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_ktrace.h" 78 #include "opt_vm.h" 79 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/kernel.h> 83 #include <sys/lock.h> 84 #include <sys/proc.h> 85 #include <sys/resourcevar.h> 86 #include <sys/rwlock.h> 87 #include <sys/sysctl.h> 88 #include <sys/vmmeter.h> 89 #include <sys/vnode.h> 90 #ifdef KTRACE 91 #include <sys/ktrace.h> 92 #endif 93 94 #include <vm/vm.h> 95 #include <vm/vm_param.h> 96 #include <vm/pmap.h> 97 #include <vm/vm_map.h> 98 #include <vm/vm_object.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_pageout.h> 101 #include <vm/vm_kern.h> 102 #include <vm/vm_pager.h> 103 #include <vm/vm_extern.h> 104 105 #define PFBAK 4 106 #define PFFOR 4 107 #define PAGEORDER_SIZE (PFBAK+PFFOR) 108 109 static int prefault_pageorder[] = { 110 -1 * PAGE_SIZE, 1 * PAGE_SIZE, 111 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 112 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 113 -4 * PAGE_SIZE, 4 * PAGE_SIZE 114 }; 115 116 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 117 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t); 118 119 #define VM_FAULT_READ_BEHIND 8 120 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 121 #define VM_FAULT_NINCR (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND) 122 #define VM_FAULT_SUM (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2) 123 #define VM_FAULT_CACHE_BEHIND (VM_FAULT_READ_BEHIND * VM_FAULT_SUM) 124 125 struct faultstate { 126 vm_page_t m; 127 vm_object_t object; 128 vm_pindex_t pindex; 129 vm_page_t first_m; 130 vm_object_t first_object; 131 vm_pindex_t first_pindex; 132 vm_map_t map; 133 vm_map_entry_t entry; 134 int lookup_still_valid; 135 struct vnode *vp; 136 }; 137 138 static void vm_fault_cache_behind(const struct faultstate *fs, int distance); 139 140 static inline void 141 release_page(struct faultstate *fs) 142 { 143 144 vm_page_wakeup(fs->m); 145 vm_page_lock(fs->m); 146 vm_page_deactivate(fs->m); 147 vm_page_unlock(fs->m); 148 fs->m = NULL; 149 } 150 151 static inline void 152 unlock_map(struct faultstate *fs) 153 { 154 155 if (fs->lookup_still_valid) { 156 vm_map_lookup_done(fs->map, fs->entry); 157 fs->lookup_still_valid = FALSE; 158 } 159 } 160 161 static void 162 unlock_and_deallocate(struct faultstate *fs) 163 { 164 165 vm_object_pip_wakeup(fs->object); 166 VM_OBJECT_WUNLOCK(fs->object); 167 if (fs->object != fs->first_object) { 168 VM_OBJECT_WLOCK(fs->first_object); 169 vm_page_lock(fs->first_m); 170 vm_page_free(fs->first_m); 171 vm_page_unlock(fs->first_m); 172 vm_object_pip_wakeup(fs->first_object); 173 VM_OBJECT_WUNLOCK(fs->first_object); 174 fs->first_m = NULL; 175 } 176 vm_object_deallocate(fs->first_object); 177 unlock_map(fs); 178 if (fs->vp != NULL) { 179 vput(fs->vp); 180 fs->vp = NULL; 181 } 182 } 183 184 /* 185 * TRYPAGER - used by vm_fault to calculate whether the pager for the 186 * current object *might* contain the page. 187 * 188 * default objects are zero-fill, there is no real pager. 189 */ 190 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 191 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired)) 192 193 /* 194 * vm_fault: 195 * 196 * Handle a page fault occurring at the given address, 197 * requiring the given permissions, in the map specified. 198 * If successful, the page is inserted into the 199 * associated physical map. 200 * 201 * NOTE: the given address should be truncated to the 202 * proper page address. 203 * 204 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 205 * a standard error specifying why the fault is fatal is returned. 206 * 207 * The map in question must be referenced, and remains so. 208 * Caller may hold no locks. 209 */ 210 int 211 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 212 int fault_flags) 213 { 214 struct thread *td; 215 int result; 216 217 td = curthread; 218 if ((td->td_pflags & TDP_NOFAULTING) != 0) 219 return (KERN_PROTECTION_FAILURE); 220 #ifdef KTRACE 221 if (map != kernel_map && KTRPOINT(td, KTR_FAULT)) 222 ktrfault(vaddr, fault_type); 223 #endif 224 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags, 225 NULL); 226 #ifdef KTRACE 227 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND)) 228 ktrfaultend(result); 229 #endif 230 return (result); 231 } 232 233 int 234 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 235 int fault_flags, vm_page_t *m_hold) 236 { 237 vm_prot_t prot; 238 long ahead, behind; 239 int alloc_req, era, faultcount, nera, reqpage, result; 240 boolean_t growstack, is_first_object_locked, wired; 241 int map_generation; 242 vm_object_t next_object; 243 vm_page_t marray[VM_FAULT_READ_MAX]; 244 int hardfault; 245 struct faultstate fs; 246 struct vnode *vp; 247 int locked, error; 248 249 hardfault = 0; 250 growstack = TRUE; 251 PCPU_INC(cnt.v_vm_faults); 252 fs.vp = NULL; 253 faultcount = reqpage = 0; 254 255 RetryFault:; 256 257 /* 258 * Find the backing store object and offset into it to begin the 259 * search. 