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 { 155 156 vm_object_pip_wakeup(fs->object); 157 VM_OBJECT_UNLOCK(fs->object); 158 if (fs->object != fs->first_object) { 159 VM_OBJECT_LOCK(fs->first_object); 160 vm_page_lock_queues(); 161 vm_page_free(fs->first_m); 162 vm_page_unlock_queues(); 163 vm_object_pip_wakeup(fs->first_object); 164 VM_OBJECT_UNLOCK(fs->first_object); 165 fs->first_m = NULL; 166 } 167 vm_object_deallocate(fs->first_object); 168 unlock_map(fs); 169 if (fs->vp != NULL) { 170 int vfslocked; 171 172 vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount); 173 vput(fs->vp); 174 fs->vp = NULL; 175 VFS_UNLOCK_GIANT(vfslocked); 176 } 177 } 178 179 /* 180 * TRYPAGER - used by vm_fault to calculate whether the pager for the 181 * current object *might* contain the page. 182 * 183 * default objects are zero-fill, there is no real pager. 184 */ 185 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 186 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 187 188 /* 189 * vm_fault: 190 * 191 * Handle a page fault occurring at the given address, 192 * requiring the given permissions, in the map specified. 193 * If successful, the page is inserted into the 194 * associated physical map. 195 * 196 * NOTE: the given address should be truncated to the 197 * proper page address. 198 * 199 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 200 * a standard error specifying why the fault is fatal is returned. 201 * 202 * 203 * The map in question must be referenced, and remains so. 204 * Caller may hold no locks. 205 */ 206 int 207 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 208 int fault_flags) 209 { 210 vm_prot_t prot; 211 int is_first_object_locked, result; 212 boolean_t growstack, wired; 213 int map_generation; 214 vm_object_t next_object; 215 vm_page_t marray[VM_FAULT_READ]; 216 int hardfault; 217 int faultcount; 218 struct faultstate fs; 219 220 hardfault = 0; 221 growstack = TRUE; 222 atomic_add_int(&cnt.v_vm_faults, 1); 223 224 RetryFault:; 225 226 /* 227 * Find the backing store object and offset into it to begin the 228 * search. 229 */ 230 fs.map = map; 231 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 232 &fs.first_object, &fs.first_pindex, &prot, &wired); 233 if (result != KERN_SUCCESS) { 234 if (result != KERN_PROTECTION_FAILURE || 235 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) { 236 if (growstack && result == KERN_INVALID_ADDRESS && 237 map != kernel_map && curproc != NULL) { 238 result = vm_map_growstack(curproc, vaddr); 239 if (result != KERN_SUCCESS) 240 return (KERN_FAILURE); 241 growstack = FALSE; 242 goto RetryFault; 243 } 244 return (result); 245 } 246 247 /* 248 * If we are user-wiring a r/w segment, and it is COW, then 249 * we need to do the COW operation. Note that we don't COW 250 * currently RO sections now, because it is NOT desirable 251 * to COW .text. We simply keep .text from ever being COW'ed 252 * and take the heat that one cannot debug wired .text sections. 253 */ 254 result = vm_map_lookup(&fs.map, vaddr, 255 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 256 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 257 if (result != KERN_SUCCESS) 258 return (result); 259 260 /* 261 * If we don't COW now, on a user wire, the user will never 262 * be able to write to the mapping. If we don't make this 263 * restriction, the bookkeeping would be nearly impossible. 264 * 265 * XXX The following assignment modifies the map without 266 * holding a write lock on it. 267 */ 268 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 269 fs.entry->max_protection &= ~VM_PROT_WRITE; 270 } 271 272 map_generation = fs.map->timestamp; 273 274 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 275 panic("vm_fault: fault on nofault entry, addr: %lx", 276 (u_long)vaddr); 277 } 278 279 /* 280 * Make a reference to this object to prevent its disposal while we 281 * are messing with it. Once we have the reference, the map is free 282 * to be diddled. Since objects reference their shadows (and copies), 283 * they will stay around as well. 284 * 285 * Bump the paging-in-progress count to prevent size changes (e.g. 286 * truncation operations) during I/O. This must be done after 287 * obtaining the vnode lock in order to avoid possible deadlocks. 