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