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, alloc_req; 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 * Unlocked read of the p_flag is harmless. At 391 * worst, the P_KILLED might be not observed 392 * there, and allocation can fail, causing 393 * restart and new reading of the p_flag. 394 */ 395 fs.m = NULL; 396 if (!vm_page_count_severe() || P_KILLED(curproc)) { 397 #if VM_NRESERVLEVEL > 0 398 if ((fs.object->flags & OBJ_COLORED) == 0) { 399 fs.object->flags |= OBJ_COLORED; 400 fs.object->pg_color = atop(vaddr) - 401 fs.pindex; 402 } 403 #endif 404 alloc_req = P_KILLED(curproc) ? 405 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 406 if (fs.object->type != OBJT_VNODE && 407 fs.object->backing_object == NULL) 408 alloc_req |= VM_ALLOC_ZERO; 409 fs.m = vm_page_alloc(fs.object, fs.pindex, 410 alloc_req); 411 } 412 if (fs.m == NULL) { 413 unlock_and_deallocate(&fs); 414 VM_WAITPFAULT; 415 goto RetryFault; 416 } else if (fs.m->valid == VM_PAGE_BITS_ALL) 417 break; 418 } 419 420 readrest: 421 /* 422 * We have found a valid page or we have allocated a new page. 423 * The page thus may not be valid or may not be entirely 424 * valid. 425 * 426 * Attempt to fault-in the page if there is a chance that the 427 * pager has it, and potentially fault in additional pages 428 * at the same time. 429 */ 430 if (TRYPAGER) { 431 int rv; 432 int reqpage = 0; 433 u_char behavior = vm_map_entry_behavior(fs.entry); 434 435 if (behavior == MAP_ENTRY_BEHAV_RANDOM || 436 P_KILLED(curproc)) { 437 ahead = 0; 438 behind = 0; 439 } else { 440 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 441 if (behind > VM_FAULT_READ_BEHIND) 442 behind = VM_FAULT_READ_BEHIND; 443 444 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 445 if (ahead > VM_FAULT_READ_AHEAD) 446 ahead = VM_FAULT_READ_AHEAD; 447 } 448 is_first_object_locked = FALSE; 449 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 450 (behavior != MAP_ENTRY_BEHAV_RANDOM && 451 fs.pindex >= fs.entry->lastr && 452 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) && 453 (fs.first_object == fs.object || 454 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) && 455 fs.first_object->type != OBJT_DEVICE && 456 fs.first_object->type != OBJT_PHYS && 457 fs.first_object->type != OBJT_SG) { 458 vm_pindex_t firstpindex, tmppindex; 459 460 if (fs.first_pindex < 2 * VM_FAULT_READ) 461 firstpindex = 0; 462 else 463 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 464 465 are_queues_locked = FALSE; 466 /* 467 * note: partially valid pages cannot be 468 * included in the lookahead - NFS piecemeal 469 * writes will barf on it badly. 470 */ 471 for (tmppindex = fs.first_pindex - 1; 472 tmppindex >= firstpindex; 473 --tmppindex) { 474 vm_page_t mt; 475 476 mt = vm_page_lookup(fs.first_object, tmppindex); 477 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 478 break; 479 if (mt->busy || 480 (mt->oflags & VPO_BUSY)) 481 continue; 482 if (!are_queues_locked) { 483 are_queues_locked = TRUE; 484 vm_page_lock_queues(); 485 } 486 if (mt->hold_count || 487 mt->wire_count) 488 continue; 489 pmap_remove_all(mt); 490 if (mt->dirty) { 491 vm_page_deactivate(mt); 492 } else { 493 vm_page_cache(mt); 494 } 495 } 496 if (are_queues_locked) 497 vm_page_unlock_queues(); 498 ahead += behind; 499 behind = 0; 500 } 501 if (is_first_object_locked) 502 VM_OBJECT_UNLOCK(fs.first_object); 503 504 /* 505 * Call the pager to retrieve the data, if any, after 506 * releasing the lock on the