vm_fault.c revision 278571
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 278571 2015-02-11 09:02:21Z 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> 104 105#define PFBAK 4 106#define PFFOR 4 107 108static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *); 109 110#define VM_FAULT_READ_BEHIND 8 111#define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 112#define VM_FAULT_NINCR (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND) 113#define VM_FAULT_SUM (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2) 114#define VM_FAULT_CACHE_BEHIND (VM_FAULT_READ_BEHIND * VM_FAULT_SUM) 115 116struct faultstate { 117 vm_page_t m; 118 vm_object_t object; 119 vm_pindex_t pindex; 120 vm_page_t first_m; 121 vm_object_t first_object; 122 vm_pindex_t first_pindex; 123 vm_map_t map; 124 vm_map_entry_t entry; 125 int lookup_still_valid; 126 struct vnode *vp; 127}; 128 129static void vm_fault_cache_behind(const struct faultstate *fs, int distance); 130static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 131 int faultcount, int reqpage); 132 133static inline void 134release_page(struct faultstate *fs) 135{ 136 137 vm_page_xunbusy(fs->m); 138 vm_page_lock(fs->m); 139 vm_page_deactivate(fs->m); 140 vm_page_unlock(fs->m); 141 fs->m = NULL; 142} 143 144static inline void 145unlock_map(struct faultstate *fs) 146{ 147 148 if (fs->lookup_still_valid) { 149 vm_map_lookup_done(fs->map, fs->entry); 150 fs->lookup_still_valid = FALSE; 151 } 152} 153 154static void 155unlock_and_deallocate(struct faultstate *fs) 156{ 157 158 vm_object_pip_wakeup(fs->object); 159 VM_OBJECT_WUNLOCK(fs->object); 160 if (fs->object != fs->first_object) { 161 VM_OBJECT_WLOCK(fs->first_object); 162 vm_page_lock(fs->first_m); 163 vm_page_free(fs->first_m); 164 vm_page_unlock(fs->first_m); 165 vm_object_pip_wakeup(fs->first_object); 166 VM_OBJECT_WUNLOCK(fs->first_object); 167 fs->first_m = NULL; 168 } 169 vm_object_deallocate(fs->first_object); 170 unlock_map(fs); 171 if (fs->vp != NULL) { 172 vput(fs->vp); 173 fs->vp = NULL; 174 } 175} 176 177static void 178vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, 179 vm_prot_t fault_type, int fault_flags, boolean_t set_wd) 180{ 181 boolean_t need_dirty; 182 183 if (((prot & VM_PROT_WRITE) == 0 && 184 (fault_flags & VM_FAULT_DIRTY) == 0) || 185 (m->oflags & VPO_UNMANAGED) != 0) 186 return; 187 188 VM_OBJECT_ASSERT_LOCKED(m->object); 189 190 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && 191 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) || 192 (fault_flags & VM_FAULT_DIRTY) != 0; 193 194 if (set_wd) 195 vm_object_set_writeable_dirty(m->object); 196 else 197 /* 198 * If two callers of vm_fault_dirty() with set_wd == 199 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC 200 * flag set, other with flag clear, race, it is 201 * possible for the no-NOSYNC thread to see m->dirty 202 * != 0 and not clear VPO_NOSYNC. Take vm_page lock 203 * around manipulation of VPO_NOSYNC and 204 * vm_page_dirty() call, to avoid the race and keep 205 * m->oflags consistent. 206 */ 207 vm_page_lock(m); 208 209 /* 210 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 211 * if the page is already dirty to prevent data written with 212 * the expectation of being synced from not being synced. 213 * Likewise if this entry does not request NOSYNC then make 214 * sure the page isn't marked NOSYNC. Applications sharing 215 * data should use the same flags to avoid ping ponging. 216 */ 217 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { 218 if (m->dirty == 0) { 219 m->oflags |= VPO_NOSYNC; 220 } 221 } else { 222 m->oflags &= ~VPO_NOSYNC; 223 } 224 225 /* 226 * If the fault is a write, we know that this page is being 227 * written NOW so dirty it explicitly to save on 228 * pmap_is_modified() calls later. 229 * 230 * Also tell the backing pager, if any, that it should remove 231 * any swap backing since the page is now dirty. 232 */ 233 if (need_dirty) 234 vm_page_dirty(m); 235 if (!set_wd) 236 vm_page_unlock(m); 237 if (need_dirty) 238 vm_pager_page_unswapped(m); 239} 240 241/* 242 * TRYPAGER - used by vm_fault to calculate whether the pager for the 243 * current object *might* contain the page. 244 * 245 * default objects are zero-fill, there is no real pager. 246 */ 247#define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 248 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired)) 249 250/* 251 * vm_fault: 252 * 253 * Handle a page fault occurring at the given address, 254 * requiring the given permissions, in the map specified. 255 * If successful, the page is inserted into the 256 * associated physical map. 257 * 258 * NOTE: the given address should be truncated to the 259 * proper page address. 260 * 261 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 262 * a standard error specifying why the fault is fatal is returned. 263 * 264 * The map in question must be referenced, and remains so. 265 * Caller may hold no locks. 