vm_fault.c revision 286362
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: stable/10/sys/vm/vm_fault.c 286362 2015-08-06 08:51:15Z kib $"); 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#include <vm/vm_reserv.h> 105 106#define PFBAK 4 107#define PFFOR 4 108 109static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 110 111#define VM_FAULT_READ_BEHIND 8 112#define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 113#define VM_FAULT_NINCR (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND) 114#define VM_FAULT_SUM (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2) 115#define VM_FAULT_CACHE_BEHIND (VM_FAULT_READ_BEHIND * VM_FAULT_SUM) 116 117struct faultstate { 118 vm_page_t m; 119 vm_object_t object; 120 vm_pindex_t pindex; 121 vm_page_t first_m; 122 vm_object_t first_object; 123 vm_pindex_t first_pindex; 124 vm_map_t map; 125 vm_map_entry_t entry; 126 int lookup_still_valid; 127 struct vnode *vp; 128}; 129 130static void vm_fault_cache_behind(const struct faultstate *fs, int distance); 131static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 132 int faultcount, int reqpage); 133 134static inline void 135release_page(struct faultstate *fs) 136{ 137 138 vm_page_xunbusy(fs->m); 139 vm_page_lock(fs->m); 140 vm_page_deactivate(fs->m); 141 vm_page_unlock(fs->m); 142 fs->m = NULL; 143} 144 145static inline void 146unlock_map(struct faultstate *fs) 147{ 148 149 if (fs->lookup_still_valid) { 150 vm_map_lookup_done(fs->map, fs->entry); 151 fs->lookup_still_valid = FALSE; 152 } 153} 154 155static void 156unlock_and_deallocate(struct faultstate *fs) 157{ 158 159 vm_object_pip_wakeup(fs->object); 160 VM_OBJECT_WUNLOCK(fs->object); 161 if (fs->object != fs->first_object) { 162 VM_OBJECT_WLOCK(fs->first_object); 163 vm_page_lock(fs->first_m); 164 vm_page_free(fs->first_m); 165 vm_page_unlock(fs->first_m); 166 vm_object_pip_wakeup(fs->first_object); 167 VM_OBJECT_WUNLOCK(fs->first_object); 168 fs->first_m = NULL; 169 } 170 vm_object_deallocate(fs->first_object); 171 unlock_map(fs); 172 if (fs->vp != NULL) { 173 vput(fs->vp); 174 fs->vp = NULL; 175 } 176} 177 178static void 179vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, 180 vm_prot_t fault_type, int fault_flags, boolean_t set_wd) 181{ 182 boolean_t need_dirty; 183 184 if (((prot & VM_PROT_WRITE) == 0 && 185 (fault_flags & VM_FAULT_DIRTY) == 0) || 186 (m->oflags & VPO_UNMANAGED) != 0) 187 return; 188 189 VM_OBJECT_ASSERT_LOCKED(m->object); 190 191 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && 192 (fault_flags & VM_FAULT_WIRE) == 0) || 193 (fault_flags & VM_FAULT_DIRTY) != 0; 194 195 if (set_wd) 196 vm_object_set_writeable_dirty(m->object); 197 else 198 /* 199 * If two callers of vm_fault_dirty() with set_wd == 200 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC 201 * flag set, other with flag clear, race, it is 202 * possible for the no-NOSYNC thread to see m->dirty 203 * != 0 and not clear VPO_NOSYNC. Take vm_page lock 204 * around manipulation of VPO_NOSYNC and 205 * vm_page_dirty() call, to avoid the race and keep 206 * m->oflags consistent. 207 */ 208 vm_page_lock(m); 209 210 /* 211 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 212 * if the page is already dirty to prevent data written with 213 * the expectation of being synced from not being synced. 214 * Likewise if this entry does not request NOSYNC then make 215 * sure the page isn't marked NOSYNC. Applications sharing 216 * data should use the same flags to avoid ping ponging. 217 */ 218 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { 219 if (m->dirty == 0) { 220 m->oflags |= VPO_NOSYNC; 221 } 222 } else { 223 m->oflags &= ~VPO_NOSYNC; 224 } 225 226 /* 227 * If the fault is a write, we know that this page is being 228 * written NOW so dirty it explicitly to save on 229 * pmap_is_modified() calls later. 230 * 231 * Also tell the backing pager, if any, that it should remove 232 * any swap backing since the page is now dirty. 233 */ 234 if (need_dirty) 235 vm_page_dirty(m); 236 if (!set_wd) 237 vm_page_unlock(m); 238 if (need_dirty) 239 vm_pager_page_unswapped(m); 240} 241 242/* 243 * vm_fault: 244 * 245 * Handle a page fault occurring at the given address, 246 * requiring the given permissions, in the map specified. 247 * If successful, the page is inserted into the 248 * associated physical map. 249 * 250 * NOTE: the given address should be truncated to the 251 * proper page address. 252 * 253 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 254 * a standard error specifying why the fault is fatal is returned. 255 * 256 * The map in question must be referenced, and remains so. 257 * Caller may hold no locks. 258 */ 259int 260vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 261 int fault_flags) 262{ 263 struct thread *td; 264 int result; 265 266 td = curthread; 267 if ((td->td_pflags & TDP_NOFAULTING) != 0) 268 return (KERN_PROTECTION_FAILURE); 269#ifdef KTRACE 270 if (map != kernel_map && KTRPOINT(td, KTR_FAULT)) 271 ktrfault(vaddr, fault_type); 272#endif 273 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags, 274 NULL); 275#ifdef KTRACE 276 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND)) 277 ktrfaultend(result); 278#endif 279 return (result); 280} 281 282int 283vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 284 int fault_flags, vm_page_t *m_hold) 285{ 286 vm_prot_t prot; 287 long ahead, behind; 288 int alloc_req, era, faultcount, nera, reqpage, result; 289 boolean_t growstack, is_first_object_locked, wired; 290 int map_generation; 291 vm_object_t next_object; 292 vm_page_t marray[VM_FAULT_READ_MAX]; 293 int hardfault; 294 struct faultstate fs; 295 struct vnode *vp; 296 vm_page_t m; 297 int locked, error; 298 299 hardfault = 0; 300 growstack = TRUE; 301 PCPU_INC(cnt.v_vm_faults); 302 fs.vp = NULL; 303 faultcount = reqpage = 0; 304 305RetryFault:; 306 307 /* 308 * Find the backing store object and offset into it to begin the 309 * search. 