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_wakeup(fs->m); 137 vm_page_lock_queues(); 138 vm_page_deactivate(fs->m); 139 vm_page_unlock_queues(); 140 fs->m = NULL; 141 } 142 143 static inline void 144 unlock_map(struct faultstate *fs) 145 { 146 if (fs->lookup_still_valid) { 147 vm_map_lookup_done(fs->map, fs->entry); 148 fs->lookup_still_valid = FALSE; 149 } 150 } 151 152 static void 153 unlock_and_deallocate(struct faultstate *fs) 154 { 155 156 vm_object_pip_wakeup(fs->object); 157 VM_OBJECT_UNLOCK(fs->object); 158 if (fs->object != fs->first_object) { 159 VM_OBJECT_LOCK(fs->first_object); 160 vm_page_lock_queues(); 161 vm_page_free(fs->first_m); 162 vm_page_unlock_queues(); 163 vm_object_pip_wakeup(fs->first_object); 164 VM_OBJECT_UNLOCK(fs->first_object); 165 fs->first_m = NULL; 166 } 167 vm_object_deallocate(fs->first_object); 168 unlock_map(fs); 169 if (fs->vp != NULL) { 170 int vfslocked; 171 172 vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount); 173 vput(fs->vp); 174 fs->vp = NULL; 175 VFS_UNLOCK_GIANT(vfslocked); 176 } 177 } 178 179 /* 180 * TRYPAGER - used by vm_fault to calculate whether the pager for the 181 * current object *might* contain the page. 182 * 183 * default objects are zero-fill, there is no real pager. 184 */ 185 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 186 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 187 188 /* 189 * vm_fault: 190 * 191 * Handle a page fault occurring at the given address, 192 * requiring the given permissions, in the map specified. 193 * If successful, the page is inserted into the 194 * associated physical map. 195 * 196 * NOTE: the given address should be truncated to the 197 * proper page address. 198 * 199 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 200 * a standard error specifying why the fault is fatal is returned. 201 * 202 * 203 * The map in question must be referenced, and remains so. 204 * Caller may hold no locks. 205 */ 206 int 207 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 208 int fault_flags) 209 { 210 vm_prot_t prot; 211 int is_first_object_locked, result; 212 boolean_t growstack, wired; 213 int map_generation; 214 vm_object_t next_object; 215 vm_page_t marray[VM_FAULT_READ]; 216 int hardfault; 217 int faultcount; 218 struct faultstate fs; 219 220 hardfault = 0; 221 growstack = TRUE; 222 PCPU_INC(cnt.v_vm_faults); 223 224 RetryFault:; 225 226 /* 227 * Find the backing store object and offset into it to begin the 228 * search. 229 */ 230 fs.map = map; 231 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 232 &fs.first_object, &fs.first_pindex, &prot, &wired); 233 if (result != KERN_SUCCESS) { 234 if (result != KERN_PROTECTION_FAILURE || 235 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) { 236 if (growstack && result == KERN_INVALID_ADDRESS && 237 map != kernel_map && curproc != NULL) { 238 result = vm_map_growstack(curproc, vaddr); 239 if (result != KERN_SUCCESS) 240 return (KERN_FAILURE); 241 growstack = FALSE; 242 goto RetryFault; 243 } 244 return (result); 245 } 246 247 /* 248 * If we are user-wiring a r/w segment, and it is COW, then 249 * we need to do the COW operation. Note that we don't COW 250 * currently RO sections now, because it is NOT desirable 251 * to COW .text. We simply keep .text from ever being COW'ed 252 * and take the heat that one cannot debug wired .text sections. 253 */ 254 result = vm_map_lookup(&fs.map, vaddr, 255 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 256 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 257 if (result != KERN_SUCCESS) 258 return (result); 259 260 /* 261 * If we don't COW now, on a user wire, the user will never 262 * be able to write to the mapping. If we don't make this 263 * restriction, the bookkeeping would be nearly impossible. 