266 */ 267int 268vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 269 int fault_flags) 270{ 271 struct thread *td; 272 int result; 273 274 td = curthread; 275 if ((td->td_pflags & TDP_NOFAULTING) != 0) 276 return (KERN_PROTECTION_FAILURE); 277#ifdef KTRACE 278 if (map != kernel_map && KTRPOINT(td, KTR_FAULT)) 279 ktrfault(vaddr, fault_type); 280#endif 281 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags, 282 NULL); 283#ifdef KTRACE 284 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND)) 285 ktrfaultend(result); 286#endif 287 return (result); 288} 289 290int 291vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 292 int fault_flags, vm_page_t *m_hold) 293{ 294 vm_prot_t prot; 295 long ahead, behind; 296 int alloc_req, era, faultcount, nera, reqpage, result; 297 boolean_t growstack, is_first_object_locked, wired; 298 int map_generation; 299 vm_object_t next_object; 300 vm_page_t marray[VM_FAULT_READ_MAX]; 301 int hardfault; 302 struct faultstate fs; 303 struct vnode *vp; 304 vm_page_t m; 305 int locked, error; 306 307 hardfault = 0; 308 growstack = TRUE; 309 PCPU_INC(cnt.v_vm_faults); 310 fs.vp = NULL; 311 faultcount = reqpage = 0; 312 313RetryFault:; 314 315 /* 316 * Find the backing store object and offset into it to begin the 317 * search. 318 */ 319 fs.map = map; 320 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry, 321 &fs.first_object, &fs.first_pindex, &prot, &wired); 322 if (result != KERN_SUCCESS) { 323 if (growstack && result == KERN_INVALID_ADDRESS && 324 map != kernel_map) { 325 result = vm_map_growstack(curproc, vaddr); 326 if (result != KERN_SUCCESS) 327 return (KERN_FAILURE); 328 growstack = FALSE; 329 goto RetryFault; 330 } 331 return (result); 332 } 333 334 map_generation = fs.map->timestamp; 335 336 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 337 panic("vm_fault: fault on nofault entry, addr: %lx", 338 (u_long)vaddr); 339 } 340 341 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 342 fs.entry->wiring_thread != curthread) { 343 vm_map_unlock_read(fs.map); 344 vm_map_lock(fs.map); 345 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 346 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 347 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 348 vm_map_unlock_and_wait(fs.map, 0); 349 } else 350 vm_map_unlock(fs.map); 351 goto RetryFault; 352 } 353 354 if (wired) 355 fault_type = prot | (fault_type & VM_PROT_COPY); 356 357 if (fs.vp == NULL /* avoid locked vnode leak */ && 358 (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 && 359 /* avoid calling vm_object_set_writeable_dirty() */ 360 ((prot & VM_PROT_WRITE) == 0 || 361 (fs.first_object->type != OBJT_VNODE && 362 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 363 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { 364 VM_OBJECT_RLOCK(fs.first_object); 365 if ((prot & VM_PROT_WRITE) != 0 && 366 (fs.first_object->type == OBJT_VNODE || 367 (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) && 368 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0) 369 goto fast_failed; 370 m = vm_page_lookup(fs.first_object, fs.first_pindex); 371 /* A busy page can be mapped for read|execute access. */ 372 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && 373 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) 374 goto fast_failed; 375 result = pmap_enter(fs.map->pmap, vaddr, m, prot, 376 fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 377 0), 0); 378 if (result != KERN_SUCCESS) 379 goto fast_failed; 380 if (m_hold != NULL) { 381 *m_hold = m; 382 vm_page_lock(m); 383 vm_page_hold(m); 384 vm_page_unlock(m); 385 } 386 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags, 387 FALSE); 388 VM_OBJECT_RUNLOCK(fs.first_object); 389 if (!wired) 390 vm_fault_prefault(&fs, vaddr, 0, 0); 391 vm_map_lookup_done(fs.map, fs.entry); 392 curthread->td_ru.ru_minflt++; 393 return (KERN_SUCCESS); 394fast_failed: 395 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 396 VM_OBJECT_RUNLOCK(fs.first_object); 397 VM_OBJECT_WLOCK(fs.first_object); 398 } 399 } else { 400 VM_OBJECT_WLOCK(fs.first_object); 401 } 402 403 /* 404 * Make a reference to this object to prevent its disposal while we 405 * are messing with it. Once we have the reference, the map is free 406 * to be diddled. Since objects reference their shadows (and copies), 407 * they will stay around as well. 408 * 409 * Bump the paging-in-progress count to prevent size changes (e.g. 410 * truncation operations) during I/O. This must be done after 411 * obtaining the vnode lock in order to avoid possible deadlocks. 412 */ 413 vm_object_reference_locked(fs.first_object); 414 vm_object_pip_add(fs.first_object, 1); 415 416 fs.lookup_still_valid = TRUE; 417 418 fs.first_m = NULL; 419 420 /* 421 * Search for the page at object/offset. 422 */ 423 fs.object = fs.first_object; 424 fs.pindex = fs.first_pindex; 425 while (TRUE) { 426 /* 427 * If the object is dead, we stop here 428 */ 429 if (fs.object->flags & OBJ_DEAD) { 430 unlock_and_deallocate(&fs); 431 return (KERN_PROTECTION_FAILURE); 432 } 433 434 /* 435 * See if page is resident 436 */ 437 fs.m = vm_page_lookup(fs.object, fs.pindex); 438 if (fs.m != NULL) { 439 /* 440 * Wait/Retry if the page is busy. We have to do this 441 * if the page is either exclusive or shared busy 442 * because the vm_pager may be using read busy for 443 * pageouts (and even pageins if it is the vnode 444 * pager), and we could end up trying to pagein and 445 * pageout the same page simultaneously. 