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 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */ 101 102 #define PFBAK 4 103 #define PFFOR 4 104 #define PAGEORDER_SIZE (PFBAK+PFFOR) 105 106 static int prefault_pageorder[] = { 107 -1 * PAGE_SIZE, 1 * PAGE_SIZE, 108 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 109 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 110 -4 * PAGE_SIZE, 4 * PAGE_SIZE 111 }; 112 113 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 114 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t); 115 116 #define VM_FAULT_READ_AHEAD 8 117 #define VM_FAULT_READ_BEHIND 7 118 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 119 120 struct faultstate { 121 vm_page_t m; 122 vm_object_t object; 123 vm_pindex_t pindex; 124 vm_page_t first_m; 125 vm_object_t first_object; 126 vm_pindex_t first_pindex; 127 vm_map_t map; 128 vm_map_entry_t entry; 129 int lookup_still_valid; 130 struct vnode *vp; 131 }; 132 133 static inline void 134 release_page(struct faultstate *fs) 135 { 136 vm_page_lock_queues(); 137 vm_page_wakeup(fs->m); 138 vm_page_deactivate(fs->m); 139 vm_page_unlock_queues(); 140 fs->m = NULL; 141 } 142 143 static inline void 144 unlock_map(struct faultstate *fs) 145 { 146 if (fs->lookup_still_valid) { 147 vm_map_lookup_done(fs->map, fs->entry); 148 fs->lookup_still_valid = FALSE; 149 } 150 } 151 152 static void 153 unlock_and_deallocate(struct faultstate *fs) 154 { 155 boolean_t firstobjneedgiant; 156 157 vm_object_pip_wakeup(fs->object); 158 VM_OBJECT_UNLOCK(fs->object); 159 if (fs->object != fs->first_object) { 160 VM_OBJECT_LOCK(fs->first_object); 161 vm_page_lock_queues(); 162 vm_page_free(fs->first_m); 163 vm_page_unlock_queues(); 164 vm_object_pip_wakeup(fs->first_object); 165 VM_OBJECT_UNLOCK(fs->first_object); 166 fs->first_m = NULL; 167 } 168 firstobjneedgiant = (fs->first_object->flags & OBJ_NEEDGIANT) != 0; 169 vm_object_deallocate(fs->first_object); 170 unlock_map(fs); 171 if (fs->vp != NULL) { 172 int vfslocked; 173 174 vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount); 175 vput(fs->vp); 176 fs->vp = NULL; 177 VFS_UNLOCK_GIANT(vfslocked); 178 } 179 if (firstobjneedgiant) 180 VM_UNLOCK_GIANT(); 181 } 182 183 /* 184 * TRYPAGER - used by vm_fault to calculate whether the pager for the 185 * current object *might* contain the page. 186 * 187 * default objects are zero-fill, there is no real pager. 188 */ 189 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 190 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 191 192 /* 193 * vm_fault: 194 * 195 * Handle a page fault occurring at the given address, 196 * requiring the given permissions, in the map specified. 197 * If successful, the page is inserted into the 198 * associated physical map. 199 * 200 * NOTE: the given address should be truncated to the 201 * proper page address. 202 * 203 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 204 * a standard error specifying why the fault is fatal is returned. 205 * 206 * 207 * The map in question must be referenced, and remains so. 208 * Caller may hold no locks. 209 */ 210 int 211 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 212 int fault_flags) 213 { 214 vm_prot_t prot; 215 int is_first_object_locked, result; 216 boolean_t growstack, wired; 217 int map_generation; 218 vm_object_t next_object; 219 vm_page_t marray[VM_FAULT_READ]; 220 int hardfault; 221 int faultcount; 222 struct faultstate fs; 223 224 hardfault = 0; 225 growstack = TRUE; 226 atomic_add_int(&cnt.v_vm_faults, 1); 227 228 RetryFault:; 229 230 /* 231 * Find the backing store object and offset into it to begin the 232 * search. 233 */ 234 fs.map = map; 235 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 236 &fs.first_object, &fs.first_pindex, &prot, &wired); 237 if (result != KERN_SUCCESS) { 238 if (result != KERN_PROTECTION_FAILURE || 239 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) { 240 if (growstack && result == KERN_INVALID_ADDRESS && 241 map != kernel_map && curproc != NULL) { 242 result = vm_map_growstack(curproc, vaddr); 243 if (result != KERN_SUCCESS) 244 return (KERN_FAILURE); 245 growstack = FALSE; 246 goto RetryFault; 247 } 248 return (result); 249 } 250 251 /* 252 * If we are user-wiring a r/w segment, and it is COW, then 253 * we need to do the COW operation. Note that we don't COW 254 * currently RO sections now, because it is NOT desirable 255 * to COW .text. We simply keep .text from ever being COW'ed 256 * and take the heat that one cannot debug wired .text sections. 