAGER || fs.object == fs.first_object) { 409 if (fs.pindex >= fs.object->size) { 410 unlock_and_deallocate(&fs); 411 return (KERN_PROTECTION_FAILURE); 412 } 413 414 /* 415 * Allocate a new page for this object/offset pair. 416 */ 417 fs.m = NULL; 418 if (!vm_page_count_severe()) { 419 fs.m = vm_page_alloc(fs.object, fs.pindex, 420 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); 421 } 422 if (fs.m == NULL) { 423 unlock_and_deallocate(&fs); 424 VM_WAITPFAULT; 425 goto RetryFault; 426 } 427 } 428 429 readrest: 430 /* 431 * We have found a valid page or we have allocated a new page. 432 * The page thus may not be valid or may not be entirely 433 * valid. 434 * 435 * Attempt to fault-in the page if there is a chance that the 436 * pager has it, and potentially fault in additional pages 437 * at the same time. 438 */ 439 if (TRYPAGER) { 440 int rv; 441 int reqpage; 442 int ahead, behind; 443 u_char behavior = vm_map_entry_behavior(fs.entry); 444 445 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 446 ahead = 0; 447 behind = 0; 448 } else { 449 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 450 if (behind > VM_FAULT_READ_BEHIND) 451 behind = VM_FAULT_READ_BEHIND; 452 453 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 454 if (ahead > VM_FAULT_READ_AHEAD) 455 ahead = VM_FAULT_READ_AHEAD; 456 } 457 is_first_object_locked = FALSE; 458 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 459 (behavior != MAP_ENTRY_BEHAV_RANDOM && 460 fs.pindex >= fs.entry->lastr && 461 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 462 (fs.first_object == fs.object || 463 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 464 fs.first_object->type != OBJT_DEVICE) { 465 vm_pindex_t firstpindex, tmppindex; 466 467 if (fs.first_pindex < 2 * VM_FAULT_READ) 468 firstpindex = 0; 469 else 470 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 471 472 vm_page_lock_queues(); 473 /* 474 * note: partially valid pages cannot be 475 * included in the lookahead - NFS piecemeal 476 * writes will barf on it badly. 477 */ 478 for (tmppindex = fs.first_pindex - 1; 479 tmppindex >= firstpindex; 480 --tmppindex) { 481 vm_page_t mt; 482 483 mt = vm_page_lookup(fs.first_object, tmppindex); 484 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 485 break; 486 if (mt->busy || 487 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 488 mt->hold_count || 489 mt->wire_count) 490 continue; 491 pmap_remove_all(mt); 492 if (mt->dirty) { 493 vm_page_deactivate(mt); 494 } else { 495 vm_page_cache(mt); 496 } 497 } 498 vm_page_unlock_queues(); 499 ahead += behind; 500 behind = 0; 501 } 502 if (is_first_object_locked) 503 VM_OBJECT_UNLOCK(fs.first_object); 504 /* 505 * now we find out if any other pages should be paged 506 * in at this time this routine checks to see if the 507 * pages surrounding this fault reside in the same 508 * object as the page for this fault. If they do, 509 * then they are faulted in also into the object. The 510 * array "marray" returned contains an array of 511 * vm_page_t structs where one of them is the 512 * vm_page_t passed to the routine. The reqpage 513 * return value is the index into the marray for the 514 * vm_page_t passed to the routine. 515 * 516 * fs.m plus the additional pages are PG_BUSY'd. 517 * 518 * XXX vm_fault_additional_pages() can block 519 * without releasing the map lock. 520 */ 521 faultcount = vm_fault_additional_pages( 522 fs.m, behind, ahead, marray, &reqpage); 523 524 /* 525 * update lastr imperfectly (we do not know how much 526 * getpages will actually read), but good enough. 