map. We hold a ref on 507 * fs.object and the pages are VPO_BUSY'd. 508 */ 509 unlock_map(&fs); 510 511 vnode_lock: 512 if (fs.object->type == OBJT_VNODE) { 513 vp = fs.object->handle; 514 if (vp == fs.vp) 515 goto vnode_locked; 516 else if (fs.vp != NULL) { 517 vput(fs.vp); 518 fs.vp = NULL; 519 } 520 locked = VOP_ISLOCKED(vp); 521 522 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) { 523 fs.vfslocked = 1; 524 if (!mtx_trylock(&Giant)) { 525 VM_OBJECT_UNLOCK(fs.object); 526 mtx_lock(&Giant); 527 VM_OBJECT_LOCK(fs.object); 528 goto vnode_lock; 529 } 530 } 531 if (locked != LK_EXCLUSIVE) 532 locked = LK_SHARED; 533 /* Do not sleep for vnode lock while fs.m is busy */ 534 error = vget(vp, locked | LK_CANRECURSE | 535 LK_NOWAIT, curthread); 536 if (error != 0) { 537 int vfslocked; 538 539 vfslocked = fs.vfslocked; 540 fs.vfslocked = 0; /* Keep Giant */ 541 vhold(vp); 542 release_page(&fs); 543 unlock_and_deallocate(&fs); 544 error = vget(vp, locked | LK_RETRY | 545 LK_CANRECURSE, curthread); 546 vdrop(vp); 547 fs.vp = vp; 548 fs.vfslocked = vfslocked; 549 KASSERT(error == 0, 550 ("vm_fault: vget failed")); 551 goto RetryFault; 552 } 553 fs.vp = vp; 554 } 555 vnode_locked: 556 KASSERT(fs.vp == NULL || !fs.map->system_map, 557 ("vm_fault: vnode-backed object mapped by system map")); 558 559 /* 560 * now we find out if any other pages should be paged 561 * in at this time this routine checks to see if the 562 * pages surrounding this fault reside in the same 563 * object as the page for this fault. If they do, 564 * then they are faulted in also into the object. The 565 * array "marray" returned contains an array of 566 * vm_page_t structs where one of them is the 567 * vm_page_t passed to the routine. The reqpage 568 * return value is the index into the marray for the 569 * vm_page_t passed to the routine. 570 * 571 * fs.m plus the additional pages are VPO_BUSY'd. 572 */ 573 faultcount = vm_fault_additional_pages( 574 fs.m, behind, ahead, marray, &reqpage); 575 576 rv = faultcount ? 577 vm_pager_get_pages(fs.object, marray, faultcount, 578 reqpage) : VM_PAGER_FAIL; 579 580 if (rv == VM_PAGER_OK) { 581 /* 582 * Found the page. Leave it busy while we play 583 * with it. 584 */ 585 586 /* 587 * Relookup in case pager changed page. Pager 588 * is responsible for disposition of old page 589 * if moved. 590 */ 591 fs.m = vm_page_lookup(fs.object, fs.pindex); 592 if (!fs.m) { 593 unlock_and_deallocate(&fs); 594 goto RetryFault; 595 } 596 597 hardfault++; 598 break; /* break to PAGE HAS BEEN FOUND */ 599 } 600 /* 601 * Remove the bogus page (which does not exist at this 602 * object/offset); before doing so, we must get back 603 * our object lock to preserve our invariant. 604 * 605 * Also wake up any other process that may want to bring 606 * in this page. 607 * 608 * If this is the top-level object, we must leave the 609 * busy page to prevent another process from rushing 610 * past us, and inserting the page in that object at 611 * the same time that we are. 612 */ 613 if (rv == VM_PAGER_ERROR) 614 printf("vm_fault: pager read error, pid %d (%s)\n", 615 curproc->p_pid, curproc->p_comm); 616 /* 617 * Data outside the range of the pager or an I/O error 618 */ 619 /* 620 * XXX - the check for kernel_map is a kludge to work 621 * around having the machine panic on a kernel space 622 * fault w/ I/O error. 