260 */ 261 fs.map = map; 262 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 263 &fs.first_object, &fs.first_pindex, &prot, &wired); 264 if (result != KERN_SUCCESS) { 265 if (growstack && result == KERN_INVALID_ADDRESS && 266 map != kernel_map) { 267 result = vm_map_growstack(curproc, vaddr); 268 if (result != KERN_SUCCESS) 269 return (KERN_FAILURE); 270 growstack = FALSE; 271 goto RetryFault; 272 } 273 return (result); 274 } 275 276 map_generation = fs.map->timestamp; 277 278 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 279 panic("vm_fault: fault on nofault entry, addr: %lx", 280 (u_long)vaddr); 281 } 282 283 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 284 fs.entry->wiring_thread != curthread) { 285 vm_map_unlock_read(fs.map); 286 vm_map_lock(fs.map); 287 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 288 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 289 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 290 vm_map_unlock_and_wait(fs.map, 0); 291 } else 292 vm_map_unlock(fs.map); 293 goto RetryFault; 294 } 295 296 /* 297 * Make a reference to this object to prevent its disposal while we 298 * are messing with it. Once we have the reference, the map is free 299 * to be diddled. Since objects reference their shadows (and copies), 300 * they will stay around as well. 301 * 302 * Bump the paging-in-progress count to prevent size changes (e.g. 303 * truncation operations) during I/O. This must be done after 304 * obtaining the vnode lock in order to avoid possible deadlocks. 305 */ 306 VM_OBJECT_WLOCK(fs.first_object); 307 vm_object_reference_locked(fs.first_object); 308 vm_object_pip_add(fs.first_object, 1); 309 310 fs.lookup_still_valid = TRUE; 311 312 if (wired) 313 fault_type = prot | (fault_type & VM_PROT_COPY); 314 315 fs.first_m = NULL; 316 317 /* 318 * Search for the page at object/offset. 319 */ 320 fs.object = fs.first_object; 321 fs.pindex = fs.first_pindex; 322 while (TRUE) { 323 /* 324 * If the object is dead, we stop here 325 */ 326 if (fs.object->flags & OBJ_DEAD) { 327 unlock_and_deallocate(&fs); 328 return (KERN_PROTECTION_FAILURE); 329 } 330 331 /* 332 * See if page is resident 333 */ 334 fs.m = vm_page_lookup(fs.object, fs.pindex); 335 if (fs.m != NULL) { 336 /* 337 * check for page-based copy on write. 338 * We check fs.object == fs.first_object so 339 * as to ensure the legacy COW mechanism is 340 * used when the page in question is part of 341 * a shadow object. Otherwise, vm_page_cowfault() 342 * removes the page from the backing object, 343 * which is not what we want. 344 */ 345 vm_page_lock(fs.m); 346 if ((fs.m->cow) && 347 (fault_type & VM_PROT_WRITE) && 348 (fs.object == fs.first_object)) { 349 vm_page_cowfault(fs.m); 350 unlock_and_deallocate(&fs); 351 goto RetryFault; 352 } 353 354 /* 355 * Wait/Retry if the page is busy. We have to do this 356 * if the page is busy via either VPO_BUSY or 357 * vm_page_t->busy because the vm_pager may be using 358 * vm_page_t->busy for pageouts ( and even pageins if 359 * it is the vnode pager ), and we could end up trying 360 * to pagein and pageout the same page simultaneously. 361 * 362 * We can theoretically allow the busy case on a read 363 * fault if the page is marked valid, but since such 364 * pages are typically already pmap'd, putting that 365 * special case in might be more effort then it is 366 * worth. We cannot under any circumstances mess 367 * around with a vm_page_t->busy page except, perhaps, 368 * to pmap it. 369 */ 370 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) { 371 /* 372 * Reference the page before unlocking and 373 * sleeping so that the page daemon is less 374 * likely to reclaim it. 375 */ 376 vm_page_aflag_set(fs.m, PGA_REFERENCED); 377 vm_page_unlock(fs.m); 378 if (fs.object != fs.first_object) { 379 if (!VM_OBJECT_TRYWLOCK( 380 fs.first_object)) { 381 VM_OBJECT_WUNLOCK(fs.object); 382 VM_OBJECT_WLOCK(fs.first_object); 383 VM_OBJECT_WLOCK(fs.object); 384 } 385 vm_page_lock(fs.first_m); 386 vm_page_free(fs.first_m); 387 vm_page_unlock(fs.first_m); 388 vm_object_pip_wakeup(fs.first_object); 389 VM_OBJECT_WUNLOCK(fs.first_object); 390 fs.first_m = NULL; 391 } 392 unlock_map(&fs); 393 if (fs.m == vm_page_lookup(fs.object, 394 fs.pindex)) { 395 vm_page_sleep_if_busy(fs.m, TRUE, 396 "vmpfw"); 397 } 398 vm_object_pip_wakeup(fs.object); 399 VM_OBJECT_WUNLOCK(fs.object); 400 PCPU_INC(cnt.v_intrans); 401 vm_object_deallocate(fs.first_object); 402 goto RetryFault; 403 } 404 vm_page_remque(fs.m); 405 vm_page_unlock(fs.m); 406 407 /* 408 * Mark page busy for other processes, and the 409 * pagedaemon. If it still isn't completely valid 410 * (readable), jump to readrest, else break-out ( we 411 * found the page ). 412 */ 413 vm_page_busy(fs.m); 414 if (fs.m->valid != VM_PAGE_BITS_ALL) 415 goto readrest; 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 * Unlocked read of the p_flag is harmless. At 433 * worst, the P_KILLED might be not observed 434 * there, and allocation can fail, causing 435 * restart and new reading of the p_flag. 436 */ 437 fs.m = NULL; 438 if (!vm_page_count_severe() || P_KILLED(curproc)) { 439 #if VM_NRESERVLEVEL > 0 440 if ((fs.object->flags & OBJ_COLORED) == 0) { 441 fs.object->flags |= OBJ_COLORED; 442 fs.object->pg_color = atop(vaddr) - 443 fs.pindex; 444 } 445 #endif 446 alloc_req = P_KILLED(curproc) ? 