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 "opt_vm.h" 78 79 #include <sys/param.h> 80 #include <sys/systm.h> 81 #include <sys/kernel.h> 82 #include <sys/lock.h> 83 #include <sys/mutex.h> 84 #include <sys/proc.h> 85 #include <sys/resourcevar.h> 86 #include <sys/sysctl.h> 87 #include <sys/vmmeter.h> 88 #include <sys/vnode.h> 89 90 #include <vm/vm.h> 91 #include <vm/vm_param.h> 92 #include <vm/pmap.h> 93 #include <vm/vm_map.h> 94 #include <vm/vm_object.h> 95 #include <vm/vm_page.h> 96 #include <vm/vm_pageout.h> 97 #include <vm/vm_kern.h> 98 #include <vm/vm_pager.h> 99 #include <vm/vm_extern.h> 100 101 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */ 102 103 #define PFBAK 4 104 #define PFFOR 4 105 #define PAGEORDER_SIZE (PFBAK+PFFOR) 106 107 static int prefault_pageorder[] = { 108 -1 * PAGE_SIZE, 1 * PAGE_SIZE, 109 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 110 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 111 -4 * PAGE_SIZE, 4 * PAGE_SIZE 112 }; 113 114 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 115 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t); 116 117 #define VM_FAULT_READ_AHEAD 8 118 #define VM_FAULT_READ_BEHIND 7 119 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 120 121 struct faultstate { 122 vm_page_t m; 123 vm_object_t object; 124 vm_pindex_t pindex; 125 vm_page_t first_m; 126 vm_object_t first_object; 127 vm_pindex_t first_pindex; 128 vm_map_t map; 129 vm_map_entry_t entry; 130 int lookup_still_valid; 131 struct vnode *vp; 132 int vfslocked; 133 }; 134 135 static inline void 136 release_page(struct faultstate *fs) 137 { 138 139 vm_page_wakeup(fs->m); 140 vm_page_lock_queues(); 141 vm_page_deactivate(fs->m); 142 vm_page_unlock_queues(); 143 fs->m = NULL; 144 } 145 146 static inline void 147 unlock_map(struct faultstate *fs) 148 { 149 150 if (fs->lookup_still_valid) { 151 vm_map_lookup_done(fs->map, fs->entry); 152 fs->lookup_still_valid = FALSE; 153 } 154 } 155 156 static void 157 unlock_and_deallocate(struct faultstate *fs) 158 { 159 160 vm_object_pip_wakeup(fs->object); 161 VM_OBJECT_UNLOCK(fs->object); 162 if (fs->object != fs->first_object) { 163 VM_OBJECT_LOCK(fs->first_object); 164 vm_page_lock_queues(); 165 vm_page_free(fs->first_m); 166 vm_page_unlock_queues(); 167 vm_object_pip_wakeup(fs->first_object); 168 VM_OBJECT_UNLOCK(fs->first_object); 169 fs->first_m = NULL; 170 } 171 vm_object_deallocate(fs->first_object); 172 unlock_map(fs); 173 if (fs->vp != NULL) { 174 vput(fs->vp); 175 fs->vp = NULL; 176 } 177 VFS_UNLOCK_GIANT(fs->vfslocked); 178 fs->vfslocked = 0; 179 } 180 181 /* 182 * TRYPAGER - used by vm_fault to calculate whether the pager for the 183 * current object *might* contain the page. 184 * 185 * default objects are zero-fill, there is no real pager. 186 */ 187 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 188 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired)) 189 190 /* 191 * vm_fault: 192 * 193 * Handle a page fault occurring at the given address, 194 * requiring the given permissions, in the map specified. 195 * If successful, the page is inserted into the 196 * associated physical map. 197 * 198 * NOTE: the given address should be truncated to the 199 * proper page address. 200 * 201 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 202 * a standard error specifying why the fault is fatal is returned. 203 * 204 * 205 * The map in question must be referenced, and remains so. 206 * Caller may hold no locks. 207 */ 208 int 209 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 210 int fault_flags) 211 { 212 vm_prot_t prot; 213 int is_first_object_locked, result; 214 boolean_t are_queues_locked, growstack, wired; 215 int map_generation; 216 vm_object_t next_object; 217 vm_page_t marray[VM_FAULT_READ]; 218 int hardfault; 219 int faultcount, ahead, behind; 220 struct faultstate fs; 221 struct vnode *vp; 222 int locked, error; 223 224 hardfault = 0; 225 growstack = TRUE; 226 PCPU_INC(cnt.v_vm_faults); 227 fs.vp = NULL; 228 fs.vfslocked = 0; 229 faultcount = behind = 0; 230 231 RetryFault:; 232 233 /* 234 * Find the backing store object and offset into it to begin the 235 * search. 