623 */ 624 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 625 (rv == VM_PAGER_BAD)) { 626 vm_page_lock_queues(); 627 vm_page_free(fs.m); 628 vm_page_unlock_queues(); 629 fs.m = NULL; 630 unlock_and_deallocate(&fs); 631 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 632 } 633 if (fs.object != fs.first_object) { 634 vm_page_lock_queues(); 635 vm_page_free(fs.m); 636 vm_page_unlock_queues(); 637 fs.m = NULL; 638 /* 639 * XXX - we cannot just fall out at this 640 * point, m has been freed and is invalid! 641 */ 642 } 643 } 644 645 /* 646 * We get here if the object has default pager (or unwiring) 647 * or the pager doesn't have the page. 648 */ 649 if (fs.object == fs.first_object) 650 fs.first_m = fs.m; 651 652 /* 653 * Move on to the next object. Lock the next object before 654 * unlocking the current one. 655 */ 656 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 657 next_object = fs.object->backing_object; 658 if (next_object == NULL) { 659 /* 660 * If there's no object left, fill the page in the top 661 * object with zeros. 662 */ 663 if (fs.object != fs.first_object) { 664 vm_object_pip_wakeup(fs.object); 665 VM_OBJECT_UNLOCK(fs.object); 666 667 fs.object = fs.first_object; 668 fs.pindex = fs.first_pindex; 669 fs.m = fs.first_m; 670 VM_OBJECT_LOCK(fs.object); 671 } 672 fs.first_m = NULL; 673 674 /* 675 * Zero the page if necessary and mark it valid. 676 */ 677 if ((fs.m->flags & PG_ZERO) == 0) { 678 pmap_zero_page(fs.m); 679 } else { 680 PCPU_INC(cnt.v_ozfod); 681 } 682 PCPU_INC(cnt.v_zfod); 683 fs.m->valid = VM_PAGE_BITS_ALL; 684 break; /* break to PAGE HAS BEEN FOUND */ 685 } else { 686 KASSERT(fs.object != next_object, 687 ("object loop %p", next_object)); 688 VM_OBJECT_LOCK(next_object); 689 vm_object_pip_add(next_object, 1); 690 if (fs.object != fs.first_object) 691 vm_object_pip_wakeup(fs.object); 692 VM_OBJECT_UNLOCK(fs.object); 693 fs.object = next_object; 694 } 695 } 696 697 KASSERT((fs.m->oflags & VPO_BUSY) != 0, 698 ("vm_fault: not busy after main loop")); 699 700 /* 701 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 702 * is held.] 703 */ 704 705 /* 706 * If the page is being written, but isn't already owned by the 707 * top-level object, we have to copy it into a new page owned by the 708 * top-level object. 709 */ 710 if (fs.object != fs.first_object) { 711 /* 712 * We only really need to copy if we want to write it. 713 */ 714 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 715 /* 716 * This allows pages to be virtually copied from a 717 * backing_object into the first_object, where the 718 * backing object has no other refs to it, and cannot 719 * gain any more refs. Instead of a bcopy, we just 720 * move the page from the backing object to the 721 * first object. Note that we must mark the page 722 * dirty in the first object so that it will go out 723 * to swap when needed. 724 */ 725 is_first_object_locked = FALSE; 726 if ( 727 /* 728 * Only one shadow object 729 */ 730 (fs.object->shadow_count == 1) && 731 /* 732 * No COW refs, except us 733 */ 734 (fs.object->ref_count == 1) && 735 /* 736 * No one else can look this object up 737 */ 738 (fs.object->handle == NULL) && 739 /* 740 * No other ways to look the object up 741 */ 742 ((fs.object->type == OBJT_DEFAULT) || 743 (fs.object->type == OBJT_SWAP)) && 744 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) && 745 /* 746 * We don't chase down the shadow chain 747 */ 748 fs.object == fs.first_object->backing_object) { 749 vm_page_lock_queues(); 750 /* 751 * get rid of the unnecessary page 752 */ 753 vm_page_free(fs.first_m); 754 /* 755 * grab the page and put it into the 756 * process'es object. The page is 757 * automatically made dirty. 