447 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 448 if (fs.object->type != OBJT_VNODE && 449 fs.object->backing_object == NULL) 450 alloc_req |= VM_ALLOC_ZERO; 451 fs.m = vm_page_alloc(fs.object, fs.pindex, 452 alloc_req); 453 } 454 if (fs.m == NULL) { 455 unlock_and_deallocate(&fs); 456 VM_WAITPFAULT; 457 goto RetryFault; 458 } else if (fs.m->valid == VM_PAGE_BITS_ALL) 459 break; 460 } 461 462 readrest: 463 /* 464 * We have found a valid page or we have allocated a new page. 465 * The page thus may not be valid or may not be entirely 466 * valid. 467 * 468 * Attempt to fault-in the page if there is a chance that the 469 * pager has it, and potentially fault in additional pages 470 * at the same time. 471 */ 472 if (TRYPAGER) { 473 int rv; 474 u_char behavior = vm_map_entry_behavior(fs.entry); 475 476 if (behavior == MAP_ENTRY_BEHAV_RANDOM || 477 P_KILLED(curproc)) { 478 behind = 0; 479 ahead = 0; 480 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 481 behind = 0; 482 ahead = atop(fs.entry->end - vaddr) - 1; 483 if (ahead > VM_FAULT_READ_AHEAD_MAX) 484 ahead = VM_FAULT_READ_AHEAD_MAX; 485 if (fs.pindex == fs.entry->next_read) 486 vm_fault_cache_behind(&fs, 487 VM_FAULT_READ_MAX); 488 } else { 489 /* 490 * If this is a sequential page fault, then 491 * arithmetically increase the number of pages 492 * in the read-ahead window. Otherwise, reset 493 * the read-ahead window to its smallest size. 494 */ 495 behind = atop(vaddr - fs.entry->start); 496 if (behind > VM_FAULT_READ_BEHIND) 497 behind = VM_FAULT_READ_BEHIND; 498 ahead = atop(fs.entry->end - vaddr) - 1; 499 era = fs.entry->read_ahead; 500 if (fs.pindex == fs.entry->next_read) { 501 nera = era + behind; 502 if (nera > VM_FAULT_READ_AHEAD_MAX) 503 nera = VM_FAULT_READ_AHEAD_MAX; 504 behind = 0; 505 if (ahead > nera) 506 ahead = nera; 507 if (era == VM_FAULT_READ_AHEAD_MAX) 508 vm_fault_cache_behind(&fs, 509 VM_FAULT_CACHE_BEHIND); 510 } else if (ahead > VM_FAULT_READ_AHEAD_MIN) 511 ahead = VM_FAULT_READ_AHEAD_MIN; 512 if (era != ahead) 513 fs.entry->read_ahead = ahead; 514 } 515 516 /* 517 * Call the pager to retrieve the data, if any, after 518 * releasing the lock on the map. We hold a ref on 519 * fs.object and the pages are VPO_BUSY'd. 520 */ 521 unlock_map(&fs); 522 523 if (fs.object->type == OBJT_VNODE) { 524 vp = fs.object->handle; 525 if (vp == fs.vp) 526 goto vnode_locked; 527 else if (fs.vp != NULL) { 528 vput(fs.vp); 529 fs.vp = NULL; 530 } 531 locked = VOP_ISLOCKED(vp); 532 533 if (locked != LK_EXCLUSIVE) 534 locked = LK_SHARED; 535 /* Do not sleep for vnode lock while fs.m is busy */ 536 error = vget(vp, locked | LK_CANRECURSE | 537 LK_NOWAIT, curthread); 538 if (error != 0) { 539 vhold(vp); 540 release_page(&fs); 541 unlock_and_deallocate(&fs); 542 error = vget(vp, locked | LK_RETRY | 543 LK_CANRECURSE, curthread); 544 vdrop(vp); 545 fs.vp = vp; 546 KASSERT(error == 0, 547 ("vm_fault: vget failed")); 548 goto RetryFault; 549 } 550 fs.vp = vp; 551 } 552 vnode_locked: 553 KASSERT(fs.vp == NULL || !fs.map->system_map, 554 ("vm_fault: vnode-backed object mapped by system map")); 555 556 /* 557 * now we find out if any other pages should be paged 558 * in at this time this routine checks to see if the 559 * pages surrounding this fault reside in the same 560 * object as the page for this fault. If they do, 561 * then they are faulted in also into the object. The 562 * array "marray" returned contains an array of 563 * vm_page_t structs where one of them is the 564 * vm_page_t passed to the routine. The reqpage 565 * return value is the index into the marray for the 566 * vm_page_t passed to the routine. 567 * 568 * fs.m plus the additional pages are VPO_BUSY'd. 569 */ 570 faultcount = vm_fault_additional_pages( 571 fs.m, behind, ahead, marray, &reqpage); 572 573 rv = faultcount ? 574 vm_pager_get_pages(fs.object, marray, faultcount, 575 reqpage) : VM_PAGER_FAIL; 576 577 if (rv == VM_PAGER_OK) { 578 /* 579 * Found the page. Leave it busy while we play 580 * with it. 581 */ 582 583 /* 584 * Relookup in case pager changed page. Pager 585 * is responsible for disposition of old page 586 * if moved. 587 */ 588 fs.m = vm_page_lookup(fs.object, fs.pindex); 589 if (!fs.m) { 590 unlock_and_deallocate(&fs); 591 goto RetryFault; 592 } 593 594 hardfault++; 595 break; /* break to PAGE HAS BEEN FOUND */ 596 } 597 /* 598 * Remove the bogus page (which does not exist at this 599 * object/offset); before doing so, we must get back 600 * our object lock to preserve our invariant. 601 * 602 * Also wake up any other process that may want to bring 603 * in this page. 604 * 605 * If this is the top-level object, we must leave the 606 * busy page to prevent another process from rushing 607 * past us, and inserting the page in that object at 608 * the same time that we are. 609 */ 610 if (rv == VM_PAGER_ERROR) 611 printf("vm_fault: pager read error, pid %d (%s)\n", 612 curproc->p_pid, curproc->p_comm); 613 /* 614 * Data outside the range of the pager or an I/O error 615 */ 616 /* 617 * XXX - the check for kernel_map is a kludge to work 618 * around having the machine panic on a kernel space 619 * fault w/ I/O error. 