288 * 289 * XXX vnode_pager_lock() can block without releasing the map lock. 290 */ 291 if (fs.first_object->flags & OBJ_NEEDGIANT) 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 KASSERT(fs.vp == NULL || !fs.map->system_map, 297 ("vm_fault: vnode-backed object mapped by system map")); 298 KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 || 299 !fs.map->system_map, 300 ("vm_fault: Object requiring giant mapped by system map")); 301 if (fs.first_object->flags & OBJ_NEEDGIANT) 302 mtx_unlock(&Giant); 303 vm_object_pip_add(fs.first_object, 1); 304 305 fs.lookup_still_valid = TRUE; 306 307 if (wired) 308 fault_type = prot; 309 310 fs.first_m = NULL; 311 312 /* 313 * Search for the page at object/offset. 314 */ 315 fs.object = fs.first_object; 316 fs.pindex = fs.first_pindex; 317 while (TRUE) { 318 /* 319 * If the object is dead, we stop here 320 */ 321 if (fs.object->flags & OBJ_DEAD) { 322 unlock_and_deallocate(&fs); 323 return (KERN_PROTECTION_FAILURE); 324 } 325 326 /* 327 * See if page is resident 328 */ 329 fs.m = vm_page_lookup(fs.object, fs.pindex); 330 if (fs.m != NULL) { 331 int queue; 332 333 /* 334 * check for page-based copy on write. 335 * We check fs.object == fs.first_object so 336 * as to ensure the legacy COW mechanism is 337 * used when the page in question is part of 338 * a shadow object. Otherwise, vm_page_cowfault() 339 * removes the page from the backing object, 340 * which is not what we want. 341 */ 342 vm_page_lock_queues(); 343 if ((fs.m->cow) && 344 (fault_type & VM_PROT_WRITE) && 345 (fs.object == fs.first_object)) { 346 vm_page_cowfault(fs.m); 347 vm_page_unlock_queues(); 348 unlock_and_deallocate(&fs); 349 goto RetryFault; 350 } 351 352 /* 353 * Wait/Retry if the page is busy. We have to do this 354 * if the page is busy via either PG_BUSY or 355 * vm_page_t->busy because the vm_pager may be using 356 * vm_page_t->busy for pageouts ( and even pageins if 357 * it is the vnode pager ), and we could end up trying 358 * to pagein and pageout the same page simultaneously. 359 * 360 * We can theoretically allow the busy case on a read 361 * fault if the page is marked valid, but since such 362 * pages are typically already pmap'd, putting that 363 * special case in might be more effort then it is 364 * worth. We cannot under any circumstances mess 365 * around with a vm_page_t->busy page except, perhaps, 366 * to pmap it. 367 */ 368 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 369 vm_page_unlock_queues(); 370 VM_OBJECT_UNLOCK(fs.object); 371 if (fs.object != fs.first_object) { 372 VM_OBJECT_LOCK(fs.first_object); 373 vm_page_lock_queues(); 374 vm_page_free(fs.first_m); 375 vm_page_unlock_queues(); 376 vm_object_pip_wakeup(fs.first_object); 377 VM_OBJECT_UNLOCK(fs.first_object); 378 fs.first_m = NULL; 379 } 380 unlock_map(&fs); 381 if (fs.vp != NULL) { 382 int vfslck; 383 384 vfslck = VFS_LOCK_GIANT(fs.vp->v_mount); 385 vput(fs.vp); 386 fs.vp = NULL; 387 VFS_UNLOCK_GIANT(vfslck); 388 } 389 VM_OBJECT_LOCK(fs.object); 390 if (fs.m == vm_page_lookup(fs.object, 391 fs.pindex)) { 392 vm_page_lock_queues(); 393 if (!vm_page_sleep_if_busy(fs.m, TRUE, 394 "vmpfw")) 395 vm_page_unlock_queues(); 396 } 397 vm_object_pip_wakeup(fs.object); 398 VM_OBJECT_UNLOCK(fs.object); 399 atomic_add_int(&cnt.v_intrans, 1); 400 vm_object_deallocate(fs.first_object); 401 goto RetryFault; 402 } 403 queue = fs.m->queue; 404 405 vm_pageq_remove_nowakeup(fs.m); 406 407 if (VM_PAGE_RESOLVEQUEUE(fs.m, queue) == PQ_CACHE && 408 vm_page_count_severe()) { 409 vm_page_activate(fs.m); 410 vm_page_unlock_queues(); 411 unlock_and_deallocate(&fs); 412 VM_WAITPFAULT; 413 goto RetryFault; 414 } 415 416 /* 417 * Mark page busy for other processes, and the 418 * pagedaemon. If it still isn't completely valid 419 * (readable), jump to readrest, else break-out ( we 420 * found the page ). 