1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 42 * 43 * 44 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 45 * All rights reserved. 46 * 47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 70 /* 71 * Page fault handling module. 72 */ 73 74 #include <sys/cdefs.h> 75 __FBSDID("$FreeBSD$"); 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/lock.h> 81 #include <sys/mutex.h> 82 #include <sys/proc.h> 83 #include <sys/resourcevar.h> 84 #include <sys/sysctl.h> 85 #include <sys/vmmeter.h> 86 #include <sys/vnode.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vnode_pager.h> 98 #include <vm/vm_extern.h> 99 100 #define PFBAK 4 101 #define PFFOR 4 102 #define PAGEORDER_SIZE (PFBAK+PFFOR) 103 104 static int prefault_pageorder[] = { 105 -1 * PAGE_SIZE, 1 * PAGE_SIZE, 106 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 107 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 108 -4 * PAGE_SIZE, 4 * PAGE_SIZE 109 }; 110 111 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 112 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t); 113 114 #define VM_FAULT_READ_AHEAD 8 115 #define VM_FAULT_READ_BEHIND 7 116 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 117 118 struct faultstate { 119 vm_page_t m; 120 vm_object_t object; 121 vm_pindex_t pindex; 122 vm_page_t first_m; 123 vm_object_t first_object; 124 vm_pindex_t first_pindex; 125 vm_map_t map; 126 vm_map_entry_t entry; 127 int lookup_still_valid; 128 struct vnode *vp; 129 }; 130 131 static __inline void 132 release_page(struct faultstate *fs) 133 { 134 vm_page_lock_queues(); 135 vm_page_wakeup(fs->m); 136 vm_page_deactivate(fs->m); 137 vm_page_unlock_queues(); 138 fs->m = NULL; 139 } 140 141 static __inline void 142 unlock_map(struct faultstate *fs) 143 { 144 if (fs->lookup_still_valid) { 145 vm_map_lookup_done(fs->map, fs->entry); 146 fs->lookup_still_valid = FALSE; 147 } 148 } 149 150 static void 151 _unlock_things(struct faultstate *fs, int dealloc) 152 { 153 154 vm_object_pip_wakeup(fs->object); 155 VM_OBJECT_UNLOCK(fs->object); 156 if (fs->object != fs->first_object) { 157 VM_OBJECT_LOCK(fs->first_object); 158 vm_page_lock_queues(); 159 vm_page_free(fs->first_m); 160 vm_page_unlock_queues(); 161 vm_object_pip_wakeup(fs->first_object); 162 VM_OBJECT_UNLOCK(fs->first_object); 163 fs->first_m = NULL; 164 } 165 if (dealloc) { 166 vm_object_deallocate(fs->first_object); 167 } 168 unlock_map(fs); 169 if (fs->vp != NULL) { 170 vput(fs->vp); 171 if (debug_mpsafevm) 172 mtx_unlock(&Giant); 173 fs->vp = NULL; 174 } 175 if (dealloc) 176 VM_UNLOCK_GIANT(); 177 } 178 179 #define unlock_things(fs) _unlock_things(fs, 0) 180 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 181 182 /* 183 * TRYPAGER - used by vm_fault to calculate whether the pager for the 184 * current object *might* contain the page. 185 * 186 * default objects are zero-fill, there is no real pager. 187 */ 188 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 189 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 190 191 /* 192 * vm_fault: 193 * 194 * Handle a page fault occurring at the given address, 195 * requiring the given permissions, in the map specified. 196 * If successful, the page is inserted into the 197 * associated physical map. 198 * 199 * NOTE: the given address should be truncated to the 200 * proper page address. 201 * 202 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 203 * a standard error specifying why the fault is fatal is returned. 204 * 205 * 206 * The map in question must be referenced, and remains so. 207 * Caller may hold no locks. 208 */ 209 int 210 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 211 int fault_flags) 212 { 213 vm_prot_t prot; 214 int is_first_object_locked, result; 215 boolean_t growstack, wired; 216 int map_generation; 217 vm_object_t next_object; 218 vm_page_t marray[VM_FAULT_READ]; 219 int hardfault; 220 int faultcount; 221 struct faultstate fs; 222 223 hardfault = 0; 224 growstack = TRUE; 225 atomic_add_int(&cnt.v_vm_faults, 1); 226 227 RetryFault:; 228 229 /* 230 * Find the backing store object and offset into it to begin the 231 * search. 