446 * 447 * We can theoretically allow the busy case on a read 448 * fault if the page is marked valid, but since such 449 * pages are typically already pmap'd, putting that 450 * special case in might be more effort then it is 451 * worth. We cannot under any circumstances mess 452 * around with a shared busied page except, perhaps, 453 * to pmap it. 454 */ 455 if (vm_page_busied(fs.m)) { 456 /* 457 * Reference the page before unlocking and 458 * sleeping so that the page daemon is less 459 * likely to reclaim it. 460 */ 461 vm_page_aflag_set(fs.m, PGA_REFERENCED); 462 if (fs.object != fs.first_object) { 463 if (!VM_OBJECT_TRYWLOCK( 464 fs.first_object)) { 465 VM_OBJECT_WUNLOCK(fs.object); 466 VM_OBJECT_WLOCK(fs.first_object); 467 VM_OBJECT_WLOCK(fs.object); 468 } 469 vm_page_lock(fs.first_m); 470 vm_page_free(fs.first_m); 471 vm_page_unlock(fs.first_m); 472 vm_object_pip_wakeup(fs.first_object); 473 VM_OBJECT_WUNLOCK(fs.first_object); 474 fs.first_m = NULL; 475 } 476 unlock_map(&fs); 477 if (fs.m == vm_page_lookup(fs.object, 478 fs.pindex)) { 479 vm_page_sleep_if_busy(fs.m, "vmpfw"); 480 } 481 vm_object_pip_wakeup(fs.object); 482 VM_OBJECT_WUNLOCK(fs.object); 483 PCPU_INC(cnt.v_intrans); 484 vm_object_deallocate(fs.first_object); 485 goto RetryFault; 486 } 487 vm_page_lock(fs.m); 488 vm_page_remque(fs.m); 489 vm_page_unlock(fs.m); 490 491 /* 492 * Mark page busy for other processes, and the 493 * pagedaemon. If it still isn't completely valid 494 * (readable), jump to readrest, else break-out ( we 495 * found the page ). 496 */ 497 vm_page_xbusy(fs.m); 498 if (fs.m->valid != VM_PAGE_BITS_ALL) 499 goto readrest; 500 break; 501 } 502 503 /* 504 * Page is not resident, If this is the search termination 505 * or the pager might contain the page, allocate a new page. 506 */ 507 if (TRYPAGER || fs.object == fs.first_object) { 508 if (fs.pindex >= fs.object->size) { 509 unlock_and_deallocate(&fs); 510 return (KERN_PROTECTION_FAILURE); 511 } 512 513 /* 514 * Allocate a new page for this object/offset pair. 515 * 516 * Unlocked read of the p_flag is harmless. At 517 * worst, the P_KILLED might be not observed 518 * there, and allocation can fail, causing 519 * restart and new reading of the p_flag. 520 */ 521 fs.m = NULL; 522 if (!vm_page_count_severe() || P_KILLED(curproc)) { 523#if VM_NRESERVLEVEL > 0 524 if ((fs.object->flags & OBJ_COLORED) == 0) { 525 fs.object->flags |= OBJ_COLORED; 526 fs.object->pg_color = atop(vaddr) - 527 fs.pindex; 528 } 529#endif 530 alloc_req = P_KILLED(curproc) ? 531 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 532 if (fs.object->type != OBJT_VNODE && 533 fs.object->backing_object == NULL) 534 alloc_req |= VM_ALLOC_ZERO; 535 fs.m = vm_page_alloc(fs.object, fs.pindex, 536 alloc_req); 537 } 538 if (fs.m == NULL) { 539 unlock_and_deallocate(&fs); 540 VM_WAITPFAULT; 541 goto RetryFault; 542 } else if (fs.m->valid == VM_PAGE_BITS_ALL) 543 break; 544 } 545 546readrest: 547 /* 548 * We have found a valid page or we have allocated a new page. 549 * The page thus may not be valid or may not be entirely 550 * valid. 551 * 552 * Attempt to fault-in the page if there is a chance that the 553 * pager has it, and potentially fault in additional pages 554 * at the same time. 555 */ 556 if (TRYPAGER) { 557 int rv; 558 u_char behavior = vm_map_entry_behavior(fs.entry); 559 560 if (behavior == MAP_ENTRY_BEHAV_RANDOM || 561 P_KILLED(curproc)) { 562 behind = 0; 563 ahead = 0; 564 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 565 behind = 0; 566 ahead = atop(fs.entry->end - vaddr) - 1; 567 if (ahead > VM_FAULT_READ_AHEAD_MAX) 568 ahead = VM_FAULT_READ_AHEAD_MAX; 569 if (fs.pindex == fs.entry->next_read) 570 vm_fault_cache_behind(&fs, 571 VM_FAULT_READ_MAX); 572 } else { 573 /* 574 * If this is a sequential page fault, then 575 * arithmetically increase the number of pages 576 * in the read-ahead window. Otherwise, reset 577 * the read-ahead window to its smallest size. 578 */ 579 behind = atop(vaddr - fs.entry->start); 580 if (behind > VM_FAULT_READ_BEHIND) 581 behind = VM_FAULT_READ_BEHIND; 582 ahead = atop(fs.entry->end - vaddr) - 1; 583 era = fs.entry->read_ahead; 584 if (fs.pindex == fs.entry->next_read) { 585 nera = era + behind; 586 if (nera > VM_FAULT_READ_AHEAD_MAX) 587 nera = VM_FAULT_READ_AHEAD_MAX; 588 behind = 0; 589 if (ahead > nera) 590 ahead = nera; 591 if (era == VM_FAULT_READ_AHEAD_MAX) 592 vm_fault_cache_behind(&fs, 593 VM_FAULT_CACHE_BEHIND); 594 } else if (ahead > VM_FAULT_READ_AHEAD_MIN) 595 ahead = VM_FAULT_READ_AHEAD_MIN; 596 if (era != ahead) 597 fs.entry->read_ahead = ahead; 598 } 599 600 /* 601 * Call the pager to retrieve the data, if any, after 602 * releasing the lock on the map. We hold a ref on 603 * fs.object and the pages are exclusive busied. 604 */ 605 unlock_map(&fs); 606 607 if (fs.object->type == OBJT_VNODE) { 608 vp = fs.object->handle; 609 if (vp == fs.vp) 610 goto vnode_locked; 611 else if (fs.vp != NULL) { 612 vput(fs.vp); 613 fs.vp = NULL; 614 } 615 locked = VOP_ISLOCKED(vp); 616 617 if (locked != LK_EXCLUSIVE) 618 locked = LK_SHARED; 619 /* Do not sleep for vnode lock while fs.m is busy */ 620 error = vget(vp, locked | LK_CANRECURSE | 621 LK_NOWAIT, curthread); 622 if (error != 0) { 623 vhold(vp); 624 release_page(&fs); 625 unlock_and_deallocate(&fs); 626 error = vget(vp, locked | LK_RETRY | 627 LK_CANRECURSE, curthread); 628 vdrop(vp); 629 fs.vp = vp; 630 KASSERT(error == 0, 631 ("vm_fault: vget failed")); 632 goto RetryFault; 633 } 634 fs.vp = vp; 635 } 636vnode_locked: 637 KASSERT(fs.vp == NULL || !fs.map->system_map, 638 ("vm_fault: vnode-backed object mapped by system map")); 639 640 /* 641 * now we find out if any other pages should be paged 642 * in at this time this routine checks to see if the 643 * pages surrounding this fault reside in the same 644 * object as the page for this fault. If they do, 645 * then they are faulted in also into the object. The 646 * array "marray" returned contains an array of 647 * vm_page_t structs where one of them is the 648 * vm_page_t passed to the routine. The reqpage 649 * return value is the index into the marray for the 650 * vm_page_t passed to the routine. 