257 */ 258 result = vm_map_lookup(&fs.map, vaddr, 259 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 260 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 261 if (result != KERN_SUCCESS) 262 return (result); 263 264 /* 265 * If we don't COW now, on a user wire, the user will never 266 * be able to write to the mapping. If we don't make this 267 * restriction, the bookkeeping would be nearly impossible. 268 * 269 * XXX The following assignment modifies the map without 270 * holding a write lock on it. 271 */ 272 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 273 fs.entry->max_protection &= ~VM_PROT_WRITE; 274 } 275 276 map_generation = fs.map->timestamp; 277 278 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 279 panic("vm_fault: fault on nofault entry, addr: %lx", 280 (u_long)vaddr); 281 } 282 283 /* 284 * Make a reference to this object to prevent its disposal while we 285 * are messing with it. Once we have the reference, the map is free 286 * to be diddled. Since objects reference their shadows (and copies), 287 * they will stay around as well. 288 * 289 * Bump the paging-in-progress count to prevent size changes (e.g. 290 * truncation operations) during I/O. This must be done after 291 * obtaining the vnode lock in order to avoid possible deadlocks. 292 * 293 * XXX vnode_pager_lock() can block without releasing the map lock. 294 */ 295 if (fs.first_object->flags & OBJ_NEEDGIANT) 296 mtx_lock(&Giant); 297 VM_OBJECT_LOCK(fs.first_object); 298 vm_object_reference_locked(fs.first_object); 299 fs.vp = vnode_pager_lock(fs.first_object); 300 KASSERT(fs.vp == NULL || !fs.map->system_map, 301 ("vm_fault: vnode-backed object mapped by system map")); 302 KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 || 303 !fs.map->system_map, 304 ("vm_fault: Object requiring giant mapped by system map")); 305 if (fs.first_object->flags & OBJ_NEEDGIANT && debug_mpsafevm) 306 mtx_unlock(&Giant); 307 vm_object_pip_add(fs.first_object, 1); 308 309 fs.lookup_still_valid = TRUE; 310 311 if (wired) 312 fault_type = prot; 313 314 fs.first_m = NULL; 315 316 /* 317 * Search for the page at object/offset. 318 */ 319 fs.object = fs.first_object; 320 fs.pindex = fs.first_pindex; 321 while (TRUE) { 322 /* 323 * If the object is dead, we stop here 324 */ 325 if (fs.object->flags & OBJ_DEAD) { 326 unlock_and_deallocate(&fs); 327 return (KERN_PROTECTION_FAILURE); 328 } 329 330 /* 331 * See if page is resident 332 */ 333 fs.m = vm_page_lookup(fs.object, fs.pindex); 334 if (fs.m != NULL) { 335 int queue; 336 337 /* 338 * check for page-based copy on write. 339 * We check fs.object == fs.first_object so 340 * as to ensure the legacy COW mechanism is 341 * used when the page in question is part of 342 * a shadow object. Otherwise, vm_page_cowfault() 343 * removes the page from the backing object, 344 * which is not what we want. 345 */ 346 vm_page_lock_queues(); 347 if ((fs.m->cow) && 348 (fault_type & VM_PROT_WRITE) && 349 (fs.object == fs.first_object)) { 350 vm_page_cowfault(fs.m); 351 vm_page_unlock_queues(); 352 unlock_and_deallocate(&fs); 353 goto RetryFault; 354 } 355 356 /* 357 * Wait/Retry if the page is busy. We have to do this 358 * if the page is busy via either PG_BUSY or 359 * vm_page_t->busy because the vm_pager may be using 360 * vm_page_t->busy for pageouts ( and even pageins if 361 * it is the vnode pager ), and we could end up trying 362 * to pagein and pageout the same page simultaneously. 363 * 364 * We can theoretically allow the busy case on a read 365 * fault if the page is marked valid, but since such 366 * pages are typically already pmap'd, putting that 367 * special case in might be more effort then it is 368 * worth. We cannot under any circumstances mess 369 * around with a vm_page_t->busy page except, perhaps, 370 * to pmap it. 371 */ 372 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 373 vm_page_unlock_queues(); 374 VM_OBJECT_UNLOCK(fs.object); 375 if (fs.object != fs.first_object) { 376 VM_OBJECT_LOCK(fs.first_object); 377 vm_page_lock_queues(); 378 vm_page_free(fs.first_m); 379 vm_page_unlock_queues(); 380 