421 */ 422 vm_page_busy(fs.m); 423 vm_page_unlock_queues(); 424 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 425 fs.m->object != kernel_object && fs.m->object != kmem_object) { 426 goto readrest; 427 } 428 429 break; 430 } 431 432 /* 433 * Page is not resident, If this is the search termination 434 * or the pager might contain the page, allocate a new page. 435 */ 436 if (TRYPAGER || fs.object == fs.first_object) { 437 if (fs.pindex >= fs.object->size) { 438 unlock_and_deallocate(&fs); 439 return (KERN_PROTECTION_FAILURE); 440 } 441 442 /* 443 * Allocate a new page for this object/offset pair. 444 */ 445 fs.m = NULL; 446 if (!vm_page_count_severe()) { 447 fs.m = vm_page_alloc(fs.object, fs.pindex, 448 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); 449 } 450 if (fs.m == NULL) { 451 unlock_and_deallocate(&fs); 452 VM_WAITPFAULT; 453 goto RetryFault; 454 } 455 } 456 457 readrest: 458 /* 459 * We have found a valid page or we have allocated a new page. 460 * The page thus may not be valid or may not be entirely 461 * valid. 462 * 463 * Attempt to fault-in the page if there is a chance that the 464 * pager has it, and potentially fault in additional pages 465 * at the same time. 466 */ 467 if (TRYPAGER) { 468 int rv; 469 int reqpage; 470 int ahead, behind; 471 u_char behavior = vm_map_entry_behavior(fs.entry); 472 473 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 474 ahead = 0; 475 behind = 0; 476 } else { 477 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 478 if (behind > VM_FAULT_READ_BEHIND) 479 behind = VM_FAULT_READ_BEHIND; 480 481 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 482 if (ahead > VM_FAULT_READ_AHEAD) 483 ahead = VM_FAULT_READ_AHEAD; 484 } 485 is_first_object_locked = FALSE; 486 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 487 (behavior != MAP_ENTRY_BEHAV_RANDOM && 488 fs.pindex >= fs.entry->lastr && 489 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 490 (fs.first_object == fs.object || 491 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 492 fs.first_object->type != OBJT_DEVICE) { 493 vm_pindex_t firstpindex, tmppindex; 494 495 if (fs.first_pindex < 2 * VM_FAULT_READ) 496 firstpindex = 0; 497 else 498 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 499 500 vm_page_lock_queues(); 501 /* 502 * note: partially valid pages cannot be 503 * included in the lookahead - NFS piecemeal 504 * writes will barf on it badly. 505 */ 506 for (tmppindex = fs.first_pindex - 1; 507 tmppindex >= firstpindex; 508 --tmppindex) { 509 vm_page_t mt; 510 511 mt = vm_page_lookup(fs.first_object, tmppindex); 512 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 513 break; 514 if (mt->busy || 515 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 516 mt->hold_count || 517 mt->wire_count) 518 continue; 519 pmap_remove_all(mt); 520 if (mt->dirty) { 521 vm_page_deactivate(mt); 522 } else { 523 vm_page_cache(mt); 524 } 525 } 526 vm_page_unlock_queues(); 527 ahead += behind; 528 behind = 0; 529 } 530 if (is_first_object_locked) 531 VM_OBJECT_UNLOCK(fs.first_object); 532 /* 533 * now we find out if any other pages should be paged 534 * in at this time this routine checks to see if the 535 * pages surrounding this fault reside in the same 536 * object as the page for this fault. If they do, 537 * then they are faulted in also into the object. The 538 * array "marray" returned contains an array of 539 * vm_page_t structs where one of them is the 540 * vm_page_t passed to the routine. The reqpage 541 * return value is the index into the marray for the 542 * vm_page_t passed to the routine. 543 * 544 * fs.m plus the additional pages are PG_BUSY'd. 545 * 546 * XXX vm_fault_additional_pages() can block 547 * without releasing the map lock. 548 */ 549 faultcount = vm_fault_additional_pages( 550 fs.m, behind, ahead, marray, &reqpage); 551 552 /* 553 * update lastr imperfectly (we do not know how much 554 * getpages will actually read), but good enough. 