527 * 528 * XXX The following assignment modifies the map 529 * without holding a write lock on it. 530 */ 531 fs.entry->lastr = fs.pindex + faultcount - behind; 532 533 /* 534 * Call the pager to retrieve the data, if any, after 535 * releasing the lock on the map. We hold a ref on 536 * fs.object and the pages are PG_BUSY'd. 537 */ 538 unlock_map(&fs); 539 540 rv = faultcount ? 541 vm_pager_get_pages(fs.object, marray, faultcount, 542 reqpage) : VM_PAGER_FAIL; 543 544 if (rv == VM_PAGER_OK) { 545 /* 546 * Found the page. Leave it busy while we play 547 * with it. 548 */ 549 550 /* 551 * Relookup in case pager changed page. Pager 552 * is responsible for disposition of old page 553 * if moved. 554 */ 555 fs.m = vm_page_lookup(fs.object, fs.pindex); 556 if (!fs.m) { 557 unlock_and_deallocate(&fs); 558 goto RetryFault; 559 } 560 561 hardfault++; 562 break; /* break to PAGE HAS BEEN FOUND */ 563 } 564 /* 565 * Remove the bogus page (which does not exist at this 566 * object/offset); before doing so, we must get back 567 * our object lock to preserve our invariant. 568 * 569 * Also wake up any other process that may want to bring 570 * in this page. 571 * 572 * If this is the top-level object, we must leave the 573 * busy page to prevent another process from rushing 574 * past us, and inserting the page in that object at 575 * the same time that we are. 576 */ 577 if (rv == VM_PAGER_ERROR) 578 printf("vm_fault: pager read error, pid %d (%s)\n", 579 curproc->p_pid, curproc->p_comm); 580 /* 581 * Data outside the range of the pager or an I/O error 582 */ 583 /* 584 * XXX - the check for kernel_map is a kludge to work 585 * around having the machine panic on a kernel space 586 * fault w/ I/O error. 587 */ 588 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 589 (rv == VM_PAGER_BAD)) { 590 vm_page_lock_queues(); 591 vm_page_free(fs.m); 592 vm_page_unlock_queues(); 593 fs.m = NULL; 594 unlock_and_deallocate(&fs); 595 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 596 } 597 if (fs.object != fs.first_object) { 598 vm_page_lock_queues(); 599 vm_page_free(fs.m); 600 vm_page_unlock_queues(); 601 fs.m = NULL; 602 /* 603 * XXX - we cannot just fall out at this 604 * point, m has been freed and is invalid! 605 */ 606 } 607 } 608 609 /* 610 * We get here if the object has default pager (or unwiring) 611 * or the pager doesn't have the page. 612 */ 613 if (fs.object == fs.first_object) 614 fs.first_m = fs.m; 615 616 /* 617 * Move on to the next object. Lock the next object before 618 * unlocking the current one. 619 */ 620 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 621 next_object = fs.object->backing_object; 622 if (next_object == NULL) { 623 /* 624 * If there's no object left, fill the page in the top 625 * object with zeros. 626 */ 627 if (fs.object != fs.first_object) { 628 vm_object_pip_wakeup(fs.object); 629 VM_OBJECT_UNLOCK(fs.object); 630 631 fs.object = fs.first_object; 632 fs.pindex = fs.first_pindex; 633 fs.m = fs.first_m; 634 VM_OBJECT_LOCK(fs.object); 635 } 636 fs.first_m = NULL; 637 638 /* 639 * Zero the page if necessary and mark it valid. 640 */ 641 if ((fs.m->flags & PG_ZERO) == 0) { 642 pmap_zero_page(fs.m); 643 } else { 644 cnt.v_ozfod++; 645 } 646 cnt.v_zfod++; 647 fs.m->valid = VM_PAGE_BITS_ALL; 648 break; /* break to PAGE HAS BEEN FOUND */ 649 } else { 650 KASSERT(fs.object != next_object, 651 ("object loop %p", next_object)); 652 VM_OBJECT_LOCK(next_object); 653 vm_object_pip_add(next_object, 1); 654 if (fs.object != fs.first_object) 655 vm_object_pip_wakeup(fs.object); 656 VM_OBJECT_UNLOCK(fs.object); 657 fs.object = next_object; 658 } 659 } 660 661 KASSERT((fs.m->flags & PG_BUSY) != 0, 662 ("vm_fault: not busy after main loop")); 663 664 /* 665 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 666 * is held.] 667 */ 668 669 /* 670 * If the page is being written, but isn't already owned by the 671 * top-level object, we have to copy it into a new page owned by the 672 * top-level object. 