620 */ 621 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 622 (rv == VM_PAGER_BAD)) { 623 vm_page_lock(fs.m); 624 vm_page_free(fs.m); 625 vm_page_unlock(fs.m); 626 fs.m = NULL; 627 unlock_and_deallocate(&fs); 628 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 629 } 630 if (fs.object != fs.first_object) { 631 vm_page_lock(fs.m); 632 vm_page_free(fs.m); 633 vm_page_unlock(fs.m); 634 fs.m = NULL; 635 /* 636 * XXX - we cannot just fall out at this 637 * point, m has been freed and is invalid! 638 */ 639 } 640 } 641 642 /* 643 * We get here if the object has default pager (or unwiring) 644 * or the pager doesn't have the page. 645 */ 646 if (fs.object == fs.first_object) 647 fs.first_m = fs.m; 648 649 /* 650 * Move on to the next object. Lock the next object before 651 * unlocking the current one. 652 */ 653 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 654 next_object = fs.object->backing_object; 655 if (next_object == NULL) { 656 /* 657 * If there's no object left, fill the page in the top 658 * object with zeros. 659 */ 660 if (fs.object != fs.first_object) { 661 vm_object_pip_wakeup(fs.object); 662 VM_OBJECT_WUNLOCK(fs.object); 663 664 fs.object = fs.first_object; 665 fs.pindex = fs.first_pindex; 666 fs.m = fs.first_m; 667 VM_OBJECT_WLOCK(fs.object); 668 } 669 fs.first_m = NULL; 670 671 /* 672 * Zero the page if necessary and mark it valid. 673 */ 674 if ((fs.m->flags & PG_ZERO) == 0) { 675 pmap_zero_page(fs.m); 676 } else { 677 PCPU_INC(cnt.v_ozfod); 678 } 679 PCPU_INC(cnt.v_zfod); 680 fs.m->valid = VM_PAGE_BITS_ALL; 681 break; /* break to PAGE HAS BEEN FOUND */ 682 } else { 683 KASSERT(fs.object != next_object, 684 ("object loop %p", next_object)); 685 VM_OBJECT_WLOCK(next_object); 686 vm_object_pip_add(next_object, 1); 687 if (fs.object != fs.first_object) 688 vm_object_pip_wakeup(fs.object); 689 VM_OBJECT_WUNLOCK(fs.object); 690 fs.object = next_object; 691 } 692 } 693 694 KASSERT((fs.m->oflags & VPO_BUSY) != 0, 695 ("vm_fault: not busy after main loop")); 696 697 /* 698 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 699 * is held.] 700 */ 701 702 /* 703 * If the page is being written, but isn't already owned by the 704 * top-level object, we have to copy it into a new page owned by the 705 * top-level object. 706 */ 707 if (fs.object != fs.first_object) { 708 /* 709 * We only really need to copy if we want to write it. 710 */ 711 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 712 /* 713 * This allows pages to be virtually copied from a 714 * backing_object into the first_object, where the 715 * backing object has no other refs to it, and cannot 716 * gain any more refs. Instead of a bcopy, we just 717 * move the page from the backing object to the 718 * first object. Note that we must mark the page 719 * dirty in the first object so that it will go out 720 * to swap when needed. 721 */ 722 is_first_object_locked = FALSE; 723 if ( 724 /* 725 * Only one shadow object 726 */ 727 (fs.object->shadow_count == 1) && 728 /* 729 * No COW refs, except us 730 */ 731 (fs.object->ref_count == 1) && 732 /* 733 * No one else can look this object up 734 */ 735 (fs.object->handle == NULL) && 736 /* 737 * No other ways to look the object up 738 */ 739 ((fs.object->type == OBJT_DEFAULT) || 740 (fs.object->type == OBJT_SWAP)) && 741 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 742 /* 743 * We don't chase down the shadow chain 744 */ 745 fs.object == fs.first_object->backing_object) { 746 /* 747 * get rid of the unnecessary page 748 */ 749 vm_page_lock(fs.first_m); 750 vm_page_free(fs.first_m); 751 vm_page_unlock(fs.first_m); 752 /* 753 * grab the page and put it into the 754 * process'es object. The page is 755 * automatically made dirty. 756 */ 757 vm_page_lock(fs.m); 758 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 759 vm_page_unlock(fs.m); 760 vm_page_busy(fs.m); 761 fs.first_m = fs.m; 762 fs.m = NULL; 763 PCPU_INC(cnt.v_cow_optim); 764 } else { 765 /* 766 * Oh, well, lets copy it. 767 */ 768 pmap_copy_page(fs.m, fs.first_m); 769 fs.first_m->valid = VM_PAGE_BITS_ALL; 770 if (wired && (fault_flags & 771 VM_FAULT_CHANGE_WIRING) == 0) { 772 vm_page_lock(fs.first_m); 773 vm_page_wire(fs.first_m); 774 vm_page_unlock(fs.first_m); 775 776 vm_page_lock(fs.m); 777 vm_page_unwire(fs.m, FALSE); 778 vm_page_unlock(fs.m); 779 } 780 /* 781 * We no longer need the old page or object. 782 */ 783 release_page(&fs); 784 } 785 /* 786 * fs.object != fs.first_object due to above 787 * conditional 788 */ 789 vm_object_pip_wakeup(fs.object); 790 VM_OBJECT_WUNLOCK(fs.object); 791 /* 792 * Only use the new page below... 793 */ 794 fs.object = fs.first_object; 795 fs.pindex = fs.first_pindex; 796 fs.m = fs.first_m; 797 if (!is_first_object_locked) 798 VM_OBJECT_WLOCK(fs.object); 799 PCPU_INC(cnt.v_cow_faults); 800 curthread->td_cow++; 801 } else { 802 prot &= ~VM_PROT_WRITE; 803 } 804 } 805 806 /* 807 * We must verify that the maps have not changed since our last 808 * lookup. 