264 * 265 * XXX The following assignment modifies the map without 266 * holding a write lock on it. 267 */ 268 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 269 fs.entry->max_protection &= ~VM_PROT_WRITE; 270 } 271 272 map_generation = fs.map->timestamp; 273 274 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 275 panic("vm_fault: fault on nofault entry, addr: %lx", 276 (u_long)vaddr); 277 } 278 279 /* 280 * Make a reference to this object to prevent its disposal while we 281 * are messing with it. Once we have the reference, the map is free 282 * to be diddled. Since objects reference their shadows (and copies), 283 * they will stay around as well. 284 * 285 * Bump the paging-in-progress count to prevent size changes (e.g. 286 * truncation operations) during I/O. This must be done after 287 * obtaining the vnode lock in order to avoid possible deadlocks. 288 * 289 * XXX vnode_pager_lock() can block without releasing the map lock. 290 */ 291 if (fs.first_object->flags & OBJ_NEEDGIANT) 292 mtx_lock(&Giant); 293 VM_OBJECT_LOCK(fs.first_object); 294 vm_object_reference_locked(fs.first_object); 295 fs.vp = vnode_pager_lock(fs.first_object); 296 KASSERT(fs.vp == NULL || !fs.map->system_map, 297 ("vm_fault: vnode-backed object mapped by system map")); 298 KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 || 299 !fs.map->system_map, 300 ("vm_fault: Object requiring giant mapped by system map")); 301 if (fs.first_object->flags & OBJ_NEEDGIANT) 302 mtx_unlock(&Giant); 303 vm_object_pip_add(fs.first_object, 1); 304 305 fs.lookup_still_valid = TRUE; 306 307 if (wired) 308 fault_type = prot; 309 310 fs.first_m = NULL; 311 312 /* 313 * Search for the page at object/offset. 314 */ 315 fs.object = fs.first_object; 316 fs.pindex = fs.first_pindex; 317 while (TRUE) { 318 /* 319 * If the object is dead, we stop here 320 */ 321 if (fs.object->flags & OBJ_DEAD) { 322 unlock_and_deallocate(&fs); 323 return (KERN_PROTECTION_FAILURE); 324 } 325 326 /* 327 * See if page is resident 328 */ 329 fs.m = vm_page_lookup(fs.object, fs.pindex); 330 if (fs.m != NULL) { 331 int queue; 332 333 /* 334 * check for page-based copy on write. 335 * We check fs.object == fs.first_object so 336 * as to ensure the legacy COW mechanism is 337 * used when the page in question is part of 338 * a shadow object. Otherwise, vm_page_cowfault() 339 * removes the page from the backing object, 340 * which is not what we want. 341 */ 342 vm_page_lock_queues(); 343 if ((fs.m->cow) && 344 (fault_type & VM_PROT_WRITE) && 345 (fs.object == fs.first_object)) { 346 vm_page_cowfault(fs.m); 347 vm_page_unlock_queues(); 348 unlock_and_deallocate(&fs); 349 goto RetryFault; 350 } 351 352 /* 353 * Wait/Retry if the page is busy. We have to do this 354 * if the page is busy via either VPO_BUSY or 355 * vm_page_t->busy because the vm_pager may be using 356 * vm_page_t->busy for pageouts ( and even pageins if 357 * it is the vnode pager ), and we could end up trying 358 * to pagein and pageout the same page simultaneously. 359 * 360 * We can theoretically allow the busy case on a read 361 * fault if the page is marked valid, but since such 362 * pages are typically already pmap'd, putting that 363 * special case in might be more effort then it is 364 * worth. We cannot under any circumstances mess 365 * around with a vm_page_t->busy page except, perhaps, 366 * to pmap it. 367 */ 368 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) { 369 vm_page_unlock_queues(); 370 VM_OBJECT_UNLOCK(fs.object); 371 if (fs.object != fs.first_object) { 372 VM_OBJECT_LOCK(fs.first_object); 373 vm_page_lock_queues(); 374 vm_page_free(fs.first_m); 375 vm_page_unlock_queues(); 376 vm_object_pip_wakeup(fs.first_object); 377 VM_OBJECT_UNLOCK(fs.first_object); 378 fs.first_m = NULL; 379 } 380 unlock_map(&fs); 381 if (fs.vp != NULL) { 382 int vfslck; 383 384 vfslck = VFS_LOCK_GIANT(fs.vp->v_mount); 385 vput(fs.vp); 386 fs.vp = NULL; 387 VFS_UNLOCK_GIANT(vfslck); 388 } 389 VM_OBJECT_LOCK(fs.object); 390 if (fs.m == vm_page_lookup(fs.object, 391 fs.pindex)) { 392 vm_page_sleep_if_busy(fs.m, TRUE, 393 "vmpfw"); 394 } 395 vm_object_pip_wakeup(fs.object); 396 VM_OBJECT_UNLOCK(fs.object); 397 PCPU_INC(cnt.v_intrans); 398 vm_object_deallocate(fs.first_object); 399 goto RetryFault; 400 } 401 queue = fs.m->queue; 402 403 vm_pageq_remove_nowakeup(fs.m); 404 405 if (VM_PAGE_RESOLVEQUEUE(fs.m, queue) == PQ_CACHE) { 406 cnt.v_reactivated++; 407 if (vm_page_count_severe()) { 408 vm_page_activate(fs.m); 409 vm_page_unlock_queues(); 410 unlock_and_deallocate(&fs); 411 VM_WAITPFAULT; 412 goto RetryFault; 413 } 414 } 415 vm_page_unlock_queues(); 416 417 /* 418 * Mark page busy for other processes, and the 419 * pagedaemon. If it still isn't completely valid 420 * (readable), jump to readrest, else break-out ( we 421 * found the page ). 