232 */ 233 fs.map = map; 234 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 235 &fs.first_object, &fs.first_pindex, &prot, &wired); 236 if (result != KERN_SUCCESS) { 237 if (result != KERN_PROTECTION_FAILURE || 238 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) { 239 if (growstack && result == KERN_INVALID_ADDRESS && 240 map != kernel_map && curproc != NULL) { 241 result = vm_map_growstack(curproc, vaddr); 242 if (result != KERN_SUCCESS) 243 return (KERN_FAILURE); 244 growstack = FALSE; 245 goto RetryFault; 246 } 247 return (result); 248 } 249 250 /* 251 * If we are user-wiring a r/w segment, and it is COW, then 252 * we need to do the COW operation. Note that we don't COW 253 * currently RO sections now, because it is NOT desirable 254 * to COW .text. We simply keep .text from ever being COW'ed 255 * and take the heat that one cannot debug wired .text sections. 256 */ 257 result = vm_map_lookup(&fs.map, vaddr, 258 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 259 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 260 if (result != KERN_SUCCESS) 261 return (result); 262 263 /* 264 * If we don't COW now, on a user wire, the user will never 265 * be able to write to the mapping. If we don't make this 266 * restriction, the bookkeeping would be nearly impossible. 267 * 268 * XXX The following assignment modifies the map without 269 * holding a write lock on it. 270 */ 271 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 272 fs.entry->max_protection &= ~VM_PROT_WRITE; 273 } 274 275 map_generation = fs.map->timestamp; 276 277 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 278 panic("vm_fault: fault on nofault entry, addr: %lx", 279 (u_long)vaddr); 280 } 281 282 /* 283 * Make a reference to this object to prevent its disposal while we 284 * are messing with it. Once we have the reference, the map is free 285 * to be diddled. Since objects reference their shadows (and copies), 286 * they will stay around as well. 287 * 288 * Bump the paging-in-progress count to prevent size changes (e.g. 289 * truncation operations) during I/O. This must be done after 290 * obtaining the vnode lock in order to avoid possible deadlocks. 291 * 292 * XXX vnode_pager_lock() can block without releasing the map lock. 293 */ 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 if (fs.vp == NULL && debug_mpsafevm) 299 mtx_unlock(&Giant); 300 vm_object_pip_add(fs.first_object, 1); 301 302 fs.lookup_still_valid = TRUE; 303 304 if (wired) 305 fault_type = prot; 306 307 fs.first_m = NULL; 308 309 /* 310 * Search for the page at object/offset. 311 */ 312 fs.object = fs.first_object; 313 fs.pindex = fs.first_pindex; 314 while (TRUE) { 315 /* 316 * If the object is dead, we stop here 317 */ 318 if (fs.object->flags & OBJ_DEAD) { 319 unlock_and_deallocate(&fs); 320 return (KERN_PROTECTION_FAILURE); 321 } 322 323 /* 324 * See if page is resident 325 */ 326 fs.m = vm_page_lookup(fs.object, fs.pindex); 327 if (fs.m != NULL) { 328 int queue; 329 330 /* 331 * check for page-based copy on write. 332 * We check fs.object == fs.first_object so 333 * as to ensure the legacy COW mechanism is 334 * used when the page in question is part of 335 * a shadow object. Otherwise, vm_page_cowfault() 336 * removes the page from the backing object, 337 * which is not what we want. 338 */ 339 vm_page_lock_queues(); 340 if ((fs.m->cow) && 341 (fault_type & VM_PROT_WRITE) && 342 (fs.object == fs.first_object)) { 343 vm_page_cowfault(fs.m); 344 vm_page_unlock_queues(); 345 unlock_and_deallocate(&fs); 346 goto RetryFault; 347 } 348 349 /* 350 * Wait/Retry if the page is busy. We have to do this 351 * if the page is busy via either PG_BUSY or 352 * vm_page_t->busy because the vm_pager may be using 353 * vm_page_t->busy for pageouts ( and even pageins if 354 * it is the vnode pager ), and we could end up trying 355 * to pagein and pageout the same page simultaneously. 356 * 357 * We can theoretically allow the busy case on a read 358 * fault if the page is marked valid, but since such 359 * pages are typically already pmap'd, putting that 360 * special case in might be more effort then it is 361 * worth. We cannot under any circumstances mess 362 * around with a vm_page_t->busy page except, perhaps, 363 * to pmap it. 364 */ 365 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 366 vm_page_unlock_queues(); 367 unlock_things(&fs); 368 vm_page_lock_queues(); 369 if (!vm_page_sleep_if_busy(fs.m, TRUE, "vmpfw")) 370 vm_page_unlock_queues(); 371 atomic_add_int(&cnt.v_intrans, 1); 372 VM_UNLOCK_GIANT(); 373 vm_object_deallocate(fs.first_object); 374 goto RetryFault; 375 } 376 queue = fs.m->queue; 377 378 vm_pageq_remove_nowakeup(fs.m); 379 380 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) { 381 vm_page_activate(fs.m); 382 vm_page_unlock_queues(); 383 unlock_and_deallocate(&fs); 384 VM_WAITPFAULT; 385 goto RetryFault; 386 } 387 388 /* 389 * Mark page busy for other processes, and the 390 * pagedaemon. If it still isn't completely valid 391 * (readable), jump to readrest, else break-out ( we 392 * found the page ). 