310 */ 311 fs.map = map; 312 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 313 &fs.first_object, &fs.first_pindex, &prot, &wired); 314 if (result != KERN_SUCCESS) { 315 if (growstack && result == KERN_INVALID_ADDRESS && 316 map != kernel_map) { 317 result = vm_map_growstack(curproc, vaddr); 318 if (result != KERN_SUCCESS) 319 return (KERN_FAILURE); 320 growstack = FALSE; 321 goto RetryFault; 322 } 323 return (result); 324 } 325 326 map_generation = fs.map->timestamp; 327 328 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 329 panic("vm_fault: fault on nofault entry, addr: %lx", 330 (u_long)vaddr); 331 } 332 333 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 334 fs.entry->wiring_thread != curthread) { 335 vm_map_unlock_read(fs.map); 336 vm_map_lock(fs.map); 337 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 338 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 339 if (fs.vp != NULL) { 340 vput(fs.vp); 341 fs.vp = NULL; 342 } 343 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 344 vm_map_unlock_and_wait(fs.map, 0); 345 } else 346 vm_map_unlock(fs.map); 347 goto RetryFault; 348 } 349 350 if (wired) 351 fault_type = prot | (fault_type & VM_PROT_COPY); 352 else 353 KASSERT((fault_flags & VM_FAULT_WIRE) == 0, 354 ("!wired && VM_FAULT_WIRE")); 355 356 if (fs.vp == NULL /* avoid locked vnode leak */ && 357 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && 358 /* avoid calling vm_object_set_writeable_dirty() */ 359 ((prot & VM_PROT_WRITE) == 0 || 360 (fs.first_object->type != OBJT_VNODE && 361 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 362 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { 363 VM_OBJECT_RLOCK(fs.first_object); 364 if ((prot & VM_PROT_WRITE) != 0 && 365 (fs.first_object->type == OBJT_VNODE || 366 (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) && 367 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0) 368 goto fast_failed; 369 m = vm_page_lookup(fs.first_object, fs.first_pindex); 370 /* A busy page can be mapped for read|execute access. */ 371 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && 372 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) 373 goto fast_failed; 374 result = pmap_enter(fs.map->pmap, vaddr, m, prot, 375 fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 376 0), 0); 377 if (result != KERN_SUCCESS) 378 goto fast_failed; 379 if (m_hold != NULL) { 380 *m_hold = m; 381 vm_page_lock(m); 382 vm_page_hold(m); 383 vm_page_unlock(m); 384 } 385 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags, 386 FALSE); 387 VM_OBJECT_RUNLOCK(fs.first_object); 388 if (!wired) 389 vm_fault_prefault(&fs, vaddr, 0, 0); 390 vm_map_lookup_done(fs.map, fs.entry); 391 curthread->td_ru.ru_minflt++; 392 return (KERN_SUCCESS); 393fast_failed: 394 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 395 VM_OBJECT_RUNLOCK(fs.first_object); 396 VM_OBJECT_WLOCK(fs.first_object); 397 } 398 } else { 399 VM_OBJECT_WLOCK(fs.first_object); 400 } 401 402 /* 403 * Make a reference to this object to prevent its disposal while we 404 * are messing with it. Once we have the reference, the map is free 405 * to be diddled. Since objects reference their shadows (and copies), 406 * they will stay around as well. 407 * 408 * Bump the paging-in-progress count to prevent size changes (e.g. 409 * truncation operations) during I/O. This must be done after 410 * obtaining the vnode lock in order to avoid possible deadlocks. 411 */ 412 vm_object_reference_locked(fs.first_object); 413 vm_object_pip_add(fs.first_object, 1); 414 415 fs.lookup_still_valid = TRUE; 416 417 fs.first_m = NULL; 418 419 /* 420 * Search for the page at object/offset. 421 */ 422 fs.object = fs.first_object; 423 fs.pindex = fs.first_pindex; 424 while (TRUE) { 425 /* 426 * If the object is dead, we stop here 427 */ 428 if (fs.object->flags & OBJ_DEAD) { 429 unlock_and_deallocate(&fs); 430 return (KERN_PROTECTION_FAILURE); 431 } 432 433 /* 434 * See if page is resident 435 */ 436 fs.m = vm_page_lookup(fs.object, fs.pindex); 437 if (fs.m != NULL) { 438 /* 439 * Wait/Retry if the page is busy. We have to do this 440 * if the page is either exclusive or shared busy 441 * because the vm_pager may be using read busy for 442 * pageouts (and even pageins if it is the vnode 443 * pager), and we could end up trying to pagein and 444 * pageout the same page simultaneously. 