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)
499 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
502 * Page is not resident. If the pager might contain the page
503 * or this is the beginning of the search, allocate a new
504 * page. (Default objects are zero-fill, so there is no real
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);
515 * Allocate a new page for this object/offset pair.
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.
522 if (!vm_page_count_severe() || P_KILLED(curproc)) {
523 #if VM_NRESERVLEVEL > 0
524 vm_object_color(fs.object, atop(vaddr) -
527 alloc_req = P_KILLED(curproc) ?
528 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
529 if (fs.object->type != OBJT_VNODE &&
530 fs.object->backing_object == NULL)
531 alloc_req |= VM_ALLOC_ZERO;
532 fs.m = vm_page_alloc(fs.object, fs.pindex,
536 unlock_and_deallocate(&fs);
539 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
545 * We have either allocated a new page or found an existing
546 * page that is only partially valid.
548 * Attempt to fault-in the page if there is a chance that the
549 * pager has it, and potentially fault in additional pages
552 if (fs.object->type != OBJT_DEFAULT) {
554 u_char behavior = vm_map_entry_behavior(fs.entry);
556 era = fs.entry->read_ahead;
557 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
562 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
564 nera = VM_FAULT_READ_AHEAD_MAX;
566 if (fs.pindex == fs.entry->next_read)
567 vm_fault_dontneed(&fs, vaddr, ahead);
568 } else if (fs.pindex == fs.entry->next_read) {
570 * This is a sequential fault. Arithmetically
571 * increase the requested number of pages in
572 * the read-ahead window. The requested
573 * number of pages is "# of sequential faults
574 * x (read ahead min + 1) + read ahead min"
577 nera = VM_FAULT_READ_AHEAD_MIN;
580 if (nera > VM_FAULT_READ_AHEAD_MAX)
581 nera = VM_FAULT_READ_AHEAD_MAX;
584 if (era == VM_FAULT_READ_AHEAD_MAX)
585 vm_fault_dontneed(&fs, vaddr, ahead);
588 * This is a non-sequential fault. Request a
589 * cluster of pages that is aligned to a
590 * VM_FAULT_READ_DEFAULT page offset boundary
591 * within the object. Alignment to a page
592 * offset boundary is more likely to coincide
593 * with the underlying file system block than
594 * alignment to a virtual address boundary.
596 cluster_offset = fs.pindex %
597 VM_FAULT_READ_DEFAULT;
598 behind = ulmin(cluster_offset,
599 atop(vaddr - fs.entry->start));
601 ahead = VM_FAULT_READ_DEFAULT - 1 -
604 ahead = ulmin(ahead, atop(fs.entry->end - vaddr) - 1);
606 fs.entry->read_ahead = nera;
609 * Call the pager to retrieve the data, if any, after
610 * releasing the lock on the map. We hold a ref on
611 * fs.object and the pages are exclusive busied.
615 if (fs.object->type == OBJT_VNODE) {
616 vp = fs.object->handle;
619 else if (fs.vp != NULL) {
623 locked = VOP_ISLOCKED(vp);
625 if (locked != LK_EXCLUSIVE)
627 /* Do not sleep for vnode lock while fs.m is busy */
628 error = vget(vp, locked | LK_CANRECURSE |
629 LK_NOWAIT, curthread);
633 unlock_and_deallocate(&fs);
634 error = vget(vp, locked | LK_RETRY |
635 LK_CANRECURSE, curthread);
639 ("vm_fault: vget failed"));
645 KASSERT(fs.vp == NULL || !fs.map->system_map,
646 ("vm_fault: vnode-backed object mapped by system map"));
649 * Page in the requested page and hint the pager,
650 * that it may bring up surrounding pages.
652 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
654 if (rv == VM_PAGER_OK) {
655 faultcount = behind + 1 + ahead;
657 break; /* break to PAGE HAS BEEN FOUND */
659 if (rv == VM_PAGER_ERROR)
660 printf("vm_fault: pager read error, pid %d (%s)\n",
661 curproc->p_pid, curproc->p_comm);
664 * If an I/O error occurred or the requested page was
665 * outside the range of the pager, clean up and return
668 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
671 vm_page_unlock(fs.m);
673 unlock_and_deallocate(&fs);
674 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
675 KERN_PROTECTION_FAILURE);
679 * The requested page does not exist at this object/
680 * offset. Remove the invalid page from the object,
681 * waking up anyone waiting for it, and continue on to
682 * the next object. However, if this is the top-level
683 * object, we must leave the busy page in place to
684 * prevent another process from rushing past us, and
685 * inserting the page in that object at the same time
688 if (fs.object != fs.first_object) {
691 vm_page_unlock(fs.m);
697 * We get here if the object has default pager (or unwiring)
698 * or the pager doesn't have the page.