393 */ 394 vm_page_busy(fs.m); 395 vm_page_unlock_queues(); 396 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 397 fs.m->object != kernel_object && fs.m->object != kmem_object) { 398 goto readrest; 399 } 400 401 break; 402 } 403 404 /* 405 * Page is not resident, If this is the search termination 406 * or the pager might contain the page, allocate a new page. 407 */ 408 if (TRYPAGER || fs.object == fs.first_object) { 409 if (fs.pindex >= fs.object->size) { 410 unlock_and_deallocate(&fs); 411 return (KERN_PROTECTION_FAILURE); 412 } 413 414 /* 415 * Allocate a new page for this object/offset pair. 416 */ 417 fs.m = NULL; 418 if (!vm_page_count_severe()) { 419 fs.m = vm_page_alloc(fs.object, fs.pindex, 420 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); 421 } 422 if (fs.m == NULL) { 423 unlock_and_deallocate(&fs); 424 VM_WAITPFAULT; 425 goto RetryFault; 426 } 427 } 428 429 readrest: 430 /* 431 * We have found a valid page or we have allocated a new page. 432 * The page thus may not be valid or may not be entirely 433 * valid. 434 * 435 * Attempt to fault-in the page if there is a chance that the 436 * pager has it, and potentially fault in additional pages 437 * at the same time. 438 */ 439 if (TRYPAGER) { 440 int rv; 441 int reqpage; 442 int ahead, behind; 443 u_char behavior = vm_map_entry_behavior(fs.entry); 444 445 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 446 ahead = 0; 447 behind = 0; 448 } else { 449 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 450 if (behind > VM_FAULT_READ_BEHIND) 451 behind = VM_FAULT_READ_BEHIND; 452 453 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 454 if (ahead > VM_FAULT_READ_AHEAD) 455 ahead = VM_FAULT_READ_AHEAD; 456 } 457 is_first_object_locked = FALSE; 458 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 459 (behavior != MAP_ENTRY_BEHAV_RANDOM && 460 fs.pindex >= fs.entry->lastr && 461 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 462 (fs.first_object == fs.object || 463 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 464 fs.first_object->type != OBJT_DEVICE) { 465 vm_pindex_t firstpindex, tmppindex; 466 467 if (fs.first_pindex < 2 * VM_FAULT_READ) 468 firstpindex = 0; 469 else 470 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 471 472 vm_page_lock_queues(); 473 /* 474 * note: partially valid pages cannot be 475 * included in the lookahead - NFS piecemeal 476 * writes will barf on it badly. 477 */ 478 for (tmppindex = fs.first_pindex - 1; 479 tmppindex >= firstpindex; 480 --tmppindex) { 481 vm_page_t mt; 482 483 mt = vm_page_lookup(fs.first_object, tmppindex); 484 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 485 break; 486 if (mt->busy || 487 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 488 mt->hold_count || 489 mt->wire_count) 490 continue; 491 pmap_remove_all(mt); 492 if (mt->dirty) { 493 vm_page_deactivate(mt); 494 } else { 495 vm_page_cache(mt); 496 } 497 } 498 vm_page_unlock_queues(); 499 ahead += behind; 500 behind = 0; 501 } 502 if (is_first_object_locked) 503 VM_OBJECT_UNLOCK(fs.first_object); 504 /* 505 * now we find out if any other pages should be paged 506 * in at this time this routine checks to see if the 507 * pages surrounding this fault reside in the same 508 * object as the page for this fault. If they do, 509 * then they are faulted in also into the object. The 510 * array "marray" returned contains an array of 511 * vm_page_t structs where one of them is the 512 * vm_page_t passed to the routine. The reqpage 513 * return value is the index into the marray for the 514 * vm_page_t passed to the routine. 515 * 516 * fs.m plus the additional pages are PG_BUSY'd. 517 * 518 * XXX vm_fault_additional_pages() can block 519 * without releasing the map lock. 520 */ 521 faultcount = vm_fault_additional_pages( 522 fs.m, behind, ahead, marray, &reqpage); 523 524 /* 525 * update lastr imperfectly (we do not know how much 526 * getpages will actually read), but good enough. 