445 * 446 * We can theoretically allow the busy case on a read 447 * fault if the page is marked valid, but since such 448 * pages are typically already pmap'd, putting that 449 * special case in might be more effort then it is 450 * worth. We cannot under any circumstances mess 451 * around with a shared busied page except, perhaps, 452 * to pmap it. 453 */ 454 if (vm_page_busied(fs.m)) { 455 /* 456 * Reference the page before unlocking and 457 * sleeping so that the page daemon is less 458 * likely to reclaim it. 459 */ 460 vm_page_aflag_set(fs.m, PGA_REFERENCED); 461 if (fs.object != fs.first_object) { 462 if (!VM_OBJECT_TRYWLOCK( 463 fs.first_object)) { 464 VM_OBJECT_WUNLOCK(fs.object); 465 VM_OBJECT_WLOCK(fs.first_object); 466 VM_OBJECT_WLOCK(fs.object); 467 } 468 vm_page_lock(fs.first_m); 469 vm_page_free(fs.first_m); 470 vm_page_unlock(fs.first_m); 471 vm_object_pip_wakeup(fs.first_object); 472 VM_OBJECT_WUNLOCK(fs.first_object); 473 fs.first_m = NULL; 474 } 475 unlock_map(&fs); 476 if (fs.m == vm_page_lookup(fs.object, 477 fs.pindex)) { 478 vm_page_sleep_if_busy(fs.m, "vmpfw"); 479 } 480 vm_object_pip_wakeup(fs.object); 481 VM_OBJECT_WUNLOCK(fs.object); 482 PCPU_INC(cnt.v_intrans); 483 vm_object_deallocate(fs.first_object); 484 goto RetryFault; 485 } 486 vm_page_lock(fs.m); 487 vm_page_remque(fs.m); 488 vm_page_unlock(fs.m); 489 490 /* 491 * Mark page busy for other processes, and the 492 * pagedaemon. If it still isn't completely valid 493 * (readable), jump to readrest, else break-out ( we 494 * found the page ). 495 */ 496 vm_page_xbusy(fs.m); 497 if (fs.m->valid != VM_PAGE_BITS_ALL) 498 goto readrest; 499 break; 500 } 501 502 /* 503 * Page is not resident. If this is the search termination 504 * or the pager might contain the page, allocate a new page. 505 * Default objects are zero-fill, there is no real pager. 506 */ 507 if (fs.object->type != OBJT_DEFAULT || 508 fs.object == fs.first_object) { 509 if (fs.pindex >= fs.object->size) { 510 unlock_and_deallocate(&fs); 511 return (KERN_PROTECTION_FAILURE); 512 } 513 514 /* 515 * Allocate a new page for this object/offset pair. 516 * 517 * Unlocked read of the p_flag is harmless. At 518 * worst, the P_KILLED might be not observed 519 * there, and allocation can fail, causing 520 * restart and new reading of the p_flag. 521 */ 522 fs.m = NULL; 523 if (!vm_page_count_severe() || P_KILLED(curproc)) { 524#if VM_NRESERVLEVEL > 0 525 if ((fs.object->flags & OBJ_COLORED) == 0) { 526 fs.object->flags |= OBJ_COLORED; 527 fs.object->pg_color = atop(vaddr) - 528 fs.pindex; 529 } 530#endif 531 alloc_req = P_KILLED(curproc) ? 532 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 533 if (fs.object->type != OBJT_VNODE && 534 fs.object->backing_object == NULL) 535 alloc_req |= VM_ALLOC_ZERO; 536 fs.m = vm_page_alloc(fs.object, fs.pindex, 537 alloc_req); 538 } 539 if (fs.m == NULL) { 540 unlock_and_deallocate(&fs); 541 VM_WAITPFAULT; 542 goto RetryFault; 543 } else if (fs.m->valid == VM_PAGE_BITS_ALL) 544 break; 545 } 546 547readrest: 548 /* 549 * We have found a valid page or we have allocated a new page. 550 * The page thus may not be valid or may not be entirely 551 * valid. 552 * 553 * Attempt to fault-in the page if there is a chance that the 554 * pager has it, and potentially fault in additional pages 555 * at the same time. For default objects simply provide 556 * zero-filled pages. 557 */ 558 if (fs.object->type != OBJT_DEFAULT) { 559 int rv; 560 u_char behavior = vm_map_entry_behavior(fs.entry); 561 562 if (behavior == MAP_ENTRY_BEHAV_RANDOM || 563 P_KILLED(curproc)) { 564 behind = 0; 565 ahead = 0; 566 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 567 behind = 0; 568 ahead = atop(fs.entry->end - vaddr) - 1; 569 if (ahead > VM_FAULT_READ_AHEAD_MAX) 570 ahead = VM_FAULT_READ_AHEAD_MAX; 571 if (fs.pindex == fs.entry->next_read) 572 vm_fault_cache_behind(&fs, 573 VM_FAULT_READ_MAX); 574 } else { 575 /* 576 * If this is a sequential page fault, then 577 * arithmetically increase the number of pages 578 * in the read-ahead window. Otherwise, reset 579 * the read-ahead window to its smallest size. 580 */ 581 behind = atop(vaddr - fs.entry->start); 582 if (behind > VM_FAULT_READ_BEHIND) 583 behind = VM_FAULT_READ_BEHIND; 584 ahead = atop(fs.entry->end - vaddr) - 1; 585 era = fs.entry->read_ahead; 586 if (fs.pindex == fs.entry->next_read) { 587 nera = era + behind; 588 if (nera > VM_FAULT_READ_AHEAD_MAX) 589 nera = VM_FAULT_READ_AHEAD_MAX; 590 behind = 0; 591 if (ahead > nera) 592 ahead = nera; 593 if (era == VM_FAULT_READ_AHEAD_MAX) 594 vm_fault_cache_behind(&fs, 595 VM_FAULT_CACHE_BEHIND); 596 } else if (ahead > VM_FAULT_READ_AHEAD_MIN) 597 ahead = VM_FAULT_READ_AHEAD_MIN; 598 if (era != ahead) 599 fs.entry->read_ahead = ahead; 600 } 601 602 /* 603 * Call the pager to retrieve the data, if any, after 604 * releasing the lock on the map. We hold a ref on 605 * fs.object and the pages are exclusive busied. 606 */ 607 unlock_map(&fs); 608 609 if (fs.object->type == OBJT_VNODE) { 610 vp = fs.object->handle; 611 if (vp == fs.vp) 612 goto vnode_locked; 613 else if (fs.vp != NULL) { 614 vput(fs.vp); 615 fs.vp = NULL; 616 } 617 locked = VOP_ISLOCKED(vp); 618 619 if (locked != LK_EXCLUSIVE) 620 locked = LK_SHARED; 621 /* Do not sleep for vnode lock while fs.m is busy */ 622 error = vget(vp, locked | LK_CANRECURSE | 623 LK_NOWAIT, curthread); 624 if (error != 0) { 625 vhold(vp); 626 release_page(&fs); 627 unlock_and_deallocate(&fs); 628 error = vget(vp, locked | LK_RETRY | 629 LK_CANRECURSE, curthread); 630 vdrop(vp); 631 fs.vp = vp; 632 KASSERT(error == 0, 633 ("vm_fault: vget failed")); 634 goto RetryFault; 635 } 636 fs.vp = vp; 637 } 638vnode_locked: 639 KASSERT(fs.vp == NULL || !fs.map->system_map, 640 ("vm_fault: vnode-backed object mapped by system map")); 641 642 /* 643 * now we find out if any other pages should be paged 644 * in at this time this routine checks to see if the 645 * pages surrounding this fault reside in the same 646 * object as the page for this fault. If they do, 647 * then they are faulted in also into the object. The 648 * array "marray" returned contains an array of 649 * vm_page_t structs where one of them is the 650 * vm_page_t passed to the routine. The reqpage 651 * return value is the index into the marray for the 652 * vm_page_t passed to the routine. 