2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
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.
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
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
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.
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.
59 * Carnegie Mellon requests users of this software to return to
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
71 * Page fault handling module.
74 #include <sys/cdefs.h>
75 __FBSDID("$FreeBSD$");
77 #include "opt_ktrace.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
86 #include <sys/racct.h>
87 #include <sys/resourcevar.h>
88 #include <sys/rwlock.h>
89 #include <sys/sysctl.h>
90 #include <sys/vmmeter.h>
91 #include <sys/vnode.h>
93 #include <sys/ktrace.h>
97 #include <vm/vm_param.h>
99 #include <vm/vm_map.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/vm_reserv.h>
111 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
112 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
114 #define VM_FAULT_DONTNEED_MIN 1048576
121 vm_object_t first_object;
122 vm_pindex_t first_pindex;
124 vm_map_entry_t entry;
125 int lookup_still_valid;
129 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
131 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
132 int backward, int forward);
135 release_page(struct faultstate *fs)
138 vm_page_xunbusy(fs->m);
140 vm_page_deactivate(fs->m);
141 vm_page_unlock(fs->m);
146 unlock_map(struct faultstate *fs)
149 if (fs->lookup_still_valid) {
150 vm_map_lookup_done(fs->map, fs->entry);
151 fs->lookup_still_valid = FALSE;
156 unlock_and_deallocate(struct faultstate *fs)
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);
170 vm_object_deallocate(fs->first_object);
172 if (fs->vp != NULL) {
179 vm_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)
182 boolean_t need_dirty;
184 if (((prot & VM_PROT_WRITE) == 0 &&
185 (fault_flags & VM_FAULT_DIRTY) == 0) ||
186 (m->oflags & VPO_UNMANAGED) != 0)
189 VM_OBJECT_ASSERT_LOCKED(m->object);
191 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
192 (fault_flags & VM_FAULT_WIRE) == 0) ||
193 (fault_flags & VM_FAULT_DIRTY) != 0;
196 vm_object_set_writeable_dirty(m->object);
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.
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.
218 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
220 m->oflags |= VPO_NOSYNC;
223 m->oflags &= ~VPO_NOSYNC;
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.
231 * Also tell the backing pager, if any, that it should remove
232 * any swap backing since the page is now dirty.
239 vm_pager_page_unswapped(m);
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.
250 * NOTE: the given address should be truncated to the
251 * proper page address.
253 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
254 * a standard error specifying why the fault is fatal is returned.
256 * The map in question must be referenced, and remains so.
257 * Caller may hold no locks.
260 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
267 if ((td->td_pflags & TDP_NOFAULTING) != 0)
268 return (KERN_PROTECTION_FAILURE);
270 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
271 ktrfault(vaddr, fault_type);
273 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
276 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
283 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
284 int fault_flags, vm_page_t *m_hold)
287 int alloc_req, era, faultcount, nera, result;
288 boolean_t growstack, is_first_object_locked, wired;
290 vm_object_t next_object;
292 struct faultstate fs;
295 int ahead, behind, cluster_offset, error, locked;
299 PCPU_INC(cnt.v_vm_faults);
306 * Find the backing store object and offset into it to begin the
310 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
311 &fs.first_object, &fs.first_pindex, &prot, &wired);
312 if (result != KERN_SUCCESS) {
313 if (growstack && result == KERN_INVALID_ADDRESS &&
315 result = vm_map_growstack(curproc, vaddr);
316 if (result != KERN_SUCCESS)
317 return (KERN_FAILURE);
324 map_generation = fs.map->timestamp;
326 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
327 panic("vm_fault: fault on nofault entry, addr: %lx",
331 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
332 fs.entry->wiring_thread != curthread) {
333 vm_map_unlock_read(fs.map);
335 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