527 * 528 * XXX The following assignment modifies the map 529 * without holding a write lock on it. 530 */ 531 fs.entry->lastr = fs.pindex + faultcount - behind; 532 533 /* 534 * Call the pager to retrieve the data, if any, after 535 * releasing the lock on the map. We hold a ref on 536 * fs.object and the pages are PG_BUSY'd. 537 */ 538 unlock_map(&fs); 539 540 rv = faultcount ? 541 vm_pager_get_pages(fs.object, marray, faultcount, 542 reqpage) : VM_PAGER_FAIL; 543 544 if (rv == VM_PAGER_OK) { 545 /* 546 * Found the page. Leave it busy while we play 547 * with it. 548 */ 549 550 /* 551 * Relookup in case pager changed page. Pager 552 * is responsible for disposition of old page 553 * if moved. 554 */ 555 fs.m = vm_page_lookup(fs.object, fs.pindex); 556 if (!fs.m) { 557 unlock_and_deallocate(&fs); 558 goto RetryFault; 559 } 560 561 hardfault++; 562 break; /* break to PAGE HAS BEEN FOUND */ 563 } 564 /* 565 * Remove the bogus page (which does not exist at this 566 * object/offset); before doing so, we must get back 567 * our object lock to preserve our invariant. 568 * 569 * Also wake up any other process that may want to bring 570 * in this page. 571 * 572 * If this is the top-level object, we must leave the 573 * busy page to prevent another process from rushing 574 * past us, and inserting the page in that object at 575 * the same time that we are. 576 */ 577 if (rv == VM_PAGER_ERROR) 578 printf("vm_fault: pager read error, pid %d (%s)\n", 579 curproc->p_pid, curproc->p_comm); 580 /* 581 * Data outside the range of the pager or an I/O error 582 */ 583 /* 584 * XXX - the check for kernel_map is a kludge to work 585 * around having the machine panic on a kernel space 586 * fault w/ I/O error. 587 */ 588 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 589 (rv == VM_PAGER_BAD)) { 590 vm_page_lock_queues(); 591 vm_page_free(fs.m); 592 vm_page_unlock_queues(); 593 fs.m = NULL; 594 unlock_and_deallocate(&fs); 595 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 596 } 597 if (fs.object != fs.first_object) { 598 vm_page_lock_queues(); 599 vm_page_free(fs.m); 600 vm_page_unlock_queues(); 601 fs.m = NULL; 602 /* 603 * XXX - we cannot just fall out at this 604 * point, m has been freed and is invalid! 605 */ 606 } 607 } 608 609 /* 610 * We get here if the object has default pager (or unwiring) 611 * or the pager doesn't have the page. 612 */ 613 if (fs.object == fs.first_object) 614 fs.first_m = fs.m; 615 616 /* 617 * Move on to the next object. Lock the next object before 618 * unlocking the current one. 619 */ 620 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 621 next_object = fs.object->backing_object; 622 if (next_object == NULL) { 623 /* 624 * If there's no object left, fill the page in the top 625 * object with zeros. 626 */ 627 if (fs.object != fs.first_object) { 628 vm_object_pip_wakeup(fs.object); 629 VM_OBJECT_UNLOCK(fs.object); 630 631 fs.object = fs.first_object; 632 fs.pindex = fs.first_pindex; 633 fs.m = fs.first_m; 634 VM_OBJECT_LOCK(fs.object); 635 } 636 fs.first_m = NULL; 637 638 /* 639 * Zero the page if necessary and mark it valid. 640 */ 641 if ((fs.m->flags & PG_ZERO) == 0) { 642 pmap_zero_page(fs.m); 643 } else { 644 atomic_add_int(&cnt.v_ozfod, 1); 645 } 646 atomic_add_int(&cnt.v_zfod, 1); 647 fs.m->valid = VM_PAGE_BITS_ALL; 648 break; /* break to PAGE HAS BEEN FOUND */ 649 } else { 650 KASSERT(fs.object != next_object, 651 ("object loop %p", next_object)); 652 VM_OBJECT_LOCK(next_object); 653 vm_object_pip_add(next_object, 1); 654 if (fs.object != fs.first_object) 655 vm_object_pip_wakeup(fs.object); 656 VM_OBJECT_UNLOCK(fs.object); 657 fs.object = next_object; 658 } 659 } 660 661 KASSERT((fs.m->flags & PG_BUSY) != 0, 662 ("vm_fault: not busy after main loop")); 663 664 /* 665 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 666 * is held.] 667 */ 668 669 /* 670 * If the page is being written, but isn't already owned by the 671 * top-level object, we have to copy it into a new page owned by the 672 * top-level object. 