653 * 654 * fs.m plus the additional pages are exclusive busied. 655 */ 656 faultcount = vm_fault_additional_pages( 657 fs.m, behind, ahead, marray, &reqpage); 658 659 rv = faultcount ? 660 vm_pager_get_pages(fs.object, marray, faultcount, 661 reqpage) : VM_PAGER_FAIL; 662 663 if (rv == VM_PAGER_OK) { 664 /* 665 * Found the page. Leave it busy while we play 666 * with it. 667 */ 668 669 /* 670 * Relookup in case pager changed page. Pager 671 * is responsible for disposition of old page 672 * if moved. 673 */ 674 fs.m = vm_page_lookup(fs.object, fs.pindex); 675 if (!fs.m) { 676 unlock_and_deallocate(&fs); 677 goto RetryFault; 678 } 679 680 hardfault++; 681 break; /* break to PAGE HAS BEEN FOUND */ 682 } 683 /* 684 * Remove the bogus page (which does not exist at this 685 * object/offset); before doing so, we must get back 686 * our object lock to preserve our invariant. 687 * 688 * Also wake up any other process that may want to bring 689 * in this page. 690 * 691 * If this is the top-level object, we must leave the 692 * busy page to prevent another process from rushing 693 * past us, and inserting the page in that object at 694 * the same time that we are. 695 */ 696 if (rv == VM_PAGER_ERROR) 697 printf("vm_fault: pager read error, pid %d (%s)\n", 698 curproc->p_pid, curproc->p_comm); 699 /* 700 * Data outside the range of the pager or an I/O error 701 */ 702 /* 703 * XXX - the check for kernel_map is a kludge to work 704 * around having the machine panic on a kernel space 705 * fault w/ I/O error. 706 */ 707 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 708 (rv == VM_PAGER_BAD)) { 709 vm_page_lock(fs.m); 710 vm_page_free(fs.m); 711 vm_page_unlock(fs.m); 712 fs.m = NULL; 713 unlock_and_deallocate(&fs); 714 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 715 } 716 if (fs.object != fs.first_object) { 717 vm_page_lock(fs.m); 718 vm_page_free(fs.m); 719 vm_page_unlock(fs.m); 720 fs.m = NULL; 721 /* 722 * XXX - we cannot just fall out at this 723 * point, m has been freed and is invalid! 724 */ 725 } 726 } 727 728 /* 729 * We get here if the object has default pager (or unwiring) 730 * or the pager doesn't have the page. 731 */ 732 if (fs.object == fs.first_object) 733 fs.first_m = fs.m; 734 735 /* 736 * Move on to the next object. Lock the next object before 737 * unlocking the current one. 738 */ 739 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 740 next_object = fs.object->backing_object; 741 if (next_object == NULL) { 742 /* 743 * If there's no object left, fill the page in the top 744 * object with zeros. 745 */ 746 if (fs.object != fs.first_object) { 747 vm_object_pip_wakeup(fs.object); 748 VM_OBJECT_WUNLOCK(fs.object); 749 750 fs.object = fs.first_object; 751 fs.pindex = fs.first_pindex; 752 fs.m = fs.first_m; 753 VM_OBJECT_WLOCK(fs.object); 754 } 755 fs.first_m = NULL; 756 757 /* 758 * Zero the page if necessary and mark it valid. 759 */ 760 if ((fs.m->flags & PG_ZERO) == 0) { 761 pmap_zero_page(fs.m); 762 } else { 763 PCPU_INC(cnt.v_ozfod); 764 } 765 PCPU_INC(cnt.v_zfod); 766 fs.m->valid = VM_PAGE_BITS_ALL; 767 /* Don't try to prefault neighboring pages. */ 768 faultcount = 1; 769 break; /* break to PAGE HAS BEEN FOUND */ 770 } else { 771 KASSERT(fs.object != next_object, 772 ("object loop %p", next_object)); 773 VM_OBJECT_WLOCK(next_object); 774 vm_object_pip_add(next_object, 1); 775 if (fs.object != fs.first_object) 776 vm_object_pip_wakeup(fs.object); 777 VM_OBJECT_WUNLOCK(fs.object); 778 fs.object = next_object; 779 } 780 } 781 782 vm_page_assert_xbusied(fs.m); 783 784 /* 785 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 786 * is held.] 787 */ 788 789 /* 790 * If the page is being written, but isn't already owned by the 791 * top-level object, we have to copy it into a new page owned by the 792 * top-level object. 793 */ 794 if (fs.object != fs.first_object) { 795 /* 796 * We only really need to copy if we want to write it. 797 */ 798 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 799 /* 800 * This allows pages to be virtually copied from a 801 * backing_object into the first_object, where the 802 * backing object has no other refs to it, and cannot 803 * gain any more refs. Instead of a bcopy, we just 804 * move the page from the backing object to the 805 * first object. Note that we must mark the page 806 * dirty in the first object so that it will go out 807 * to swap when needed. 808 */ 809 is_first_object_locked = FALSE; 810 if ( 811 /* 812 * Only one shadow object 813 */ 814 (fs.object->shadow_count == 1) && 815 /* 816 * No COW refs, except us 817 */ 818 (fs.object->ref_count == 1) && 819 /* 820 * No one else can look this object up 821 */ 822 (fs.object->handle == NULL) && 823 /* 824 * No other ways to look the object up 825 */ 826 ((fs.object->type == OBJT_DEFAULT) || 827 (fs.object->type == OBJT_SWAP)) && 828 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 829 /* 830 * We don't chase down the shadow chain 831 */ 832 fs.object == fs.first_object->backing_object) { 833 /* 834 * get rid of the unnecessary page 835 */ 836 vm_page_lock(fs.first_m); 837 vm_page_free(fs.first_m); 838 vm_page_unlock(fs.first_m); 839 /* 840 * grab the page and put it into the 841 * process'es object. The page is 842 * automatically made dirty. 