422 */ 423 vm_page_busy(fs.m); 424 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 425 fs.m->object != kernel_object && fs.m->object != kmem_object) { 426 goto readrest; 427 } 428 429 break; 430 } 431 432 /* 433 * Page is not resident, If this is the search termination 434 * or the pager might contain the page, allocate a new page. 435 */ 436 if (TRYPAGER || fs.object == fs.first_object) { 437 if (fs.pindex >= fs.object->size) { 438 unlock_and_deallocate(&fs); 439 return (KERN_PROTECTION_FAILURE); 440 } 441 442 /* 443 * Allocate a new page for this object/offset pair. 444 */ 445 fs.m = NULL; 446 if (!vm_page_count_severe()) { 447 fs.m = vm_page_alloc(fs.object, fs.pindex, 448 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO); 449 } 450 if (fs.m == NULL) { 451 unlock_and_deallocate(&fs); 452 VM_WAITPFAULT; 453 goto RetryFault; 454 } 455 } 456 457 readrest: 458 /* 459 * We have found a valid page or we have allocated a new page. 460 * The page thus may not be valid or may not be entirely 461 * valid. 462 * 463 * Attempt to fault-in the page if there is a chance that the 464 * pager has it, and potentially fault in additional pages 465 * at the same time. 466 */ 467 if (TRYPAGER) { 468 int rv; 469 int reqpage = 0; 470 int ahead, behind; 471 u_char behavior = vm_map_entry_behavior(fs.entry); 472 473 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 474 ahead = 0; 475 behind = 0; 476 } else { 477 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 478 if (behind > VM_FAULT_READ_BEHIND) 479 behind = VM_FAULT_READ_BEHIND; 480 481 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 482 if (ahead > VM_FAULT_READ_AHEAD) 483 ahead = VM_FAULT_READ_AHEAD; 484 } 485 is_first_object_locked = FALSE; 486 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 487 (behavior != MAP_ENTRY_BEHAV_RANDOM && 488 fs.pindex >= fs.entry->lastr && 489 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 490 (fs.first_object == fs.object || 491 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 492 fs.first_object->type != OBJT_DEVICE) { 493 vm_pindex_t firstpindex, tmppindex; 494 495 if (fs.first_pindex < 2 * VM_FAULT_READ) 496 firstpindex = 0; 497 else 498 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 499 500 vm_page_lock_queues(); 501 /* 502 * note: partially valid pages cannot be 503 * included in the lookahead - NFS piecemeal 504 * writes will barf on it badly. 505 */ 506 for (tmppindex = fs.first_pindex - 1; 507 tmppindex >= firstpindex; 508 --tmppindex) { 509 vm_page_t mt; 510 511 mt = vm_page_lookup(fs.first_object, tmppindex); 512 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 513 break; 514 if (mt->busy || 515 (mt->oflags & VPO_BUSY) || 516 (mt->flags & (PG_FICTITIOUS | PG_UNMANAGED)) || 517 mt->hold_count || 518 mt->wire_count) 519 continue; 520 pmap_remove_all(mt); 521 if (mt->dirty) { 522 vm_page_deactivate(mt); 523 } else { 524 vm_page_cache(mt); 525 } 526 } 527 vm_page_unlock_queues(); 528 ahead += behind; 529 behind = 0; 530 } 531 if (is_first_object_locked) 532 VM_OBJECT_UNLOCK(fs.first_object); 533 /* 534 * now we find out if any other pages should be paged 535 * in at this time this routine checks to see if the 536 * pages surrounding this fault reside in the same 537 * object as the page for this fault. If they do, 538 * then they are faulted in also into the object. The 539 * array "marray" returned contains an array of 540 * vm_page_t structs where one of them is the 541 * vm_page_t passed to the routine. The reqpage 542 * return value is the index into the marray for the 543 * vm_page_t passed to the routine. 544 * 545 * fs.m plus the additional pages are VPO_BUSY'd. 546 * 547 * XXX vm_fault_additional_pages() can block 548 * without releasing the map lock. 549 */ 550 faultcount = vm_fault_additional_pages( 551 fs.m, behind, ahead, marray, &reqpage); 552 553 /* 554 * update lastr imperfectly (we do not know how much 555 * getpages will actually read), but good enough. 