809 */ 810 if (!fs.lookup_still_valid) { 811 vm_object_t retry_object; 812 vm_pindex_t retry_pindex; 813 vm_prot_t retry_prot; 814 815 if (!vm_map_trylock_read(fs.map)) { 816 release_page(&fs); 817 unlock_and_deallocate(&fs); 818 goto RetryFault; 819 } 820 fs.lookup_still_valid = TRUE; 821 if (fs.map->timestamp != map_generation) { 822 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 823 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 824 825 /* 826 * If we don't need the page any longer, put it on the inactive 827 * list (the easiest thing to do here). If no one needs it, 828 * pageout will grab it eventually. 829 */ 830 if (result != KERN_SUCCESS) { 831 release_page(&fs); 832 unlock_and_deallocate(&fs); 833 834 /* 835 * If retry of map lookup would have blocked then 836 * retry fault from start. 837 */ 838 if (result == KERN_FAILURE) 839 goto RetryFault; 840 return (result); 841 } 842 if ((retry_object != fs.first_object) || 843 (retry_pindex != fs.first_pindex)) { 844 release_page(&fs); 845 unlock_and_deallocate(&fs); 846 goto RetryFault; 847 } 848 849 /* 850 * Check whether the protection has changed or the object has 851 * been copied while we left the map unlocked. Changing from 852 * read to write permission is OK - we leave the page 853 * write-protected, and catch the write fault. Changing from 854 * write to read permission means that we can't mark the page 855 * write-enabled after all. 856 */ 857 prot &= retry_prot; 858 } 859 } 860 /* 861 * If the page was filled by a pager, update the map entry's 862 * last read offset. Since the pager does not return the 863 * actual set of pages that it read, this update is based on 864 * the requested set. Typically, the requested and actual 865 * sets are the same. 866 * 867 * XXX The following assignment modifies the map 868 * without holding a write lock on it. 869 */ 870 if (hardfault) 871 fs.entry->next_read = fs.pindex + faultcount - reqpage; 872 873 if ((prot & VM_PROT_WRITE) != 0 || 874 (fault_flags & VM_FAULT_DIRTY) != 0) { 875 vm_object_set_writeable_dirty(fs.object); 876 877 /* 878 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 879 * if the page is already dirty to prevent data written with 880 * the expectation of being synced from not being synced. 881 * Likewise if this entry does not request NOSYNC then make 882 * sure the page isn't marked NOSYNC. Applications sharing 883 * data should use the same flags to avoid ping ponging. 884 */ 885 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 886 if (fs.m->dirty == 0) 887 fs.m->oflags |= VPO_NOSYNC; 888 } else { 889 fs.m->oflags &= ~VPO_NOSYNC; 890 } 891 892 /* 893 * If the fault is a write, we know that this page is being 894 * written NOW so dirty it explicitly to save on 895 * pmap_is_modified() calls later. 896 * 897 * Also tell the backing pager, if any, that it should remove 898 * any swap backing since the page is now dirty. 899 */ 900 if (((fault_type & VM_PROT_WRITE) != 0 && 901 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) || 902 (fault_flags & VM_FAULT_DIRTY) != 0) { 903 vm_page_dirty(fs.m); 904 vm_pager_page_unswapped(fs.m); 905 } 906 } 907 908 /* 909 * Page had better still be busy 910 */ 911 KASSERT(fs.m->oflags & VPO_BUSY, 912 ("vm_fault: page %p not busy!", fs.m)); 913 /* 914 * Page must be completely valid or it is not fit to 915 * map into user space. vm_pager_get_pages() ensures this. 916 */ 917 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 918 ("vm_fault: page %p partially invalid", fs.m)); 919 VM_OBJECT_WUNLOCK(fs.object); 920 921 /* 922 * Put this page into the physical map. We had to do the unlock above 923 * because pmap_enter() may sleep. We don't put the page 924 * back on the active queue until later so that the pageout daemon 925 * won't find it (yet). 926 */ 927 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired); 928 if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0) 929 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry); 930 VM_OBJECT_WLOCK(fs.object); 931 vm_page_lock(fs.m); 932 933 /* 934 * If the page is not wired down, then put it where the pageout daemon 935 * can find it. 936 */ 937 if (fault_flags & VM_FAULT_CHANGE_WIRING) { 938 if (wired) 939 vm_page_wire(fs.m); 940 else 941 vm_page_unwire(fs.m, 1); 942 } else 943 vm_page_activate(fs.m); 944 if (m_hold != NULL) { 945 *m_hold = fs.m; 946 vm_page_hold(fs.m); 947 } 948 vm_page_unlock(fs.m); 949 vm_page_wakeup(fs.m); 950 951 /* 952 * Unlock everything, and return 953 */ 954 unlock_and_deallocate(&fs); 955 if (hardfault) { 956 PCPU_INC(cnt.v_io_faults); 957 curthread->td_ru.ru_majflt++; 958 } else 959 curthread->td_ru.ru_minflt++; 960 961 return (KERN_SUCCESS); 962 } 963 964 /* 965 * Speed up the reclamation of up to "distance" pages that precede the 966 * faulting pindex within the first object of the shadow chain. 