236 */ 237 fs.map = map; 238 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 239 &fs.first_object, &fs.first_pindex, &prot, &wired); 240 if (result != KERN_SUCCESS) { 241 if (growstack && result == KERN_INVALID_ADDRESS && 242 map != kernel_map) { 243 result = vm_map_growstack(curproc, vaddr); 244 if (result != KERN_SUCCESS) 245 return (KERN_FAILURE); 246 growstack = FALSE; 247 goto RetryFault; 248 } 249 return (result); 250 } 251 252 map_generation = fs.map->timestamp; 253 254 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 255 panic("vm_fault: fault on nofault entry, addr: %lx", 256 (u_long)vaddr); 257 } 258 259 /* 260 * Make a reference to this object to prevent its disposal while we 261 * are messing with it. Once we have the reference, the map is free 262 * to be diddled. Since objects reference their shadows (and copies), 263 * they will stay around as well. 264 * 265 * Bump the paging-in-progress count to prevent size changes (e.g. 266 * truncation operations) during I/O. This must be done after 267 * obtaining the vnode lock in order to avoid possible deadlocks. 268 */ 269 VM_OBJECT_LOCK(fs.first_object); 270 vm_object_reference_locked(fs.first_object); 271 vm_object_pip_add(fs.first_object, 1); 272 273 fs.lookup_still_valid = TRUE; 274 275 if (wired) 276 fault_type = prot | (fault_type & VM_PROT_COPY); 277 278 fs.first_m = NULL; 279 280 /* 281 * Search for the page at object/offset. 282 */ 283 fs.object = fs.first_object; 284 fs.pindex = fs.first_pindex; 285 while (TRUE) { 286 /* 287 * If the object is dead, we stop here 288 */ 289 if (fs.object->flags & OBJ_DEAD) { 290 unlock_and_deallocate(&fs); 291 return (KERN_PROTECTION_FAILURE); 292 } 293 294 /* 295 * See if page is resident 296 */ 297 fs.m = vm_page_lookup(fs.object, fs.pindex); 298 if (fs.m != NULL) { 299 /* 300 * check for page-based copy on write. 301 * We check fs.object == fs.first_object so 302 * as to ensure the legacy COW mechanism is 303 * used when the page in question is part of 304 * a shadow object. Otherwise, vm_page_cowfault() 305 * removes the page from the backing object, 306 * which is not what we want. 307 */ 308 vm_page_lock_queues(); 309 if ((fs.m->cow) && 310 (fault_type & VM_PROT_WRITE) && 311 (fs.object == fs.first_object)) { 312 vm_page_cowfault(fs.m); 313 vm_page_unlock_queues(); 314 unlock_and_deallocate(&fs); 315 goto RetryFault; 316 } 317 318 /* 319 * Wait/Retry if the page is busy. We have to do this 320 * if the page is busy via either VPO_BUSY or 321 * vm_page_t->busy because the vm_pager may be using 322 * vm_page_t->busy for pageouts ( and even pageins if 323 * it is the vnode pager ), and we could end up trying 324 * to pagein and pageout the same page simultaneously. 325 * 326 * We can theoretically allow the busy case on a read 327 * fault if the page is marked valid, but since such 328 * pages are typically already pmap'd, putting that 329 * special case in might be more effort then it is 330 * worth. We cannot under any circumstances mess 331 * around with a vm_page_t->busy page except, perhaps, 332 * to pmap it. 333 */ 334 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) { 335 vm_page_unlock_queues(); 336 VM_OBJECT_UNLOCK(fs.object); 337 if (fs.object != fs.first_object) { 338 VM_OBJECT_LOCK(fs.first_object); 339 vm_page_lock_queues(); 340 vm_page_free(fs.first_m); 341 vm_page_unlock_queues(); 342 vm_object_pip_wakeup(fs.first_object); 343 VM_OBJECT_UNLOCK(fs.first_object); 344 fs.first_m = NULL; 345 } 346 unlock_map(&fs); 347 VM_OBJECT_LOCK(fs.object); 348 if (fs.m == vm_page_lookup(fs.object, 349 fs.pindex)) { 350 vm_page_sleep_if_busy(fs.m, TRUE, 351 "vmpfw"); 352 } 353 vm_object_pip_wakeup(fs.object); 354 VM_OBJECT_UNLOCK(fs.object); 355 PCPU_INC(cnt.v_intrans); 356 vm_object_deallocate(fs.first_object); 357 goto RetryFault; 358 } 359 vm_pageq_remove(fs.m); 360 vm_page_unlock_queues(); 361 362 /* 363 * Mark page busy for other processes, and the 364 * pagedaemon. If it still isn't completely valid 365 * (readable), jump to readrest, else break-out ( we 366 * found the page ). 367 */ 368 vm_page_busy(fs.m); 369 if (fs.m->valid != VM_PAGE_BITS_ALL && 370 fs.m->object != kernel_object && fs.m->object != kmem_object) { 371 goto readrest; 372 } 373 374 break; 375 } 376 377 /* 378 * Page is not resident, If this is the search termination 379 * or the pager might contain the page, allocate a new page. 380 */ 381 if (TRYPAGER || fs.object == fs.first_object) { 382 if (fs.pindex >= fs.object->size) { 383 unlock_and_deallocate(&fs); 384 return (KERN_PROTECTION_FAILURE); 385 } 386 387 /* 388 * Allocate a new page for this object/offset pair. 389 */ 390 fs.m = NULL; 391 if (!vm_page_count_severe()) { 392 #if VM_NRESERVLEVEL > 0 393 if ((fs.object->flags & OBJ_COLORED) == 0) { 394 fs.object->flags |= OBJ_COLORED; 395 fs.object->pg_color = atop(vaddr) - 396 fs.pindex; 397 } 398 #endif 399 fs.m = vm_page_alloc(fs.object, fs.pindex, 400 (fs.object->type == OBJT_VNODE || 401 fs.object->backing_object != NULL) ? 