555 * 556 * XXX The following assignment modifies the map 557 * without holding a write lock on it. 558 */ 559 fs.entry->lastr = fs.pindex + faultcount - behind; 560 561 /* 562 * Call the pager to retrieve the data, if any, after 563 * releasing the lock on the map. We hold a ref on 564 * fs.object and the pages are PG_BUSY'd. 565 */ 566 unlock_map(&fs); 567 568 rv = faultcount ? 569 vm_pager_get_pages(fs.object, marray, faultcount, 570 reqpage) : VM_PAGER_FAIL; 571 572 if (rv == VM_PAGER_OK) { 573 /* 574 * Found the page. Leave it busy while we play 575 * with it. 576 */ 577 578 /* 579 * Relookup in case pager changed page. Pager 580 * is responsible for disposition of old page 581 * if moved. 582 */ 583 fs.m = vm_page_lookup(fs.object, fs.pindex); 584 if (!fs.m) { 585 unlock_and_deallocate(&fs); 586 goto RetryFault; 587 } 588 589 hardfault++; 590 break; /* break to PAGE HAS BEEN FOUND */ 591 } 592 /* 593 * Remove the bogus page (which does not exist at this 594 * object/offset); before doing so, we must get back 595 * our object lock to preserve our invariant. 596 * 597 * Also wake up any other process that may want to bring 598 * in this page. 599 * 600 * If this is the top-level object, we must leave the 601 * busy page to prevent another process from rushing 602 * past us, and inserting the page in that object at 603 * the same time that we are. 604 */ 605 if (rv == VM_PAGER_ERROR) 606 printf("vm_fault: pager read error, pid %d (%s)\n", 607 curproc->p_pid, curproc->p_comm); 608 /* 609 * Data outside the range of the pager or an I/O error 610 */ 611 /* 612 * XXX - the check for kernel_map is a kludge to work 613 * around having the machine panic on a kernel space 614 * fault w/ I/O error. 615 */ 616 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 617 (rv == VM_PAGER_BAD)) { 618 vm_page_lock_queues(); 619 vm_page_free(fs.m); 620 vm_page_unlock_queues(); 621 fs.m = NULL; 622 unlock_and_deallocate(&fs); 623 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 624 } 625 if (fs.object != fs.first_object) { 626 vm_page_lock_queues(); 627 vm_page_free(fs.m); 628 vm_page_unlock_queues(); 629 fs.m = NULL; 630 /* 631 * XXX - we cannot just fall out at this 632 * point, m has been freed and is invalid! 633 */ 634 } 635 } 636 637 /* 638 * We get here if the object has default pager (or unwiring) 639 * or the pager doesn't have the page. 640 */ 641 if (fs.object == fs.first_object) 642 fs.first_m = fs.m; 643 644 /* 645 * Move on to the next object. Lock the next object before 646 * unlocking the current one. 647 */ 648 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 649 next_object = fs.object->backing_object; 650 if (next_object == NULL) { 651 /* 652 * If there's no object left, fill the page in the top 653 * object with zeros. 654 */ 655 if (fs.object != fs.first_object) { 656 vm_object_pip_wakeup(fs.object); 657 VM_OBJECT_UNLOCK(fs.object); 658 659 fs.object = fs.first_object; 660 fs.pindex = fs.first_pindex; 661 fs.m = fs.first_m; 662 VM_OBJECT_LOCK(fs.object); 663 } 664 fs.first_m = NULL; 665 666 /* 667 * Zero the page if necessary and mark it valid. 668 */ 669 if ((fs.m->flags & PG_ZERO) == 0) { 670 pmap_zero_page(fs.m); 671 } else { 672 atomic_add_int(&cnt.v_ozfod, 1); 673 } 674 atomic_add_int(&cnt.v_zfod, 1); 675 fs.m->valid = VM_PAGE_BITS_ALL; 676 break; /* break to PAGE HAS BEEN FOUND */ 677 } else { 678 KASSERT(fs.object != next_object, 679 ("object loop %p", next_object)); 680 VM_OBJECT_LOCK(next_object); 681 vm_object_pip_add(next_object, 1); 682 if (fs.object != fs.first_object) 683 vm_object_pip_wakeup(fs.object); 684 VM_OBJECT_UNLOCK(fs.object); 685 fs.object = next_object; 686 } 687 } 688 689 KASSERT((fs.m->flags & PG_BUSY) != 0, 690 ("vm_fault: not busy after main loop")); 691 692 /* 693 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 694 * is held.] 695 */ 696 697 /* 698 * If the page is being written, but isn't already owned by the 699 * top-level object, we have to copy it into a new page owned by the 700 * top-level object. 