758 */ 759 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 760 vm_page_unlock_queues(); 761 vm_page_busy(fs.m); 762 fs.first_m = fs.m; 763 fs.m = NULL; 764 PCPU_INC(cnt.v_cow_optim); 765 } else { 766 /* 767 * Oh, well, lets copy it. 768 */ 769 pmap_copy_page(fs.m, fs.first_m); 770 fs.first_m->valid = VM_PAGE_BITS_ALL; 771 if (wired && (fault_flags & 772 VM_FAULT_CHANGE_WIRING) == 0) { 773 vm_page_lock_queues(); 774 vm_page_wire(fs.first_m); 775 vm_page_unwire(fs.m, FALSE); 776 vm_page_unlock_queues(); 777 } 778 /* 779 * We no longer need the old page or object. 780 */ 781 release_page(&fs); 782 } 783 /* 784 * fs.object != fs.first_object due to above 785 * conditional 786 */ 787 vm_object_pip_wakeup(fs.object); 788 VM_OBJECT_UNLOCK(fs.object); 789 /* 790 * Only use the new page below... 791 */ 792 fs.object = fs.first_object; 793 fs.pindex = fs.first_pindex; 794 fs.m = fs.first_m; 795 if (!is_first_object_locked) 796 VM_OBJECT_LOCK(fs.object); 797 PCPU_INC(cnt.v_cow_faults); 798 } else { 799 prot &= ~VM_PROT_WRITE; 800 } 801 } 802 803 /* 804 * We must verify that the maps have not changed since our last 805 * lookup. 806 */ 807 if (!fs.lookup_still_valid) { 808 vm_object_t retry_object; 809 vm_pindex_t retry_pindex; 810 vm_prot_t retry_prot; 811 812 if (!vm_map_trylock_read(fs.map)) { 813 release_page(&fs); 814 unlock_and_deallocate(&fs); 815 goto RetryFault; 816 } 817 fs.lookup_still_valid = TRUE; 818 if (fs.map->timestamp != map_generation) { 819 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 820 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 821 822 /* 823 * If we don't need the page any longer, put it on the inactive 824 * list (the easiest thing to do here). If no one needs it, 825 * pageout will grab it eventually. 826 */ 827 if (result != KERN_SUCCESS) { 828 release_page(&fs); 829 unlock_and_deallocate(&fs); 830 831 /* 832 * If retry of map lookup would have blocked then 833 * retry fault from start. 834 */ 835 if (result == KERN_FAILURE) 836 goto RetryFault; 837 return (result); 838 } 839 if ((retry_object != fs.first_object) || 840 (retry_pindex != fs.first_pindex)) { 841 release_page(&fs); 842 unlock_and_deallocate(&fs); 843 goto RetryFault; 844 } 845 846 /* 847 * Check whether the protection has changed or the object has 848 * been copied while we left the map unlocked. Changing from 849 * read to write permission is OK - we leave the page 850 * write-protected, and catch the write fault. Changing from 851 * write to read permission means that we can't mark the page 852 * write-enabled after all. 853 */ 854 prot &= retry_prot; 855 } 856 } 857 /* 858 * If the page was filled by a pager, update the map entry's 859 * last read offset. Since the pager does not return the 860 * actual set of pages that it read, this update is based on 861 * the requested set. Typically, the requested and actual 862 * sets are the same. 863 * 864 * XXX The following assignment modifies the map 865 * without holding a write lock on it. 866 */ 867 if (hardfault) 868 fs.entry->lastr = fs.pindex + faultcount - behind; 869 870 if (prot & VM_PROT_WRITE) { 871 vm_object_set_writeable_dirty(fs.object); 872 873 /* 874 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 875 * if the page is already dirty to prevent data written with 876 * the expectation of being synced from not being synced. 