336 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
341 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
342 vm_map_unlock_and_wait(fs.map, 0);
344 vm_map_unlock(fs.map);
349 fault_type = prot | (fault_type & VM_PROT_COPY);
351 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
352 ("!wired && VM_FAULT_WIRE"));
354 if (fs.vp == NULL /* avoid locked vnode leak */ &&
355 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
356 /* avoid calling vm_object_set_writeable_dirty() */
357 ((prot & VM_PROT_WRITE) == 0 ||
358 (fs.first_object->type != OBJT_VNODE &&
359 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
360 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
361 VM_OBJECT_RLOCK(fs.first_object);
362 if ((prot & VM_PROT_WRITE) != 0 &&
363 (fs.first_object->type == OBJT_VNODE ||
364 (fs.first_object->flags & OBJ_TMPFS_NODE) != 0) &&
365 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
367 m = vm_page_lookup(fs.first_object, fs.first_pindex);
368 /* A busy page can be mapped for read|execute access. */
369 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
370 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
372 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
373 fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
375 if (result != KERN_SUCCESS)
377 if (m_hold != NULL) {
383 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
385 VM_OBJECT_RUNLOCK(fs.first_object);
387 vm_fault_prefault(&fs, vaddr, PFBAK, PFFOR);
388 vm_map_lookup_done(fs.map, fs.entry);
389 curthread->td_ru.ru_minflt++;
390 return (KERN_SUCCESS);
392 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
393 VM_OBJECT_RUNLOCK(fs.first_object);
394 VM_OBJECT_WLOCK(fs.first_object);
397 VM_OBJECT_WLOCK(fs.first_object);
401 * Make a reference to this object to prevent its disposal while we
402 * are messing with it. Once we have the reference, the map is free
403 * to be diddled. Since objects reference their shadows (and copies),
404 * they will stay around as well.
406 * Bump the paging-in-progress count to prevent size changes (e.g.
407 * truncation operations) during I/O. This must be done after
408 * obtaining the vnode lock in order to avoid possible deadlocks.
410 vm_object_reference_locked(fs.first_object);
411 vm_object_pip_add(fs.first_object, 1);
413 fs.lookup_still_valid = TRUE;
418 * Search for the page at object/offset.
420 fs.object = fs.first_object;
421 fs.pindex = fs.first_pindex;
424 * If the object is dead, we stop here
426 if (fs.object->flags & OBJ_DEAD) {
427 unlock_and_deallocate(&fs);
428 return (KERN_PROTECTION_FAILURE);
432 * See if page is resident
434 fs.m = vm_page_lookup(fs.object, fs.pindex);
437 * Wait/Retry if the page is busy. We have to do this
438 * if the page is either exclusive or shared busy
439 * because the vm_pager may be using read busy for
440 * pageouts (and even pageins if it is the vnode
441 * pager), and we could end up trying to pagein and
442 * pageout the same page simultaneously.
444 * We can theoretically allow the busy case on a read
445 * fault if the page is marked valid, but since such
446 * pages are typically already pmap'd, putting that
447 * special case in might be more effort then it is
448 * worth. We cannot under any circumstances mess
449 * around with a shared busied page except, perhaps,
452 if (vm_page_busied(fs.m)) {
454 * Reference the page before unlocking and
455 * sleeping so that the page daemon is less
456 * likely to reclaim it.
458 vm_page_aflag_set(fs.m, PGA_REFERENCED);
459 if (fs.object != fs.first_object) {
460 if (!VM_OBJECT_TRYWLOCK(
462 VM_OBJECT_WUNLOCK(fs.object);
463 VM_OBJECT_WLOCK(fs.first_object);
464 VM_OBJECT_WLOCK(fs.object);
466 vm_page_lock(fs.first_m);
467 vm_page_free(fs.first_m);
468 vm_page_unlock(fs.first_m);
469 vm_object_pip_wakeup(fs.first_object);
470 VM_OBJECT_WUNLOCK(fs.first_object);
474 if (fs.m == vm_page_lookup(fs.object,
476 vm_page_sleep_if_busy(fs.m, "vmpfw");
478 vm_object_pip_wakeup(fs.object);
479 VM_OBJECT_WUNLOCK(fs.object);
480 PCPU_INC(cnt.v_intrans);
481 vm_object_deallocate(fs.first_object);
485 vm_page_remque(fs.m);
486 vm_page_unlock(fs.m);
489 * Mark page busy for other processes, and the
490 * pagedaemon. If it still isn't completely valid
491 * (readable), jump to readrest, else break-out ( we
495 if (fs.m->valid != VM_PAGE_BITS_ALL)
501 * Page is not resident. If this is the search termination
502 * or the pager might contain the page, allocate a new page.