673 */ 674 if (fs.object != fs.first_object) { 675 /* 676 * We only really need to copy if we want to write it. 677 */ 678 if (fault_type & VM_PROT_WRITE) { 679 /* 680 * This allows pages to be virtually copied from a 681 * backing_object into the first_object, where the 682 * backing object has no other refs to it, and cannot 683 * gain any more refs. Instead of a bcopy, we just 684 * move the page from the backing object to the 685 * first object. Note that we must mark the page 686 * dirty in the first object so that it will go out 687 * to swap when needed. 688 */ 689 is_first_object_locked = FALSE; 690 if ( 691 /* 692 * Only one shadow object 693 */ 694 (fs.object->shadow_count == 1) && 695 /* 696 * No COW refs, except us 697 */ 698 (fs.object->ref_count == 1) && 699 /* 700 * No one else can look this object up 701 */ 702 (fs.object->handle == NULL) && 703 /* 704 * No other ways to look the object up 705 */ 706 ((fs.object->type == OBJT_DEFAULT) || 707 (fs.object->type == OBJT_SWAP)) && 708 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 709 /* 710 * We don't chase down the shadow chain 711 */ 712 fs.object == fs.first_object->backing_object) { 713 vm_page_lock_queues(); 714 /* 715 * get rid of the unnecessary page 716 */ 717 pmap_remove_all(fs.first_m); 718 vm_page_free(fs.first_m); 719 /* 720 * grab the page and put it into the 721 * process'es object. The page is 722 * automatically made dirty. 723 */ 724 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 725 vm_page_busy(fs.m); 726 vm_page_unlock_queues(); 727 fs.first_m = fs.m; 728 fs.m = NULL; 729 cnt.v_cow_optim++; 730 } else { 731 /* 732 * Oh, well, lets copy it. 733 */ 734 pmap_copy_page(fs.m, fs.first_m); 735 fs.first_m->valid = VM_PAGE_BITS_ALL; 736 } 737 if (fs.m) { 738 /* 739 * We no longer need the old page or object. 740 */ 741 release_page(&fs); 742 } 743 /* 744 * fs.object != fs.first_object due to above 745 * conditional 746 */ 747 vm_object_pip_wakeup(fs.object); 748 VM_OBJECT_UNLOCK(fs.object); 749 /* 750 * Only use the new page below... 751 */ 752 fs.object = fs.first_object; 753 fs.pindex = fs.first_pindex; 754 fs.m = fs.first_m; 755 if (!is_first_object_locked) 756 VM_OBJECT_LOCK(fs.object); 757 cnt.v_cow_faults++; 758 } else { 759 prot &= ~VM_PROT_WRITE; 760 } 761 } 762 763 /* 764 * We must verify that the maps have not changed since our last 765 * lookup. 766 */ 767 if (!fs.lookup_still_valid && 768 (fs.map->timestamp != map_generation)) { 769 vm_object_t retry_object; 770 vm_pindex_t retry_pindex; 771 vm_prot_t retry_prot; 772 773 /* 774 * Since map entries may be pageable, make sure we can take a 775 * page fault on them. 776 */ 777 778 /* 779 * Unlock vnode before the lookup to avoid deadlock. E.G. 780 * avoid a deadlock between the inode and exec_map that can 781 * occur due to locks being obtained in different orders. 782 */ 783 if (fs.vp != NULL) { 784 vput(fs.vp); 785 fs.vp = NULL; 786 } 787 788 if (fs.map->infork) { 789 release_page(&fs); 790 unlock_and_deallocate(&fs); 791 goto RetryFault; 792 } 793 VM_OBJECT_UNLOCK(fs.object); 794 795 /* 796 * To avoid trying to write_lock the map while another process 797 * has it read_locked (in vm_map_wire), we do not try for 798 * write permission. If the page is still writable, we will 799 * get write permission. If it is not, or has been marked 800 * needs_copy, we enter the mapping without write permission, 801 * and will merely take another fault. 