vm_object_pip_wakeup(fs.first_object); 381 VM_OBJECT_UNLOCK(fs.first_object); 382 fs.first_m = NULL; 383 } 384 unlock_map(&fs); 385 if (fs.vp != NULL) { 386 int vfslck; 387 388 vfslck = VFS_LOCK_GIANT(fs.vp->v_mount); 389 vput(fs.vp); 390 fs.vp = NULL; 391 VFS_UNLOCK_GIANT(vfslck); 392 } 393 VM_OBJECT_LOCK(fs.object); 394 if (fs.m == vm_page_lookup(fs.object, 395 fs.pindex)) { 396 vm_page_lock_queues(); 397 if (!vm_page_sleep_if_busy(fs.m, TRUE, 398 "vmpfw")) 399 vm_page_unlock_queues(); 400 } 401 vm_object_pip_wakeup(fs.object); 402 VM_OBJECT_UNLOCK(fs.object); 403 atomic_add_int(&cnt.v_intrans, 1); 404 if (fs.first_object->flags & OBJ_NEEDGIANT) 405 VM_UNLOCK_GIANT(); 406 vm_object_deallocate(fs.first_object); 407 goto RetryFault; 408 } 409 queue = fs.m->queue; 410 411 vm_pageq_remove_nowakeup(fs.m); 412 413 if (VM_PAGE_RESOLVEQUEUE(fs.m, queue) == PQ_CACHE && 414 vm_page_count_severe()) { 415 vm_page_activate(fs.m); 416 vm_page_unlock_queues(); 417 unlock_and_deallocate(&fs); 418 VM_WAITPFAULT; 419 goto RetryFault; 420 } 421 422 /* 423 * Mark page busy for other processes, and the 424 * pagedaemon. If it still isn't completely valid 425 * (readable), jump to readrest, else break-out ( we 426 * found the page ). 427 */ 428 vm_page_busy(fs.m); 429 vm_page_unlock_queues(); 430 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 431 fs.m->object != kernel_object && fs.m->object != kmem_object) { 432 goto readrest; 433 } 434 435 break; 436 } 437 438 /* 439 * Page is not resident, If this is the search termination 440 * or the pager might contain the page, allocate a new page. 441 */ 442 if (TRYPAGER || fs.object == fs.first_object) { 443 if (fs.pindex >= fs.object->size) { 444 unlock_and_deallocate(&fs); 445 return (KERN_PROTECTION_FAILURE); 446 } 447 448 /* 449 * Allocate a new page for this object/offset pair. 450 */ 451 fs.m = NULL; 452 if (!vm_page_count_severe()) { 453 fs.m = vm_page_alloc(fs.object, fs.pindex, 454 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); 455 } 456 if (fs.m == NULL) { 457 unlock_and_deallocate(&fs); 458 VM_WAITPFAULT; 459 goto RetryFault; 460 } 461 } 462 463 readrest: 464 /* 465 * We have found a valid page or we have allocated a new page. 466 * The page thus may not be valid or may not be entirely 467 * valid. 468 * 469 * Attempt to fault-in the page if there is a chance that the 470 * pager has it, and potentially fault in additional pages 471 * at the same time. 472 */ 473 if (TRYPAGER) { 474 int rv; 475 int reqpage; 476 int ahead, behind; 477 u_char behavior = vm_map_entry_behavior(fs.entry); 478 479 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 480 ahead = 0; 481 behind = 0; 482 } else { 483 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 484 if (behind > VM_FAULT_READ_BEHIND) 485 behind = VM_FAULT_READ_BEHIND; 486 487 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 488 if (ahead > VM_FAULT_READ_AHEAD) 489 ahead = VM_FAULT_READ_AHEAD; 490 } 491 is_first_object_locked = FALSE; 492 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 493 (behavior != MAP_ENTRY_BEHAV_RANDOM && 494 fs.pindex >= fs.entry->lastr && 495 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 496 (fs.first_object == fs.object || 497 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 498 fs.first_object->type != OBJT_DEVICE) { 499 vm_pindex_t firstpindex, tmppindex; 500 501 if (fs.first_pindex < 2 * VM_FAULT_READ) 502 firstpindex = 0; 503 else 504 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 505 506 vm_page_lock_queues(); 507 /* 508 * note: partially valid pages cannot be 509 * included in the lookahead - NFS piecemeal 510 * writes will barf on it badly. 511 */ 512 for (tmppindex = fs.first_pindex - 1; 513 tmppindex >= firstpindex; 514 --tmppindex) { 515 vm_page_t mt; 516 517 mt = vm_page_lookup(fs.first_object, tmppindex); 518 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 519 break; 520 if (mt->busy || 521 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 522 mt->hold_count || 523 mt->wire_count) 524 continue; 525 pmap_remove_all(mt); 526 if (mt->dirty) { 527 vm_page_deactivate(mt); 528 } else { 529 vm_page_cache(mt); 530 } 531 } 532 vm_page_unlock_queues(); 533 ahead += behind; 534 behind = 0; 535 } 536 if (is_first_object_locked) 537 VM_OBJECT_UNLOCK(fs.first_object); 538 /* 539 * now we find out if any other pages should be paged 540 * in at this time this routine checks to see if the 541 * pages surrounding this fault reside in the same 542 * object as the page for this fault. If they do, 543 * then they are faulted in also into the object. The 544 * array "marray" returned contains an array of 545 * vm_page_t structs where one of them is the 546 * vm_page_t passed to the routine. The reqpage 547 * return value is the index into the marray for the 548 * vm_page_t passed to the routine. 