503 * Default objects are zero-fill, there is no real pager.
505 if (fs.object->type != OBJT_DEFAULT ||
506 fs.object == fs.first_object) {
507 if (fs.pindex >= fs.object->size) {
508 unlock_and_deallocate(&fs);
509 return (KERN_PROTECTION_FAILURE);
513 * Allocate a new page for this object/offset pair.
515 * Unlocked read of the p_flag is harmless. At
516 * worst, the P_KILLED might be not observed
517 * there, and allocation can fail, causing
518 * restart and new reading of the p_flag.
521 if (!vm_page_count_severe() || P_KILLED(curproc)) {
522 #if VM_NRESERVLEVEL > 0
523 vm_object_color(fs.object, atop(vaddr) -
526 alloc_req = P_KILLED(curproc) ?
527 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
528 if (fs.object->type != OBJT_VNODE &&
529 fs.object->backing_object == NULL)
530 alloc_req |= VM_ALLOC_ZERO;
531 fs.m = vm_page_alloc(fs.object, fs.pindex,
535 unlock_and_deallocate(&fs);
538 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
544 * We have found a valid page or we have allocated a new page.
545 * The page thus may not be valid or may not be entirely
548 * Attempt to fault-in the page if there is a chance that the
549 * pager has it, and potentially fault in additional pages
550 * at the same time. For default objects simply provide
553 if (fs.object->type != OBJT_DEFAULT) {
555 u_char behavior = vm_map_entry_behavior(fs.entry);
557 era = fs.entry->read_ahead;
558 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
563 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
565 nera = VM_FAULT_READ_AHEAD_MAX;
567 if (fs.pindex == fs.entry->next_read)
568 vm_fault_dontneed(&fs, vaddr, ahead);
569 } else if (fs.pindex == fs.entry->next_read) {
571 * This is a sequential fault. Arithmetically
572 * increase the requested number of pages in
573 * the read-ahead window. The requested
574 * number of pages is "# of sequential faults
575 * x (read ahead min + 1) + read ahead min"
578 nera = VM_FAULT_READ_AHEAD_MIN;
581 if (nera > VM_FAULT_READ_AHEAD_MAX)
582 nera = VM_FAULT_READ_AHEAD_MAX;
585 if (era == VM_FAULT_READ_AHEAD_MAX)
586 vm_fault_dontneed(&fs, vaddr, ahead);
589 * This is a non-sequential fault. Request a
590 * cluster of pages that is aligned to a
591 * VM_FAULT_READ_DEFAULT page offset boundary
592 * within the object. Alignment to a page
593 * offset boundary is more likely to coincide
594 * with the underlying file system block than
595 * alignment to a virtual address boundary.
597 cluster_offset = fs.pindex %
598 VM_FAULT_READ_DEFAULT;
599 behind = ulmin(cluster_offset,
600 atop(vaddr - fs.entry->start));
602 ahead = VM_FAULT_READ_DEFAULT - 1 -
605 ahead = ulmin(ahead, atop(fs.entry->end - vaddr) - 1);
607 fs.entry->read_ahead = nera;
610 * Call the pager to retrieve the data, if any, after
611 * releasing the lock on the map. We hold a ref on
612 * fs.object and the pages are exclusive busied.
616 if (fs.object->type == OBJT_VNODE) {
617 vp = fs.object->handle;
620 else if (fs.vp != NULL) {
624 locked = VOP_ISLOCKED(vp);
626 if (locked != LK_EXCLUSIVE)
628 /* Do not sleep for vnode lock while fs.m is busy */
629 error = vget(vp, locked | LK_CANRECURSE |
630 LK_NOWAIT, curthread);
634 unlock_and_deallocate(&fs);
635 error = vget(vp, locked | LK_RETRY |
636 LK_CANRECURSE, curthread);
640 ("vm_fault: vget failed"));
646 KASSERT(fs.vp == NULL || !fs.map->system_map,
647 ("vm_fault: vnode-backed object mapped by system map"));
650 * Page in the requested page and hint the pager,
651 * that it may bring up surrounding pages.