556 * 557 * XXX The following assignment modifies the map 558 * without holding a write lock on it. 559 */ 560 fs.entry->lastr = fs.pindex + faultcount - behind; 561 562 /* 563 * Call the pager to retrieve the data, if any, after 564 * releasing the lock on the map. We hold a ref on 565 * fs.object and the pages are VPO_BUSY'd. 566 */ 567 unlock_map(&fs); 568 569 rv = faultcount ? 570 vm_pager_get_pages(fs.object, marray, faultcount, 571 reqpage) : VM_PAGER_FAIL; 572 573 if (rv == VM_PAGER_OK) { 574 /* 575 * Found the page. Leave it busy while we play 576 * with it. 577 */ 578 579 /* 580 * Relookup in case pager changed page. Pager 581 * is responsible for disposition of old page 582 * if moved. 583 */ 584 fs.m = vm_page_lookup(fs.object, fs.pindex); 585 if (!fs.m) { 586 unlock_and_deallocate(&fs); 587 goto RetryFault; 588 } 589 590 hardfault++; 591 break; /* break to PAGE HAS BEEN FOUND */ 592 } 593 /* 594 * Remove the bogus page (which does not exist at this 595 * object/offset); before doing so, we must get back 596 * our object lock to preserve our invariant. 597 * 598 * Also wake up any other process that may want to bring 599 * in this page. 600 * 601 * If this is the top-level object, we must leave the 602 * busy page to prevent another process from rushing 603 * past us, and inserting the page in that object at 604 * the same time that we are. 605 */ 606 if (rv == VM_PAGER_ERROR) 607 printf("vm_fault: pager read error, pid %d (%s)\n", 608 curproc->p_pid, curproc->p_comm); 609 /* 610 * Data outside the range of the pager or an I/O error 611 */ 612 /* 613 * XXX - the check for kernel_map is a kludge to work 614 * around having the machine panic on a kernel space 615 * fault w/ I/O error. 616 */ 617 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 618 (rv == VM_PAGER_BAD)) { 619 vm_page_lock_queues(); 620 vm_page_free(fs.m); 621 vm_page_unlock_queues(); 622 fs.m = NULL; 623 unlock_and_deallocate(&fs); 624 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 625 } 626 if (fs.object != fs.first_object) { 627 vm_page_lock_queues(); 628 vm_page_free(fs.m); 629 vm_page_unlock_queues(); 630 fs.m = NULL; 631 /* 632 * XXX - we cannot just fall out at this 633 * point, m has been freed and is invalid! 634 */ 635 } 636 } 637 638 /* 639 * We get here if the object has default pager (or unwiring) 640 * or the pager doesn't have the page. 641 */ 642 if (fs.object == fs.first_object) 643 fs.first_m = fs.m; 644 645 /* 646 * Move on to the next object. Lock the next object before 647 * unlocking the current one. 648 */ 649 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 650 next_object = fs.object->backing_object; 651 if (next_object == NULL) { 652 /* 653 * If there's no object left, fill the page in the top 654 * object with zeros. 655 */ 656 if (fs.object != fs.first_object) { 657 vm_object_pip_wakeup(fs.object); 658 VM_OBJECT_UNLOCK(fs.object); 659 660 fs.object = fs.first_object; 661 fs.pindex = fs.first_pindex; 662 fs.m = fs.first_m; 663 VM_OBJECT_LOCK(fs.object); 664 } 665 fs.first_m = NULL; 666 667 /* 668 * Zero the page if necessary and mark it valid. 669 */ 670 if ((fs.m->flags & PG_ZERO) == 0) { 671 pmap_zero_page(fs.m); 672 } else { 673 PCPU_INC(cnt.v_ozfod); 674 } 675 PCPU_INC(cnt.v_zfod); 676 fs.m->valid = VM_PAGE_BITS_ALL; 677 break; /* break to PAGE HAS BEEN FOUND */ 678 } else { 679 KASSERT(fs.object != next_object, 680 ("object loop %p", next_object)); 681 VM_OBJECT_LOCK(next_object); 682 vm_object_pip_add(next_object, 1); 683 if (fs.object != fs.first_object) 684 vm_object_pip_wakeup(fs.object); 685 VM_OBJECT_UNLOCK(fs.object); 686 fs.object = next_object; 687 } 688 } 689 690 KASSERT((fs.m->oflags & VPO_BUSY) != 0, 691 ("vm_fault: not busy after main loop")); 692 693 /* 694 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 695 * is held.] 696 */ 697 698 /* 699 * If the page is being written, but isn't already owned by the 700 * top-level object, we have to copy it into a new page owned by the 701 * top-level object. 