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