843 */ 844 if (vm_page_rename(fs.m, fs.first_object, 845 fs.first_pindex)) { 846 unlock_and_deallocate(&fs); 847 goto RetryFault; 848 } 849#if VM_NRESERVLEVEL > 0 850 /* 851 * Rename the reservation. 852 */ 853 vm_reserv_rename(fs.m, fs.first_object, 854 fs.object, OFF_TO_IDX( 855 fs.first_object->backing_object_offset)); 856#endif 857 vm_page_xbusy(fs.m); 858 fs.first_m = fs.m; 859 fs.m = NULL; 860 PCPU_INC(cnt.v_cow_optim); 861 } else { 862 /* 863 * Oh, well, lets copy it. 864 */ 865 pmap_copy_page(fs.m, fs.first_m); 866 fs.first_m->valid = VM_PAGE_BITS_ALL; 867 if (wired && (fault_flags & 868 VM_FAULT_WIRE) == 0) { 869 vm_page_lock(fs.first_m); 870 vm_page_wire(fs.first_m); 871 vm_page_unlock(fs.first_m); 872 873 vm_page_lock(fs.m); 874 vm_page_unwire(fs.m, FALSE); 875 vm_page_unlock(fs.m); 876 } 877 /* 878 * We no longer need the old page or object. 879 */ 880 release_page(&fs); 881 } 882 /* 883 * fs.object != fs.first_object due to above 884 * conditional 885 */ 886 vm_object_pip_wakeup(fs.object); 887 VM_OBJECT_WUNLOCK(fs.object); 888 /* 889 * Only use the new page below... 890 */ 891 fs.object = fs.first_object; 892 fs.pindex = fs.first_pindex; 893 fs.m = fs.first_m; 894 if (!is_first_object_locked) 895 VM_OBJECT_WLOCK(fs.object); 896 PCPU_INC(cnt.v_cow_faults); 897 curthread->td_cow++; 898 } else { 899 prot &= ~VM_PROT_WRITE; 900 } 901 } 902 903 /* 904 * We must verify that the maps have not changed since our last 905 * lookup. 906 */ 907 if (!fs.lookup_still_valid) { 908 vm_object_t retry_object; 909 vm_pindex_t retry_pindex; 910 vm_prot_t retry_prot; 911 912 if (!vm_map_trylock_read(fs.map)) { 913 release_page(&fs); 914 unlock_and_deallocate(&fs); 915 goto RetryFault; 916 } 917 fs.lookup_still_valid = TRUE; 918 if (fs.map->timestamp != map_generation) { 919 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 920 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 921 922 /* 923 * If we don't need the page any longer, put it on the inactive 924 * list (the easiest thing to do here). If no one needs it, 925 * pageout will grab it eventually. 926 */ 927 if (result != KERN_SUCCESS) { 928 release_page(&fs); 929 unlock_and_deallocate(&fs); 930 931 /* 932 * If retry of map lookup would have blocked then 933 * retry fault from start. 934 */ 935 if (result == KERN_FAILURE) 936 goto RetryFault; 937 return (result); 938 } 939 if ((retry_object != fs.first_object) || 940 (retry_pindex != fs.first_pindex)) { 941 release_page(&fs); 942 unlock_and_deallocate(&fs); 943 goto RetryFault; 944 } 945 946 /* 947 * Check whether the protection has changed or the object has 948 * been copied while we left the map unlocked. Changing from 949 * read to write permission is OK - we leave the page 950 * write-protected, and catch the write fault. Changing from 951 * write to read permission means that we can't mark the page 952 * write-enabled after all. 953 */ 954 prot &= retry_prot; 955 } 956 } 957 /* 958 * If the page was filled by a pager, update the map entry's 959 * last read offset. Since the pager does not return the 960 * actual set of pages that it read, this update is based on 961 * the requested set. Typically, the requested and actual 962 * sets are the same. 963 * 964 * XXX The following assignment modifies the map 965 * without holding a write lock on it. 966 */ 967 if (hardfault) 968 fs.entry->next_read = fs.pindex + faultcount - reqpage; 969 970 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE); 971 vm_page_assert_xbusied(fs.m); 972 973 /* 974 * Page must be completely valid or it is not fit to 975 * map into user space. vm_pager_get_pages() ensures this. 976 */ 977 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 978 ("vm_fault: page %p partially invalid", fs.m)); 979 VM_OBJECT_WUNLOCK(fs.object); 980 981 /* 982 * Put this page into the physical map. We had to do the unlock above 983 * because pmap_enter() may sleep. We don't put the page 984 * back on the active queue until later so that the pageout daemon 985 * won't find it (yet). 986 */ 987 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, 988 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); 989 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && 990 wired == 0) 991 vm_fault_prefault(&fs, vaddr, faultcount, reqpage); 992 VM_OBJECT_WLOCK(fs.object); 993 vm_page_lock(fs.m); 994 995 /* 996 * If the page is not wired down, then put it where the pageout daemon 997 * can find it. 998 */ 999 if ((fault_flags & VM_FAULT_WIRE) != 0) { 1000 KASSERT(wired, ("VM_FAULT_WIRE && !wired")); 1001 vm_page_wire(fs.m); 1002 } else 1003 vm_page_activate(fs.m); 1004 if (m_hold != NULL) { 1005 *m_hold = fs.m; 1006 vm_page_hold(fs.m); 1007 } 1008 vm_page_unlock(fs.m); 1009 vm_page_xunbusy(fs.m); 1010 1011 /* 1012 * Unlock everything, and return 1013 */ 1014 unlock_and_deallocate(&fs); 1015 if (hardfault) { 1016 PCPU_INC(cnt.v_io_faults); 1017 curthread->td_ru.ru_majflt++; 1018 } else 1019 curthread->td_ru.ru_minflt++; 1020 1021 return (KERN_SUCCESS); 1022} 1023 1024/* 1025 * Speed up the reclamation of up to "distance" pages that precede the 1026 * faulting pindex within the first object of the shadow chain. 