673 */ 674 if (fs.object != fs.first_object) { 675 /* 676 * We only really need to copy if we want to write it. 677 */ 678 if (fault_type & VM_PROT_WRITE) { 679 /* 680 * This allows pages to be virtually copied from a 681 * backing_object into the first_object, where the 682 * backing object has no other refs to it, and cannot 683 * gain any more refs. Instead of a bcopy, we just 684 * move the page from the backing object to the 685 * first object. Note that we must mark the page 686 * dirty in the first object so that it will go out 687 * to swap when needed. 688 */ 689 is_first_object_locked = FALSE; 690 if ( 691 /* 692 * Only one shadow object 693 */ 694 (fs.object->shadow_count == 1) && 695 /* 696 * No COW refs, except us 697 */ 698 (fs.object->ref_count == 1) && 699 /* 700 * No one else can look this object up 701 */ 702 (fs.object->handle == NULL) && 703 /* 704 * No other ways to look the object up 705 */ 706 ((fs.object->type == OBJT_DEFAULT) || 707 (fs.object->type == OBJT_SWAP)) && 708 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 709 /* 710 * We don't chase down the shadow chain 711 */ 712 fs.object == fs.first_object->backing_object) { 713 vm_page_lock_queues(); 714 /* 715 * get rid of the unnecessary page 716 */ 717 pmap_remove_all(fs.first_m); 718 vm_page_free(fs.first_m); 719 /* 720 * grab the page and put it into the 721 * process'es object. The page is 722 * automatically made dirty. 723 */ 724 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 725 vm_page_busy(fs.m); 726 vm_page_unlock_queues(); 727 fs.first_m = fs.m; 728 fs.m = NULL; 729 atomic_add_int(&cnt.v_cow_optim, 1); 730 } else { 731 /* 732 * Oh, well, lets copy it. 733 */ 734 pmap_copy_page(fs.m, fs.first_m); 735 fs.first_m->valid = VM_PAGE_BITS_ALL; 736 } 737 if (fs.m) { 738 /* 739 * We no longer need the old page or object. 740 */ 741 release_page(&fs); 742 } 743 /* 744 * fs.object != fs.first_object due to above 745 * conditional 746 */ 747 vm_object_pip_wakeup(fs.object); 748 VM_OBJECT_UNLOCK(fs.object); 749 /* 750 * Only use the new page below... 751 */ 752 fs.object = fs.first_object; 753 fs.pindex = fs.first_pindex; 754 fs.m = fs.first_m; 755 if (!is_first_object_locked) 756 VM_OBJECT_LOCK(fs.object); 757 atomic_add_int(&cnt.v_cow_faults, 1); 758 } else { 759 prot &= ~VM_PROT_WRITE; 760 } 761 } 762 763 /* 764 * We must verify that the maps have not changed since our last 765 * lookup. 766 */ 767 if (!fs.lookup_still_valid) { 768 vm_object_t retry_object; 769 vm_pindex_t retry_pindex; 770 vm_prot_t retry_prot; 771 772 if (!vm_map_trylock_read(fs.map)) { 773 release_page(&fs); 774 unlock_and_deallocate(&fs); 775 goto RetryFault; 776 } 777 fs.lookup_still_valid = TRUE; 778 if (fs.map->timestamp != map_generation) { 779 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 780 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 781 782 /* 783 * If we don't need the page any longer, put it on the active 784 * list (the easiest thing to do here). If no one needs it, 785 * pageout will grab it eventually. 786 */ 787 if (result != KERN_SUCCESS) { 788 release_page(&fs); 789 unlock_and_deallocate(&fs); 790 791 /* 792 * If retry of map lookup would have blocked then 793 * retry fault from start. 794 */ 795 if (result == KERN_FAILURE) 796 goto RetryFault; 797 return (result); 798 } 799 if ((retry_object != fs.first_object) || 800 (retry_pindex != fs.first_pindex)) { 801 release_page(&fs); 802 unlock_and_deallocate(&fs); 803 goto RetryFault; 804 } 805 806 /* 807 * Check whether the protection has changed or the object has 808 * been copied while we left the map unlocked. Changing from 809 * read to write permission is OK - we leave the page 810 * write-protected, and catch the write fault. Changing from 811 * write to read permission means that we can't mark the page 812 * write-enabled after all. 813 */ 814 prot &= retry_prot; 815 } 816 } 817 if (prot & VM_PROT_WRITE) { 818 vm_page_lock_queues(); 819 vm_page_flag_set(fs.m, PG_WRITEABLE); 820 vm_object_set_writeable_dirty(fs.m->object); 821 822 /* 823 * If the fault is a write, we know that this page is being 824 * written NOW so dirty it explicitly to save on 825 * pmap_is_modified() calls later. 826 * 827 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 828 * if the page is already dirty to prevent data written with 829 * the expectation of being synced from not being synced. 830 * Likewise if this entry does not request NOSYNC then make 831 * sure the page isn't marked NOSYNC. Applications sharing 832 * data should use the same flags to avoid ping ponging. 