651 * 652 * fs.m plus the additional pages are exclusive busied. 653 */ 654 faultcount = vm_fault_additional_pages( 655 fs.m, behind, ahead, marray, &reqpage); 656 657 rv = faultcount ? 658 vm_pager_get_pages(fs.object, marray, faultcount, 659 reqpage) : VM_PAGER_FAIL; 660 661 if (rv == VM_PAGER_OK) { 662 /* 663 * Found the page. Leave it busy while we play 664 * with it. 665 */ 666 667 /* 668 * Relookup in case pager changed page. Pager 669 * is responsible for disposition of old page 670 * if moved. 671 */ 672 fs.m = vm_page_lookup(fs.object, fs.pindex); 673 if (!fs.m) { 674 unlock_and_deallocate(&fs); 675 goto RetryFault; 676 } 677 678 hardfault++; 679 break; /* break to PAGE HAS BEEN FOUND */ 680 } 681 /* 682 * Remove the bogus page (which does not exist at this 683 * object/offset); before doing so, we must get back 684 * our object lock to preserve our invariant. 685 * 686 * Also wake up any other process that may want to bring 687 * in this page. 688 * 689 * If this is the top-level object, we must leave the 690 * busy page to prevent another process from rushing 691 * past us, and inserting the page in that object at 692 * the same time that we are. 693 */ 694 if (rv == VM_PAGER_ERROR) 695 printf("vm_fault: pager read error, pid %d (%s)\n", 696 curproc->p_pid, curproc->p_comm); 697 /* 698 * Data outside the range of the pager or an I/O error 699 */ 700 /* 701 * XXX - the check for kernel_map is a kludge to work 702 * around having the machine panic on a kernel space 703 * fault w/ I/O error. 704 */ 705 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 706 (rv == VM_PAGER_BAD)) { 707 vm_page_lock(fs.m); 708 vm_page_free(fs.m); 709 vm_page_unlock(fs.m); 710 fs.m = NULL; 711 unlock_and_deallocate(&fs); 712 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 713 } 714 if (fs.object != fs.first_object) { 715 vm_page_lock(fs.m); 716 vm_page_free(fs.m); 717 vm_page_unlock(fs.m); 718 fs.m = NULL; 719 /* 720 * XXX - we cannot just fall out at this 721 * point, m has been freed and is invalid! 722 */ 723 } 724 } 725 726 /* 727 * We get here if the object has default pager (or unwiring) 728 * or the pager doesn't have the page. 729 */ 730 if (fs.object == fs.first_object) 731 fs.first_m = fs.m; 732 733 /* 734 * Move on to the next object. Lock the next object before 735 * unlocking the current one. 736 */ 737 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 738 next_object = fs.object->backing_object; 739 if (next_object == NULL) { 740 /* 741 * If there's no object left, fill the page in the top 742 * object with zeros. 743 */ 744 if (fs.object != fs.first_object) { 745 vm_object_pip_wakeup(fs.object); 746 VM_OBJECT_WUNLOCK(fs.object); 747 748 fs.object = fs.first_object; 749 fs.pindex = fs.first_pindex; 750 fs.m = fs.first_m; 751 VM_OBJECT_WLOCK(fs.object); 752 } 753 fs.first_m = NULL; 754 755 /* 756 * Zero the page if necessary and mark it valid. 757 */ 758 if ((fs.m->flags & PG_ZERO) == 0) { 759 pmap_zero_page(fs.m); 760 } else { 761 PCPU_INC(cnt.v_ozfod); 762 } 763 PCPU_INC(cnt.v_zfod); 764 fs.m->valid = VM_PAGE_BITS_ALL; 765 /* Don't try to prefault neighboring pages. */ 766 faultcount = 1; 767 break; /* break to PAGE HAS BEEN FOUND */ 768 } else { 769 KASSERT(fs.object != next_object, 770 ("object loop %p", next_object)); 771 VM_OBJECT_WLOCK(next_object); 772 vm_object_pip_add(next_object, 1); 773 if (fs.object != fs.first_object) 774 vm_object_pip_wakeup(fs.object); 775 VM_OBJECT_WUNLOCK(fs.object); 776 fs.object = next_object; 777 } 778 } 779 780 vm_page_assert_xbusied(fs.m); 781 782 /* 783 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 784 * is held.] 785 */ 786 787 /* 788 * If the page is being written, but isn't already owned by the 789 * top-level object, we have to copy it into a new page owned by the 790 * top-level object. 791 */ 792 if (fs.object != fs.first_object) { 793 /* 794 * We only really need to copy if we want to write it. 795 */ 796 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 797 /* 798 * This allows pages to be virtually copied from a 799 * backing_object into the first_object, where the 800 * backing object has no other refs to it, and cannot 801 * gain any more refs. Instead of a bcopy, we just 802 * move the page from the backing object to the 803 * first object. Note that we must mark the page 804 * dirty in the first object so that it will go out 805 * to swap when needed. 806 */ 807 is_first_object_locked = FALSE; 808 if ( 809 /* 810 * Only one shadow object 811 */ 812 (fs.object->shadow_count == 1) && 813 /* 814 * No COW refs, except us 815 */ 816 (fs.object->ref_count == 1) && 817 /* 818 * No one else can look this object up 819 */ 820 (fs.object->handle == NULL) && 821 /* 822 * No other ways to look the object up 823 */ 824 ((fs.object->type == OBJT_DEFAULT) || 825 (fs.object->type == OBJT_SWAP)) && 826 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 827 /* 828 * We don't chase down the shadow chain 829 */ 830 fs.object == fs.first_object->backing_object) { 831 /* 832 * get rid of the unnecessary page 833 */ 834 vm_page_lock(fs.first_m); 835 vm_page_free(fs.first_m); 836 vm_page_unlock(fs.first_m); 837 /* 838 * grab the page and put it into the 839 * process'es object. The page is 840 * automatically made dirty. 