402 VM_ALLOC_NORMAL : VM_ALLOC_ZERO); 403 } 404 if (fs.m == NULL) { 405 unlock_and_deallocate(&fs); 406 VM_WAITPFAULT; 407 goto RetryFault; 408 } else if (fs.m->valid == VM_PAGE_BITS_ALL) 409 break; 410 } 411 412 readrest: 413 /* 414 * We have found a valid page or we have allocated a new page. 415 * The page thus may not be valid or may not be entirely 416 * valid. 417 * 418 * Attempt to fault-in the page if there is a chance that the 419 * pager has it, and potentially fault in additional pages 420 * at the same time. 421 */ 422 if (TRYPAGER) { 423 int rv; 424 int reqpage = 0; 425 u_char behavior = vm_map_entry_behavior(fs.entry); 426 427 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 428 ahead = 0; 429 behind = 0; 430 } else { 431 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 432 if (behind > VM_FAULT_READ_BEHIND) 433 behind = VM_FAULT_READ_BEHIND; 434 435 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 436 if (ahead > VM_FAULT_READ_AHEAD) 437 ahead = VM_FAULT_READ_AHEAD; 438 } 439 is_first_object_locked = FALSE; 440 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 441 (behavior != MAP_ENTRY_BEHAV_RANDOM && 442 fs.pindex >= fs.entry->lastr && 443 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 444 (fs.first_object == fs.object || 445 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 446 fs.first_object->type != OBJT_DEVICE && 447 fs.first_object->type != OBJT_PHYS && 448 fs.first_object->type != OBJT_SG) { 449 vm_pindex_t firstpindex, tmppindex; 450 451 if (fs.first_pindex < 2 * VM_FAULT_READ) 452 firstpindex = 0; 453 else 454 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 455 456 are_queues_locked = FALSE; 457 /* 458 * note: partially valid pages cannot be 459 * included in the lookahead - NFS piecemeal 460 * writes will barf on it badly. 461 */ 462 for (tmppindex = fs.first_pindex - 1; 463 tmppindex >= firstpindex; 464 --tmppindex) { 465 vm_page_t mt; 466 467 mt = vm_page_lookup(fs.first_object, tmppindex); 468 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 469 break; 470 if (mt->busy || 471 (mt->oflags & VPO_BUSY)) 472 continue; 473 if (!are_queues_locked) { 474 are_queues_locked = TRUE; 475 vm_page_lock_queues(); 476 } 477 if (mt->hold_count || 478 mt->wire_count) 479 continue; 480 pmap_remove_all(mt); 481 if (mt->dirty) { 482 vm_page_deactivate(mt); 483 } else { 484 vm_page_cache(mt); 485 } 486 } 487 if (are_queues_locked) 488 vm_page_unlock_queues(); 489 ahead += behind; 490 behind = 0; 491 } 492 if (is_first_object_locked) 493 VM_OBJECT_UNLOCK(fs.first_object); 494 495 /* 496 * Call the pager to retrieve the data, if any, after 497 * releasing the lock on the map. We hold a ref on 498 * fs.object and the pages are VPO_BUSY'd. 499 */ 500 unlock_map(&fs); 501 502 vnode_lock: 503 if (fs.object->type == OBJT_VNODE) { 504 vp = fs.object->handle; 505 if (vp == fs.vp) 506 goto vnode_locked; 507 else if (fs.vp != NULL) { 508 vput(fs.vp); 509 fs.vp = NULL; 510 } 511 locked = VOP_ISLOCKED(vp); 512 513 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) { 514 fs.vfslocked = 1; 515 if (!mtx_trylock(&Giant)) { 516 VM_OBJECT_UNLOCK(fs.object); 517 mtx_lock(&Giant); 518 VM_OBJECT_LOCK(fs.object); 519 goto vnode_lock; 520 } 521 } 522 if (locked != LK_EXCLUSIVE) 523 locked = LK_SHARED; 524 /* Do not sleep for vnode lock while fs.m is busy */ 525 error = vget(vp, locked | LK_CANRECURSE | 526 LK_NOWAIT, curthread); 527 if (error != 0) { 528 int vfslocked; 529 530 vfslocked = fs.vfslocked; 531 fs.vfslocked = 0; /* Keep Giant */ 532 vhold(vp); 533 release_page(&fs); 534 unlock_and_deallocate(&fs); 535 error = vget(vp, locked | LK_RETRY | 536 LK_CANRECURSE, curthread); 537 vdrop(vp); 538 fs.vp = vp; 539 fs.vfslocked = vfslocked; 540 KASSERT(error == 0, 541 ("vm_fault: vget failed")); 542 goto RetryFault; 543 } 544 fs.vp = vp; 545 } 546 vnode_locked: 547 KASSERT(fs.vp == NULL || !fs.map->system_map, 548 ("vm_fault: vnode-backed object mapped by system map")); 549 550 /* 551 * now we find out if any other pages should be paged 552 * in at this time this routine checks to see if the 553 * pages surrounding this fault reside in the same 554 * object as the page for this fault. If they do, 555 * then they are faulted in also into the object. The 556 * array "marray" returned contains an array of 557 * vm_page_t structs where one of them is the 558 * vm_page_t passed to the routine. The reqpage 559 * return value is the index into the marray for the 560 * vm_page_t passed to the routine. 