802 */ 803 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE, 804 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 805 map_generation = fs.map->timestamp; 806 807 VM_OBJECT_LOCK(fs.object); 808 /* 809 * If we don't need the page any longer, put it on the active 810 * list (the easiest thing to do here). If no one needs it, 811 * pageout will grab it eventually. 812 */ 813 if (result != KERN_SUCCESS) { 814 release_page(&fs); 815 unlock_and_deallocate(&fs); 816 return (result); 817 } 818 fs.lookup_still_valid = TRUE; 819 820 if ((retry_object != fs.first_object) || 821 (retry_pindex != fs.first_pindex)) { 822 release_page(&fs); 823 unlock_and_deallocate(&fs); 824 goto RetryFault; 825 } 826 /* 827 * Check whether the protection has changed or the object has 828 * been copied while we left the map unlocked. Changing from 829 * read to write permission is OK - we leave the page 830 * write-protected, and catch the write fault. Changing from 831 * write to read permission means that we can't mark the page 832 * write-enabled after all. 833 */ 834 prot &= retry_prot; 835 } 836 837 /* 838 * Put this page into the physical map. We had to do the unlock above 839 * because pmap_enter may cause other faults. We don't put the page 840 * back on the active queue until later so that the page-out daemon 841 * won't find us (yet). 842 */ 843 844 if (prot & VM_PROT_WRITE) { 845 vm_page_lock_queues(); 846 vm_page_flag_set(fs.m, PG_WRITEABLE); 847 vm_object_set_writeable_dirty(fs.m->object); 848 849 /* 850 * If the fault is a write, we know that this page is being 851 * written NOW so dirty it explicitly to save on 852 * pmap_is_modified() calls later. 853 * 854 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 855 * if the page is already dirty to prevent data written with 856 * the expectation of being synced from not being synced. 857 * Likewise if this entry does not request NOSYNC then make 858 * sure the page isn't marked NOSYNC. Applications sharing 859 * data should use the same flags to avoid ping ponging. 860 * 861 * Also tell the backing pager, if any, that it should remove 862 * any swap backing since the page is now dirty. 863 */ 864 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 865 if (fs.m->dirty == 0) 866 vm_page_flag_set(fs.m, PG_NOSYNC); 867 } else { 868 vm_page_flag_clear(fs.m, PG_NOSYNC); 869 } 870 vm_page_unlock_queues(); 871 if (fault_flags & VM_FAULT_DIRTY) { 872 int s; 873 vm_page_dirty(fs.m); 874 s = splvm(); 875 vm_pager_page_unswapped(fs.m); 876 splx(s); 877 } 878 } 879 880 /* 881 * Page had better still be busy 882 */ 883 KASSERT(fs.m->flags & PG_BUSY, 884 ("vm_fault: page %p not busy!", fs.m)); 885 /* 886 * Sanity check: page must be completely valid or it is not fit to 887 * map into user space. vm_pager_get_pages() ensures this. 888 */ 889 if (fs.m->valid != VM_PAGE_BITS_ALL) { 890 vm_page_zero_invalid(fs.m, TRUE); 891 printf("Warning: page %p partially invalid on fault\n", fs.m); 892 } 893 unlock_things(&fs); 894 895 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 896 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 897 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 898 } 899 mtx_unlock(&Giant); 900 vm_page_lock_queues(); 901 vm_page_flag_clear(fs.m, PG_ZERO); 902 vm_page_flag_set(fs.m, PG_REFERENCED); 903 904 /* 905 * If the page is not wired down, then put it where the pageout daemon 906 * can find it. 907 */ 908 if (fault_flags & VM_FAULT_WIRE_MASK) { 909 if (wired) 910 vm_page_wire(fs.m); 911 else 912 vm_page_unwire(fs.m, 1); 913 } else { 914 vm_page_activate(fs.m); 915 } 916 vm_page_wakeup(fs.m); 917 vm_page_unlock_queues(); 918 919 PROC_LOCK(curproc); 920 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) { 921 if (hardfault) { 922 curproc->p_stats->p_ru.ru_majflt++; 