702 */ 703 if (fs.object != fs.first_object) { 704 /* 705 * We only really need to copy if we want to write it. 706 */ 707 if (fault_type & VM_PROT_WRITE) { 708 /* 709 * This allows pages to be virtually copied from a 710 * backing_object into the first_object, where the 711 * backing object has no other refs to it, and cannot 712 * gain any more refs. Instead of a bcopy, we just 713 * move the page from the backing object to the 714 * first object. Note that we must mark the page 715 * dirty in the first object so that it will go out 716 * to swap when needed. 717 */ 718 is_first_object_locked = FALSE; 719 if ( 720 /* 721 * Only one shadow object 722 */ 723 (fs.object->shadow_count == 1) && 724 /* 725 * No COW refs, except us 726 */ 727 (fs.object->ref_count == 1) && 728 /* 729 * No one else can look this object up 730 */ 731 (fs.object->handle == NULL) && 732 /* 733 * No other ways to look the object up 734 */ 735 ((fs.object->type == OBJT_DEFAULT) || 736 (fs.object->type == OBJT_SWAP)) && 737 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 738 /* 739 * We don't chase down the shadow chain 740 */ 741 fs.object == fs.first_object->backing_object) { 742 vm_page_lock_queues(); 743 /* 744 * get rid of the unnecessary page 745 */ 746 vm_page_free(fs.first_m); 747 /* 748 * grab the page and put it into the 749 * process'es object. The page is 750 * automatically made dirty. 751 */ 752 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 753 vm_page_unlock_queues(); 754 vm_page_busy(fs.m); 755 fs.first_m = fs.m; 756 fs.m = NULL; 757 PCPU_INC(cnt.v_cow_optim); 758 } else { 759 /* 760 * Oh, well, lets copy it. 761 */ 762 pmap_copy_page(fs.m, fs.first_m); 763 fs.first_m->valid = VM_PAGE_BITS_ALL; 764 } 765 if (fs.m) { 766 /* 767 * We no longer need the old page or object. 768 */ 769 release_page(&fs); 770 } 771 /* 772 * fs.object != fs.first_object due to above 773 * conditional 774 */ 775 vm_object_pip_wakeup(fs.object); 776 VM_OBJECT_UNLOCK(fs.object); 777 /* 778 * Only use the new page below... 779 */ 780 fs.object = fs.first_object; 781 fs.pindex = fs.first_pindex; 782 fs.m = fs.first_m; 783 if (!is_first_object_locked) 784 VM_OBJECT_LOCK(fs.object); 785 PCPU_INC(cnt.v_cow_faults); 786 } else { 787 prot &= ~VM_PROT_WRITE; 788 } 789 } 790 791 /* 792 * We must verify that the maps have not changed since our last 793 * lookup. 794 */ 795 if (!fs.lookup_still_valid) { 796 vm_object_t retry_object; 797 vm_pindex_t retry_pindex; 798 vm_prot_t retry_prot; 799 800 if (!vm_map_trylock_read(fs.map)) { 801 release_page(&fs); 802 unlock_and_deallocate(&fs); 803 goto RetryFault; 804 } 805 fs.lookup_still_valid = TRUE; 806 if (fs.map->timestamp != map_generation) { 807 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 808 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 809 810 /* 811 * If we don't need the page any longer, put it on the inactive 812 * list (the easiest thing to do here). If no one needs it, 813 * pageout will grab it eventually. 814 */ 815 if (result != KERN_SUCCESS) { 816 release_page(&fs); 817 unlock_and_deallocate(&fs); 818 819 /* 820 * If retry of map lookup would have blocked then 821 * retry fault from start. 822 */ 823 if (result == KERN_FAILURE) 824 goto RetryFault; 825 return (result); 826 } 827 if ((retry_object != fs.first_object) || 828 (retry_pindex != fs.first_pindex)) { 829 release_page(&fs); 830 unlock_and_deallocate(&fs); 831 goto RetryFault; 832 } 833 834 /* 835 * Check whether the protection has changed or the object has 836 * been copied while we left the map unlocked. Changing from 837 * read to write permission is OK - we leave the page 838 * write-protected, and catch the write fault. Changing from 839 * write to read permission means that we can't mark the page 840 * write-enabled after all. 841 */ 842 prot &= retry_prot; 843 } 844 } 845 if (prot & VM_PROT_WRITE) { 846 vm_object_set_writeable_dirty(fs.object); 847 848 /* 849 * If the fault is a write, we know that this page is being 850 * written NOW so dirty it explicitly to save on 851 * pmap_is_modified() calls later. 