549 * 550 * fs.m plus the additional pages are PG_BUSY'd. 551 * 552 * XXX vm_fault_additional_pages() can block 553 * without releasing the map lock. 554 */ 555 faultcount = vm_fault_additional_pages( 556 fs.m, behind, ahead, marray, &reqpage); 557 558 /* 559 * update lastr imperfectly (we do not know how much 560 * getpages will actually read), but good enough. 561 * 562 * XXX The following assignment modifies the map 563 * without holding a write lock on it. 564 */ 565 fs.entry->lastr = fs.pindex + faultcount - behind; 566 567 /* 568 * Call the pager to retrieve the data, if any, after 569 * releasing the lock on the map. We hold a ref on 570 * fs.object and the pages are PG_BUSY'd. 571 */ 572 unlock_map(&fs); 573 574 rv = faultcount ? 575 vm_pager_get_pages(fs.object, marray, faultcount, 576 reqpage) : VM_PAGER_FAIL; 577 578 if (rv == VM_PAGER_OK) { 579 /* 580 * Found the page. Leave it busy while we play 581 * with it. 582 */ 583 584 /* 585 * Relookup in case pager changed page. Pager 586 * is responsible for disposition of old page 587 * if moved. 588 */ 589 fs.m = vm_page_lookup(fs.object, fs.pindex); 590 if (!fs.m) { 591 unlock_and_deallocate(&fs); 592 goto RetryFault; 593 } 594 595 hardfault++; 596 break; /* break to PAGE HAS BEEN FOUND */ 597 } 598 /* 599 * Remove the bogus page (which does not exist at this 600 * object/offset); before doing so, we must get back 601 * our object lock to preserve our invariant. 602 * 603 * Also wake up any other process that may want to bring 604 * in this page. 605 * 606 * If this is the top-level object, we must leave the 607 * busy page to prevent another process from rushing 608 * past us, and inserting the page in that object at 609 * the same time that we are. 610 */ 611 if (rv == VM_PAGER_ERROR) 612 printf("vm_fault: pager read error, pid %d (%s)\n", 613 curproc->p_pid, curproc->p_comm); 614 /* 615 * Data outside the range of the pager or an I/O error 616 */ 617 /* 618 * XXX - the check for kernel_map is a kludge to work 619 * around having the machine panic on a kernel space 620 * fault w/ I/O error. 621 */ 622 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 623 (rv == VM_PAGER_BAD)) { 624 vm_page_lock_queues(); 625 vm_page_free(fs.m); 626 vm_page_unlock_queues(); 627 fs.m = NULL; 628 unlock_and_deallocate(&fs); 629 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 630 } 631 if (fs.object != fs.first_object) { 632 vm_page_lock_queues(); 633 vm_page_free(fs.m); 634 vm_page_unlock_queues(); 635 fs.m = NULL; 636 /* 637 * XXX - we cannot just fall out at this 638 * point, m has been freed and is invalid! 639 */ 640 } 641 } 642 643 /* 644 * We get here if the object has default pager (or unwiring) 645 * or the pager doesn't have the page. 646 */ 647 if (fs.object == fs.first_object) 648 fs.first_m = fs.m; 649 650 /* 651 * Move on to the next object. Lock the next object before 652 * unlocking the current one. 653 */ 654 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 655 next_object = fs.object->backing_object; 656 if (next_object == NULL) { 657 /* 658 * If there's no object left, fill the page in the top 659 * object with zeros. 660 */ 661 if (fs.object != fs.first_object) { 662 vm_object_pip_wakeup(fs.object); 663 VM_OBJECT_UNLOCK(fs.object); 664 665 fs.object = fs.first_object; 666 fs.pindex = fs.first_pindex; 667 fs.m = fs.first_m; 668 VM_OBJECT_LOCK(fs.object); 669 } 670 fs.first_m = NULL; 671 672 /* 673 * Zero the page if necessary and mark it valid. 