833 * 834 * Also tell the backing pager, if any, that it should remove 835 * any swap backing since the page is now dirty. 836 */ 837 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 838 if (fs.m->dirty == 0) 839 vm_page_flag_set(fs.m, PG_NOSYNC); 840 } else { 841 vm_page_flag_clear(fs.m, PG_NOSYNC); 842 } 843 vm_page_unlock_queues(); 844 if (fault_flags & VM_FAULT_DIRTY) { 845 vm_page_dirty(fs.m); 846 vm_pager_page_unswapped(fs.m); 847 } 848 } 849 850 /* 851 * Page had better still be busy 852 */ 853 KASSERT(fs.m->flags & PG_BUSY, 854 ("vm_fault: page %p not busy!", fs.m)); 855 /* 856 * Sanity check: page must be completely valid or it is not fit to 857 * map into user space. vm_pager_get_pages() ensures this. 858 */ 859 if (fs.m->valid != VM_PAGE_BITS_ALL) { 860 vm_page_zero_invalid(fs.m, TRUE); 861 printf("Warning: page %p partially invalid on fault\n", fs.m); 862 } 863 VM_OBJECT_UNLOCK(fs.object); 864 865 /* 866 * Put this page into the physical map. We had to do the unlock above 867 * because pmap_enter() may sleep. We don't put the page 868 * back on the active queue until later so that the pageout daemon 869 * won't find it (yet). 870 */ 871 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 872 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 873 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 874 } 875 VM_OBJECT_LOCK(fs.object); 876 vm_page_lock_queues(); 877 vm_page_flag_set(fs.m, PG_REFERENCED); 878 879 /* 880 * If the page is not wired down, then put it where the pageout daemon 881 * can find it. 882 */ 883 if (fault_flags & VM_FAULT_WIRE_MASK) { 884 if (wired) 885 vm_page_wire(fs.m); 886 else 887 vm_page_unwire(fs.m, 1); 888 } else { 889 vm_page_activate(fs.m); 890 } 891 vm_page_wakeup(fs.m); 892 vm_page_unlock_queues(); 893 894 /* 895 * Unlock everything, and return 896 */ 897 unlock_and_deallocate(&fs); 898 PROC_LOCK(curproc); 899 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) { 900 if (hardfault) { 901 curproc->p_stats->p_ru.ru_majflt++; 902 } else { 903 curproc->p_stats->p_ru.ru_minflt++; 904 } 905 } 906 PROC_UNLOCK(curproc); 907 908 return (KERN_SUCCESS); 909 } 910 911 /* 912 * vm_fault_prefault provides a quick way of clustering 913 * pagefaults into a processes address space. It is a "cousin" 914 * of vm_map_pmap_enter, except it runs at page fault time instead 915 * of mmap time. 916 */ 917 static void 918 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 919 { 920 int i; 921 vm_offset_t addr, starta; 922 vm_pindex_t pindex; 923 vm_page_t m, mpte; 924 vm_object_t object; 925 926 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) 927 return; 928 929 object = entry->object.vm_object; 930 931 starta = addra - PFBAK * PAGE_SIZE; 932 if (starta < entry->start) { 933 starta = entry->start; 934 } else if (starta > addra) { 935 starta = 0; 936 } 937 938 mpte = NULL; 939 for (i = 0; i < PAGEORDER_SIZE; i++) { 940 vm_object_t backing_object, lobject; 941 942 addr = addra + prefault_pageorder[i]; 943 if (addr > addra + (PFFOR * PAGE_SIZE)) 944 addr = 0; 945 946 if (addr < starta || addr >= entry->end) 947 continue; 948 949 if (!pmap_is_prefaultable(pmap, addr)) 950 continue; 951 952 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 953 lobject = object; 954 VM_OBJECT_LOCK(lobject); 955 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 956 lobject->type == OBJT_DEFAULT && 957 (backing_object = lobject->backing_object) != NULL) { 958 if (lobject->backing_object_offset & PAGE_MASK) 959 break; 960 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 961 VM_OBJECT_LOCK(backing_object); 962 VM_OBJECT_UNLOCK(lobject); 963 lobject = backing_object; 964 } 965 /* 966 * give-up when a page is not in memory 967 */ 968 if (m == NULL) { 969 VM_OBJECT_UNLOCK(lobject); 970 break; 971 } 972 vm_page_lock_queues(); 973 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 974 (m->busy == 0) && 975 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 976 977 if ((m->queue - m->pc) == PQ_CACHE) { 978 vm_page_deactivate(m); 979 } 980 vm_page_busy(m); 981 vm_page_unlock_queues(); 982 VM_OBJECT_UNLOCK(lobject); 983 mpte = pmap_enter_quick(pmap, addr, m, mpte); 984 VM_OBJECT_LOCK(lobject); 985 vm_page_lock_queues(); 986 vm_page_wakeup(m); 987 } 988 vm_page_unlock_queues(); 989 VM_OBJECT_UNLOCK(lobject); 990 } 991 } 992 993 /* 994 * vm_fault_quick: 995 * 996 * Ensure that the requested virtual address, which may be in userland, 997 * is valid. Fault-in the page if necessary. Return -1 on failure. 