841 */ 842 if (vm_page_rename(fs.m, fs.first_object, 843 fs.first_pindex)) { 844 unlock_and_deallocate(&fs); 845 goto RetryFault; 846 } 847 vm_page_xbusy(fs.m); 848 fs.first_m = fs.m; 849 fs.m = NULL; 850 PCPU_INC(cnt.v_cow_optim); 851 } else { 852 /* 853 * Oh, well, lets copy it. 854 */ 855 pmap_copy_page(fs.m, fs.first_m); 856 fs.first_m->valid = VM_PAGE_BITS_ALL; 857 if (wired && (fault_flags & 858 VM_FAULT_CHANGE_WIRING) == 0) { 859 vm_page_lock(fs.first_m); 860 vm_page_wire(fs.first_m); 861 vm_page_unlock(fs.first_m); 862 863 vm_page_lock(fs.m); 864 vm_page_unwire(fs.m, FALSE); 865 vm_page_unlock(fs.m); 866 } 867 /* 868 * We no longer need the old page or object. 869 */ 870 release_page(&fs); 871 } 872 /* 873 * fs.object != fs.first_object due to above 874 * conditional 875 */ 876 vm_object_pip_wakeup(fs.object); 877 VM_OBJECT_WUNLOCK(fs.object); 878 /* 879 * Only use the new page below... 880 */ 881 fs.object = fs.first_object; 882 fs.pindex = fs.first_pindex; 883 fs.m = fs.first_m; 884 if (!is_first_object_locked) 885 VM_OBJECT_WLOCK(fs.object); 886 PCPU_INC(cnt.v_cow_faults); 887 curthread->td_cow++; 888 } else { 889 prot &= ~VM_PROT_WRITE; 890 } 891 } 892 893 /* 894 * We must verify that the maps have not changed since our last 895 * lookup. 896 */ 897 if (!fs.lookup_still_valid) { 898 vm_object_t retry_object; 899 vm_pindex_t retry_pindex; 900 vm_prot_t retry_prot; 901 902 if (!vm_map_trylock_read(fs.map)) { 903 release_page(&fs); 904 unlock_and_deallocate(&fs); 905 goto RetryFault; 906 } 907 fs.lookup_still_valid = TRUE; 908 if (fs.map->timestamp != map_generation) { 909 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 910 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 911 912 /* 913 * If we don't need the page any longer, put it on the inactive 914 * list (the easiest thing to do here). If no one needs it, 915 * pageout will grab it eventually. 916 */ 917 if (result != KERN_SUCCESS) { 918 release_page(&fs); 919 unlock_and_deallocate(&fs); 920 921 /* 922 * If retry of map lookup would have blocked then 923 * retry fault from start. 924 */ 925 if (result == KERN_FAILURE) 926 goto RetryFault; 927 return (result); 928 } 929 if ((retry_object != fs.first_object) || 930 (retry_pindex != fs.first_pindex)) { 931 release_page(&fs); 932 unlock_and_deallocate(&fs); 933 goto RetryFault; 934 } 935 936 /* 937 * Check whether the protection has changed or the object has 938 * been copied while we left the map unlocked. Changing from 939 * read to write permission is OK - we leave the page 940 * write-protected, and catch the write fault. Changing from 941 * write to read permission means that we can't mark the page 942 * write-enabled after all. 943 */ 944 prot &= retry_prot; 945 } 946 } 947 /* 948 * If the page was filled by a pager, update the map entry's 949 * last read offset. Since the pager does not return the 950 * actual set of pages that it read, this update is based on 951 * the requested set. Typically, the requested and actual 952 * sets are the same. 953 * 954 * XXX The following assignment modifies the map 955 * without holding a write lock on it. 956 */ 957 if (hardfault) 958 fs.entry->next_read = fs.pindex + faultcount - reqpage; 959 960 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE); 961 vm_page_assert_xbusied(fs.m); 962 963 /* 964 * Page must be completely valid or it is not fit to 965 * map into user space. vm_pager_get_pages() ensures this. 966 */ 967 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 968 ("vm_fault: page %p partially invalid", fs.m)); 969 VM_OBJECT_WUNLOCK(fs.object); 970 971 /* 972 * Put this page into the physical map. We had to do the unlock above 973 * because pmap_enter() may sleep. We don't put the page 974 * back on the active queue until later so that the pageout daemon 975 * won't find it (yet). 976 */ 977 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, 978 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); 979 if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && 980 wired == 0) 981 vm_fault_prefault(&fs, vaddr, faultcount, reqpage); 982 VM_OBJECT_WLOCK(fs.object); 983 vm_page_lock(fs.m); 984 985 /* 986 * If the page is not wired down, then put it where the pageout daemon 987 * can find it. 988 */ 989 if (fault_flags & VM_FAULT_CHANGE_WIRING) { 990 if (wired) 991 vm_page_wire(fs.m); 992 else 993 vm_page_unwire(fs.m, 1); 994 } else 995 vm_page_activate(fs.m); 996 if (m_hold != NULL) { 997 *m_hold = fs.m; 998 vm_page_hold(fs.m); 999 } 1000 vm_page_unlock(fs.m); 1001 vm_page_xunbusy(fs.m); 1002 1003 /* 1004 * Unlock everything, and return 1005 */ 1006 unlock_and_deallocate(&fs); 1007 if (hardfault) { 1008 PCPU_INC(cnt.v_io_faults); 1009 curthread->td_ru.ru_majflt++; 1010 } else 1011 curthread->td_ru.ru_minflt++; 1012 1013 return (KERN_SUCCESS); 1014} 1015 1016/* 1017 * Speed up the reclamation of up to "distance" pages that precede the 1018 * faulting pindex within the first object of the shadow chain. 