653 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
655 if (rv == VM_PAGER_OK) {
656 faultcount = behind + 1 + ahead;
658 break; /* break to PAGE HAS BEEN FOUND */
661 * Remove the bogus page (which does not exist at this
662 * object/offset); before doing so, we must get back
663 * our object lock to preserve our invariant.
665 * Also wake up any other process that may want to bring
668 * If this is the top-level object, we must leave the
669 * busy page to prevent another process from rushing
670 * past us, and inserting the page in that object at
671 * the same time that we are.
673 if (rv == VM_PAGER_ERROR)
674 printf("vm_fault: pager read error, pid %d (%s)\n",
675 curproc->p_pid, curproc->p_comm);
677 * Data outside the range of the pager or an I/O error
680 * XXX - the check for kernel_map is a kludge to work
681 * around having the machine panic on a kernel space
682 * fault w/ I/O error.
684 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
685 (rv == VM_PAGER_BAD)) {
688 vm_page_unlock(fs.m);
690 unlock_and_deallocate(&fs);
691 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
693 if (fs.object != fs.first_object) {
696 vm_page_unlock(fs.m);
699 * XXX - we cannot just fall out at this
700 * point, m has been freed and is invalid!
706 * We get here if the object has default pager (or unwiring)
707 * or the pager doesn't have the page.
709 if (fs.object == fs.first_object)
713 * Move on to the next object. Lock the next object before
714 * unlocking the current one.
716 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
717 next_object = fs.object->backing_object;
718 if (next_object == NULL) {
720 * If there's no object left, fill the page in the top
723 if (fs.object != fs.first_object) {
724 vm_object_pip_wakeup(fs.object);
725 VM_OBJECT_WUNLOCK(fs.object);
727 fs.object = fs.first_object;
728 fs.pindex = fs.first_pindex;
730 VM_OBJECT_WLOCK(fs.object);
735 * Zero the page if necessary and mark it valid.
737 if ((fs.m->flags & PG_ZERO) == 0) {
738 pmap_zero_page(fs.m);
740 PCPU_INC(cnt.v_ozfod);
742 PCPU_INC(cnt.v_zfod);
743 fs.m->valid = VM_PAGE_BITS_ALL;
744 /* Don't try to prefault neighboring pages. */
746 break; /* break to PAGE HAS BEEN FOUND */
748 KASSERT(fs.object != next_object,
749 ("object loop %p", next_object));
750 VM_OBJECT_WLOCK(next_object);
751 vm_object_pip_add(next_object, 1);
752 if (fs.object != fs.first_object)
753 vm_object_pip_wakeup(fs.object);
754 VM_OBJECT_WUNLOCK(fs.object);
755 fs.object = next_object;
759 vm_page_assert_xbusied(fs.m);
762 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
767 * If the page is being written, but isn't already owned by the
768 * top-level object, we have to copy it into a new page owned by the
771 if (fs.object != fs.first_object) {
773 * We only really need to copy if we want to write it.
775 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
777 * This allows pages to be virtually copied from a
778 * backing_object into the first_object, where the
779 * backing object has no other refs to it, and cannot
780 * gain any more refs. Instead of a bcopy, we just
781 * move the page from the backing object to the
782 * first object. Note that we must mark the page
783 * dirty in the first object so that it will go out
784 * to swap when needed.
786 is_first_object_locked = FALSE;
789 * Only one shadow object
791 (fs.object->shadow_count == 1) &&
793 * No COW refs, except us
795 (fs.object->ref_count == 1) &&
797 * No one else can look this object up
799 (fs.object->handle == NULL) &&
801 * No other ways to look the object up
803 ((fs.object->type == OBJT_DEFAULT) ||
804 (fs.object->type == OBJT_SWAP)) &&
805 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
807 * We don't chase down the shadow chain
809 fs.object == fs.first_object->backing_object) {
811 * get rid of the unnecessary page
813 vm_page_lock(fs.first_m);
814 vm_page_remove(fs.first_m);
815 vm_page_unlock(fs.first_m);
817 * grab the page and put it into the
818 * process'es object. The page is
819 * automatically made dirty.