877 * Likewise if this entry does not request NOSYNC then make 878 * sure the page isn't marked NOSYNC. Applications sharing 879 * data should use the same flags to avoid ping ponging. 880 */ 881 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 882 if (fs.m->dirty == 0) 883 fs.m->oflags |= VPO_NOSYNC; 884 } else { 885 fs.m->oflags &= ~VPO_NOSYNC; 886 } 887 888 /* 889 * If the fault is a write, we know that this page is being 890 * written NOW so dirty it explicitly to save on 891 * pmap_is_modified() calls later. 892 * 893 * Also tell the backing pager, if any, that it should remove 894 * any swap backing since the page is now dirty. 895 */ 896 if ((fault_type & VM_PROT_WRITE) != 0 && 897 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) { 898 vm_page_dirty(fs.m); 899 vm_pager_page_unswapped(fs.m); 900 } 901 } 902 903 /* 904 * Page had better still be busy 905 */ 906 KASSERT(fs.m->oflags & VPO_BUSY, 907 ("vm_fault: page %p not busy!", fs.m)); 908 /* 909 * Page must be completely valid or it is not fit to 910 * map into user space. vm_pager_get_pages() ensures this. 911 */ 912 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 913 ("vm_fault: page %p partially invalid", fs.m)); 914 VM_OBJECT_UNLOCK(fs.object); 915 916 /* 917 * Put this page into the physical map. We had to do the unlock above 918 * because pmap_enter() may sleep. We don't put the page 919 * back on the active queue until later so that the pageout daemon 920 * won't find it (yet). 921 */ 922 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired); 923 if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0) 924 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 925 VM_OBJECT_LOCK(fs.object); 926 vm_page_lock_queues(); 927 vm_page_flag_set(fs.m, PG_REFERENCED); 928 929 /* 930 * If the page is not wired down, then put it where the pageout daemon 931 * can find it. 932 */ 933 if (fault_flags & VM_FAULT_CHANGE_WIRING) { 934 if (wired) 935 vm_page_wire(fs.m); 936 else 937 vm_page_unwire(fs.m, 1); 938 } else { 939 vm_page_activate(fs.m); 940 } 941 vm_page_unlock_queues(); 942 vm_page_wakeup(fs.m); 943 944 /* 945 * Unlock everything, and return 946 */ 947 unlock_and_deallocate(&fs); 948 if (hardfault) 949 curthread->td_ru.ru_majflt++; 950 else 951 curthread->td_ru.ru_minflt++; 952 953 return (KERN_SUCCESS); 954 } 955 956 /* 957 * vm_fault_prefault provides a quick way of clustering 958 * pagefaults into a processes address space. It is a "cousin" 959 * of vm_map_pmap_enter, except it runs at page fault time instead 960 * of mmap time. 961 */ 962 static void 963 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 964 { 965 int i; 966 vm_offset_t addr, starta; 967 vm_pindex_t pindex; 968 vm_page_t m; 969 vm_object_t object; 970 971 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 972 return; 973 974 object = entry->object.vm_object; 975 976 starta = addra - PFBAK * PAGE_SIZE; 977 if (starta < entry->start) { 978 starta = entry->start; 979 } else if (starta > addra) { 980 starta = 0; 981 } 982 983 for (i = 0; i < PAGEORDER_SIZE; i++) { 984 vm_object_t backing_object, lobject; 985 986 addr = addra + prefault_pageorder[i]; 987 if (addr > addra + (PFFOR * PAGE_SIZE)) 988 addr = 0; 989 990 if (addr < starta || addr >= entry->end) 991 continue; 992 993 if (!pmap_is_prefaultable(pmap, addr)) 994 continue; 995 996 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 997 lobject = object; 998 VM_OBJECT_LOCK(lobject); 999 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1000 lobject->type == OBJT_DEFAULT && 1001 (backing_object = lobject->backing_object) != NULL) { 1002 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1003 0, ("vm_fault_prefault: unaligned object offset")); 1004 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1005 VM_OBJECT_LOCK(backing_object); 1006 VM_OBJECT_UNLOCK(lobject); 1007 lobject = backing_object; 1008 } 1009 /* 1010 * give-up when a page is not in memory 1011 */ 1012 if (m == NULL) { 1013 VM_OBJECT_UNLOCK(lobject); 1014 break; 1015 } 1016 if (m->valid == VM_PAGE_BITS_ALL && 1017 (m->flags & PG_FICTITIOUS) == 0) { 1018 vm_page_lock_queues(); 1019 pmap_enter_quick(pmap, addr, m, entry->protection); 1020 vm_page_unlock_queues(); 1021 } 1022 VM_OBJECT_UNLOCK(lobject); 1023 } 1024 } 1025 1026 /* 1027 * vm_fault_quick: 1028 * 1029 * Ensure that the requested virtual address, which may be in userland, 1030 * is valid. Fault-in the page if necessary. Return -1 on failure. 1031 */ 1032 int 1033 vm_fault_quick(caddr_t v, int prot) 1034 { 1035 int r; 1036 1037 if (prot & VM_PROT_WRITE) 1038 r = subyte(v, fubyte(v)); 1039 else 1040 r = fubyte(v); 1041 return(r); 1042 } 1043 1044 /* 1045 * vm_fault_wire: 1046 * 1047 * Wire down a range of virtual addresses in a map. 1048 */ 1049 int 1050 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1051 boolean_t fictitious) 1052 { 1053 vm_offset_t va; 1054 int rv; 1055 1056 /* 1057 * We simulate a fault to get the page and enter it in the physical 1058 * map. For user wiring, we only ask for read access on currently 1059 * read-only sections. 1060 */ 1061 for (va = start; va < end; va += PAGE_SIZE) { 1062 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING); 1063 if (rv) { 1064 if (va != start) 1065 vm_fault_unwire(map, start, va, fictitious); 1066 return (rv); 1067 } 1068 } 1069 return (KERN_SUCCESS); 1070 } 1071 1072 /* 1073 * vm_fault_unwire: 1074 * 1075 * Unwire a range of virtual addresses in a map. 1076 */ 1077 void 1078 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1079 boolean_t fictitious) 1080 { 1081 vm_paddr_t pa; 1082 vm_offset_t va; 1083 pmap_t pmap; 1084 1085 pmap = vm_map_pmap(map); 1086 1087 /* 1088 * Since the pages are wired down, we must be able to get their 1089 * mappings from the physical map system. 1090 */ 1091 for (va = start; va < end; va += PAGE_SIZE) { 1092 pa = pmap_extract(pmap, va); 1093 if (pa != 0) { 1094 pmap_change_wiring(pmap, va, FALSE); 1095 if (!fictitious) { 1096 vm_page_lock_queues(); 1097 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1098 vm_page_unlock_queues(); 1099 } 1100 } 1101 } 1102 } 1103 1104 /* 1105 * Routine: 1106 * vm_fault_copy_entry 1107 * Function: 1108 * Create new shadow object backing dst_entry with private copy of 1109 * all underlying pages. When src_entry is equal to dst_entry, 1110 * function implements COW for wired-down map entry. Otherwise, 1111 * it forks wired entry into dst_map. 1112 * 1113 * In/out conditions: 1114 * The source and destination maps must be locked for write. 1115 * The source map entry must be wired down (or be a sharing map 1116 * entry corresponding to a main map entry that is wired down). 