561 * 562 * fs.m plus the additional pages are VPO_BUSY'd. 563 */ 564 faultcount = vm_fault_additional_pages( 565 fs.m, behind, ahead, marray, &reqpage); 566 567 rv = faultcount ? 568 vm_pager_get_pages(fs.object, marray, faultcount, 569 reqpage) : VM_PAGER_FAIL; 570 571 if (rv == VM_PAGER_OK) { 572 /* 573 * Found the page. Leave it busy while we play 574 * with it. 575 */ 576 577 /* 578 * Relookup in case pager changed page. Pager 579 * is responsible for disposition of old page 580 * if moved. 581 */ 582 fs.m = vm_page_lookup(fs.object, fs.pindex); 583 if (!fs.m) { 584 unlock_and_deallocate(&fs); 585 goto RetryFault; 586 } 587 588 hardfault++; 589 break; /* break to PAGE HAS BEEN FOUND */ 590 } 591 /* 592 * Remove the bogus page (which does not exist at this 593 * object/offset); before doing so, we must get back 594 * our object lock to preserve our invariant. 595 * 596 * Also wake up any other process that may want to bring 597 * in this page. 598 * 599 * If this is the top-level object, we must leave the 600 * busy page to prevent another process from rushing 601 * past us, and inserting the page in that object at 602 * the same time that we are. 603 */ 604 if (rv == VM_PAGER_ERROR) 605 printf("vm_fault: pager read error, pid %d (%s)\n", 606 curproc->p_pid, curproc->p_comm); 607 /* 608 * Data outside the range of the pager or an I/O error 609 */ 610 /* 611 * XXX - the check for kernel_map is a kludge to work 612 * around having the machine panic on a kernel space 613 * fault w/ I/O error. 614 */ 615 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 616 (rv == VM_PAGER_BAD)) { 617 vm_page_lock_queues(); 618 vm_page_free(fs.m); 619 vm_page_unlock_queues(); 620 fs.m = NULL; 621 unlock_and_deallocate(&fs); 622 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 623 } 624 if (fs.object != fs.first_object) { 625 vm_page_lock_queues(); 626 vm_page_free(fs.m); 627 vm_page_unlock_queues(); 628 fs.m = NULL; 629 /* 630 * XXX - we cannot just fall out at this 631 * point, m has been freed and is invalid! 632 */ 633 } 634 } 635 636 /* 637 * We get here if the object has default pager (or unwiring) 638 * or the pager doesn't have the page. 639 */ 640 if (fs.object == fs.first_object) 641 fs.first_m = fs.m; 642 643 /* 644 * Move on to the next object. Lock the next object before 645 * unlocking the current one. 646 */ 647 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 648 next_object = fs.object->backing_object; 649 if (next_object == NULL) { 650 /* 651 * If there's no object left, fill the page in the top 652 * object with zeros. 653 */ 654 if (fs.object != fs.first_object) { 655 vm_object_pip_wakeup(fs.object); 656 VM_OBJECT_UNLOCK(fs.object); 657 658 fs.object = fs.first_object; 659 fs.pindex = fs.first_pindex; 660 fs.m = fs.first_m; 661 VM_OBJECT_LOCK(fs.object); 662 } 663 fs.first_m = NULL; 664 665 /* 666 * Zero the page if necessary and mark it valid. 667 */ 668 if ((fs.m->flags & PG_ZERO) == 0) { 669 pmap_zero_page(fs.m); 670 } else { 671 PCPU_INC(cnt.v_ozfod); 672 } 673 PCPU_INC(cnt.v_zfod); 674 fs.m->valid = VM_PAGE_BITS_ALL; 675 break; /* break to PAGE HAS BEEN FOUND */ 676 } else { 677 KASSERT(fs.object != next_object, 678 ("object loop %p", next_object)); 679 VM_OBJECT_LOCK(next_object); 680 vm_object_pip_add(next_object, 1); 681 if (fs.object != fs.first_object) 682 vm_object_pip_wakeup(fs.object); 683 VM_OBJECT_UNLOCK(fs.object); 684 fs.object = next_object; 685 } 686 } 687 688 KASSERT((fs.m->oflags & VPO_BUSY) != 0, 689 ("vm_fault: not busy after main loop")); 690 691 /* 692 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 693 * is held.] 694 */ 695 696 /* 697 * If the page is being written, but isn't already owned by the 698 * top-level object, we have to copy it into a new page owned by the 699 * top-level object. 700 */ 701 if (fs.object != fs.first_object) { 702 /* 703 * We only really need to copy if we want to write it. 704 */ 705 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 706 /* 707 * This allows pages to be virtually copied from a 708 * backing_object into the first_object, where the 709 * backing object has no other refs to it, and cannot 710 * gain any more refs. Instead of a bcopy, we just 711 * move the page from the backing object to the 712 * first object. Note that we must mark the page 713 * dirty in the first object so that it will go out 714 * to swap when needed. 