674 */ 675 if ((fs.m->flags & PG_ZERO) == 0) { 676 pmap_zero_page(fs.m); 677 } else { 678 atomic_add_int(&cnt.v_ozfod, 1); 679 } 680 atomic_add_int(&cnt.v_zfod, 1); 681 fs.m->valid = VM_PAGE_BITS_ALL; 682 break; /* break to PAGE HAS BEEN FOUND */ 683 } else { 684 KASSERT(fs.object != next_object, 685 ("object loop %p", next_object)); 686 VM_OBJECT_LOCK(next_object); 687 vm_object_pip_add(next_object, 1); 688 if (fs.object != fs.first_object) 689 vm_object_pip_wakeup(fs.object); 690 VM_OBJECT_UNLOCK(fs.object); 691 fs.object = next_object; 692 } 693 } 694 695 KASSERT((fs.m->flags & PG_BUSY) != 0, 696 ("vm_fault: not busy after main loop")); 697 698 /* 699 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 700 * is held.] 701 */ 702 703 /* 704 * If the page is being written, but isn't already owned by the 705 * top-level object, we have to copy it into a new page owned by the 706 * top-level object. 707 */ 708 if (fs.object != fs.first_object) { 709 /* 710 * We only really need to copy if we want to write it. 711 */ 712 if (fault_type & VM_PROT_WRITE) { 713 /* 714 * This allows pages to be virtually copied from a 715 * backing_object into the first_object, where the 716 * backing object has no other refs to it, and cannot 717 * gain any more refs. Instead of a bcopy, we just 718 * move the page from the backing object to the 719 * first object. Note that we must mark the page 720 * dirty in the first object so that it will go out 721 * to swap when needed. 722 */ 723 is_first_object_locked = FALSE; 724 if ( 725 /* 726 * Only one shadow object 727 */ 728 (fs.object->shadow_count == 1) && 729 /* 730 * No COW refs, except us 731 */ 732 (fs.object->ref_count == 1) && 733 /* 734 * No one else can look this object up 735 */ 736 (fs.object->handle == NULL) && 737 /* 738 * No other ways to look the object up 739 */ 740 ((fs.object->type == OBJT_DEFAULT) || 741 (fs.object->type == OBJT_SWAP)) && 742 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 743 /* 744 * We don't chase down the shadow chain 745 */ 746 fs.object == fs.first_object->backing_object) { 747 vm_page_lock_queues(); 748 /* 749 * get rid of the unnecessary page 750 */ 751 vm_page_free(fs.first_m); 752 /* 753 * grab the page and put it into the 754 * process'es object. The page is 755 * automatically made dirty. 756 */ 757 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 758 vm_page_busy(fs.m); 759 vm_page_unlock_queues(); 760 fs.first_m = fs.m; 761 fs.m = NULL; 762 atomic_add_int(&cnt.v_cow_optim, 1); 763 } else { 764 /* 765 * Oh, well, lets copy it. 766 */ 767 pmap_copy_page(fs.m, fs.first_m); 768 fs.first_m->valid = VM_PAGE_BITS_ALL; 769 } 770 if (fs.m) { 771 /* 772 * We no longer need the old page or object. 773 */ 774 release_page(&fs); 775 } 776 /* 777 * fs.object != fs.first_object due to above 778 * conditional 779 */ 780 vm_object_pip_wakeup(fs.object); 781 VM_OBJECT_UNLOCK(fs.object); 782 /* 783 * Only use the new page below... 784 */ 785 fs.object = fs.first_object; 786 fs.pindex = fs.first_pindex; 787 fs.m = fs.first_m; 788 if (!is_first_object_locked) 789 VM_OBJECT_LOCK(fs.object); 790 atomic_add_int(&cnt.v_cow_faults, 1); 791 } else { 792 prot &= ~VM_PROT_WRITE; 793 } 794 } 795 796 /* 797 * We must verify that the maps have not changed since our last 798 * lookup. 799 */ 800 if (!fs.lookup_still_valid) { 801 vm_object_t retry_object; 802 vm_pindex_t retry_pindex; 803 vm_prot_t retry_prot; 804 805 if (!vm_map_trylock_read(fs.map)) { 806 release_page(&fs); 807 unlock_and_deallocate(&fs); 808 goto RetryFault; 809 } 810 fs.lookup_still_valid = TRUE; 811 if (fs.map->timestamp != map_generation) { 812 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 813 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 814 815 /* 816 * If we don't need the page any longer, put it on the inactive 817 * list (the easiest thing to do here). If no one needs it, 818 * pageout will grab it eventually. 819 */ 820 if (result != KERN_SUCCESS) { 821 release_page(&fs); 822 unlock_and_deallocate(&fs); 823 824 /* 825 * If retry of map lookup would have blocked then 826 * retry fault from start. 827 */ 828 if (result == KERN_FAILURE) 829 goto RetryFault; 830 return (result); 831 } 832 if ((retry_object != fs.first_object) || 833 (retry_pindex != fs.first_pindex)) { 834 release_page(&fs); 835 unlock_and_deallocate(&fs); 836 goto RetryFault; 837 } 838 839 /* 840 * Check whether the protection has changed or the object has 841 * been copied while we left the map unlocked. Changing from 842 * read to write permission is OK - we leave the page 843 * write-protected, and catch the write fault. Changing from 844 * write to read permission means that we can't mark the page 845 * write-enabled after all. 846 */ 847 prot &= retry_prot; 848 } 849 } 850 if (prot & VM_PROT_WRITE) { 851 vm_page_lock_queues(); 852 vm_page_flag_set(fs.m, PG_WRITEABLE); 853 vm_object_set_writeable_dirty(fs.m->object); 854 855 /* 856 * If the fault is a write, we know that this page is being 857 * written NOW so dirty it explicitly to save on 858 * pmap_is_modified() calls later. 