1027 */ 1028static void 1029vm_fault_cache_behind(const struct faultstate *fs, int distance) 1030{ 1031 vm_object_t first_object, object; 1032 vm_page_t m, m_prev; 1033 vm_pindex_t pindex; 1034 1035 object = fs->object; 1036 VM_OBJECT_ASSERT_WLOCKED(object); 1037 first_object = fs->first_object; 1038 if (first_object != object) { 1039 if (!VM_OBJECT_TRYWLOCK(first_object)) { 1040 VM_OBJECT_WUNLOCK(object); 1041 VM_OBJECT_WLOCK(first_object); 1042 VM_OBJECT_WLOCK(object); 1043 } 1044 } 1045 /* Neither fictitious nor unmanaged pages can be cached. */ 1046 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1047 if (fs->first_pindex < distance) 1048 pindex = 0; 1049 else 1050 pindex = fs->first_pindex - distance; 1051 if (pindex < OFF_TO_IDX(fs->entry->offset)) 1052 pindex = OFF_TO_IDX(fs->entry->offset); 1053 m = first_object != object ? fs->first_m : fs->m; 1054 vm_page_assert_xbusied(m); 1055 m_prev = vm_page_prev(m); 1056 while ((m = m_prev) != NULL && m->pindex >= pindex && 1057 m->valid == VM_PAGE_BITS_ALL) { 1058 m_prev = vm_page_prev(m); 1059 if (vm_page_busied(m)) 1060 continue; 1061 vm_page_lock(m); 1062 if (m->hold_count == 0 && m->wire_count == 0) { 1063 pmap_remove_all(m); 1064 vm_page_aflag_clear(m, PGA_REFERENCED); 1065 if (m->dirty != 0) 1066 vm_page_deactivate(m); 1067 else 1068 vm_page_cache(m); 1069 } 1070 vm_page_unlock(m); 1071 } 1072 } 1073 if (first_object != object) 1074 VM_OBJECT_WUNLOCK(first_object); 1075} 1076 1077/* 1078 * vm_fault_prefault provides a quick way of clustering 1079 * pagefaults into a processes address space. It is a "cousin" 1080 * of vm_map_pmap_enter, except it runs at page fault time instead 1081 * of mmap time. 1082 */ 1083static void 1084vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1085 int faultcount, int reqpage) 1086{ 1087 pmap_t pmap; 1088 vm_map_entry_t entry; 1089 vm_object_t backing_object, lobject; 1090 vm_offset_t addr, starta; 1091 vm_pindex_t pindex; 1092 vm_page_t m; 1093 int backward, forward, i; 1094 1095 pmap = fs->map->pmap; 1096 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1097 return; 1098 1099 if (faultcount > 0) { 1100 backward = reqpage; 1101 forward = faultcount - reqpage - 1; 1102 } else { 1103 backward = PFBAK; 1104 forward = PFFOR; 1105 } 1106 entry = fs->entry; 1107 1108 starta = addra - backward * PAGE_SIZE; 1109 if (starta < entry->start) { 1110 starta = entry->start; 1111 } else if (starta > addra) { 1112 starta = 0; 1113 } 1114 1115 /* 1116 * Generate the sequence of virtual addresses that are candidates for 1117 * prefaulting in an outward spiral from the faulting virtual address, 1118 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1119 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1120 * If the candidate address doesn't have a backing physical page, then 1121 * the loop immediately terminates. 1122 */ 1123 for (i = 0; i < 2 * imax(backward, forward); i++) { 1124 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1125 PAGE_SIZE); 1126 if (addr > addra + forward * PAGE_SIZE) 1127 addr = 0; 1128 1129 if (addr < starta || addr >= entry->end) 1130 continue; 1131 1132 if (!pmap_is_prefaultable(pmap, addr)) 1133 continue; 1134 1135 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1136 lobject = entry->object.vm_object; 1137 VM_OBJECT_RLOCK(lobject); 1138 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1139 lobject->type == OBJT_DEFAULT && 1140 (backing_object = lobject->backing_object) != NULL) { 1141 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1142 0, ("vm_fault_prefault: unaligned object offset")); 1143 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1144 VM_OBJECT_RLOCK(backing_object); 1145 VM_OBJECT_RUNLOCK(lobject); 1146 lobject = backing_object; 1147 } 1148 if (m == NULL) { 1149 VM_OBJECT_RUNLOCK(lobject); 1150 break; 1151 } 1152 if (m->valid == VM_PAGE_BITS_ALL && 1153 (m->flags & PG_FICTITIOUS) == 0) 1154 pmap_enter_quick(pmap, addr, m, entry->protection); 1155 VM_OBJECT_RUNLOCK(lobject); 1156 } 1157} 1158 1159/* 1160 * Hold each of the physical pages that are mapped by the specified range of 1161 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1162 * and allow the specified types of access, "prot". If all of the implied 1163 * pages are successfully held, then the number of held pages is returned 1164 * together with pointers to those pages in the array "ma". However, if any 1165 * of the pages cannot be held, -1 is returned. 1166 */ 1167int 1168vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1169 vm_prot_t prot, vm_page_t *ma, int max_count) 1170{ 1171 vm_offset_t end, va; 1172 vm_page_t *mp; 1173 int count; 1174 boolean_t pmap_failed; 1175 1176 if (len == 0) 1177 return (0); 1178 end = round_page(addr + len); 1179 addr = trunc_page(addr); 1180 1181 /* 1182 * Check for illegal addresses. 1183 */ 1184 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1185 return (-1); 1186 1187 if (atop(end - addr) > max_count) 1188 panic("vm_fault_quick_hold_pages: count > max_count"); 1189 count = atop(end - addr); 1190 1191 /* 1192 * Most likely, the physical pages are resident in the pmap, so it is 1193 * faster to try pmap_extract_and_hold() first. 1194 */ 1195 pmap_failed = FALSE; 1196 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1197 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1198 if (*mp == NULL) 1199 pmap_failed = TRUE; 1200 else if ((prot & VM_PROT_WRITE) != 0 && 1201 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1202 /* 1203 * Explicitly dirty the physical page. Otherwise, the 1204 * caller's changes may go unnoticed because they are 1205 * performed through an unmanaged mapping or by a DMA 1206 * operation. 