1019 */ 1020static void 1021vm_fault_cache_behind(const struct faultstate *fs, int distance) 1022{ 1023 vm_object_t first_object, object; 1024 vm_page_t m, m_prev; 1025 vm_pindex_t pindex; 1026 1027 object = fs->object; 1028 VM_OBJECT_ASSERT_WLOCKED(object); 1029 first_object = fs->first_object; 1030 if (first_object != object) { 1031 if (!VM_OBJECT_TRYWLOCK(first_object)) { 1032 VM_OBJECT_WUNLOCK(object); 1033 VM_OBJECT_WLOCK(first_object); 1034 VM_OBJECT_WLOCK(object); 1035 } 1036 } 1037 /* Neither fictitious nor unmanaged pages can be cached. */ 1038 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1039 if (fs->first_pindex < distance) 1040 pindex = 0; 1041 else 1042 pindex = fs->first_pindex - distance; 1043 if (pindex < OFF_TO_IDX(fs->entry->offset)) 1044 pindex = OFF_TO_IDX(fs->entry->offset); 1045 m = first_object != object ? fs->first_m : fs->m; 1046 vm_page_assert_xbusied(m); 1047 m_prev = vm_page_prev(m); 1048 while ((m = m_prev) != NULL && m->pindex >= pindex && 1049 m->valid == VM_PAGE_BITS_ALL) { 1050 m_prev = vm_page_prev(m); 1051 if (vm_page_busied(m)) 1052 continue; 1053 vm_page_lock(m); 1054 if (m->hold_count == 0 && m->wire_count == 0) { 1055 pmap_remove_all(m); 1056 vm_page_aflag_clear(m, PGA_REFERENCED); 1057 if (m->dirty != 0) 1058 vm_page_deactivate(m); 1059 else 1060 vm_page_cache(m); 1061 } 1062 vm_page_unlock(m); 1063 } 1064 } 1065 if (first_object != object) 1066 VM_OBJECT_WUNLOCK(first_object); 1067} 1068 1069/* 1070 * vm_fault_prefault provides a quick way of clustering 1071 * pagefaults into a processes address space. It is a "cousin" 1072 * of vm_map_pmap_enter, except it runs at page fault time instead 1073 * of mmap time. 1074 */ 1075static void 1076vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1077 int faultcount, int reqpage) 1078{ 1079 pmap_t pmap; 1080 vm_map_entry_t entry; 1081 vm_object_t backing_object, lobject; 1082 vm_offset_t addr, starta; 1083 vm_pindex_t pindex; 1084 vm_page_t m; 1085 int backward, forward, i; 1086 1087 pmap = fs->map->pmap; 1088 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1089 return; 1090 1091 if (faultcount > 0) { 1092 backward = reqpage; 1093 forward = faultcount - reqpage - 1; 1094 } else { 1095 backward = PFBAK; 1096 forward = PFFOR; 1097 } 1098 entry = fs->entry; 1099 1100 starta = addra - backward * PAGE_SIZE; 1101 if (starta < entry->start) { 1102 starta = entry->start; 1103 } else if (starta > addra) { 1104 starta = 0; 1105 } 1106 1107 /* 1108 * Generate the sequence of virtual addresses that are candidates for 1109 * prefaulting in an outward spiral from the faulting virtual address, 1110 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1111 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1112 * If the candidate address doesn't have a backing physical page, then 1113 * the loop immediately terminates. 1114 */ 1115 for (i = 0; i < 2 * imax(backward, forward); i++) { 1116 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1117 PAGE_SIZE); 1118 if (addr > addra + forward * PAGE_SIZE) 1119 addr = 0; 1120 1121 if (addr < starta || addr >= entry->end) 1122 continue; 1123 1124 if (!pmap_is_prefaultable(pmap, addr)) 1125 continue; 1126 1127 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1128 lobject = entry->object.vm_object; 1129 VM_OBJECT_RLOCK(lobject); 1130 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1131 lobject->type == OBJT_DEFAULT && 1132 (backing_object = lobject->backing_object) != NULL) { 1133 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1134 0, ("vm_fault_prefault: unaligned object offset")); 1135 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1136 VM_OBJECT_RLOCK(backing_object); 1137 VM_OBJECT_RUNLOCK(lobject); 1138 lobject = backing_object; 1139 } 1140 if (m == NULL) { 1141 VM_OBJECT_RUNLOCK(lobject); 1142 break; 1143 } 1144 if (m->valid == VM_PAGE_BITS_ALL && 1145 (m->flags & PG_FICTITIOUS) == 0) 1146 pmap_enter_quick(pmap, addr, m, entry->protection); 1147 VM_OBJECT_RUNLOCK(lobject); 1148 } 1149} 1150 1151/* 1152 * Hold each of the physical pages that are mapped by the specified range of 1153 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1154 * and allow the specified types of access, "prot". If all of the implied 1155 * pages are successfully held, then the number of held pages is returned 1156 * together with pointers to those pages in the array "ma". However, if any 1157 * of the pages cannot be held, -1 is returned. 1158 */ 1159int 1160vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1161 vm_prot_t prot, vm_page_t *ma, int max_count) 1162{ 1163 vm_offset_t end, va; 1164 vm_page_t *mp; 1165 int count; 1166 boolean_t pmap_failed; 1167 1168 if (len == 0) 1169 return (0); 1170 end = round_page(addr + len); 1171 addr = trunc_page(addr); 1172 1173 /* 1174 * Check for illegal addresses. 1175 */ 1176 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1177 return (-1); 1178 1179 if (atop(end - addr) > max_count) 1180 panic("vm_fault_quick_hold_pages: count > max_count"); 1181 count = atop(end - addr); 1182 1183 /* 1184 * Most likely, the physical pages are resident in the pmap, so it is 1185 * faster to try pmap_extract_and_hold() first. 1186 */ 1187 pmap_failed = FALSE; 1188 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1189 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1190 if (*mp == NULL) 1191 pmap_failed = TRUE; 1192 else if ((prot & VM_PROT_WRITE) != 0 && 1193 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1194 /* 1195 * Explicitly dirty the physical page. Otherwise, the 1196 * caller's changes may go unnoticed because they are 1197 * performed through an unmanaged mapping or by a DMA 1198 * operation. 