859 * 860 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 861 * if the page is already dirty to prevent data written with 862 * the expectation of being synced from not being synced. 863 * Likewise if this entry does not request NOSYNC then make 864 * sure the page isn't marked NOSYNC. Applications sharing 865 * data should use the same flags to avoid ping ponging. 866 * 867 * Also tell the backing pager, if any, that it should remove 868 * any swap backing since the page is now dirty. 869 */ 870 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 871 if (fs.m->dirty == 0) 872 vm_page_flag_set(fs.m, PG_NOSYNC); 873 } else { 874 vm_page_flag_clear(fs.m, PG_NOSYNC); 875 } 876 vm_page_unlock_queues(); 877 if (fault_flags & VM_FAULT_DIRTY) { 878 vm_page_dirty(fs.m); 879 vm_pager_page_unswapped(fs.m); 880 } 881 } 882 883 /* 884 * Page had better still be busy 885 */ 886 KASSERT(fs.m->flags & PG_BUSY, 887 ("vm_fault: page %p not busy!", fs.m)); 888 /* 889 * Sanity check: page must be completely valid or it is not fit to 890 * map into user space. vm_pager_get_pages() ensures this. 891 */ 892 if (fs.m->valid != VM_PAGE_BITS_ALL) { 893 vm_page_zero_invalid(fs.m, TRUE); 894 printf("Warning: page %p partially invalid on fault\n", fs.m); 895 } 896 VM_OBJECT_UNLOCK(fs.object); 897 898 /* 899 * Put this page into the physical map. We had to do the unlock above 900 * because pmap_enter() may sleep. We don't put the page 901 * back on the active queue until later so that the pageout daemon 902 * won't find it (yet). 903 */ 904 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 905 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 906 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 907 } 908 VM_OBJECT_LOCK(fs.object); 909 vm_page_lock_queues(); 910 vm_page_flag_set(fs.m, PG_REFERENCED); 911 912 /* 913 * If the page is not wired down, then put it where the pageout daemon 914 * can find it. 915 */ 916 if (fault_flags & VM_FAULT_WIRE_MASK) { 917 if (wired) 918 vm_page_wire(fs.m); 919 else 920 vm_page_unwire(fs.m, 1); 921 } else { 922 vm_page_activate(fs.m); 923 } 924 vm_page_wakeup(fs.m); 925 vm_page_unlock_queues(); 926 927 /* 928 * Unlock everything, and return 929 */ 930 unlock_and_deallocate(&fs); 931 PROC_LOCK(curproc); 932 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) { 933 if (hardfault) { 934 curproc->p_stats->p_ru.ru_majflt++; 935 } else { 936 curproc->p_stats->p_ru.ru_minflt++; 937 } 938 } 939 PROC_UNLOCK(curproc); 940 941 return (KERN_SUCCESS); 942 } 943 944 /* 945 * vm_fault_prefault provides a quick way of clustering 946 * pagefaults into a processes address space. It is a "cousin" 947 * of vm_map_pmap_enter, except it runs at page fault time instead 948 * of mmap time. 949 */ 950 static void 951 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 952 { 953 int i; 954 vm_offset_t addr, starta; 955 vm_pindex_t pindex; 956 vm_page_t m, mpte; 957 vm_object_t object; 958 959 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 960 return; 961 962 object = entry->object.vm_object; 963 964 starta = addra - PFBAK * PAGE_SIZE; 965 if (starta < entry->start) { 966 starta = entry->start; 967 } else if (starta > addra) { 968 starta = 0; 969 } 970 971 mpte = NULL; 972 for (i = 0; i < PAGEORDER_SIZE; i++) { 973 vm_object_t backing_object, lobject; 974 975 addr = addra + prefault_pageorder[i]; 976 if (addr > addra + (PFFOR * PAGE_SIZE)) 977 addr = 0; 978 979 if (addr < starta || addr >= entry->end) 980 continue; 981 982 if (!pmap_is_prefaultable(pmap, addr)) 983 continue; 984 985 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 986 lobject = object; 987 VM_OBJECT_LOCK(lobject); 988 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 989 lobject->type == OBJT_DEFAULT && 990 (backing_object = lobject->backing_object) != NULL) { 991 if (lobject->backing_object_offset & PAGE_MASK) 992 break; 993 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 