1117 */ 1118 void 1119 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1120 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1121 vm_ooffset_t *fork_charge) 1122 { 1123 vm_object_t backing_object, dst_object, object, src_object; 1124 vm_pindex_t dst_pindex, pindex, src_pindex; 1125 vm_prot_t access, prot; 1126 vm_offset_t vaddr; 1127 vm_page_t dst_m; 1128 vm_page_t src_m; 1129 boolean_t src_readonly, upgrade; 1130 1131 #ifdef lint 1132 src_map++; 1133 #endif /* lint */ 1134 1135 upgrade = src_entry == dst_entry; 1136 1137 src_object = src_entry->object.vm_object; 1138 src_pindex = OFF_TO_IDX(src_entry->offset); 1139 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0; 1140 1141 /* 1142 * Create the top-level object for the destination entry. (Doesn't 1143 * actually shadow anything - we copy the pages directly.) 1144 */ 1145 dst_object = vm_object_allocate(OBJT_DEFAULT, 1146 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1147 #if VM_NRESERVLEVEL > 0 1148 dst_object->flags |= OBJ_COLORED; 1149 dst_object->pg_color = atop(dst_entry->start); 1150 #endif 1151 1152 VM_OBJECT_LOCK(dst_object); 1153 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1154 ("vm_fault_copy_entry: vm_object not NULL")); 1155 dst_entry->object.vm_object = dst_object; 1156 dst_entry->offset = 0; 1157 dst_object->charge = dst_entry->end - dst_entry->start; 1158 if (fork_charge != NULL) { 1159 KASSERT(dst_entry->uip == NULL, 1160 ("vm_fault_copy_entry: leaked swp charge")); 1161 dst_object->uip = curthread->td_ucred->cr_ruidinfo; 1162 uihold(dst_object->uip); 1163 *fork_charge += dst_object->charge; 1164 } else { 1165 dst_object->uip = dst_entry->uip; 1166 dst_entry->uip = NULL; 1167 } 1168 access = prot = dst_entry->protection; 1169 /* 1170 * If not an upgrade, then enter the mappings in the pmap as 1171 * read and/or execute accesses. Otherwise, enter them as 1172 * write accesses. 1173 * 1174 * A writeable large page mapping is only created if all of 1175 * the constituent small page mappings are modified. Marking 1176 * PTEs as modified on inception allows promotion to happen 1177 * without taking potentially large number of soft faults. 1178 */ 1179 if (!upgrade) 1180 access &= ~VM_PROT_WRITE; 1181 1182 /* 1183 * Loop through all of the pages in the entry's range, copying each 1184 * one from the source object (it should be there) to the destination 1185 * object. 1186 */ 1187 for (vaddr = dst_entry->start, dst_pindex = 0; 1188 vaddr < dst_entry->end; 1189 vaddr += PAGE_SIZE, dst_pindex++) { 1190 1191 /* 1192 * Allocate a page in the destination object. 1193 */ 1194 do { 1195 dst_m = vm_page_alloc(dst_object, dst_pindex, 1196 VM_ALLOC_NORMAL); 1197 if (dst_m == NULL) { 1198 VM_OBJECT_UNLOCK(dst_object); 1199 VM_WAIT; 1200 VM_OBJECT_LOCK(dst_object); 1201 } 1202 } while (dst_m == NULL); 1203 1204 /* 1205 * Find the page in the source object, and copy it in. 1206 * (Because the source is wired down, the page will be in 1207 * memory.) 1208 */ 1209 VM_OBJECT_LOCK(src_object); 1210 object = src_object; 1211 pindex = src_pindex + dst_pindex; 1212 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1213 src_readonly && 1214 (backing_object = object->backing_object) != NULL) { 1215 /* 1216 * Allow fallback to backing objects if we are reading. 1217 */ 1218 VM_OBJECT_LOCK(backing_object); 1219 pindex += OFF_TO_IDX(object->backing_object_offset); 1220 VM_OBJECT_UNLOCK(object); 1221 object = backing_object; 1222 } 1223 if (src_m == NULL) 1224 panic("vm_fault_copy_wired: page missing"); 1225 pmap_copy_page(src_m, dst_m); 1226 VM_OBJECT_UNLOCK(object); 1227 dst_m->valid = VM_PAGE_BITS_ALL; 1228 VM_OBJECT_UNLOCK(dst_object); 1229 1230 /* 1231 * Enter it in the pmap. If a wired, copy-on-write 1232 * mapping is being replaced by a write-enabled 1233 * mapping, then wire that new mapping. 