967 */ 968 static void 969 vm_fault_cache_behind(const struct faultstate *fs, int distance) 970 { 971 vm_object_t first_object, object; 972 vm_page_t m, m_prev; 973 vm_pindex_t pindex; 974 975 object = fs->object; 976 VM_OBJECT_ASSERT_WLOCKED(object); 977 first_object = fs->first_object; 978 if (first_object != object) { 979 if (!VM_OBJECT_TRYWLOCK(first_object)) { 980 VM_OBJECT_WUNLOCK(object); 981 VM_OBJECT_WLOCK(first_object); 982 VM_OBJECT_WLOCK(object); 983 } 984 } 985 /* Neither fictitious nor unmanaged pages can be cached. */ 986 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 987 if (fs->first_pindex < distance) 988 pindex = 0; 989 else 990 pindex = fs->first_pindex - distance; 991 if (pindex < OFF_TO_IDX(fs->entry->offset)) 992 pindex = OFF_TO_IDX(fs->entry->offset); 993 m = first_object != object ? fs->first_m : fs->m; 994 KASSERT((m->oflags & VPO_BUSY) != 0, 995 ("vm_fault_cache_behind: page %p is not busy", m)); 996 m_prev = vm_page_prev(m); 997 while ((m = m_prev) != NULL && m->pindex >= pindex && 998 m->valid == VM_PAGE_BITS_ALL) { 999 m_prev = vm_page_prev(m); 1000 if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0) 1001 continue; 1002 vm_page_lock(m); 1003 if (m->hold_count == 0 && m->wire_count == 0) { 1004 pmap_remove_all(m); 1005 vm_page_aflag_clear(m, PGA_REFERENCED); 1006 if (m->dirty != 0) 1007 vm_page_deactivate(m); 1008 else 1009 vm_page_cache(m); 1010 } 1011 vm_page_unlock(m); 1012 } 1013 } 1014 if (first_object != object) 1015 VM_OBJECT_WUNLOCK(first_object); 1016 } 1017 1018 /* 1019 * vm_fault_prefault provides a quick way of clustering 1020 * pagefaults into a processes address space. It is a "cousin" 1021 * of vm_map_pmap_enter, except it runs at page fault time instead 1022 * of mmap time. 1023 */ 1024 static void 1025 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry) 1026 { 1027 int i; 1028 vm_offset_t addr, starta; 1029 vm_pindex_t pindex; 1030 vm_page_t m; 1031 vm_object_t object; 1032 1033 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1034 return; 1035 1036 object = entry->object.vm_object; 1037 1038 starta = addra - PFBAK * PAGE_SIZE; 1039 if (starta < entry->start) { 1040 starta = entry->start; 1041 } else if (starta > addra) { 1042 starta = 0; 1043 } 1044 1045 for (i = 0; i < PAGEORDER_SIZE; i++) { 1046 vm_object_t backing_object, lobject; 1047 1048 addr = addra + prefault_pageorder[i]; 1049 if (addr > addra + (PFFOR * PAGE_SIZE)) 1050 addr = 0; 1051 1052 if (addr < starta || addr >= entry->end) 1053 continue; 1054 1055 if (!pmap_is_prefaultable(pmap, addr)) 1056 continue; 1057 1058 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1059 lobject = object; 1060 VM_OBJECT_RLOCK(lobject); 1061 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1062 lobject->type == OBJT_DEFAULT && 1063 (backing_object = lobject->backing_object) != NULL) { 1064 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1065 0, ("vm_fault_prefault: unaligned object offset")); 1066 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1067 VM_OBJECT_RLOCK(backing_object); 1068 VM_OBJECT_RUNLOCK(lobject); 1069 lobject = backing_object; 1070 } 1071 /* 1072 * give-up when a page is not in memory 1073 */ 1074 if (m == NULL) { 1075 VM_OBJECT_RUNLOCK(lobject); 1076 break; 1077 } 1078 if (m->valid == VM_PAGE_BITS_ALL && 1079 (m->flags & PG_FICTITIOUS) == 0) 1080 pmap_enter_quick(pmap, addr, m, entry->protection); 1081 VM_OBJECT_RUNLOCK(lobject); 1082 } 1083 } 1084 1085 /* 1086 * Hold each of the physical pages that are mapped by the specified range of 1087 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1088 * and allow the specified types of access, "prot". If all of the implied 1089 * pages are successfully held, then the number of held pages is returned 1090 * together with pointers to those pages in the array "ma". However, if any 1091 * of the pages cannot be held, -1 is returned. 1092 */ 1093 int 1094 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1095 vm_prot_t prot, vm_page_t *ma, int max_count) 1096 { 1097 vm_offset_t end, va; 1098 vm_page_t *mp; 1099 int count; 1100 boolean_t pmap_failed; 1101 1102 if (len == 0) 1103 return (0); 1104 end = round_page(addr + len); 1105 addr = trunc_page(addr); 1106 1107 /* 1108 * Check for illegal addresses. 1109 */ 1110 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1111 return (-1); 1112 1113 count = howmany(end - addr, PAGE_SIZE); 1114 if (count > max_count) 1115 panic("vm_fault_quick_hold_pages: count > max_count"); 1116 1117 /* 1118 * Most likely, the physical pages are resident in the pmap, so it is 1119 * faster to try pmap_extract_and_hold() first. 1120 */ 1121 pmap_failed = FALSE; 1122 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1123 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1124 if (*mp == NULL) 1125 pmap_failed = TRUE; 1126 else if ((prot & VM_PROT_WRITE) != 0 && 1127 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1128 /* 1129 * Explicitly dirty the physical page. Otherwise, the 1130 * caller's changes may go unnoticed because they are 1131 * performed through an unmanaged mapping or by a DMA 1132 * operation. 1133 * 1134 * The object lock is not held here. 1135 * See vm_page_clear_dirty_mask(). 1136 */ 1137 vm_page_dirty(*mp); 1138 } 1139 } 1140 if (pmap_failed) { 1141 /* 1142 * One or more pages could not be held by the pmap. Either no 1143 * page was mapped at the specified virtual address or that 1144 * mapping had insufficient permissions. Attempt to fault in 1145 * and hold these pages. 