701 */ 702 if (fs.object != fs.first_object) { 703 /* 704 * We only really need to copy if we want to write it. 705 */ 706 if (fault_type & VM_PROT_WRITE) { 707 /* 708 * This allows pages to be virtually copied from a 709 * backing_object into the first_object, where the 710 * backing object has no other refs to it, and cannot 711 * gain any more refs. Instead of a bcopy, we just 712 * move the page from the backing object to the 713 * first object. Note that we must mark the page 714 * dirty in the first object so that it will go out 715 * to swap when needed. 716 */ 717 is_first_object_locked = FALSE; 718 if ( 719 /* 720 * Only one shadow object 721 */ 722 (fs.object->shadow_count == 1) && 723 /* 724 * No COW refs, except us 725 */ 726 (fs.object->ref_count == 1) && 727 /* 728 * No one else can look this object up 729 */ 730 (fs.object->handle == NULL) && 731 /* 732 * No other ways to look the object up 733 */ 734 ((fs.object->type == OBJT_DEFAULT) || 735 (fs.object->type == OBJT_SWAP)) && 736 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 737 /* 738 * We don't chase down the shadow chain 739 */ 740 fs.object == fs.first_object->backing_object) { 741 vm_page_lock_queues(); 742 /* 743 * get rid of the unnecessary page 744 */ 745 vm_page_free(fs.first_m); 746 /* 747 * grab the page and put it into the 748 * process'es object. The page is 749 * automatically made dirty. 750 */ 751 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 752 vm_page_busy(fs.m); 753 vm_page_unlock_queues(); 754 fs.first_m = fs.m; 755 fs.m = NULL; 756 atomic_add_int(&cnt.v_cow_optim, 1); 757 } else { 758 /* 759 * Oh, well, lets copy it. 760 */ 761 pmap_copy_page(fs.m, fs.first_m); 762 fs.first_m->valid = VM_PAGE_BITS_ALL; 763 } 764 if (fs.m) { 765 /* 766 * We no longer need the old page or object. 767 */ 768 release_page(&fs); 769 } 770 /* 771 * fs.object != fs.first_object due to above 772 * conditional 773 */ 774 vm_object_pip_wakeup(fs.object); 775 VM_OBJECT_UNLOCK(fs.object); 776 /* 777 * Only use the new page below... 778 */ 779 fs.object = fs.first_object; 780 fs.pindex = fs.first_pindex; 781 fs.m = fs.first_m; 782 if (!is_first_object_locked) 783 VM_OBJECT_LOCK(fs.object); 784 atomic_add_int(&cnt.v_cow_faults, 1); 785 } else { 786 prot &= ~VM_PROT_WRITE; 787 } 788 } 789 790 /* 791 * We must verify that the maps have not changed since our last 792 * lookup. 793 */ 794 if (!fs.lookup_still_valid) { 795 vm_object_t retry_object; 796 vm_pindex_t retry_pindex; 797 vm_prot_t retry_prot; 798 799 if (!vm_map_trylock_read(fs.map)) { 800 release_page(&fs); 801 unlock_and_deallocate(&fs); 802 goto RetryFault; 803 } 804 fs.lookup_still_valid = TRUE; 805 if (fs.map->timestamp != map_generation) { 806 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 807 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 808 809 /* 810 * If we don't need the page any longer, put it on the inactive 811 * list (the easiest thing to do here). If no one needs it, 812 * pageout will grab it eventually. 813 */ 814 if (result != KERN_SUCCESS) { 815 release_page(&fs); 816 unlock_and_deallocate(&fs); 817 818 /* 819 * If retry of map lookup would have blocked then 820 * retry fault from start. 821 */ 822 if (result == KERN_FAILURE) 823 goto RetryFault; 824 return (result); 825 } 826 if ((retry_object != fs.first_object) || 827 (retry_pindex != fs.first_pindex)) { 828 release_page(&fs); 829 unlock_and_deallocate(&fs); 830 goto RetryFault; 831 } 832 833 /* 834 * Check whether the protection has changed or the object has 835 * been copied while we left the map unlocked. Changing from 836 * read to write permission is OK - we leave the page 837 * write-protected, and catch the write fault. Changing from 838 * write to read permission means that we can't mark the page 839 * write-enabled after all. 840 */ 841 prot &= retry_prot; 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 vm_page_dirty(fs.m); 873 vm_pager_page_unswapped(fs.m); 874 } 875 } 876 877 /* 878 * Page had better still be