1199 * 1200 * The object lock is not held here. 1201 * See vm_page_clear_dirty_mask(). 1202 */ 1203 vm_page_dirty(*mp); 1204 } 1205 } 1206 if (pmap_failed) { 1207 /* 1208 * One or more pages could not be held by the pmap. Either no 1209 * page was mapped at the specified virtual address or that 1210 * mapping had insufficient permissions. Attempt to fault in 1211 * and hold these pages. 1212 */ 1213 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1214 if (*mp == NULL && vm_fault_hold(map, va, prot, 1215 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1216 goto error; 1217 } 1218 return (count); 1219error: 1220 for (mp = ma; mp < ma + count; mp++) 1221 if (*mp != NULL) { 1222 vm_page_lock(*mp); 1223 vm_page_unhold(*mp); 1224 vm_page_unlock(*mp); 1225 } 1226 return (-1); 1227} 1228 1229/* 1230 * Routine: 1231 * vm_fault_copy_entry 1232 * Function: 1233 * Create new shadow object backing dst_entry with private copy of 1234 * all underlying pages. When src_entry is equal to dst_entry, 1235 * function implements COW for wired-down map entry. Otherwise, 1236 * it forks wired entry into dst_map. 1237 * 1238 * In/out conditions: 1239 * The source and destination maps must be locked for write. 1240 * The source map entry must be wired down (or be a sharing map 1241 * entry corresponding to a main map entry that is wired down). 1242 */ 1243void 1244vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1245 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1246 vm_ooffset_t *fork_charge) 1247{ 1248 vm_object_t backing_object, dst_object, object, src_object; 1249 vm_pindex_t dst_pindex, pindex, src_pindex; 1250 vm_prot_t access, prot; 1251 vm_offset_t vaddr; 1252 vm_page_t dst_m; 1253 vm_page_t src_m; 1254 boolean_t upgrade; 1255 1256#ifdef lint 1257 src_map++; 1258#endif /* lint */ 1259 1260 upgrade = src_entry == dst_entry; 1261 access = prot = dst_entry->protection; 1262 1263 src_object = src_entry->object.vm_object; 1264 src_pindex = OFF_TO_IDX(src_entry->offset); 1265 1266 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1267 dst_object = src_object; 1268 vm_object_reference(dst_object); 1269 } else { 1270 /* 1271 * Create the top-level object for the destination entry. (Doesn't 1272 * actually shadow anything - we copy the pages directly.) 1273 */ 1274 dst_object = vm_object_allocate(OBJT_DEFAULT, 1275 OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1276#if VM_NRESERVLEVEL > 0 1277 dst_object->flags |= OBJ_COLORED; 1278 dst_object->pg_color = atop(dst_entry->start); 1279#endif 1280 } 1281 1282 VM_OBJECT_WLOCK(dst_object); 1283 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1284 ("vm_fault_copy_entry: vm_object not NULL")); 1285 if (src_object != dst_object) { 1286 dst_entry->object.vm_object = dst_object; 1287 dst_entry->offset = 0; 1288 dst_object->charge = dst_entry->end - dst_entry->start; 1289 } 1290 if (fork_charge != NULL) { 1291 KASSERT(dst_entry->cred == NULL, 1292 ("vm_fault_copy_entry: leaked swp charge")); 1293 dst_object->cred = curthread->td_ucred; 1294 crhold(dst_object->cred); 1295 *fork_charge += dst_object->charge; 1296 } else if (dst_object->cred == NULL) { 1297 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1298 dst_entry)); 1299 dst_object->cred = dst_entry->cred; 1300 dst_entry->cred = NULL; 1301 } 1302 1303 /* 1304 * If not an upgrade, then enter the mappings in the pmap as 1305 * read and/or execute accesses. Otherwise, enter them as 1306 * write accesses. 1307 * 1308 * A writeable large page mapping is only created if all of 1309 * the constituent small page mappings are modified. Marking 1310 * PTEs as modified on inception allows promotion to happen 1311 * without taking potentially large number of soft faults. 1312 */ 1313 if (!upgrade) 1314 access &= ~VM_PROT_WRITE; 1315 1316 /* 1317 * Loop through all of the virtual pages within the entry's 1318 * range, copying each page from the source object to the 1319 * destination object. Since the source is wired, those pages 1320 * must exist. In contrast, the destination is pageable. 1321 * Since the destination object does share any backing storage 1322 * with the source object, all of its pages must be dirtied, 1323 * regardless of whether they can be written. 1324 */ 1325 for (vaddr = dst_entry->start, dst_pindex = 0; 1326 vaddr < dst_entry->end; 1327 vaddr += PAGE_SIZE, dst_pindex++) { 1328again: 1329 /* 1330 * Find the page in the source object, and copy it in. 1331 * Because the source is wired down, the page will be 1332 * in memory. 1333 */ 1334 if (src_object != dst_object) 1335 VM_OBJECT_RLOCK(src_object); 1336 object = src_object; 1337 pindex = src_pindex + dst_pindex; 1338 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1339 (backing_object = object->backing_object) != NULL) { 1340 /* 1341 * Unless the source mapping is read-only or 1342 * it is presently being upgraded from 1343 * read-only, the first object in the shadow 1344 * chain should provide all of the pages. In 1345 * other words, this loop body should never be 1346 * executed when the source mapping is already 1347 * read/write. 