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