923 } else { 924 curproc->p_stats->p_ru.ru_minflt++; 925 } 926 } 927 PROC_UNLOCK(curproc); 928 929 /* 930 * Unlock everything, and return 931 */ 932 vm_object_deallocate(fs.first_object); 933 return (KERN_SUCCESS); 934 } 935 936 /* 937 * vm_fault_prefault provides a quick way of clustering 938 * pagefaults into a processes address space. It is a "cousin" 939 * of vm_map_pmap_enter, except it runs at page fault time instead 940 * of mmap time. 941 */ 942 static void 943 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 944 { 945 int i; 946 vm_offset_t addr, starta; 947 vm_pindex_t pindex; 948 vm_page_t m, mpte; 949 vm_object_t object; 950 951 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) 952 return; 953 954 object = entry->object.vm_object; 955 956 starta = addra - PFBAK * PAGE_SIZE; 957 if (starta < entry->start) { 958 starta = entry->start; 959 } else if (starta > addra) { 960 starta = 0; 961 } 962 963 mpte = NULL; 964 for (i = 0; i < PAGEORDER_SIZE; i++) { 965 vm_object_t backing_object, lobject; 966 967 addr = addra + prefault_pageorder[i]; 968 if (addr > addra + (PFFOR * PAGE_SIZE)) 969 addr = 0; 970 971 if (addr < starta || addr >= entry->end) 972 continue; 973 974 if (!pmap_is_prefaultable(pmap, addr)) 975 continue; 976 977 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 978 lobject = object; 979 VM_OBJECT_LOCK(lobject); 980 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 981 lobject->type == OBJT_DEFAULT && 982 (backing_object = lobject->backing_object) != NULL) { 983 if (lobject->backing_object_offset & PAGE_MASK) 984 break; 985 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 986 VM_OBJECT_LOCK(backing_object); 987 VM_OBJECT_UNLOCK(lobject); 988 lobject = backing_object; 989 } 990 /* 991 * give-up when a page is not in memory 992 */ 993 if (m == NULL) { 994 VM_OBJECT_UNLOCK(lobject); 995 break; 996 } 997 vm_page_lock_queues(); 998 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 999 (m->busy == 0) && 1000 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 1001 1002 if ((m->queue - m->pc) == PQ_CACHE) { 1003 vm_page_deactivate(m); 1004 } 1005 vm_page_busy(m); 1006 vm_page_unlock_queues(); 1007 VM_OBJECT_UNLOCK(lobject); 1008 mpte = pmap_enter_quick(pmap, addr, m, mpte); 1009 VM_OBJECT_LOCK(lobject); 1010 vm_page_lock_queues(); 1011 vm_page_wakeup(m); 1012 } 1013 vm_page_unlock_queues(); 1014 VM_OBJECT_UNLOCK(lobject); 1015 } 1016 } 1017 1018 /* 1019 * vm_fault_quick: 1020 * 1021 * Ensure that the requested virtual address, which may be in userland, 1022 * is valid. Fault-in the page if necessary. Return -1 on failure. 1023 */ 1024 int 1025 vm_fault_quick(caddr_t v, int prot) 1026 { 1027 int r; 1028 1029 if (prot & VM_PROT_WRITE) 1030 r = subyte(v, fubyte(v)); 1031 else 1032 r = fubyte(v); 1033 return(r); 1034 } 1035 1036 /* 1037 * vm_fault_wire: 1038 * 1039 * Wire down a range of virtual addresses in a map. 1040 */ 1041 int 1042 vm_fault_wire(map, start, end, user_wire) 1043 vm_map_t map; 1044 vm_offset_t start, end; 1045 boolean_t user_wire; 1046 { 1047 vm_offset_t va; 1048 int rv; 1049 1050 /* 1051 * We simulate a fault to get the page and enter it in the physical 1052 * map. For user wiring, we only ask for read access on currently 1053 * read-only sections. 1054 */ 1055 for (va = start; va < end; va += PAGE_SIZE) { 1056 rv = vm_fault(map, va, 1057 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE, 1058 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING); 1059 if (rv) { 1060 if (va != start) 1061 vm_fault_unwire(map, start, va); 1062 return (rv); 1063 } 1064 } 1065 return (KERN_SUCCESS); 1066 } 1067 1068 /* 1069 * vm_fault_unwire: 1070 * 1071 * Unwire a range of virtual addresses in a map. 1072 */ 1073 void 1074 vm_fault_unwire(map, start, end) 1075 vm_map_t map; 1076 vm_offset_t start, end; 1077 { 1078 vm_paddr_t pa; 1079 vm_offset_t va; 1080 pmap_t pmap; 1081 1082 pmap = vm_map_pmap(map); 1083 1084 if (pmap != kernel_pmap) 1085 mtx_lock(&Giant); 1086 /* 1087 * Since the pages are wired down, we must be able to get their 1088 * mappings from the physical map system. 