821 if (vm_page_rename(fs.m, fs.first_object,
823 VM_OBJECT_WUNLOCK(fs.first_object);
824 unlock_and_deallocate(&fs);
827 vm_page_lock(fs.first_m);
828 vm_page_free(fs.first_m);
829 vm_page_unlock(fs.first_m);
830 #if VM_NRESERVLEVEL > 0
832 * Rename the reservation.
834 vm_reserv_rename(fs.m, fs.first_object,
835 fs.object, OFF_TO_IDX(
836 fs.first_object->backing_object_offset));
841 PCPU_INC(cnt.v_cow_optim);
844 * Oh, well, lets copy it.
846 pmap_copy_page(fs.m, fs.first_m);
847 fs.first_m->valid = VM_PAGE_BITS_ALL;
848 if (wired && (fault_flags &
849 VM_FAULT_WIRE) == 0) {
850 vm_page_lock(fs.first_m);
851 vm_page_wire(fs.first_m);
852 vm_page_unlock(fs.first_m);
855 vm_page_unwire(fs.m, PQ_INACTIVE);
856 vm_page_unlock(fs.m);
859 * We no longer need the old page or object.
864 * fs.object != fs.first_object due to above
867 vm_object_pip_wakeup(fs.object);
868 VM_OBJECT_WUNLOCK(fs.object);
870 * Only use the new page below...
872 fs.object = fs.first_object;
873 fs.pindex = fs.first_pindex;
875 if (!is_first_object_locked)
876 VM_OBJECT_WLOCK(fs.object);
877 PCPU_INC(cnt.v_cow_faults);
880 prot &= ~VM_PROT_WRITE;
885 * We must verify that the maps have not changed since our last
888 if (!fs.lookup_still_valid) {
889 vm_object_t retry_object;
890 vm_pindex_t retry_pindex;
891 vm_prot_t retry_prot;
893 if (!vm_map_trylock_read(fs.map)) {
895 unlock_and_deallocate(&fs);
898 fs.lookup_still_valid = TRUE;
899 if (fs.map->timestamp != map_generation) {
900 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
901 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
904 * If we don't need the page any longer, put it on the inactive
905 * list (the easiest thing to do here). If no one needs it,
906 * pageout will grab it eventually.
908 if (result != KERN_SUCCESS) {
910 unlock_and_deallocate(&fs);
913 * If retry of map lookup would have blocked then
914 * retry fault from start.
916 if (result == KERN_FAILURE)
920 if ((retry_object != fs.first_object) ||
921 (retry_pindex != fs.first_pindex)) {
923 unlock_and_deallocate(&fs);
928 * Check whether the protection has changed or the object has
929 * been copied while we left the map unlocked. Changing from
930 * read to write permission is OK - we leave the page
931 * write-protected, and catch the write fault. Changing from
932 * write to read permission means that we can't mark the page
933 * write-enabled after all.
939 * If the page was filled by a pager, update the map entry's
942 * XXX The following assignment modifies the map
943 * without holding a write lock on it.
946 fs.entry->next_read = fs.pindex + ahead + 1;
948 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
949 vm_page_assert_xbusied(fs.m);
952 * Page must be completely valid or it is not fit to
953 * map into user space. vm_pager_get_pages() ensures this.
955 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
956 ("vm_fault: page %p partially invalid", fs.m));
957 VM_OBJECT_WUNLOCK(fs.object);
960 * Put this page into the physical map. We had to do the unlock above
961 * because pmap_enter() may sleep. We don't put the page
962 * back on the active queue until later so that the pageout daemon
963 * won't find it (yet).