852 * 853 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 854 * if the page is already dirty to prevent data written with 855 * the expectation of being synced from not being synced. 856 * Likewise if this entry does not request NOSYNC then make 857 * sure the page isn't marked NOSYNC. Applications sharing 858 * data should use the same flags to avoid ping ponging. 859 * 860 * Also tell the backing pager, if any, that it should remove 861 * any swap backing since the page is now dirty. 862 */ 863 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 864 if (fs.m->dirty == 0) 865 fs.m->oflags |= VPO_NOSYNC; 866 } else { 867 fs.m->oflags &= ~VPO_NOSYNC; 868 } 869 if (fault_flags & VM_FAULT_DIRTY) { 870 vm_page_dirty(fs.m); 871 vm_pager_page_unswapped(fs.m); 872 } 873 } 874 875 /* 876 * Page had better still be busy 877 */ 878 KASSERT(fs.m->oflags & VPO_BUSY, 879 ("vm_fault: page %p not busy!", fs.m)); 880 /* 881 * Sanity check: page must be completely valid or it is not fit to 882 * map into user space. vm_pager_get_pages() ensures this. 883 */ 884 if (fs.m->valid != VM_PAGE_BITS_ALL) { 885 vm_page_zero_invalid(fs.m, TRUE); 886 printf("Warning: page %p partially invalid on fault\n", fs.m); 887 } 888 VM_OBJECT_UNLOCK(fs.object); 889 890 /* 891 * Put this page into the physical map. We had to do the unlock above 892 * because pmap_enter() may sleep. We don't put the page 893 * back on the active queue until later so that the pageout daemon 894 * won't find it (yet). 895 */ 896 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 897 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 898 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 899 } 900 VM_OBJECT_LOCK(fs.object); 901 vm_page_lock_queues(); 902 vm_page_flag_set(fs.m, PG_REFERENCED); 903 904 /* 905 * If the page is not wired down, then put it where the pageout daemon 906 * can find it. 907 */ 908 if (fault_flags & VM_FAULT_WIRE_MASK) { 909 if (wired) 910 vm_page_wire(fs.m); 911 else 912 vm_page_unwire(fs.m, 1); 913 } else { 914 vm_page_activate(fs.m); 915 } 916 vm_page_unlock_queues(); 917 vm_page_wakeup(fs.m); 918 919 /* 920 * Unlock everything, and return 921 */ 922 unlock_and_deallocate(&fs); 923 if (hardfault) 924 curthread->td_ru.ru_majflt++; 925 else 926 curthread->td_ru.ru_minflt++; 927 928 return (KERN_SUCCESS); 929 } 930 931 /* 932 * vm_fault_prefault provides a quick way of clustering 933 * pagefaults into a processes address space. It is a "cousin" 934 * of vm_map_pmap_enter, except it runs at page fault time instead 935 * of mmap time. 936 */ 937 static void 938 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 939 { 940 int i; 941 vm_offset_t addr, starta; 942 vm_pindex_t pindex; 943 vm_page_t m; 944 vm_object_t object; 945 946 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 947 return; 948 949 object = entry->object.vm_object; 950 951 starta = addra - PFBAK * PAGE_SIZE; 952 if (starta < entry->start) { 953 starta = entry->start; 954 } else if (starta > addra) { 955 starta = 0; 956 } 957 958 for (i = 0; i < PAGEORDER_SIZE; i++) { 959 vm_object_t backing_object, lobject; 960 961 addr = addra + prefault_pageorder[i]; 962 if (addr > addra + (PFFOR * PAGE_SIZE)) 963 addr = 0; 964 965 if (addr < starta || addr >= entry->end) 966 continue; 967 968 if (!pmap_is_prefaultable(pmap, addr)) 969 continue; 970 971 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 972 lobject = object; 973 VM_OBJECT_LOCK(lobject); 974 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 975 lobject->type == OBJT_DEFAULT && 976 (backing_object = lobject->backing_object) != NULL) { 977 if (lobject->backing_object_offset & PAGE_MASK) 978 break; 979 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 980 VM_OBJECT_LOCK(backing_object); 981 VM_OBJECT_UNLOCK(lobject); 982 lobject = backing_object; 983 } 984 /* 985 * give-up when a page is not in memory 986 */ 987 if (m == NULL) { 988 VM_OBJECT_UNLOCK(lobject); 989 break; 990 } 991 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 992 (m->busy == 0) && 993 (m->flags & PG_FICTITIOUS) == 0) { 994 995 vm_page_lock_queues(); 996 if (!VM_PAGE_INQUEUE1(m, PQ_CACHE)) 997 pmap_enter_quick(pmap, addr, m, 998 entry->protection); 999 vm_page_unlock_queues(); 1000 } 1001 VM_OBJECT_UNLOCK(lobject); 1002 } 1003 } 1004 1005 /* 1006 * vm_fault_quick: 1007 * 1008 * Ensure that the requested virtual address, which may be in userland, 1009 * is valid. Fault-in the page if necessary. Return -1 on failure. 