715 */ 716 is_first_object_locked = FALSE; 717 if ( 718 /* 719 * Only one shadow object 720 */ 721 (fs.object->shadow_count == 1) && 722 /* 723 * No COW refs, except us 724 */ 725 (fs.object->ref_count == 1) && 726 /* 727 * No one else can look this object up 728 */ 729 (fs.object->handle == NULL) && 730 /* 731 * No other ways to look the object up 732 */ 733 ((fs.object->type == OBJT_DEFAULT) || 734 (fs.object->type == OBJT_SWAP)) && 735 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 736 /* 737 * We don't chase down the shadow chain 738 */ 739 fs.object == fs.first_object->backing_object) { 740 vm_page_lock_queues(); 741 /* 742 * get rid of the unnecessary page 743 */ 744 vm_page_free(fs.first_m); 745 /* 746 * grab the page and put it into the 747 * process'es object. The page is 748 * automatically made dirty. 749 */ 750 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 751 vm_page_unlock_queues(); 752 vm_page_busy(fs.m); 753 fs.first_m = fs.m; 754 fs.m = NULL; 755 PCPU_INC(cnt.v_cow_optim); 756 } else { 757 /* 758 * Oh, well, lets copy it. 759 */ 760 pmap_copy_page(fs.m, fs.first_m); 761 fs.first_m->valid = VM_PAGE_BITS_ALL; 762 if (wired && (fault_flags & 763 VM_FAULT_CHANGE_WIRING) == 0) { 764 vm_page_lock_queues(); 765 vm_page_wire(fs.first_m); 766 vm_page_unwire(fs.m, FALSE); 767 vm_page_unlock_queues(); 768 } 769 /* 770 * We no longer need the old page or object. 771 */ 772 release_page(&fs); 773 } 774 /* 775 * fs.object != fs.first_object due to above 776 * conditional 777 */ 778 vm_object_pip_wakeup(fs.object); 779 VM_OBJECT_UNLOCK(fs.object); 780 /* 781 * Only use the new page below... 782 */ 783 fs.object = fs.first_object; 784 fs.pindex = fs.first_pindex; 785 fs.m = fs.first_m; 786 if (!is_first_object_locked) 787 VM_OBJECT_LOCK(fs.object); 788 PCPU_INC(cnt.v_cow_faults); 789 } else { 790 prot &= ~VM_PROT_WRITE; 791 } 792 } 793 794 /* 795 * We must verify that the maps have not changed since our last 796 * lookup. 797 */ 798 if (!fs.lookup_still_valid) { 799 vm_object_t retry_object; 800 vm_pindex_t retry_pindex; 801 vm_prot_t retry_prot; 802 803 if (!vm_map_trylock_read(fs.map)) { 804 release_page(&fs); 805 unlock_and_deallocate(&fs); 806 goto RetryFault; 807 } 808 fs.lookup_still_valid = TRUE; 809 if (fs.map->timestamp != map_generation) { 810 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 811 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 812 813 /* 814 * If we don't need the page any longer, put it on the inactive 815 * list (the easiest thing to do here). If no one needs it, 816 * pageout will grab it eventually. 817 */ 818 if (result != KERN_SUCCESS) { 819 release_page(&fs); 820 unlock_and_deallocate(&fs); 821 822 /* 823 * If retry of map lookup would have blocked then 824 * retry fault from start. 825 */ 826 if (result == KERN_FAILURE) 827 goto RetryFault; 828 return (result); 829 } 830 if ((retry_object != fs.first_object) || 831 (retry_pindex != fs.first_pindex)) { 832 release_page(&fs); 833 unlock_and_deallocate(&fs); 834 goto RetryFault; 835 } 836 837 /* 838 * Check whether the protection has changed or the object has 839 * been copied while we left the map unlocked. Changing from 840 * read to write permission is OK - we leave the page 841 * write-protected, and catch the write fault. Changing from 842 * write to read permission means that we can't mark the page 843 * write-enabled after all. 844 */ 845 prot &= retry_prot; 846 } 847 } 848 /* 849 * If the page was filled by a pager, update the map entry's 850 * last read offset. Since the pager does not return the 851 * actual set of pages that it read, this update is based on 852 * the requested set. Typically, the requested and actual 853 * sets are the same. 854 * 855 * XXX The following assignment modifies the map 856 * without holding a write lock on it. 857 */ 858 if (hardfault) 859 fs.entry->lastr = fs.pindex + faultcount - behind; 860 861 if (prot & VM_PROT_WRITE) { 862 vm_object_set_writeable_dirty(fs.object); 863 864 /* 865 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 866 * if the page is already dirty to prevent data written with 867 * the expectation of being synced from not being synced. 868 * Likewise if this entry does not request NOSYNC then make 869 * sure the page isn't marked NOSYNC. Applications sharing 870 * data should use the same flags to avoid ping ponging. 871 */ 872 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 873 if (fs.m->dirty == 0) 874 fs.m->oflags |= VPO_NOSYNC; 875 } else { 876 fs.m->oflags &= ~VPO_NOSYNC; 877 } 878 879 /* 880 * If the fault is a write, we know that this page is being 881 * written NOW so dirty it explicitly to save on 882 * pmap_is_modified() calls later. 