965 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
966 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
967 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
969 vm_fault_prefault(&fs, vaddr,
970 faultcount > 0 ? behind : PFBAK,
971 faultcount > 0 ? ahead : PFFOR);
972 VM_OBJECT_WLOCK(fs.object);
976 * If the page is not wired down, then put it where the pageout daemon
979 if ((fault_flags & VM_FAULT_WIRE) != 0) {
980 KASSERT(wired, ("VM_FAULT_WIRE && !wired"));
983 vm_page_activate(fs.m);
984 if (m_hold != NULL) {
988 vm_page_unlock(fs.m);
989 vm_page_xunbusy(fs.m);
992 * Unlock everything, and return
994 unlock_and_deallocate(&fs);
996 PCPU_INC(cnt.v_io_faults);
997 curthread->td_ru.ru_majflt++;
999 if (racct_enable && fs.object->type == OBJT_VNODE) {
1001 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1002 racct_add_force(curproc, RACCT_WRITEBPS,
1003 PAGE_SIZE + behind * PAGE_SIZE);
1004 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1006 racct_add_force(curproc, RACCT_READBPS,
1007 PAGE_SIZE + ahead * PAGE_SIZE);
1008 racct_add_force(curproc, RACCT_READIOPS, 1);
1010 PROC_UNLOCK(curproc);
1014 curthread->td_ru.ru_minflt++;
1016 return (KERN_SUCCESS);
1020 * Speed up the reclamation of pages that precede the faulting pindex within
1021 * the first object of the shadow chain. Essentially, perform the equivalent
1022 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1023 * the faulting pindex by the cluster size when the pages read by vm_fault()
1024 * cross a cluster-size boundary. The cluster size is the greater of the
1025 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1027 * When "fs->first_object" is a shadow object, the pages in the backing object
1028 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1029 * function must only be concerned with pages in the first object.
1032 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1034 vm_map_entry_t entry;
1035 vm_object_t first_object, object;
1036 vm_offset_t end, start;
1037 vm_page_t m, m_next;
1038 vm_pindex_t pend, pstart;
1041 object = fs->object;
1042 VM_OBJECT_ASSERT_WLOCKED(object);
1043 first_object = fs->first_object;
1044 if (first_object != object) {
1045 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1046 VM_OBJECT_WUNLOCK(object);
1047 VM_OBJECT_WLOCK(first_object);
1048 VM_OBJECT_WLOCK(object);
1051 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1052 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1053 size = VM_FAULT_DONTNEED_MIN;
1054 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1055 size = pagesizes[1];
1056 end = rounddown2(vaddr, size);
1057 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1058 (entry = fs->entry)->start < end) {
1059 if (end - entry->start < size)
1060 start = entry->start;
1063 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1064 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1066 m_next = vm_page_find_least(first_object, pstart);
1067 pend = OFF_TO_IDX(entry->offset) + atop(end -
1069 while ((m = m_next) != NULL && m->pindex < pend) {
1070 m_next = TAILQ_NEXT(m, listq);
1071 if (m->valid != VM_PAGE_BITS_ALL ||
1076 * Don't clear PGA_REFERENCED, since it would
1077 * likely represent a reference by a different
1080 * Typically, at this point, prefetched pages
1081 * are still in the inactive queue. Only
1082 * pages that triggered page faults are in the
1086 vm_page_deactivate(m);
1091 if (first_object != object)
1092 VM_OBJECT_WUNLOCK(first_object);
1096 * vm_fault_prefault provides a quick way of clustering
1097 * pagefaults into a processes address space. It is a "cousin"
1098 * of vm_map_pmap_enter, except it runs at page fault time instead
1102 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1103 int backward, int forward)
1106 vm_map_entry_t entry;
1107 vm_object_t backing_object, lobject;
1108 vm_offset_t addr, starta;
1113 pmap = fs->map->pmap;
1114 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1119 starta = addra - backward * PAGE_SIZE;
1120 if (starta < entry->start) {
1121 starta = entry->start;
1122 } else if (starta > addra) {
1127 * Generate the sequence of virtual addresses that are candidates for
1128 * prefaulting in an outward spiral from the faulting virtual address,
1129 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1130 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1131 * If the candidate address doesn't have a backing physical page, then
1132 * the loop immediately terminates.
1134 for (i = 0; i < 2 * imax(backward, forward); i++) {
1135 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1137 if (addr > addra + forward * PAGE_SIZE)
1140 if (addr < starta || addr >= entry->end)
1143 if (!pmap_is_prefaultable(pmap, addr))
1146 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1147 lobject = entry->object.vm_object;
1148 VM_OBJECT_RLOCK(lobject);
1149 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1150 lobject->type == OBJT_DEFAULT &&
1151 (backing_object = lobject->backing_object) != NULL) {
1152 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1153 0, ("vm_fault_prefault: unaligned object offset"));
1154 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1155 VM_OBJECT_RLOCK(backing_object);
1156 VM_OBJECT_RUNLOCK(lobject);
1157 lobject = backing_object;
1160 VM_OBJECT_RUNLOCK(lobject);
1163 if (m->valid == VM_PAGE_BITS_ALL &&
1164 (m->flags & PG_FICTITIOUS) == 0)
1165 pmap_enter_quick(pmap, addr, m, entry->protection);
1166 VM_OBJECT_RUNLOCK(lobject);
1171 * Hold each of the physical pages that are mapped by the specified range of
1172 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1173 * and allow the specified types of access, "prot". If all of the implied
1174 * pages are successfully held, then the number of held pages is returned
1175 * together with pointers to those pages in the array "ma". However, if any
1176 * of the pages cannot be held, -1 is returned.