1010 */ 1011 int 1012 vm_fault_quick(caddr_t v, int prot) 1013 { 1014 int r; 1015 1016 if (prot & VM_PROT_WRITE) 1017 r = subyte(v, fubyte(v)); 1018 else 1019 r = fubyte(v); 1020 return(r); 1021 } 1022 1023 /* 1024 * vm_fault_wire: 1025 * 1026 * Wire down a range of virtual addresses in a map. 1027 */ 1028 int 1029 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1030 boolean_t user_wire, boolean_t fictitious) 1031 { 1032 vm_offset_t va; 1033 int rv; 1034 1035 /* 1036 * We simulate a fault to get the page and enter it in the physical 1037 * map. For user wiring, we only ask for read access on currently 1038 * read-only sections. 1039 */ 1040 for (va = start; va < end; va += PAGE_SIZE) { 1041 rv = vm_fault(map, va, 1042 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE, 1043 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING); 1044 if (rv) { 1045 if (va != start) 1046 vm_fault_unwire(map, start, va, fictitious); 1047 return (rv); 1048 } 1049 } 1050 return (KERN_SUCCESS); 1051 } 1052 1053 /* 1054 * vm_fault_unwire: 1055 * 1056 * Unwire a range of virtual addresses in a map. 1057 */ 1058 void 1059 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1060 boolean_t fictitious) 1061 { 1062 vm_paddr_t pa; 1063 vm_offset_t va; 1064 pmap_t pmap; 1065 1066 pmap = vm_map_pmap(map); 1067 1068 /* 1069 * Since the pages are wired down, we must be able to get their 1070 * mappings from the physical map system. 1071 */ 1072 for (va = start; va < end; va += PAGE_SIZE) { 1073 pa = pmap_extract(pmap, va); 1074 if (pa != 0) { 1075 pmap_change_wiring(pmap, va, FALSE); 1076 if (!fictitious) { 1077 vm_page_lock_queues(); 1078 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1079 vm_page_unlock_queues(); 1080 } 1081 } 1082 } 1083 } 1084 1085 /* 1086 * Routine: 1087 * vm_fault_copy_entry 1088 * Function: 1089 * Copy all of the pages from a wired-down map entry to another. 1090 * 1091 * In/out conditions: 1092 * The source and destination maps must be locked for write. 1093 * The source map entry must be wired down (or be a sharing map 1094 * entry corresponding to a main map entry that is wired down). 1095 */ 1096 void 1097 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry) 1098 vm_map_t dst_map; 1099 vm_map_t src_map; 1100 vm_map_entry_t dst_entry; 1101 vm_map_entry_t src_entry; 1102 { 1103 vm_object_t backing_object, dst_object, object; 1104 vm_object_t src_object; 1105 vm_ooffset_t dst_offset; 1106 vm_ooffset_t src_offset; 1107 vm_pindex_t pindex; 1108 vm_prot_t prot; 1109 vm_offset_t vaddr; 1110 vm_page_t dst_m; 1111 vm_page_t src_m; 1112 1113 #ifdef lint 1114 src_map++; 1115 #endif /* lint */ 1116 1117 src_object = src_entry->object.vm_object; 1118 src_offset = src_entry->offset; 1119 1120 /* 1121 * Create the top-level object for the destination entry. (Doesn't 1122 * actually shadow anything - we copy the pages directly.) 1123 */ 1124 dst_object = vm_object_allocate(OBJT_DEFAULT, 1125 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1126 1127 VM_OBJECT_LOCK(dst_object); 1128 dst_entry->object.vm_object = dst_object; 1129 dst_entry->offset = 0; 1130 1131 prot = dst_entry->max_protection; 1132 1133 /* 1134 * Loop through all of the pages in the entry's range, copying each 1135 * one from the source object (it should be there) to the destination 1136 * object. 1137 */ 1138 for (vaddr = dst_entry->start, dst_offset = 0; 1139 vaddr < dst_entry->end; 1140 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1141 1142 /* 1143 * Allocate a page in the destination object 1144 */ 1145 do { 1146 dst_m = vm_page_alloc(dst_object, 1147 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1148 if (dst_m == NULL) { 1149 VM_OBJECT_UNLOCK(dst_object); 1150 VM_WAIT; 1151 VM_OBJECT_LOCK(dst_object); 1152 } 1153 } while (dst_m == NULL); 1154 1155 /* 1156 * Find the page in the source object, and copy it in. 1157 * (Because the source is wired down, the page will be in 1158 * memory.) 