1089 */ 1090 for (va = start; va < end; va += PAGE_SIZE) { 1091 pa = pmap_extract(pmap, va); 1092 if (pa != 0) { 1093 pmap_change_wiring(pmap, va, FALSE); 1094 vm_page_lock_queues(); 1095 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1096 vm_page_unlock_queues(); 1097 } 1098 } 1099 if (pmap != kernel_pmap) 1100 mtx_unlock(&Giant); 1101 } 1102 1103 /* 1104 * Routine: 1105 * vm_fault_copy_entry 1106 * Function: 1107 * Copy all of the pages from a wired-down map entry to another. 1108 * 1109 * In/out conditions: 1110 * The source and destination maps must be locked for write. 1111 * The source map entry must be wired down (or be a sharing map 1112 * entry corresponding to a main map entry that is wired down). 1113 */ 1114 void 1115 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) 1116 vm_map_t dst_map; 1117 vm_map_t src_map; 1118 vm_map_entry_t dst_entry; 1119 vm_map_entry_t src_entry; 1120 { 1121 vm_object_t backing_object, dst_object, object; 1122 vm_object_t src_object; 1123 vm_ooffset_t dst_offset; 1124 vm_ooffset_t src_offset; 1125 vm_pindex_t pindex; 1126 vm_prot_t prot; 1127 vm_offset_t vaddr; 1128 vm_page_t dst_m; 1129 vm_page_t src_m; 1130 1131 #ifdef lint 1132 src_map++; 1133 #endif /* lint */ 1134 1135 src_object = src_entry->object.vm_object; 1136 src_offset = src_entry->offset; 1137 1138 /* 1139 * Create the top-level object for the destination entry. (Doesn't 1140 * actually shadow anything - we copy the pages directly.) 1141 */ 1142 dst_object = vm_object_allocate(OBJT_DEFAULT, 1143 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1144 1145 VM_OBJECT_LOCK(dst_object); 1146 dst_entry->object.vm_object = dst_object; 1147 dst_entry->offset = 0; 1148 1149 prot = dst_entry->max_protection; 1150 1151 /* 1152 * Loop through all of the pages in the entry's range, copying each 1153 * one from the source object (it should be there) to the destination 1154 * object. 1155 */ 1156 for (vaddr = dst_entry->start, dst_offset = 0; 1157 vaddr < dst_entry->end; 1158 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1159 1160 /* 1161 * Allocate a page in the destination object 1162 */ 1163 do { 1164 dst_m = vm_page_alloc(dst_object, 1165 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1166 if (dst_m == NULL) { 1167 VM_OBJECT_UNLOCK(dst_object); 1168 VM_WAIT; 1169 VM_OBJECT_LOCK(dst_object); 1170 } 1171 } while (dst_m == NULL); 1172 1173 /* 1174 * Find the page in the source object, and copy it in. 1175 * (Because the source is wired down, the page will be in 1176 * memory.) 1177 */ 1178 VM_OBJECT_LOCK(src_object); 1179 object = src_object; 1180 pindex = 0; 1181 while ((src_m = vm_page_lookup(object, pindex + 1182 OFF_TO_IDX(dst_offset + src_offset))) == NULL && 1183 (src_entry->protection & VM_PROT_WRITE) == 0 && 1184 (backing_object = object->backing_object) != NULL) { 1185 /* 1186 * Allow fallback to backing objects if we are reading. 