1207 * 1208 * The object lock is not held here. 1209 * See vm_page_clear_dirty_mask(). 1210 */ 1211 vm_page_dirty(*mp); 1212 } 1213 } 1214 if (pmap_failed) { 1215 /* 1216 * One or more pages could not be held by the pmap. Either no 1217 * page was mapped at the specified virtual address or that 1218 * mapping had insufficient permissions. Attempt to fault in 1219 * and hold these pages. 1220 */ 1221 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1222 if (*mp == NULL && vm_fault_hold(map, va, prot, 1223 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1224 goto error; 1225 } 1226 return (count); 1227error: 1228 for (mp = ma; mp < ma + count; mp++) 1229 if (*mp != NULL) { 1230 vm_page_lock(*mp); 1231 vm_page_unhold(*mp); 1232 vm_page_unlock(*mp); 1233 } 1234 return (-1); 1235} 1236 1237/* 1238 * Routine: 1239 * vm_fault_copy_entry 1240 * Function: 1241 * Create new shadow object backing dst_entry with private copy of 1242 * all underlying pages. When src_entry is equal to dst_entry, 1243 * function implements COW for wired-down map entry. Otherwise, 1244 * it forks wired entry into dst_map. 1245 * 1246 * In/out conditions: 1247 * The source and destination maps must be locked for write. 1248 * The source map entry must be wired down (or be a sharing map 1249 * entry corresponding to a main map entry that is wired down). 1250 */ 1251void 1252vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1253 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1254 vm_ooffset_t *fork_charge) 1255{ 1256 vm_object_t backing_object, dst_object, object, src_object; 1257 vm_pindex_t dst_pindex, pindex, src_pindex; 1258 vm_prot_t access, prot; 1259 vm_offset_t vaddr; 1260 vm_page_t dst_m; 1261 vm_page_t src_m; 1262 boolean_t upgrade; 1263 1264#ifdef lint 1265 src_map++; 1266#endif /* lint */ 1267 1268 upgrade = src_entry == dst_entry; 1269 access = prot = dst_entry->protection; 1270 1271 src_object = src_entry->object.vm_object; 1272 src_pindex = OFF_TO_IDX(src_entry->offset); 1273 1274 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1275 dst_object = src_object; 1276 vm_object_reference(dst_object); 1277 } else { 1278 /* 1279 * Create the top-level object for the destination entry. (Doesn't 1280 * actually shadow anything - we copy the pages directly.) 1281 */ 1282 dst_object = vm_object_allocate(OBJT_DEFAULT, 1283 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1284#if VM_NRESERVLEVEL > 0 1285 dst_object->flags |= OBJ_COLORED; 1286 dst_object->pg_color = atop(dst_entry->start); 1287#endif 1288 } 1289 1290 VM_OBJECT_WLOCK(dst_object); 1291 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1292 ("vm_fault_copy_entry: vm_object not NULL")); 1293 if (src_object != dst_object) { 1294 dst_entry->object.vm_object = dst_object; 1295 dst_entry->offset = 0; 1296 dst_object->charge = dst_entry->end - dst_entry->start; 1297 } 1298 if (fork_charge != NULL) { 1299 KASSERT(dst_entry->cred == NULL, 1300 ("vm_fault_copy_entry: leaked swp charge")); 1301 dst_object->cred = curthread->td_ucred; 1302 crhold(dst_object->cred); 1303 *fork_charge += dst_object->charge; 1304 } else if (dst_object->cred == NULL) { 1305 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1306 dst_entry)); 1307 dst_object->cred = dst_entry->cred; 1308 dst_entry->cred = NULL; 1309 } 1310 1311 /* 1312 * If not an upgrade, then enter the mappings in the pmap as 1313 * read and/or execute accesses. Otherwise, enter them as 1314 * write accesses. 1315 * 1316 * A writeable large page mapping is only created if all of 1317 * the constituent small page mappings are modified. Marking 1318 * PTEs as modified on inception allows promotion to happen 1319 * without taking potentially large number of soft faults. 1320 */ 1321 if (!upgrade) 1322 access &= ~VM_PROT_WRITE; 1323 1324 /* 1325 * Loop through all of the virtual pages within the entry's 1326 * range, copying each page from the source object to the 1327 * destination object. Since the source is wired, those pages 1328 * must exist. In contrast, the destination is pageable. 1329 * Since the destination object does share any backing storage 1330 * with the source object, all of its pages must be dirtied, 1331 * regardless of whether they can be written. 1332 */ 1333 for (vaddr = dst_entry->start, dst_pindex = 0; 1334 vaddr < dst_entry->end; 1335 vaddr += PAGE_SIZE, dst_pindex++) { 1336again: 1337 /* 1338 * Find the page in the source object, and copy it in. 1339 * Because the source is wired down, the page will be 1340 * in memory. 1341 */ 1342 if (src_object != dst_object) 1343 VM_OBJECT_RLOCK(src_object); 1344 object = src_object; 1345 pindex = src_pindex + dst_pindex; 1346 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1347 (backing_object = object->backing_object) != NULL) { 1348 /* 1349 * Unless the source mapping is read-only or 1350 * it is presently being upgraded from 1351 * read-only, the first object in the shadow 1352 * chain should provide all of the pages. In 1353 * other words, this loop body should never be 1354 * executed when the source mapping is already 1355 * read/write. 