busy 879 */ 880 KASSERT(fs.m->flags & PG_BUSY, 881 ("vm_fault: page %p not busy!", fs.m)); 882 /* 883 * Sanity check: page must be completely valid or it is not fit to 884 * map into user space. vm_pager_get_pages() ensures this. 885 */ 886 if (fs.m->valid != VM_PAGE_BITS_ALL) { 887 vm_page_zero_invalid(fs.m, TRUE); 888 printf("Warning: page %p partially invalid on fault\n", fs.m); 889 } 890 VM_OBJECT_UNLOCK(fs.object); 891 892 /* 893 * Put this page into the physical map. We had to do the unlock above 894 * because pmap_enter() may sleep. We don't put the page 895 * back on the active queue until later so that the pageout daemon 896 * won't find it (yet). 897 */ 898 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 899 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 900 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 901 } 902 VM_OBJECT_LOCK(fs.object); 903 vm_page_lock_queues(); 904 vm_page_flag_set(fs.m, PG_REFERENCED); 905 906 /* 907 * If the page is not wired down, then put it where the pageout daemon 908 * can find it. 909 */ 910 if (fault_flags & VM_FAULT_WIRE_MASK) { 911 if (wired) 912 vm_page_wire(fs.m); 913 else 914 vm_page_unwire(fs.m, 1); 915 } else { 916 vm_page_activate(fs.m); 917 } 918 vm_page_wakeup(fs.m); 919 vm_page_unlock_queues(); 920 921 /* 922 * Unlock everything, and return 923 */ 924 unlock_and_deallocate(&fs); 925 PROC_LOCK(curproc); 926 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) { 927 if (hardfault) { 928 curproc->p_stats->p_ru.ru_majflt++; 929 } else { 930 curproc->p_stats->p_ru.ru_minflt++; 931 } 932 } 933 PROC_UNLOCK(curproc); 934 935 return (KERN_SUCCESS); 936 } 937 938 /* 939 * vm_fault_prefault provides a quick way of clustering 940 * pagefaults into a processes address space. It is a "cousin" 941 * of vm_map_pmap_enter, except it runs at page fault time instead 942 * of mmap time. 943 */ 944 static void 945 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 946 { 947 int i; 948 vm_offset_t addr, starta; 949 vm_pindex_t pindex; 950 vm_page_t m; 951 vm_object_t object; 952 953 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 954 return; 955 956 object = entry->object.vm_object; 957 958 starta = addra - PFBAK * PAGE_SIZE; 959 if (starta < entry->start) { 960 starta = entry->start; 961 } else if (starta > addra) { 962 starta = 0; 963 } 964 965 for (i = 0; i < PAGEORDER_SIZE; i++) { 966 vm_object_t backing_object, lobject; 967 968 addr = addra + prefault_pageorder[i]; 969 if (addr > addra + (PFFOR * PAGE_SIZE)) 970 addr = 0; 971 972 if (addr < starta || addr >= entry->end) 973 continue; 974 975 if (!pmap_is_prefaultable(pmap, addr)) 976 continue; 977 978 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 979 lobject = object; 980 VM_OBJECT_LOCK(lobject); 981 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 982 lobject->type == OBJT_DEFAULT && 983 (backing_object = lobject->backing_object) != NULL) { 984 if (lobject->backing_object_offset & PAGE_MASK) 985 break; 986 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 987 VM_OBJECT_LOCK(backing_object); 988 VM_OBJECT_UNLOCK(lobject); 989 lobject = backing_object; 990 } 991 /* 992 * give-up when a page is not in memory 993 */ 994 if (m == NULL) { 995 VM_OBJECT_UNLOCK(lobject); 996 break; 997 } 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 vm_page_lock_queues(); 1003 if (VM_PAGE_INQUEUE1(m, PQ_CACHE)) 1004 vm_page_deactivate(m); 1005 pmap_enter_quick(pmap, addr, m, entry->protection); 1006 vm_page_unlock_queues(); 1007 } 1008 VM_OBJECT_UNLOCK(lobject); 1009 } 1010 } 1011 1012 /* 1013 * vm_fault_quick: 1014 * 1015 * Ensure that the requested virtual address, which may be in userland, 1016 * is valid. Fault-in the page if necessary. Return -1 on failure. 