994 VM_OBJECT_LOCK(backing_object); 995 VM_OBJECT_UNLOCK(lobject); 996 lobject = backing_object; 997 } 998 /* 999 * give-up when a page is not in memory 1000 */ 1001 if (m == NULL) { 1002 VM_OBJECT_UNLOCK(lobject); 1003 break; 1004 } 1005 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 1006 (m->busy == 0) && 1007 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 1008 1009 vm_page_lock_queues(); 1010 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1011 vm_page_deactivate(m); 1012 mpte = pmap_enter_quick(pmap, addr, m, 1013 entry->protection, mpte); 1014 vm_page_unlock_queues(); 1015 } 1016 VM_OBJECT_UNLOCK(lobject); 1017 } 1018 } 1019 1020 /* 1021 * vm_fault_quick: 1022 * 1023 * Ensure that the requested virtual address, which may be in userland, 1024 * is valid. Fault-in the page if necessary. Return -1 on failure. 1025 */ 1026 int 1027 vm_fault_quick(caddr_t v, int prot) 1028 { 1029 int r; 1030 1031 if (prot & VM_PROT_WRITE) 1032 r = subyte(v, fubyte(v)); 1033 else 1034 r = fubyte(v); 1035 return(r); 1036 } 1037 1038 /* 1039 * vm_fault_wire: 1040 * 1041 * Wire down a range of virtual addresses in a map. 1042 */ 1043 int 1044 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1045 boolean_t user_wire, boolean_t fictitious) 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, fictitious); 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(vm_map_t map, vm_offset_t start, vm_offset_t end, 1075 boolean_t fictitious) 1076 { 1077 vm_paddr_t pa; 1078 vm_offset_t va; 1079 pmap_t pmap; 1080 1081 pmap = vm_map_pmap(map); 1082 1083 /* 1084 * Since the pages are wired down, we must be able to get their 1085 * mappings from the physical map system. 1086 */ 1087 for (va = start; va < end; va += PAGE_SIZE) { 1088 pa = pmap_extract(pmap, va); 1089 if (pa != 0) { 1090 pmap_change_wiring(pmap, va, FALSE); 1091 if (!fictitious) { 1092 vm_page_lock_queues(); 1093 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1094 vm_page_unlock_queues(); 1095 } 1096 } 1097 } 1098 } 1099 1100 /* 1101 * Routine: 1102 * vm_fault_copy_entry 1103 * Function: 1104 * Copy all of the pages from a wired-down map entry to another. 1105 * 1106 * In/out conditions: 1107 * The source and destination maps must be locked for write. 1108 * The source map entry must be wired down (or be a sharing map 1109 * entry corresponding to a main map entry that is wired down). 1110 */ 1111 void 1112 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) 1113 vm_map_t dst_map; 1114 vm_map_t src_map; 1115 vm_map_entry_t dst_entry; 1116 vm_map_entry_t src_entry; 1117 { 1118 vm_object_t backing_object, dst_object, object; 1119 vm_object_t src_object; 1120 vm_ooffset_t dst_offset; 1121 vm_ooffset_t src_offset; 1122 vm_pindex_t pindex; 1123 vm_prot_t prot; 1124 vm_offset_t vaddr; 1125 vm_page_t dst_m; 1126 vm_page_t src_m; 1127 1128 #ifdef lint 1129 src_map++; 1130 #endif /* lint */ 1131 1132 src_object = src_entry->object.vm_object; 1133 src_offset = src_entry->offset; 1134 1135 /* 1136 * Create the top-level object for the destination entry. (Doesn't 1137 * actually shadow anything - we copy the pages directly.) 1138 */ 1139 dst_object = vm_object_allocate(OBJT_DEFAULT, 1140 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1141 1142 VM_OBJECT_LOCK(dst_object); 1143 dst_entry->object.vm_object = dst_object; 1144 dst_entry->offset = 0; 1145 1146 prot = dst_entry->max_protection; 1147 1148 /* 1149 * Loop through all of the pages in the entry's range, copying each 1150 * one from the source object (it should be there) to the destination 1151 * object. 1152 */ 1153 for (vaddr = dst_entry->start, dst_offset = 0; 1154 vaddr < dst_entry->end; 1155 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1156 1157 /* 1158 * Allocate a page in the destination object 1159 */ 1160 do { 1161 dst_m = vm_page_alloc(dst_object, 1162 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1163 if (dst_m == NULL) { 1164 VM_OBJECT_UNLOCK(dst_object); 1165 VM_WAIT; 1166 VM_OBJECT_LOCK(dst_object); 1167 } 1168 } while (dst_m == NULL); 1169 1170 /* 1171 * Find the page in the source object, and copy it in. 1172 * (Because the source is wired down, the page will be in 1173 * memory.) 