1146 */ 1147 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1148 if (*mp == NULL && vm_fault_hold(map, va, prot, 1149 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1150 goto error; 1151 } 1152 return (count); 1153 error: 1154 for (mp = ma; mp < ma + count; mp++) 1155 if (*mp != NULL) { 1156 vm_page_lock(*mp); 1157 vm_page_unhold(*mp); 1158 vm_page_unlock(*mp); 1159 } 1160 return (-1); 1161 } 1162 1163 /* 1164 * vm_fault_wire: 1165 * 1166 * Wire down a range of virtual addresses in a map. 1167 */ 1168 int 1169 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1170 boolean_t fictitious) 1171 { 1172 vm_offset_t va; 1173 int rv; 1174 1175 /* 1176 * We simulate a fault to get the page and enter it in the physical 1177 * map. For user wiring, we only ask for read access on currently 1178 * read-only sections. 1179 */ 1180 for (va = start; va < end; va += PAGE_SIZE) { 1181 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING); 1182 if (rv) { 1183 if (va != start) 1184 vm_fault_unwire(map, start, va, fictitious); 1185 return (rv); 1186 } 1187 } 1188 return (KERN_SUCCESS); 1189 } 1190 1191 /* 1192 * vm_fault_unwire: 1193 * 1194 * Unwire a range of virtual addresses in a map. 1195 */ 1196 void 1197 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 1198 boolean_t fictitious) 1199 { 1200 vm_paddr_t pa; 1201 vm_offset_t va; 1202 vm_page_t m; 1203 pmap_t pmap; 1204 1205 pmap = vm_map_pmap(map); 1206 1207 /* 1208 * Since the pages are wired down, we must be able to get their 1209 * mappings from the physical map system. 1210 */ 1211 for (va = start; va < end; va += PAGE_SIZE) { 1212 pa = pmap_extract(pmap, va); 1213 if (pa != 0) { 1214 pmap_change_wiring(pmap, va, FALSE); 1215 if (!fictitious) { 1216 m = PHYS_TO_VM_PAGE(pa); 1217 vm_page_lock(m); 1218 vm_page_unwire(m, TRUE); 1219 vm_page_unlock(m); 1220 } 1221 } 1222 } 1223 } 1224 1225 /* 1226 * Routine: 1227 * vm_fault_copy_entry 1228 * Function: 1229 * Create new shadow object backing dst_entry with private copy of 1230 * all underlying pages. When src_entry is equal to dst_entry, 1231 * function implements COW for wired-down map entry. Otherwise, 1232 * it forks wired entry into dst_map. 1233 * 1234 * In/out conditions: 1235 * The source and destination maps must be locked for write. 1236 * The source map entry must be wired down (or be a sharing map 1237 * entry corresponding to a main map entry that is wired down). 1238 */ 1239 void 1240 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1241 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1242 vm_ooffset_t *fork_charge) 1243 { 1244 vm_object_t backing_object, dst_object, object, src_object; 1245 vm_pindex_t dst_pindex, pindex, src_pindex; 1246 vm_prot_t access, prot; 1247 vm_offset_t vaddr; 1248 vm_page_t dst_m; 1249 vm_page_t src_m; 1250 boolean_t src_readonly, upgrade; 1251 1252 #ifdef lint 1253 src_map++; 1254 #endif /* lint */ 1255 1256 upgrade = src_entry == dst_entry; 1257 1258 src_object = src_entry->object.vm_object; 1259 src_pindex = OFF_TO_IDX(src_entry->offset); 1260 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0; 1261 1262 /* 1263 * Create the top-level object for the destination entry. (Doesn't 1264 * actually shadow anything - we copy the pages directly.) 1265 */ 1266 dst_object = vm_object_allocate(OBJT_DEFAULT, 1267 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1268 #if VM_NRESERVLEVEL > 0 1269 dst_object->flags |= OBJ_COLORED; 1270 dst_object->pg_color = atop(dst_entry->start); 1271 #endif 1272 1273 VM_OBJECT_WLOCK(dst_object); 1274 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1275 ("vm_fault_copy_entry: vm_object not NULL")); 1276 dst_entry->object.vm_object = dst_object; 1277 dst_entry->offset = 0; 1278 dst_object->charge = dst_entry->end - dst_entry->start; 1279 if (fork_charge != NULL) { 1280 KASSERT(dst_entry->cred == NULL, 1281 ("vm_fault_copy_entry: leaked swp charge")); 1282 dst_object->cred = curthread->td_ucred; 1283 crhold(dst_object->cred); 1284 *fork_charge += dst_object->charge; 1285 } else { 1286 dst_object->cred = dst_entry->cred; 1287 dst_entry->cred = NULL; 1288 } 1289 access = prot = dst_entry->protection; 1290 /* 1291 * If not an upgrade, then enter the mappings in the pmap as 1292 * read and/or execute accesses. Otherwise, enter them as 1293 * write accesses. 1294 * 1295 * A writeable large page mapping is only created if all of 1296 * the constituent small page mappings are modified. Marking 1297 * PTEs as modified on inception allows promotion to happen 1298 * without taking potentially large number of soft faults. 1299 */ 1300 if (!upgrade) 1301 access &= ~VM_PROT_WRITE; 1302 1303 /* 1304 * Loop through all of the virtual pages within the entry's 1305 * range, copying each page from the source object to the 1306 * destination object. Since the source is wired, those pages 1307 * must exist. In contrast, the destination is pageable. 1308 * Since the destination object does share any backing storage 1309 * with the source object, all of its pages must be dirtied, 1310 * regardless of whether they can be written. 1311 */ 1312 for (vaddr = dst_entry->start, dst_pindex = 0; 1313 vaddr < dst_entry->end; 1314 vaddr += PAGE_SIZE, dst_pindex++) { 1315 1316 /* 1317 * Allocate a page in the destination object. 1318 */ 1319 do { 1320 dst_m = vm_page_alloc(dst_object, dst_pindex, 1321 VM_ALLOC_NORMAL); 1322 if (dst_m == NULL) { 1323 VM_OBJECT_WUNLOCK(dst_object); 1324 VM_WAIT; 1325 VM_OBJECT_WLOCK(dst_object); 1326 } 1327 } while (dst_m == NULL); 1328 1329 /* 1330 * Find the page in the source object, and copy it in. 1331 * (Because the source is wired down, the page will be in 1332 * memory.) 