883 * 884 * Also tell the backing pager, if any, that it should remove 885 * any swap backing since the page is now dirty. 886 */ 887 if ((fault_type & VM_PROT_WRITE) != 0 && 888 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) { 889 vm_page_dirty(fs.m); 890 vm_pager_page_unswapped(fs.m); 891 } 892 } 893 894 /* 895 * Page had better still be busy 896 */ 897 KASSERT(fs.m->oflags & VPO_BUSY, 898 ("vm_fault: page %p not busy!", fs.m)); 899 /* 900 * Page must be completely valid or it is not fit to 901 * map into user space. vm_pager_get_pages() ensures this. 902 */ 903 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 904 ("vm_fault: page %p partially invalid", fs.m)); 905 VM_OBJECT_UNLOCK(fs.object); 906 907 /* 908 * Put this page into the physical map. We had to do the unlock above 909 * because pmap_enter() may sleep. We don't put the page 910 * back on the active queue until later so that the pageout daemon 911 * won't find it (yet). 912 */ 913 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired); 914 if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0) 915 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 916 VM_OBJECT_LOCK(fs.object); 917 vm_page_lock_queues(); 918 vm_page_flag_set(fs.m, PG_REFERENCED); 919 920 /* 921 * If the page is not wired down, then put it where the pageout daemon 922 * can find it. 923 */ 924 if (fault_flags & VM_FAULT_CHANGE_WIRING) { 925 if (wired) 926 vm_page_wire(fs.m); 927 else 928 vm_page_unwire(fs.m, 1); 929 } else { 930 vm_page_activate(fs.m); 931 } 932 vm_page_unlock_queues(); 933 vm_page_wakeup(fs.m); 934 935 /* 936 * Unlock everything, and return 937 */ 938 unlock_and_deallocate(&fs); 939 if (hardfault) 940 curthread->td_ru.ru_majflt++; 941 else 942 curthread->td_ru.ru_minflt++; 943 944 return (KERN_SUCCESS); 945 } 946 947 /* 948 * vm_fault_prefault provides a quick way of clustering 949 * pagefaults into a processes address space. It is a "cousin" 950 * of vm_map_pmap_enter, except it runs at page fault time instead 951 * of mmap time. 952 */ 953 static void 954 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 955 { 956 int i; 957 vm_offset_t addr, starta; 958 vm_pindex_t pindex; 959 vm_page_t m; 960 vm_object_t object; 961 962 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 963 return; 964 965 object = entry->object.vm_object; 966 967 starta = addra - PFBAK * PAGE_SIZE; 968 if (starta < entry->start) { 969 starta = entry->start; 970 } else if (starta > addra) { 971 starta = 0; 972 } 973 974 for (i = 0; i < PAGEORDER_SIZE; i++) { 975 vm_object_t backing_object, lobject; 976 977 addr = addra + prefault_pageorder[i]; 978 if (addr > addra + (PFFOR * PAGE_SIZE)) 979 addr = 0; 980 981 if (addr < starta || addr >= entry->end) 982 continue; 983 984 if (!pmap_is_prefaultable(pmap, addr)) 985 continue; 986 987 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 988 lobject = object; 989 VM_OBJECT_LOCK(lobject); 990 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 991 lobject->type == OBJT_DEFAULT && 992 (backing_object = lobject->backing_object) != NULL) { 993 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 994 0, ("vm_fault_prefault: unaligned object offset")); 995 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 996 VM_OBJECT_LOCK(backing_object); 997 VM_OBJECT_UNLOCK(lobject); 998 lobject = backing_object; 999 } 1000 /* 1001 * give-up when a page is not in memory 1002 */ 1003 if (m == NULL) { 1004 VM_OBJECT_UNLOCK(lobject); 1005 break; 1006 } 1007 if (m->valid == VM_PAGE_BITS_ALL && 1008 (m->flags & PG_FICTITIOUS) == 0) { 1009 vm_page_lock_queues(); 1010 pmap_enter_quick(pmap, addr, m, entry->protection); 1011 vm_page_unlock_queues(); 1012 } 1013 VM_OBJECT_UNLOCK(lobject); 1014 } 1015 } 1016 1017 /* 1018 * vm_fault_quick: 1019 * 1020 * Ensure that the requested virtual address, which may be in userland, 1021 * is valid. Fault-in the page if necessary. Return -1 on failure. 1022 */ 1023 int 1024 vm_fault_quick(caddr_t v, int prot) 1025 { 1026 int r; 1027 1028 if (prot & VM_PROT_WRITE) 1029 r = subyte(v, fubyte(v)); 1030 else 1031 r = fubyte(v); 1032 return(r); 1033 } 1034 1035 /* 1036 * vm_fault_wire: 1037 * 1038 * Wire down a range of virtual addresses in a map. 1039 */ 1040 int 1041 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1042 boolean_t fictitious) 1043 { 1044 vm_offset_t va; 1045 int rv; 1046 1047 /* 1048 * We simulate a fault to get the page and enter it in the physical 1049 * map. For user wiring, we only ask for read access on currently 1050 * read-only sections. 