1348 */ 1349 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1350 upgrade, 1351 ("vm_fault_copy_entry: main object missing page")); 1352 1353 VM_OBJECT_RLOCK(backing_object); 1354 pindex += OFF_TO_IDX(object->backing_object_offset); 1355 if (object != dst_object) 1356 VM_OBJECT_RUNLOCK(object); 1357 object = backing_object; 1358 } 1359 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1360 1361 if (object != dst_object) { 1362 /* 1363 * Allocate a page in the destination object. 1364 */ 1365 dst_m = vm_page_alloc(dst_object, (src_object == 1366 dst_object ? src_pindex : 0) + dst_pindex, 1367 VM_ALLOC_NORMAL); 1368 if (dst_m == NULL) { 1369 VM_OBJECT_WUNLOCK(dst_object); 1370 VM_OBJECT_RUNLOCK(object); 1371 VM_WAIT; 1372 VM_OBJECT_WLOCK(dst_object); 1373 goto again; 1374 } 1375 pmap_copy_page(src_m, dst_m); 1376 VM_OBJECT_RUNLOCK(object); 1377 dst_m->valid = VM_PAGE_BITS_ALL; 1378 dst_m->dirty = VM_PAGE_BITS_ALL; 1379 } else { 1380 dst_m = src_m; 1381 if (vm_page_sleep_if_busy(dst_m, "fltupg")) 1382 goto again; 1383 vm_page_xbusy(dst_m); 1384 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL, 1385 ("invalid dst page %p", dst_m)); 1386 } 1387 VM_OBJECT_WUNLOCK(dst_object); 1388 1389 /* 1390 * Enter it in the pmap. If a wired, copy-on-write 1391 * mapping is being replaced by a write-enabled 1392 * mapping, then wire that new mapping. 1393 */ 1394 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 1395 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 1396 1397 /* 1398 * Mark it no longer busy, and put it on the active list. 1399 */ 1400 VM_OBJECT_WLOCK(dst_object); 1401 1402 if (upgrade) { 1403 if (src_m != dst_m) { 1404 vm_page_lock(src_m); 1405 vm_page_unwire(src_m, 0); 1406 vm_page_unlock(src_m); 1407 vm_page_lock(dst_m); 1408 vm_page_wire(dst_m); 1409 vm_page_unlock(dst_m); 1410 } else { 1411 KASSERT(dst_m->wire_count > 0, 1412 ("dst_m %p is not wired", dst_m)); 1413 } 1414 } else { 1415 vm_page_lock(dst_m); 1416 vm_page_activate(dst_m); 1417 vm_page_unlock(dst_m); 1418 } 1419 vm_page_xunbusy(dst_m); 1420 } 1421 VM_OBJECT_WUNLOCK(dst_object); 1422 if (upgrade) { 1423 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1424 vm_object_deallocate(src_object); 1425 } 1426} 1427 1428 1429/* 1430 * This routine checks around the requested page for other pages that 1431 * might be able to be faulted in. This routine brackets the viable 1432 * pages for the pages to be paged in. 1433 * 1434 * Inputs: 1435 * m, rbehind, rahead 1436 * 1437 * Outputs: 1438 * marray (array of vm_page_t), reqpage (index of requested page) 1439 * 1440 * Return value: 1441 * number of pages in marray 1442 */ 1443static int 1444vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage) 1445 vm_page_t m; 1446 int rbehind; 1447 int rahead; 1448 vm_page_t *marray; 1449 int *reqpage; 1450{ 1451 int i,j; 1452 vm_object_t object; 1453 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1454 vm_page_t rtm; 1455 int cbehind, cahead; 1456 1457 VM_OBJECT_ASSERT_WLOCKED(m->object); 1458 1459 object = m->object; 1460 pindex = m->pindex; 1461 cbehind = cahead = 0; 1462 1463 /* 1464 * if the requested page is not available, then give up now 1465 */ 1466 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1467 return 0; 1468 } 1469 1470 if ((cbehind == 0) && (cahead == 0)) { 1471 *reqpage = 0; 1472 marray[0] = m; 1473 return 1; 1474 } 1475 1476 if (rahead > cahead) { 1477 rahead = cahead; 1478 } 1479 1480 if (rbehind > cbehind) { 1481 rbehind = cbehind; 1482 } 1483 1484 /* 1485 * scan backward for the read behind pages -- in memory 1486 */ 1487 if (pindex > 0) { 1488 if (rbehind > pindex) { 1489 rbehind = pindex; 1490 startpindex = 0; 1491 } else { 1492 startpindex = pindex - rbehind; 1493 } 1494 1495 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL && 1496 rtm->pindex >= startpindex) 1497 startpindex = rtm->pindex + 1; 1498 1499 /* tpindex is unsigned; beware of numeric underflow. */ 1500 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex && 1501 tpindex < pindex; i++, tpindex--) { 1502 1503 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1504 VM_ALLOC_IFNOTCACHED); 1505 if (rtm == NULL) { 1506 /* 1507 * Shift the allocated pages to the 1508 * beginning of the array. 1509 */ 1510 for (j = 0; j < i; j++) { 1511 marray[j] = marray[j + tpindex + 1 - 1512 startpindex]; 1513 } 1514 break; 1515 } 1516 1517 marray[tpindex - startpindex] = rtm; 1518 } 1519 } else { 1520 startpindex = 0; 1521 i = 0; 1522 } 1523 1524 marray[i] = m; 1525 /* page offset of the required page */ 1526 *reqpage = i; 1527 1528 tpindex = pindex + 1; 1529 i++; 1530 1531 /* 1532 * scan forward for the read ahead pages 1533 */ 1534 endpindex = tpindex + rahead; 1535 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex) 1536 endpindex = rtm->pindex; 1537 if (endpindex > object->size) 1538 endpindex = object->size; 1539 1540 for (; tpindex < endpindex; i++, tpindex++) { 1541 1542 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL | 1543 VM_ALLOC_IFNOTCACHED); 1544 if (rtm == NULL) { 1545 break; 1546 } 1547 1548 marray[i] = rtm; 1549 } 1550 1551 /* return number of pages */ 1552 return i; 1553} 1554 1555/* 1556 * Block entry into the machine-independent layer's page fault handler by 1557 * the calling thread. Subsequent calls to vm_fault() by that thread will 1558 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1559 * spurious page faults. 1560 */ 1561int 1562vm_fault_disable_pagefaults(void) 1563{ 1564 1565 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1566} 1567 1568void 1569vm_fault_enable_pagefaults(int save) 1570{ 1571 1572 curthread_pflags_restore(save); 1573} 1574