998 */ 999 int 1000 vm_fault_quick(caddr_t v, int prot) 1001 { 1002 int r; 1003 1004 if (prot & VM_PROT_WRITE) 1005 r = subyte(v, fubyte(v)); 1006 else 1007 r = fubyte(v); 1008 return(r); 1009 } 1010 1011 /* 1012 * vm_fault_wire: 1013 * 1014 * Wire down a range of virtual addresses in a map. 1015 */ 1016 int 1017 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1018 boolean_t user_wire, boolean_t fictitious) 1019 { 1020 vm_offset_t va; 1021 int rv; 1022 1023 /* 1024 * We simulate a fault to get the page and enter it in the physical 1025 * map. For user wiring, we only ask for read access on currently 1026 * read-only sections. 1027 */ 1028 for (va = start; va < end; va += PAGE_SIZE) { 1029 rv = vm_fault(map, va, 1030 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE, 1031 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING); 1032 if (rv) { 1033 if (va != start) 1034 vm_fault_unwire(map, start, va, fictitious); 1035 return (rv); 1036 } 1037 } 1038 return (KERN_SUCCESS); 1039 } 1040 1041 /* 1042 * vm_fault_unwire: 1043 * 1044 * Unwire a range of virtual addresses in a map. 1045 */ 1046 void 1047 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1048 boolean_t fictitious) 1049 { 1050 vm_paddr_t pa; 1051 vm_offset_t va; 1052 pmap_t pmap; 1053 1054 pmap = vm_map_pmap(map); 1055 1056 if (pmap != kernel_pmap) 1057 mtx_lock(&Giant); 1058 /* 1059 * Since the pages are wired down, we must be able to get their 1060 * mappings from the physical map system. 1061 */ 1062 for (va = start; va < end; va += PAGE_SIZE) { 1063 pa = pmap_extract(pmap, va); 1064 if (pa != 0) { 1065 pmap_change_wiring(pmap, va, FALSE); 1066 if (!fictitious) { 1067 vm_page_lock_queues(); 1068 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1069 vm_page_unlock_queues(); 1070 } 1071 } 1072 } 1073 if (pmap != kernel_pmap) 1074 mtx_unlock(&Giant); 1075 } 1076 1077 /* 1078 * Routine: 1079 * vm_fault_copy_entry 1080 * Function: 1081 * Copy all of the pages from a wired-down map entry to another. 1082 * 1083 * In/out conditions: 1084 * The source and destination maps must be locked for write. 1085 * The source map entry must be wired down (or be a sharing map 1086 * entry corresponding to a main map entry that is wired down). 1087 */ 1088 void 1089 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) 1090 vm_map_t dst_map; 1091 vm_map_t src_map; 1092 vm_map_entry_t dst_entry; 1093 vm_map_entry_t src_entry; 1094 { 1095 vm_object_t backing_object, dst_object, object; 1096 vm_object_t src_object; 1097 vm_ooffset_t dst_offset; 1098 vm_ooffset_t src_offset; 1099 vm_pindex_t pindex; 1100 vm_prot_t prot; 1101 vm_offset_t vaddr; 1102 vm_page_t dst_m; 1103 vm_page_t src_m; 1104 1105 #ifdef lint 1106 src_map++; 1107 #endif /* lint */ 1108 1109 src_object = src_entry->object.vm_object; 1110 src_offset = src_entry->offset; 1111 1112 /* 1113 * Create the top-level object for the destination entry. (Doesn't 1114 * actually shadow anything - we copy the pages directly.) 1115 */ 1116 dst_object = vm_object_allocate(OBJT_DEFAULT, 1117 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1118 1119 VM_OBJECT_LOCK(dst_object); 1120 dst_entry->object.vm_object = dst_object; 1121 dst_entry->offset = 0; 1122 1123 prot = dst_entry->max_protection; 1124 1125 /* 1126 * Loop through all of the pages in the entry's range, copying each 1127 * one from the source object (it should be there) to the destination 1128 * object. 1129 */ 1130 for (vaddr = dst_entry->start, dst_offset = 0; 1131 vaddr < dst_entry->end; 1132 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1133 1134 /* 1135 * Allocate a page in the destination object 1136 */ 1137 do { 1138 dst_m = vm_page_alloc(dst_object, 1139 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1140 if (dst_m == NULL) { 1141 VM_OBJECT_UNLOCK(dst_object); 1142 VM_WAIT; 1143 VM_OBJECT_LOCK(dst_object); 1144 } 1145 } while (dst_m == NULL); 1146 1147 /* 1148 * Find the page in the source object, and copy it in. 1149 * (Because the source is wired down, the page will be in 1150 * memory.) 