1179 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1180 vm_prot_t prot, vm_page_t *ma, int max_count)
1182 vm_offset_t end, va;
1185 boolean_t pmap_failed;
1189 end = round_page(addr + len);
1190 addr = trunc_page(addr);
1193 * Check for illegal addresses.
1195 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1198 if (atop(end - addr) > max_count)
1199 panic("vm_fault_quick_hold_pages: count > max_count");
1200 count = atop(end - addr);
1203 * Most likely, the physical pages are resident in the pmap, so it is
1204 * faster to try pmap_extract_and_hold() first.
1206 pmap_failed = FALSE;
1207 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1208 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1211 else if ((prot & VM_PROT_WRITE) != 0 &&
1212 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1214 * Explicitly dirty the physical page. Otherwise, the
1215 * caller's changes may go unnoticed because they are
1216 * performed through an unmanaged mapping or by a DMA
1219 * The object lock is not held here.
1220 * See vm_page_clear_dirty_mask().
1227 * One or more pages could not be held by the pmap. Either no
1228 * page was mapped at the specified virtual address or that
1229 * mapping had insufficient permissions. Attempt to fault in
1230 * and hold these pages.
1232 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1233 if (*mp == NULL && vm_fault_hold(map, va, prot,
1234 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1239 for (mp = ma; mp < ma + count; mp++)
1242 vm_page_unhold(*mp);
1243 vm_page_unlock(*mp);
1250 * vm_fault_copy_entry
1252 * Create new shadow object backing dst_entry with private copy of
1253 * all underlying pages. When src_entry is equal to dst_entry,
1254 * function implements COW for wired-down map entry. Otherwise,
1255 * it forks wired entry into dst_map.
1257 * In/out conditions:
1258 * The source and destination maps must be locked for write.
1259 * The source map entry must be wired down (or be a sharing map
1260 * entry corresponding to a main map entry that is wired down).
1263 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1264 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1265 vm_ooffset_t *fork_charge)
1267 vm_object_t backing_object, dst_object, object, src_object;
1268 vm_pindex_t dst_pindex, pindex, src_pindex;
1269 vm_prot_t access, prot;
1279 upgrade = src_entry == dst_entry;
1280 access = prot = dst_entry->protection;
1282 src_object = src_entry->object.vm_object;
1283 src_pindex = OFF_TO_IDX(src_entry->offset);
1285 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1286 dst_object = src_object;
1287 vm_object_reference(dst_object);
1290 * Create the top-level object for the destination entry. (Doesn't
1291 * actually shadow anything - we copy the pages directly.)
1293 dst_object = vm_object_allocate(OBJT_DEFAULT,
1294 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1295 #if VM_NRESERVLEVEL > 0
1296 dst_object->flags |= OBJ_COLORED;
1297 dst_object->pg_color = atop(dst_entry->start);
1301 VM_OBJECT_WLOCK(dst_object);
1302 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1303 ("vm_fault_copy_entry: vm_object not NULL"));
1304 if (src_object != dst_object) {
1305 dst_entry->object.vm_object = dst_object;
1306 dst_entry->offset = 0;
1307 dst_object->charge = dst_entry->end - dst_entry->start;
1309 if (fork_charge != NULL) {
1310 KASSERT(dst_entry->cred == NULL,
1311 ("vm_fault_copy_entry: leaked swp charge"));
1312 dst_object->cred = curthread->td_ucred;
1313 crhold(dst_object->cred);
1314 *fork_charge += dst_object->charge;
1315 } else if (dst_object->cred == NULL) {
1316 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1318 dst_object->cred = dst_entry->cred;
1319 dst_entry->cred = NULL;
1323 * If not an upgrade, then enter the mappings in the pmap as
1324 * read and/or execute accesses. Otherwise, enter them as
1327 * A writeable large page mapping is only created if all of
1328 * the constituent small page mappings are modified. Marking
1329 * PTEs as modified on inception allows promotion to happen
1330 * without taking potentially large number of soft faults.