1234 */ 1235 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade); 1236 1237 /* 1238 * Mark it no longer busy, and put it on the active list. 1239 */ 1240 VM_OBJECT_LOCK(dst_object); 1241 vm_page_lock_queues(); 1242 if (upgrade) { 1243 vm_page_unwire(src_m, 0); 1244 vm_page_wire(dst_m); 1245 } else 1246 vm_page_activate(dst_m); 1247 vm_page_unlock_queues(); 1248 vm_page_wakeup(dst_m); 1249 } 1250 VM_OBJECT_UNLOCK(dst_object); 1251 if (upgrade) { 1252 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1253 vm_object_deallocate(src_object); 1254 } 1255 } 1256 1257 1258 /* 1259 * This routine checks around the requested page for other pages that 1260 * might be able to be faulted in. This routine brackets the viable 1261 * pages for the pages to be paged in. 1262 * 1263 * Inputs: 1264 * m, rbehind, rahead 1265 * 1266 * Outputs: 1267 * marray (array of vm_page_t), reqpage (index of requested page) 1268 * 1269 * Return value: 1270 * number of pages in marray 1271 */ 1272 static int 1273 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1274 vm_page_t m; 1275 int rbehind; 1276 int rahead; 1277 vm_page_t *marray; 1278 int *reqpage; 1279 { 1280 int i,j; 1281 vm_object_t object; 1282 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1283 vm_page_t rtm; 1284 int cbehind, cahead; 1285 1286 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1287 1288 object = m->object; 1289 pindex = m->pindex; 1290 cbehind = cahead = 0; 1291 1292 /* 1293 * if the requested page is not available, then give up now 1294 */ 1295 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1296 return 0; 1297 } 1298 1299 if ((cbehind == 0) && (cahead == 0)) { 1300 *reqpage = 0; 1301 marray[0] = m; 1302 return 1; 1303 } 1304 1305 if (rahead > cahead) { 1306 rahead = cahead; 1307 } 1308 1309 if (rbehind > cbehind) { 1310 rbehind = cbehind; 1311 } 1312 1313 /* 1314 * scan backward for the read behind pages -- in memory 1315 */ 1316 if (pindex > 0) { 1317 if (rbehind > pindex) { 1318 rbehind = pindex; 1319 startpindex = 0; 1320 } else { 1321 startpindex = pindex - rbehind; 1322 } 1323 1324 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1325 rtm->pindex >= startpindex) 1326 startpindex = rtm->pindex + 1; 1327 1328 /* tpindex is unsigned; beware of numeric underflow. */ 1329 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex && 1330 tpindex < pindex; i++, tpindex--) { 1331 1332 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1333 VM_ALLOC_IFNOTCACHED); 1334 if (rtm == NULL) { 1335 /* 1336 * Shift the allocated pages to the 1337 * beginning of the array. 1338 */ 1339 for (j = 0; j < i; j++) { 1340 marray[j] = marray[j + tpindex + 1 - 1341 startpindex]; 1342 } 1343 break; 1344 } 1345 1346 marray[tpindex - startpindex] = rtm; 1347 } 1348 } else { 1349 startpindex = 0; 1350 i = 0; 1351 } 1352 1353 marray[i] = m; 1354 /* page offset of the required page */ 1355 *reqpage = i; 1356 1357 tpindex = pindex + 1; 1358 i++; 1359 1360 /* 1361 * scan forward for the read ahead pages 1362 */ 1363 endpindex = tpindex + rahead; 1364 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1365 endpindex = rtm->pindex; 1366 if (endpindex > object->size) 1367 endpindex = object->size; 1368 1369 for (; tpindex < endpindex; i++, tpindex++) { 1370 1371 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1372 VM_ALLOC_IFNOTCACHED); 1373 if (rtm == NULL) { 1374 break; 1375 } 1376 1377 marray[i] = rtm; 1378 } 1379 1380 /* return number of pages */ 1381 return i; 1382 } 1383