1017 */ 1018 int 1019 vm_fault_quick(caddr_t v, int prot) 1020 { 1021 int r; 1022 1023 if (prot & VM_PROT_WRITE) 1024 r = subyte(v, fubyte(v)); 1025 else 1026 r = fubyte(v); 1027 return(r); 1028 } 1029 1030 /* 1031 * vm_fault_wire: 1032 * 1033 * Wire down a range of virtual addresses in a map. 1034 */ 1035 int 1036 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1037 boolean_t user_wire, boolean_t fictitious) 1038 { 1039 vm_offset_t va; 1040 int rv; 1041 1042 /* 1043 * We simulate a fault to get the page and enter it in the physical 1044 * map. For user wiring, we only ask for read access on currently 1045 * read-only sections. 1046 */ 1047 for (va = start; va < end; va += PAGE_SIZE) { 1048 rv = vm_fault(map, va, 1049 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE, 1050 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING); 1051 if (rv) { 1052 if (va != start) 1053 vm_fault_unwire(map, start, va, fictitious); 1054 return (rv); 1055 } 1056 } 1057 return (KERN_SUCCESS); 1058 } 1059 1060 /* 1061 * vm_fault_unwire: 1062 * 1063 * Unwire a range of virtual addresses in a map. 1064 */ 1065 void 1066 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1067 boolean_t fictitious) 1068 { 1069 vm_paddr_t pa; 1070 vm_offset_t va; 1071 pmap_t pmap; 1072 1073 pmap = vm_map_pmap(map); 1074 1075 /* 1076 * Since the pages are wired down, we must be able to get their 1077 * mappings from the physical map system. 1078 */ 1079 for (va = start; va < end; va += PAGE_SIZE) { 1080 pa = pmap_extract(pmap, va); 1081 if (pa != 0) { 1082 pmap_change_wiring(pmap, va, FALSE); 1083 if (!fictitious) { 1084 vm_page_lock_queues(); 1085 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1086 vm_page_unlock_queues(); 1087 } 1088 } 1089 } 1090 } 1091 1092 /* 1093 * Routine: 1094 * vm_fault_copy_entry 1095 * Function: 1096 * Copy all of the pages from a wired-down map entry to another. 1097 * 1098 * In/out conditions: 1099 * The source and destination maps must be locked for write. 1100 * The source map entry must be wired down (or be a sharing map 1101 * entry corresponding to a main map entry that is wired down). 1102 */ 1103 void 1104 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) 1105 vm_map_t dst_map; 1106 vm_map_t src_map; 1107 vm_map_entry_t dst_entry; 1108 vm_map_entry_t src_entry; 1109 { 1110 vm_object_t backing_object, dst_object, object; 1111 vm_object_t src_object; 1112 vm_ooffset_t dst_offset; 1113 vm_ooffset_t src_offset; 1114 vm_pindex_t pindex; 1115 vm_prot_t prot; 1116 vm_offset_t vaddr; 1117 vm_page_t dst_m; 1118 vm_page_t src_m; 1119 1120 #ifdef lint 1121 src_map++; 1122 #endif /* lint */ 1123 1124 src_object = src_entry->object.vm_object; 1125 src_offset = src_entry->offset; 1126 1127 /* 1128 * Create the top-level object for the destination entry. (Doesn't 1129 * actually shadow anything - we copy the pages directly.) 1130 */ 1131 dst_object = vm_object_allocate(OBJT_DEFAULT, 1132 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1133 1134 VM_OBJECT_LOCK(dst_object); 1135 dst_entry->object.vm_object = dst_object; 1136 dst_entry->offset = 0; 1137 1138 prot = dst_entry->max_protection; 1139 1140 /* 1141 * Loop through all of the pages in the entry's range, copying each 1142 * one from the source object (it should be there) to the destination 1143 * object. 1144 */ 1145 for (vaddr = dst_entry->start, dst_offset = 0; 1146 vaddr < dst_entry->end; 1147 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1148 1149 /* 1150 * Allocate a page in the destination object 1151 */ 1152 do { 1153 dst_m = vm_page_alloc(dst_object, 1154 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1155 if (dst_m == NULL) { 1156 VM_OBJECT_UNLOCK(dst_object); 1157 VM_WAIT; 1158 VM_OBJECT_LOCK(dst_object); 1159 } 1160 } while (dst_m == NULL); 1161 1162 /* 1163 * Find the page in the source object, and copy it in. 1164 * (Because the source is wired down, the page will be in 1165 * memory.) 1166 */ 1167 VM_OBJECT_LOCK(src_object); 1168 object = src_object; 1169 pindex = 0; 1170 while ((src_m = vm_page_lookup(object, pindex + 1171 OFF_TO_IDX(dst_offset + src_offset))) == NULL && 1172 (src_entry->protection & VM_PROT_WRITE) == 0 && 1173 (backing_object = object->backing_object) != NULL) { 1174 /* 1175 * Allow fallback to backing objects if we are reading. 