1333 access &= ~VM_PROT_WRITE;
1336 * Loop through all of the virtual pages within the entry's
1337 * range, copying each page from the source object to the
1338 * destination object. Since the source is wired, those pages
1339 * must exist. In contrast, the destination is pageable.
1340 * Since the destination object does share any backing storage
1341 * with the source object, all of its pages must be dirtied,
1342 * regardless of whether they can be written.
1344 for (vaddr = dst_entry->start, dst_pindex = 0;
1345 vaddr < dst_entry->end;
1346 vaddr += PAGE_SIZE, dst_pindex++) {
1349 * Find the page in the source object, and copy it in.
1350 * Because the source is wired down, the page will be
1353 if (src_object != dst_object)
1354 VM_OBJECT_RLOCK(src_object);
1355 object = src_object;
1356 pindex = src_pindex + dst_pindex;
1357 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1358 (backing_object = object->backing_object) != NULL) {
1360 * Unless the source mapping is read-only or
1361 * it is presently being upgraded from
1362 * read-only, the first object in the shadow
1363 * chain should provide all of the pages. In
1364 * other words, this loop body should never be
1365 * executed when the source mapping is already
1368 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1370 ("vm_fault_copy_entry: main object missing page"));
1372 VM_OBJECT_RLOCK(backing_object);
1373 pindex += OFF_TO_IDX(object->backing_object_offset);
1374 if (object != dst_object)
1375 VM_OBJECT_RUNLOCK(object);
1376 object = backing_object;
1378 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1380 if (object != dst_object) {
1382 * Allocate a page in the destination object.
1384 dst_m = vm_page_alloc(dst_object, (src_object ==
1385 dst_object ? src_pindex : 0) + dst_pindex,
1387 if (dst_m == NULL) {
1388 VM_OBJECT_WUNLOCK(dst_object);
1389 VM_OBJECT_RUNLOCK(object);
1391 VM_OBJECT_WLOCK(dst_object);
1394 pmap_copy_page(src_m, dst_m);
1395 VM_OBJECT_RUNLOCK(object);
1396 dst_m->valid = VM_PAGE_BITS_ALL;
1397 dst_m->dirty = VM_PAGE_BITS_ALL;
1400 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1402 vm_page_xbusy(dst_m);
1403 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1404 ("invalid dst page %p", dst_m));
1406 VM_OBJECT_WUNLOCK(dst_object);
1409 * Enter it in the pmap. If a wired, copy-on-write
1410 * mapping is being replaced by a write-enabled
1411 * mapping, then wire that new mapping.
1413 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1414 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1417 * Mark it no longer busy, and put it on the active list.
1419 VM_OBJECT_WLOCK(dst_object);
1422 if (src_m != dst_m) {
1423 vm_page_lock(src_m);
1424 vm_page_unwire(src_m, PQ_INACTIVE);
1425 vm_page_unlock(src_m);
1426 vm_page_lock(dst_m);
1427 vm_page_wire(dst_m);
1428 vm_page_unlock(dst_m);
1430 KASSERT(dst_m->wire_count > 0,
1431 ("dst_m %p is not wired", dst_m));
1434 vm_page_lock(dst_m);
1435 vm_page_activate(dst_m);
1436 vm_page_unlock(dst_m);
1438 vm_page_xunbusy(dst_m);
1440 VM_OBJECT_WUNLOCK(dst_object);
1442 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1443 vm_object_deallocate(src_object);
1448 * Block entry into the machine-independent layer's page fault handler by
1449 * the calling thread. Subsequent calls to vm_fault() by that thread will
1450 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1451 * spurious page faults.
1454 vm_fault_disable_pagefaults(void)
1457 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1461 vm_fault_enable_pagefaults(int save)
1464 curthread_pflags_restore(save);