1187 */ 1188 VM_OBJECT_LOCK(backing_object); 1189 pindex += OFF_TO_IDX(object->backing_object_offset); 1190 VM_OBJECT_UNLOCK(object); 1191 object = backing_object; 1192 } 1193 if (src_m == NULL) 1194 panic("vm_fault_copy_wired: page missing"); 1195 pmap_copy_page(src_m, dst_m); 1196 VM_OBJECT_UNLOCK(object); 1197 dst_m->valid = VM_PAGE_BITS_ALL; 1198 VM_OBJECT_UNLOCK(dst_object); 1199 1200 /* 1201 * Enter it in the pmap... 1202 */ 1203 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1204 VM_OBJECT_LOCK(dst_object); 1205 vm_page_lock_queues(); 1206 if ((prot & VM_PROT_WRITE) != 0) 1207 vm_page_flag_set(dst_m, PG_WRITEABLE); 1208 1209 /* 1210 * Mark it no longer busy, and put it on the active list. 1211 */ 1212 vm_page_activate(dst_m); 1213 vm_page_wakeup(dst_m); 1214 vm_page_unlock_queues(); 1215 } 1216 VM_OBJECT_UNLOCK(dst_object); 1217 } 1218 1219 1220 /* 1221 * This routine checks around the requested page for other pages that 1222 * might be able to be faulted in. This routine brackets the viable 1223 * pages for the pages to be paged in. 1224 * 1225 * Inputs: 1226 * m, rbehind, rahead 1227 * 1228 * Outputs: 1229 * marray (array of vm_page_t), reqpage (index of requested page) 1230 * 1231 * Return value: 1232 * number of pages in marray 1233 * 1234 * This routine can't block. 1235 */ 1236 static int 1237 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1238 vm_page_t m; 1239 int rbehind; 1240 int rahead; 1241 vm_page_t *marray; 1242 int *reqpage; 1243 { 1244 int i,j; 1245 vm_object_t object; 1246 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1247 vm_page_t rtm; 1248 int cbehind, cahead; 1249 1250 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1251 1252 object = m->object; 1253 pindex = m->pindex; 1254 1255 /* 1256 * we don't fault-ahead for device pager 1257 */ 1258 if (object->type == OBJT_DEVICE) { 1259 *reqpage = 0; 1260 marray[0] = m; 1261 return 1; 1262 } 1263 1264 /* 1265 * if the requested page is not available, then give up now 1266 */ 1267 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1268 return 0; 1269 } 1270 1271 if ((cbehind == 0) && (cahead == 0)) { 1272 *reqpage = 0; 1273 marray[0] = m; 1274 return 1; 1275 } 1276 1277 if (rahead > cahead) { 1278 rahead = cahead; 1279 } 1280 1281 if (rbehind > cbehind) { 1282 rbehind = cbehind; 1283 } 1284 1285 /* 1286 * try to do any readahead that we might have free pages for. 1287 */ 1288 if ((rahead + rbehind) > 1289 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) { 1290 pagedaemon_wakeup(); 1291 marray[0] = m; 1292 *reqpage = 0; 1293 return 1; 1294 } 1295 1296 /* 1297 * scan backward for the read behind pages -- in memory 1298 */ 1299 if (pindex > 0) { 1300 if (rbehind > pindex) { 1301 rbehind = pindex; 1302 startpindex = 0; 1303 } else { 1304 startpindex = pindex - rbehind; 1305 } 1306 1307 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1308 if (vm_page_lookup(object, tpindex)) { 1309 startpindex = tpindex + 1; 1310 break; 1311 } 1312 if (tpindex == 0) 1313 break; 1314 } 1315 1316 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1317 1318 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1319 if (rtm == NULL) { 1320 vm_page_lock_queues(); 1321 for (j = 0; j < i; j++) { 1322 vm_page_free(marray[j]); 1323 } 1324 vm_page_unlock_queues(); 1325 marray[0] = m; 1326 *reqpage = 0; 1327 return 1; 1328 } 1329 1330 marray[i] = rtm; 1331 } 1332 } else { 1333 startpindex = 0; 1334 i = 0; 1335 } 1336 1337 marray[i] = m; 1338 /* page offset of the required page */ 1339 *reqpage = i; 1340 1341 tpindex = pindex + 1; 1342 i++; 1343 1344 /* 1345 * scan forward for the read ahead pages 1346 */ 1347 endpindex = tpindex + rahead; 1348 if (endpindex > object->size) 1349 endpindex = object->size; 1350 1351 for (; tpindex < endpindex; i++, tpindex++) { 1352 1353 if (vm_page_lookup(object, tpindex)) { 1354 break; 1355 } 1356 1357 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1358 if (rtm == NULL) { 1359 break; 1360 } 1361 1362 marray[i] = rtm; 1363 } 1364 1365 /* return number of bytes of pages */ 1366 return i; 1367 } 1368