1174 */ 1175 VM_OBJECT_LOCK(src_object); 1176 object = src_object; 1177 pindex = 0; 1178 while ((src_m = vm_page_lookup(object, pindex + 1179 OFF_TO_IDX(dst_offset + src_offset))) == NULL && 1180 (src_entry->protection & VM_PROT_WRITE) == 0 && 1181 (backing_object = object->backing_object) != NULL) { 1182 /* 1183 * Allow fallback to backing objects if we are reading. 1184 */ 1185 VM_OBJECT_LOCK(backing_object); 1186 pindex += OFF_TO_IDX(object->backing_object_offset); 1187 VM_OBJECT_UNLOCK(object); 1188 object = backing_object; 1189 } 1190 if (src_m == NULL) 1191 panic("vm_fault_copy_wired: page missing"); 1192 pmap_copy_page(src_m, dst_m); 1193 VM_OBJECT_UNLOCK(object); 1194 dst_m->valid = VM_PAGE_BITS_ALL; 1195 VM_OBJECT_UNLOCK(dst_object); 1196 1197 /* 1198 * Enter it in the pmap... 1199 */ 1200 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1201 VM_OBJECT_LOCK(dst_object); 1202 vm_page_lock_queues(); 1203 if ((prot & VM_PROT_WRITE) != 0) 1204 vm_page_flag_set(dst_m, PG_WRITEABLE); 1205 1206 /* 1207 * Mark it no longer busy, and put it on the active list. 1208 */ 1209 vm_page_activate(dst_m); 1210 vm_page_wakeup(dst_m); 1211 vm_page_unlock_queues(); 1212 } 1213 VM_OBJECT_UNLOCK(dst_object); 1214 } 1215 1216 1217 /* 1218 * This routine checks around the requested page for other pages that 1219 * might be able to be faulted in. This routine brackets the viable 1220 * pages for the pages to be paged in. 1221 * 1222 * Inputs: 1223 * m, rbehind, rahead 1224 * 1225 * Outputs: 1226 * marray (array of vm_page_t), reqpage (index of requested page) 1227 * 1228 * Return value: 1229 * number of pages in marray 1230 * 1231 * This routine can't block. 1232 */ 1233 static int 1234 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1235 vm_page_t m; 1236 int rbehind; 1237 int rahead; 1238 vm_page_t *marray; 1239 int *reqpage; 1240 { 1241 int i,j; 1242 vm_object_t object; 1243 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1244 vm_page_t rtm; 1245 int cbehind, cahead; 1246 1247 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1248 1249 object = m->object; 1250 pindex = m->pindex; 1251 1252 /* 1253 * we don't fault-ahead for device pager 1254 */ 1255 if (object->type == OBJT_DEVICE) { 1256 *reqpage = 0; 1257 marray[0] = m; 1258 return 1; 1259 } 1260 1261 /* 1262 * if the requested page is not available, then give up now 1263 */ 1264 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1265 return 0; 1266 } 1267 1268 if ((cbehind == 0) && (cahead == 0)) { 1269 *reqpage = 0; 1270 marray[0] = m; 1271 return 1; 1272 } 1273 1274 if (rahead > cahead) { 1275 rahead = cahead; 1276 } 1277 1278 if (rbehind > cbehind) { 1279 rbehind = cbehind; 1280 } 1281 1282 /* 1283 * try to do any readahead that we might have free pages for. 1284 */ 1285 if ((rahead + rbehind) > 1286 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) { 1287 pagedaemon_wakeup(); 1288 marray[0] = m; 1289 *reqpage = 0; 1290 return 1; 1291 } 1292 1293 /* 1294 * scan backward for the read behind pages -- in memory 1295 */ 1296 if (pindex > 0) { 1297 if (rbehind > pindex) { 1298 rbehind = pindex; 1299 startpindex = 0; 1300 } else { 1301 startpindex = pindex - rbehind; 1302 } 1303 1304 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1305 if (vm_page_lookup(object, tpindex)) { 1306 startpindex = tpindex + 1; 1307 break; 1308 } 1309 if (tpindex == 0) 1310 break; 1311 } 1312 1313 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1314 1315 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1316 if (rtm == NULL) { 1317 vm_page_lock_queues(); 1318 for (j = 0; j < i; j++) { 1319 vm_page_free(marray[j]); 1320 } 1321 vm_page_unlock_queues(); 1322 marray[0] = m; 1323 *reqpage = 0; 1324 return 1; 1325 } 1326 1327 marray[i] = rtm; 1328 } 1329 } else { 1330 startpindex = 0; 1331 i = 0; 1332 } 1333 1334 marray[i] = m; 1335 /* page offset of the required page */ 1336 *reqpage = i; 1337 1338 tpindex = pindex + 1; 1339 i++; 1340 1341 /* 1342 * scan forward for the read ahead pages 1343 */ 1344 endpindex = tpindex + rahead; 1345 if (endpindex > object->size) 1346 endpindex = object->size; 1347 1348 for (; tpindex < endpindex; i++, tpindex++) { 1349 1350 if (vm_page_lookup(object, tpindex)) { 1351 break; 1352 } 1353 1354 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1355 if (rtm == NULL) { 1356 break; 1357 } 1358 1359 marray[i] = rtm; 1360 } 1361 1362 /* return number of bytes of pages */ 1363 return i; 1364 } 1365