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