1356 */ 1357 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1358 upgrade, 1359 ("vm_fault_copy_entry: main object missing page")); 1360 1361 VM_OBJECT_RLOCK(backing_object); 1362 pindex += OFF_TO_IDX(object->backing_object_offset); 1363 if (object != dst_object) 1364 VM_OBJECT_RUNLOCK(object); 1365 object = backing_object; 1366 } 1367 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1368 1369 if (object != dst_object) { 1370 /* 1371 * Allocate a page in the destination object. 1372 */ 1373 dst_m = vm_page_alloc(dst_object, (src_object == 1374 dst_object ? src_pindex : 0) + dst_pindex, 1375 VM_ALLOC_NORMAL); 1376 if (dst_m == NULL) { 1377 VM_OBJECT_WUNLOCK(dst_object); 1378 VM_OBJECT_RUNLOCK(object); 1379 VM_WAIT; 1380 VM_OBJECT_WLOCK(dst_object); 1381 goto again; 1382 } 1383 pmap_copy_page(src_m, dst_m); 1384 VM_OBJECT_RUNLOCK(object); 1385 dst_m->valid = VM_PAGE_BITS_ALL; 1386 dst_m->dirty = VM_PAGE_BITS_ALL; 1387 } else { 1388 dst_m = src_m; 1389 if (vm_page_sleep_if_busy(dst_m, "fltupg")) 1390 goto again; 1391 vm_page_xbusy(dst_m); 1392 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL, 1393 ("invalid dst page %p", dst_m)); 1394 } 1395 VM_OBJECT_WUNLOCK(dst_object); 1396 1397 /* 1398 * Enter it in the pmap. If a wired, copy-on-write 1399 * mapping is being replaced by a write-enabled 1400 * mapping, then wire that new mapping. 1401 */ 1402 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 1403 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 1404 1405 /* 1406 * Mark it no longer busy, and put it on the active list. 1407 */ 1408 VM_OBJECT_WLOCK(dst_object); 1409 1410 if (upgrade) { 1411 if (src_m != dst_m) { 1412 vm_page_lock(src_m); 1413 vm_page_unwire(src_m, 0); 1414 vm_page_unlock(src_m); 1415 vm_page_lock(dst_m); 1416 vm_page_wire(dst_m); 1417 vm_page_unlock(dst_m); 1418 } else { 1419 KASSERT(dst_m->wire_count > 0, 1420 ("dst_m %p is not wired", dst_m)); 1421 } 1422 } else { 1423 vm_page_lock(dst_m); 1424 vm_page_activate(dst_m); 1425 vm_page_unlock(dst_m); 1426 } 1427 vm_page_xunbusy(dst_m); 1428 } 1429 VM_OBJECT_WUNLOCK(dst_object); 1430 if (upgrade) { 1431 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1432 vm_object_deallocate(src_object); 1433 } 1434} 1435 1436 1437/* 1438 * This routine checks around the requested page for other pages that 1439 * might be able to be faulted in. This routine brackets the viable 1440 * pages for the pages to be paged in. 1441 * 1442 * Inputs: 1443 * m, rbehind, rahead 1444 * 1445 * Outputs: 1446 * marray (array of vm_page_t), reqpage (index of requested page) 1447 * 1448 * Return value: 1449 * number of pages in marray 1450 */ 1451static int 1452vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1453 vm_page_t m; 1454 int rbehind; 1455 int rahead; 1456 vm_page_t *marray; 1457 int *reqpage; 1458{ 1459 int i,j; 1460 vm_object_t object; 1461 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1462 vm_page_t rtm; 1463 int cbehind, cahead; 1464 1465 VM_OBJECT_ASSERT_WLOCKED(m->object); 1466 1467 object = m->object; 1468 pindex = m->pindex; 1469 cbehind = cahead = 0; 1470 1471 /* 1472 * if the requested page is not available, then give up now 1473 */ 1474 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1475 return 0; 1476 } 1477 1478 if ((cbehind == 0) && (cahead == 0)) { 1479 *reqpage = 0; 1480 marray[0] = m; 1481 return 1; 1482 } 1483 1484 if (rahead > cahead) { 1485 rahead = cahead; 1486 } 1487 1488 if (rbehind > cbehind) { 1489 rbehind = cbehind; 1490 } 1491 1492 /* 1493 * scan backward for the read behind pages -- in memory 1494 */ 1495 if (pindex > 0) { 1496 if (rbehind > pindex) { 1497 rbehind = pindex; 1498 startpindex = 0; 1499 } else { 1500 startpindex = pindex - rbehind; 1501 } 1502 1503 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1504 rtm->pindex >= startpindex) 1505 startpindex = rtm->pindex + 1; 1506 1507 /* tpindex is unsigned; beware of numeric underflow. */ 1508 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex && 1509 tpindex < pindex; i++, tpindex--) { 1510 1511 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1512 VM_ALLOC_IFNOTCACHED); 1513 if (rtm == NULL) { 1514 /* 1515 * Shift the allocated pages to the 1516 * beginning of the array. 1517 */ 1518 for (j = 0; j < i; j++) { 1519 marray[j] = marray[j + tpindex + 1 - 1520 startpindex]; 1521 } 1522 break; 1523 } 1524 1525 marray[tpindex - startpindex] = rtm; 1526 } 1527 } else { 1528 startpindex = 0; 1529 i = 0; 1530 } 1531 1532 marray[i] = m; 1533 /* page offset of the required page */ 1534 *reqpage = i; 1535 1536 tpindex = pindex + 1; 1537 i++; 1538 1539 /* 1540 * scan forward for the read ahead pages 1541 */ 1542 endpindex = tpindex + rahead; 1543 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1544 endpindex = rtm->pindex; 1545 if (endpindex > object->size) 1546 endpindex = object->size; 1547 1548 for (; tpindex < endpindex; i++, tpindex++) { 1549 1550 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1551 VM_ALLOC_IFNOTCACHED); 1552 if (rtm == NULL) { 1553 break; 1554 } 1555 1556 marray[i] = rtm; 1557 } 1558 1559 /* return number of pages */ 1560 return i; 1561} 1562 1563/* 1564 * Block entry into the machine-independent layer's page fault handler by 1565 * the calling thread. Subsequent calls to vm_fault() by that thread will 1566 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1567 * spurious page faults. 1568 */ 1569int 1570vm_fault_disable_pagefaults(void) 1571{ 1572 1573 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1574} 1575 1576void 1577vm_fault_enable_pagefaults(int save) 1578{ 1579 1580 curthread_pflags_restore(save); 1581} 1582