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