1333 */ 1334 VM_OBJECT_RLOCK(src_object); 1335 object = src_object; 1336 pindex = src_pindex + dst_pindex; 1337 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1338 src_readonly && 1339 (backing_object = object->backing_object) != NULL) { 1340 /* 1341 * Allow fallback to backing objects if we are reading. 1342 */ 1343 VM_OBJECT_RLOCK(backing_object); 1344 pindex += OFF_TO_IDX(object->backing_object_offset); 1345 VM_OBJECT_RUNLOCK(object); 1346 object = backing_object; 1347 } 1348 if (src_m == NULL) 1349 panic("vm_fault_copy_wired: page missing"); 1350 pmap_copy_page(src_m, dst_m); 1351 VM_OBJECT_RUNLOCK(object); 1352 dst_m->valid = VM_PAGE_BITS_ALL; 1353 dst_m->dirty = VM_PAGE_BITS_ALL; 1354 VM_OBJECT_WUNLOCK(dst_object); 1355 1356 /* 1357 * Enter it in the pmap. If a wired, copy-on-write 1358 * mapping is being replaced by a write-enabled 1359 * mapping, then wire that new mapping. 1360 */ 1361 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade); 1362 1363 /* 1364 * Mark it no longer busy, and put it on the active list. 1365 */ 1366 VM_OBJECT_WLOCK(dst_object); 1367 1368 if (upgrade) { 1369 vm_page_lock(src_m); 1370 vm_page_unwire(src_m, 0); 1371 vm_page_unlock(src_m); 1372 1373 vm_page_lock(dst_m); 1374 vm_page_wire(dst_m); 1375 vm_page_unlock(dst_m); 1376 } else { 1377 vm_page_lock(dst_m); 1378 vm_page_activate(dst_m); 1379 vm_page_unlock(dst_m); 1380 } 1381 vm_page_wakeup(dst_m); 1382 } 1383 VM_OBJECT_WUNLOCK(dst_object); 1384 if (upgrade) { 1385 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1386 vm_object_deallocate(src_object); 1387 } 1388 } 1389 1390 1391 /* 1392 * This routine checks around the requested page for other pages that 1393 * might be able to be faulted in. This routine brackets the viable 1394 * pages for the pages to be paged in. 1395 * 1396 * Inputs: 1397 * m, rbehind, rahead 1398 * 1399 * Outputs: 1400 * marray (array of vm_page_t), reqpage (index of requested page) 1401 * 1402 * Return value: 1403 * number of pages in marray 1404 */ 1405 static int 1406 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1407 vm_page_t m; 1408 int rbehind; 1409 int rahead; 1410 vm_page_t *marray; 1411 int *reqpage; 1412 { 1413 int i,j; 1414 vm_object_t object; 1415 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1416 vm_page_t rtm; 1417 int cbehind, cahead; 1418 1419 VM_OBJECT_ASSERT_WLOCKED(m->object); 1420 1421 object = m->object; 1422 pindex = m->pindex; 1423 cbehind = cahead = 0; 1424 1425 /* 1426 * if the requested page is not available, then give up now 1427 */ 1428 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1429 return 0; 1430 } 1431 1432 if ((cbehind == 0) && (cahead == 0)) { 1433 *reqpage = 0; 1434 marray[0] = m; 1435 return 1; 1436 } 1437 1438 if (rahead > cahead) { 1439 rahead = cahead; 1440 } 1441 1442 if (rbehind > cbehind) { 1443 rbehind = cbehind; 1444 } 1445 1446 /* 1447 * scan backward for the read behind pages -- in memory 1448 */ 1449 if (pindex > 0) { 1450 if (rbehind > pindex) { 1451 rbehind = pindex; 1452 startpindex = 0; 1453 } else { 1454 startpindex = pindex - rbehind; 1455 } 1456 1457 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1458 rtm->pindex >= startpindex) 1459 startpindex = rtm->pindex + 1; 1460 1461 /* tpindex is unsigned; beware of numeric underflow. */ 1462 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex && 1463 tpindex < pindex; i++, tpindex--) { 1464 1465 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1466 VM_ALLOC_IFNOTCACHED); 1467 if (rtm == NULL) { 1468 /* 1469 * Shift the allocated pages to the 1470 * beginning of the array. 1471 */ 1472 for (j = 0; j < i; j++) { 1473 marray[j] = marray[j + tpindex + 1 - 1474 startpindex]; 1475 } 1476 break; 1477 } 1478 1479 marray[tpindex - startpindex] = rtm; 1480 } 1481 } else { 1482 startpindex = 0; 1483 i = 0; 1484 } 1485 1486 marray[i] = m; 1487 /* page offset of the required page */ 1488 *reqpage = i; 1489 1490 tpindex = pindex + 1; 1491 i++; 1492 1493 /* 1494 * scan forward for the read ahead pages 1495 */ 1496 endpindex = tpindex + rahead; 1497 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1498 endpindex = rtm->pindex; 1499 if (endpindex > object->size) 1500 endpindex = object->size; 1501 1502 for (; tpindex < endpindex; i++, tpindex++) { 1503 1504 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1505 VM_ALLOC_IFNOTCACHED); 1506 if (rtm == NULL) { 1507 break; 1508 } 1509 1510 marray[i] = rtm; 1511 } 1512 1513 /* return number of pages */ 1514 return i; 1515 } 1516 1517 /* 1518 * Block entry into the machine-independent layer's page fault handler by 1519 * the calling thread. Subsequent calls to vm_fault() by that thread will 1520 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1521 * spurious page faults. 1522 */ 1523 int 1524 vm_fault_disable_pagefaults(void) 1525 { 1526 1527 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1528 } 1529 1530 void 1531 vm_fault_enable_pagefaults(int save) 1532 { 1533 1534 curthread_pflags_restore(save); 1535 } 1536