1176 */ 1177 VM_OBJECT_LOCK(backing_object); 1178 pindex += OFF_TO_IDX(object->backing_object_offset); 1179 VM_OBJECT_UNLOCK(object); 1180 object = backing_object; 1181 } 1182 if (src_m == NULL) 1183 panic("vm_fault_copy_wired: page missing"); 1184 pmap_copy_page(src_m, dst_m); 1185 VM_OBJECT_UNLOCK(object); 1186 dst_m->valid = VM_PAGE_BITS_ALL; 1187 VM_OBJECT_UNLOCK(dst_object); 1188 1189 /* 1190 * Enter it in the pmap... 1191 */ 1192 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1193 VM_OBJECT_LOCK(dst_object); 1194 vm_page_lock_queues(); 1195 if ((prot & VM_PROT_WRITE) != 0) 1196 vm_page_flag_set(dst_m, PG_WRITEABLE); 1197 1198 /* 1199 * Mark it no longer busy, and put it on the active list. 1200 */ 1201 vm_page_activate(dst_m); 1202 vm_page_wakeup(dst_m); 1203 vm_page_unlock_queues(); 1204 } 1205 VM_OBJECT_UNLOCK(dst_object); 1206 } 1207 1208 1209 /* 1210 * This routine checks around the requested page for other pages that 1211 * might be able to be faulted in. This routine brackets the viable 1212 * pages for the pages to be paged in. 1213 * 1214 * Inputs: 1215 * m, rbehind, rahead 1216 * 1217 * Outputs: 1218 * marray (array of vm_page_t), reqpage (index of requested page) 1219 * 1220 * Return value: 1221 * number of pages in marray 1222 * 1223 * This routine can't block. 1224 */ 1225 static int 1226 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1227 vm_page_t m; 1228 int rbehind; 1229 int rahead; 1230 vm_page_t *marray; 1231 int *reqpage; 1232 { 1233 int i,j; 1234 vm_object_t object; 1235 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1236 vm_page_t rtm; 1237 int cbehind, cahead; 1238 1239 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1240 1241 object = m->object; 1242 pindex = m->pindex; 1243 1244 /* 1245 * we don't fault-ahead for device pager 1246 */ 1247 if (object->type == OBJT_DEVICE) { 1248 *reqpage = 0; 1249 marray[0] = m; 1250 return 1; 1251 } 1252 1253 /* 1254 * if the requested page is not available, then give up now 1255 */ 1256 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1257 return 0; 1258 } 1259 1260 if ((cbehind == 0) && (cahead == 0)) { 1261 *reqpage = 0; 1262 marray[0] = m; 1263 return 1; 1264 } 1265 1266 if (rahead > cahead) { 1267 rahead = cahead; 1268 } 1269 1270 if (rbehind > cbehind) { 1271 rbehind = cbehind; 1272 } 1273 1274 /* 1275 * try to do any readahead that we might have free pages for. 1276 */ 1277 if ((rahead + rbehind) > 1278 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) { 1279 pagedaemon_wakeup(); 1280 marray[0] = m; 1281 *reqpage = 0; 1282 return 1; 1283 } 1284 1285 /* 1286 * scan backward for the read behind pages -- in memory 1287 */ 1288 if (pindex > 0) { 1289 if (rbehind > pindex) { 1290 rbehind = pindex; 1291 startpindex = 0; 1292 } else { 1293 startpindex = pindex - rbehind; 1294 } 1295 1296 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1297 rtm->pindex >= startpindex) 1298 startpindex = rtm->pindex + 1; 1299 1300 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1301 1302 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1303 if (rtm == NULL) { 1304 vm_page_lock_queues(); 1305 for (j = 0; j < i; j++) { 1306 vm_page_free(marray[j]); 1307 } 1308 vm_page_unlock_queues(); 1309 marray[0] = m; 1310 *reqpage = 0; 1311 return 1; 1312 } 1313 1314 marray[i] = rtm; 1315 } 1316 } else { 1317 startpindex = 0; 1318 i = 0; 1319 } 1320 1321 marray[i] = m; 1322 /* page offset of the required page */ 1323 *reqpage = i; 1324 1325 tpindex = pindex + 1; 1326 i++; 1327 1328 /* 1329 * scan forward for the read ahead pages 1330 */ 1331 endpindex = tpindex + rahead; 1332 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1333 endpindex = rtm->pindex; 1334 if (endpindex > object->size) 1335 endpindex = object->size; 1336 1337 for (; tpindex < endpindex; i++, tpindex++) { 1338 1339 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1340 if (rtm == NULL) { 1341 break; 1342 } 1343 1344 marray[i] = rtm; 1345 } 1346 1347 /* return number of bytes of pages */ 1348 return i; 1349 } 1350