1051 */ 1052 for (va = start; va < end; va += PAGE_SIZE) { 1053 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING); 1054 if (rv) { 1055 if (va != start) 1056 vm_fault_unwire(map, start, va, fictitious); 1057 return (rv); 1058 } 1059 } 1060 return (KERN_SUCCESS); 1061 } 1062 1063 /* 1064 * vm_fault_unwire: 1065 * 1066 * Unwire a range of virtual addresses in a map. 1067 */ 1068 void 1069 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1070 boolean_t fictitious) 1071 { 1072 vm_paddr_t pa; 1073 vm_offset_t va; 1074 pmap_t pmap; 1075 1076 pmap = vm_map_pmap(map); 1077 1078 /* 1079 * Since the pages are wired down, we must be able to get their 1080 * mappings from the physical map system. 1081 */ 1082 for (va = start; va < end; va += PAGE_SIZE) { 1083 pa = pmap_extract(pmap, va); 1084 if (pa != 0) { 1085 pmap_change_wiring(pmap, va, FALSE); 1086 if (!fictitious) { 1087 vm_page_lock_queues(); 1088 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1089 vm_page_unlock_queues(); 1090 } 1091 } 1092 } 1093 } 1094 1095 /* 1096 * Routine: 1097 * vm_fault_copy_entry 1098 * Function: 1099 * Create new shadow object backing dst_entry with private copy of 1100 * all underlying pages. When src_entry is equal to dst_entry, 1101 * function implements COW for wired-down map entry. Otherwise, 1102 * it forks wired entry into dst_map. 1103 * 1104 * In/out conditions: 1105 * The source and destination maps must be locked for write. 1106 * The source map entry must be wired down (or be a sharing map 1107 * entry corresponding to a main map entry that is wired down). 1108 */ 1109 void 1110 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1111 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1112 vm_ooffset_t *fork_charge) 1113 { 1114 vm_object_t backing_object, dst_object, object, src_object; 1115 vm_pindex_t dst_pindex, pindex, src_pindex; 1116 vm_prot_t access, prot; 1117 vm_offset_t vaddr; 1118 vm_page_t dst_m; 1119 vm_page_t src_m; 1120 boolean_t src_readonly, upgrade; 1121 1122 #ifdef lint 1123 src_map++; 1124 #endif /* lint */ 1125 1126 upgrade = src_entry == dst_entry; 1127 1128 src_object = src_entry->object.vm_object; 1129 src_pindex = OFF_TO_IDX(src_entry->offset); 1130 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0; 1131 1132 /* 1133 * Create the top-level object for the destination entry. (Doesn't 1134 * actually shadow anything - we copy the pages directly.) 1135 */ 1136 dst_object = vm_object_allocate(OBJT_DEFAULT, 1137 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1138 #if VM_NRESERVLEVEL > 0 1139 dst_object->flags |= OBJ_COLORED; 1140 dst_object->pg_color = atop(dst_entry->start); 1141 #endif 1142 1143 VM_OBJECT_LOCK(dst_object); 1144 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1145 ("vm_fault_copy_entry: vm_object not NULL")); 1146 dst_entry->object.vm_object = dst_object; 1147 dst_entry->offset = 0; 1148 dst_object->charge = dst_entry->end - dst_entry->start; 1149 if (fork_charge != NULL) { 1150 KASSERT(dst_entry->uip == NULL, 1151 ("vm_fault_copy_entry: leaked swp charge")); 1152 dst_object->uip = curthread->td_ucred->cr_ruidinfo; 1153 uihold(dst_object->uip); 1154 *fork_charge += dst_object->charge; 1155 } else { 1156 dst_object->uip = dst_entry->uip; 1157 dst_entry->uip = NULL; 1158 } 1159 access = prot = dst_entry->protection; 1160 /* 1161 * If not an upgrade, then enter the mappings in the pmap as 1162 * read and/or execute accesses. Otherwise, enter them as 1163 * write accesses. 1164 * 1165 * A writeable large page mapping is only created if all of 1166 * the constituent small page mappings are modified. Marking 1167 * PTEs as modified on inception allows promotion to happen 1168 * without taking potentially large number of soft faults. 1169 */ 1170 if (!upgrade) 1171 access &= ~VM_PROT_WRITE; 1172 1173 /* 1174 * Loop through all of the pages in the entry's range, copying each 1175 * one from the source object (it should be there) to the destination 1176 * object. 1177 */ 1178 for (vaddr = dst_entry->start, dst_pindex = 0; 1179 vaddr < dst_entry->end; 1180 vaddr += PAGE_SIZE, dst_pindex++) { 1181 1182 /* 1183 * Allocate a page in the destination object. 1184 */ 1185 do { 1186 dst_m = vm_page_alloc(dst_object, dst_pindex, 1187 VM_ALLOC_NORMAL); 1188 if (dst_m == NULL) { 1189 VM_OBJECT_UNLOCK(dst_object); 1190 VM_WAIT; 1191 VM_OBJECT_LOCK(dst_object); 1192 } 1193 } while (dst_m == NULL); 1194 1195 /* 1196 * Find the page in the source object, and copy it in. 1197 * (Because the source is wired down, the page will be in 1198 * memory.) 