1159 */ 1160 VM_OBJECT_LOCK(src_object); 1161 object = src_object; 1162 pindex = 0; 1163 while ((src_m = vm_page_lookup(object, pindex + 1164 OFF_TO_IDX(dst_offset + src_offset))) == NULL && 1165 (src_entry->protection & VM_PROT_WRITE) == 0 && 1166 (backing_object = object->backing_object) != NULL) { 1167 /* 1168 * Allow fallback to backing objects if we are reading. 1169 */ 1170 VM_OBJECT_LOCK(backing_object); 1171 pindex += OFF_TO_IDX(object->backing_object_offset); 1172 VM_OBJECT_UNLOCK(object); 1173 object = backing_object; 1174 } 1175 if (src_m == NULL) 1176 panic("vm_fault_copy_wired: page missing"); 1177 pmap_copy_page(src_m, dst_m); 1178 VM_OBJECT_UNLOCK(object); 1179 dst_m->valid = VM_PAGE_BITS_ALL; 1180 VM_OBJECT_UNLOCK(dst_object); 1181 1182 /* 1183 * Enter it in the pmap... 1184 */ 1185 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1186 1187 /* 1188 * Mark it no longer busy, and put it on the active list. 1189 */ 1190 VM_OBJECT_LOCK(dst_object); 1191 vm_page_lock_queues(); 1192 vm_page_activate(dst_m); 1193 vm_page_unlock_queues(); 1194 vm_page_wakeup(dst_m); 1195 } 1196 VM_OBJECT_UNLOCK(dst_object); 1197 } 1198 1199 1200 /* 1201 * This routine checks around the requested page for other pages that 1202 * might be able to be faulted in. This routine brackets the viable 1203 * pages for the pages to be paged in. 1204 * 1205 * Inputs: 1206 * m, rbehind, rahead 1207 * 1208 * Outputs: 1209 * marray (array of vm_page_t), reqpage (index of requested page) 1210 * 1211 * Return value: 1212 * number of pages in marray 1213 * 1214 * This routine can't block. 1215 */ 1216 static int 1217 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1218 vm_page_t m; 1219 int rbehind; 1220 int rahead; 1221 vm_page_t *marray; 1222 int *reqpage; 1223 { 1224 int i,j; 1225 vm_object_t object; 1226 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1227 vm_page_t rtm; 1228 int cbehind, cahead; 1229 1230 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1231 1232 object = m->object; 1233 pindex = m->pindex; 1234 cbehind = cahead = 0; 1235 1236 /* 1237 * if the requested page is not available, then give up now 1238 */ 1239 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1240 return 0; 1241 } 1242 1243 if ((cbehind == 0) && (cahead == 0)) { 1244 *reqpage = 0; 1245 marray[0] = m; 1246 return 1; 1247 } 1248 1249 if (rahead > cahead) { 1250 rahead = cahead; 1251 } 1252 1253 if (rbehind > cbehind) { 1254 rbehind = cbehind; 1255 } 1256 1257 /* 1258 * scan backward for the read behind pages -- in memory 1259 */ 1260 if (pindex > 0) { 1261 if (rbehind > pindex) { 1262 rbehind = pindex; 1263 startpindex = 0; 1264 } else { 1265 startpindex = pindex - rbehind; 1266 } 1267 1268 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1269 rtm->pindex >= startpindex) 1270 startpindex = rtm->pindex + 1; 1271 1272 /* tpindex is unsigned; beware of numeric underflow. */ 1273 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex && 1274 tpindex < pindex; i++, tpindex--) { 1275 1276 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1277 if (rtm == NULL) { 1278 /* 1279 * Shift the allocated pages to the 1280 * beginning of the array. 1281 */ 1282 for (j = 0; j < i; j++) { 1283 marray[j] = marray[j + tpindex + 1 - 1284 startpindex]; 1285 } 1286 break; 1287 } 1288 1289 marray[tpindex - startpindex] = rtm; 1290 } 1291 } else { 1292 startpindex = 0; 1293 i = 0; 1294 } 1295 1296 marray[i] = m; 1297 /* page offset of the required page */ 1298 *reqpage = i; 1299 1300 tpindex = pindex + 1; 1301 i++; 1302 1303 /* 1304 * scan forward for the read ahead pages 1305 */ 1306 endpindex = tpindex + rahead; 1307 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1308 endpindex = rtm->pindex; 1309 if (endpindex > object->size) 1310 endpindex = object->size; 1311 1312 for (; tpindex < endpindex; i++, tpindex++) { 1313 1314 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1315 if (rtm == NULL) { 1316 break; 1317 } 1318 1319 marray[i] = rtm; 1320 } 1321 1322 /* return number of pages */ 1323 return i; 1324 } 1325