1151 */ 1152 VM_OBJECT_LOCK(src_object); 1153 object = src_object; 1154 pindex = 0; 1155 while ((src_m = vm_page_lookup(object, pindex + 1156 OFF_TO_IDX(dst_offset + src_offset))) == NULL && 1157 (src_entry->protection & VM_PROT_WRITE) == 0 && 1158 (backing_object = object->backing_object) != NULL) { 1159 /* 1160 * Allow fallback to backing objects if we are reading. 1161 */ 1162 VM_OBJECT_LOCK(backing_object); 1163 pindex += OFF_TO_IDX(object->backing_object_offset); 1164 VM_OBJECT_UNLOCK(object); 1165 object = backing_object; 1166 } 1167 if (src_m == NULL) 1168 panic("vm_fault_copy_wired: page missing"); 1169 pmap_copy_page(src_m, dst_m); 1170 VM_OBJECT_UNLOCK(object); 1171 dst_m->valid = VM_PAGE_BITS_ALL; 1172 VM_OBJECT_UNLOCK(dst_object); 1173 1174 /* 1175 * Enter it in the pmap... 1176 */ 1177 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1178 VM_OBJECT_LOCK(dst_object); 1179 vm_page_lock_queues(); 1180 if ((prot & VM_PROT_WRITE) != 0) 1181 vm_page_flag_set(dst_m, PG_WRITEABLE); 1182 1183 /* 1184 * Mark it no longer busy, and put it on the active list. 1185 */ 1186 vm_page_activate(dst_m); 1187 vm_page_wakeup(dst_m); 1188 vm_page_unlock_queues(); 1189 } 1190 VM_OBJECT_UNLOCK(dst_object); 1191 } 1192 1193 1194 /* 1195 * This routine checks around the requested page for other pages that 1196 * might be able to be faulted in. This routine brackets the viable 1197 * pages for the pages to be paged in. 1198 * 1199 * Inputs: 1200 * m, rbehind, rahead 1201 * 1202 * Outputs: 1203 * marray (array of vm_page_t), reqpage (index of requested page) 1204 * 1205 * Return value: 1206 * number of pages in marray 1207 * 1208 * This routine can't block. 1209 */ 1210 static int 1211 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1212 vm_page_t m; 1213 int rbehind; 1214 int rahead; 1215 vm_page_t *marray; 1216 int *reqpage; 1217 { 1218 int i,j; 1219 vm_object_t object; 1220 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1221 vm_page_t rtm; 1222 int cbehind, cahead; 1223 1224 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1225 1226 object = m->object; 1227 pindex = m->pindex; 1228 1229 /* 1230 * we don't fault-ahead for device pager 1231 */ 1232 if (object->type == OBJT_DEVICE) { 1233 *reqpage = 0; 1234 marray[0] = m; 1235 return 1; 1236 } 1237 1238 /* 1239 * if the requested page is not available, then give up now 1240 */ 1241 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1242 return 0; 1243 } 1244 1245 if ((cbehind == 0) && (cahead == 0)) { 1246 *reqpage = 0; 1247 marray[0] = m; 1248 return 1; 1249 } 1250 1251 if (rahead > cahead) { 1252 rahead = cahead; 1253 } 1254 1255 if (rbehind > cbehind) { 1256 rbehind = cbehind; 1257 } 1258 1259 /* 1260 * try to do any readahead that we might have free pages for. 1261 */ 1262 if ((rahead + rbehind) > 1263 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) { 1264 pagedaemon_wakeup(); 1265 marray[0] = m; 1266 *reqpage = 0; 1267 return 1; 1268 } 1269 1270 /* 1271 * scan backward for the read behind pages -- in memory 1272 */ 1273 if (pindex > 0) { 1274 if (rbehind > pindex) { 1275 rbehind = pindex; 1276 startpindex = 0; 1277 } else { 1278 startpindex = pindex - rbehind; 1279 } 1280 1281 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1282 if (vm_page_lookup(object, tpindex)) { 1283 startpindex = tpindex + 1; 1284 break; 1285 } 1286 if (tpindex == 0) 1287 break; 1288 } 1289 1290 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1291 1292 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1293 if (rtm == NULL) { 1294 vm_page_lock_queues(); 1295 for (j = 0; j < i; j++) { 1296 vm_page_free(marray[j]); 1297 } 1298 vm_page_unlock_queues(); 1299 marray[0] = m; 1300 *reqpage = 0; 1301 return 1; 1302 } 1303 1304 marray[i] = rtm; 1305 } 1306 } else { 1307 startpindex = 0; 1308 i = 0; 1309 } 1310 1311 marray[i] = m; 1312 /* page offset of the required page */ 1313 *reqpage = i; 1314 1315 tpindex = pindex + 1; 1316 i++; 1317 1318 /* 1319 * scan forward for the read ahead pages 1320 */ 1321 endpindex = tpindex + rahead; 1322 if (endpindex > object->size) 1323 endpindex = object->size; 1324 1325 for (; tpindex < endpindex; i++, tpindex++) { 1326 1327 if (vm_page_lookup(object, tpindex)) { 1328 break; 1329 } 1330 1331 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1332 if (rtm == NULL) { 1333 break; 1334 } 1335 1336 marray[i] = rtm; 1337 } 1338 1339 /* return number of bytes of pages */ 1340 return i; 1341 } 1342