1199 */ 1200 VM_OBJECT_LOCK(src_object); 1201 object = src_object; 1202 pindex = src_pindex + dst_pindex; 1203 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1204 src_readonly && 1205 (backing_object = object->backing_object) != NULL) { 1206 /* 1207 * Allow fallback to backing objects if we are reading. 1208 */ 1209 VM_OBJECT_LOCK(backing_object); 1210 pindex += OFF_TO_IDX(object->backing_object_offset); 1211 VM_OBJECT_UNLOCK(object); 1212 object = backing_object; 1213 } 1214 if (src_m == NULL) 1215 panic("vm_fault_copy_wired: page missing"); 1216 pmap_copy_page(src_m, dst_m); 1217 VM_OBJECT_UNLOCK(object); 1218 dst_m->valid = VM_PAGE_BITS_ALL; 1219 VM_OBJECT_UNLOCK(dst_object); 1220 1221 /* 1222 * Enter it in the pmap. If a wired, copy-on-write 1223 * mapping is being replaced by a write-enabled 1224 * mapping, then wire that new mapping. 1225 */ 1226 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade); 1227 1228 /* 1229 * Mark it no longer busy, and put it on the active list. 1230 */ 1231 VM_OBJECT_LOCK(dst_object); 1232 vm_page_lock_queues(); 1233 if (upgrade) { 1234 vm_page_unwire(src_m, 0); 1235 vm_page_wire(dst_m); 1236 } else 1237 vm_page_activate(dst_m); 1238 vm_page_unlock_queues(); 1239 vm_page_wakeup(dst_m); 1240 } 1241 VM_OBJECT_UNLOCK(dst_object); 1242 if (upgrade) { 1243 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1244 vm_object_deallocate(src_object); 1245 } 1246 } 1247 1248 1249 /* 1250 * This routine checks around the requested page for other pages that 1251 * might be able to be faulted in. This routine brackets the viable 1252 * pages for the pages to be paged in. 1253 * 1254 * Inputs: 1255 * m, rbehind, rahead 1256 * 1257 * Outputs: 1258 * marray (array of vm_page_t), reqpage (index of requested page) 1259 * 1260 * Return value: 1261 * number of pages in marray 1262 */ 1263 static int 1264 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1265 vm_page_t m; 1266 int rbehind; 1267 int rahead; 1268 vm_page_t *marray; 1269 int *reqpage; 1270 { 1271 int i,j; 1272 vm_object_t object; 1273 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1274 vm_page_t rtm; 1275 int cbehind, cahead; 1276 1277 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1278 1279 object = m->object; 1280 pindex = m->pindex; 1281 cbehind = cahead = 0; 1282 1283 /* 1284 * if the requested page is not available, then give up now 1285 */ 1286 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1287 return 0; 1288 } 1289 1290 if ((cbehind == 0) && (cahead == 0)) { 1291 *reqpage = 0; 1292 marray[0] = m; 1293 return 1; 1294 } 1295 1296 if (rahead > cahead) { 1297 rahead = cahead; 1298 } 1299 1300 if (rbehind > cbehind) { 1301 rbehind = cbehind; 1302 } 1303 1304 /* 1305 * scan backward for the read behind pages -- in memory 1306 */ 1307 if (pindex > 0) { 1308 if (rbehind > pindex) { 1309 rbehind = pindex; 1310 startpindex = 0; 1311 } else { 1312 startpindex = pindex - rbehind; 1313 } 1314 1315 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1316 rtm->pindex >= startpindex) 1317 startpindex = rtm->pindex + 1; 1318 1319 /* tpindex is unsigned; beware of numeric underflow. */ 1320 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex && 1321 tpindex < pindex; i++, tpindex--) { 1322 1323 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1324 VM_ALLOC_IFNOTCACHED); 1325 if (rtm == NULL) { 1326 /* 1327 * Shift the allocated pages to the 1328 * beginning of the array. 1329 */ 1330 for (j = 0; j < i; j++) { 1331 marray[j] = marray[j + tpindex + 1 - 1332 startpindex]; 1333 } 1334 break; 1335 } 1336 1337 marray[tpindex - startpindex] = rtm; 1338 } 1339 } else { 1340 startpindex = 0; 1341 i = 0; 1342 } 1343 1344 marray[i] = m; 1345 /* page offset of the required page */ 1346 *reqpage = i; 1347 1348 tpindex = pindex + 1; 1349 i++; 1350 1351 /* 1352 * scan forward for the read ahead pages 1353 */ 1354 endpindex = tpindex + rahead; 1355 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1356 endpindex = rtm->pindex; 1357 if (endpindex > object->size) 1358 endpindex = object->size; 1359 1360 for (; tpindex < endpindex; i++, tpindex++) { 1361 1362 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1363 VM_ALLOC_IFNOTCACHED); 1364 if (rtm == NULL) { 1365 break; 1366 } 1367 1368 marray[i] = rtm; 1369 } 1370 1371 /* return number of pages */ 1372 return i; 1373 } 1374