700 if (fs.object == fs.first_object)
704 * Move on to the next object. Lock the next object before
705 * unlocking the current one.
707 next_object = fs.object->backing_object;
708 if (next_object == NULL) {
710 * If there's no object left, fill the page in the top
713 if (fs.object != fs.first_object) {
714 vm_object_pip_wakeup(fs.object);
715 VM_OBJECT_WUNLOCK(fs.object);
717 fs.object = fs.first_object;
718 fs.pindex = fs.first_pindex;
720 VM_OBJECT_WLOCK(fs.object);
725 * Zero the page if necessary and mark it valid.
727 if ((fs.m->flags & PG_ZERO) == 0) {
728 pmap_zero_page(fs.m);
730 PCPU_INC(cnt.v_ozfod);
732 PCPU_INC(cnt.v_zfod);
733 fs.m->valid = VM_PAGE_BITS_ALL;
734 /* Don't try to prefault neighboring pages. */
736 break; /* break to PAGE HAS BEEN FOUND */
738 KASSERT(fs.object != next_object,
739 ("object loop %p", next_object));
740 VM_OBJECT_WLOCK(next_object);
741 vm_object_pip_add(next_object, 1);
742 if (fs.object != fs.first_object)
743 vm_object_pip_wakeup(fs.object);
745 OFF_TO_IDX(fs.object->backing_object_offset);
746 VM_OBJECT_WUNLOCK(fs.object);
747 fs.object = next_object;
751 vm_page_assert_xbusied(fs.m);
754 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
759 * If the page is being written, but isn't already owned by the
760 * top-level object, we have to copy it into a new page owned by the
763 if (fs.object != fs.first_object) {
765 * We only really need to copy if we want to write it.
767 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
769 * This allows pages to be virtually copied from a
770 * backing_object into the first_object, where the
771 * backing object has no other refs to it, and cannot
772 * gain any more refs. Instead of a bcopy, we just
773 * move the page from the backing object to the
774 * first object. Note that we must mark the page
775 * dirty in the first object so that it will go out
776 * to swap when needed.
778 is_first_object_locked = FALSE;
781 * Only one shadow object
783 (fs.object->shadow_count == 1) &&
785 * No COW refs, except us
787 (fs.object->ref_count == 1) &&
789 * No one else can look this object up
791 (fs.object->handle == NULL) &&
793 * No other ways to look the object up
795 ((fs.object->type == OBJT_DEFAULT) ||
796 (fs.object->type == OBJT_SWAP)) &&
797 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
799 * We don't chase down the shadow chain
801 fs.object == fs.first_object->backing_object) {
803 vm_page_remove(fs.m);
804 vm_page_unlock(fs.m);
805 vm_page_lock(fs.first_m);
806 vm_page_replace_checked(fs.m, fs.first_object,
807 fs.first_pindex, fs.first_m);
808 vm_page_free(fs.first_m);
809 vm_page_unlock(fs.first_m);
811 #if VM_NRESERVLEVEL > 0
813 * Rename the reservation.
815 vm_reserv_rename(fs.m, fs.first_object,
816 fs.object, OFF_TO_IDX(
817 fs.first_object->backing_object_offset));
820 * Removing the page from the backing object
826 PCPU_INC(cnt.v_cow_optim);
829 * Oh, well, lets copy it.
831 pmap_copy_page(fs.m, fs.first_m);
832 fs.first_m->valid = VM_PAGE_BITS_ALL;
833 if (wired && (fault_flags &
834 VM_FAULT_WIRE) == 0) {
835 vm_page_lock(fs.first_m);
836 vm_page_wire(fs.first_m);
837 vm_page_unlock(fs.first_m);
840 vm_page_unwire(fs.m, PQ_INACTIVE);
841 vm_page_unlock(fs.m);
844 * We no longer need the old page or object.
849 * fs.object != fs.first_object due to above
852 vm_object_pip_wakeup(fs.object);
853 VM_OBJECT_WUNLOCK(fs.object);
855 * Only use the new page below...
857 fs.object = fs.first_object;
858 fs.pindex = fs.first_pindex;
860 if (!is_first_object_locked)
861 VM_OBJECT_WLOCK(fs.object);
862 PCPU_INC(cnt.v_cow_faults);
865 prot &= ~VM_PROT_WRITE;
870 * We must verify that the maps have not changed since our last
873 if (!fs.lookup_still_valid) {
874 vm_object_t retry_object;
875 vm_pindex_t retry_pindex;
876 vm_prot_t retry_prot;
878 if (!vm_map_trylock_read(fs.map)) {
880 unlock_and_deallocate(&fs);
883 fs.lookup_still_valid = TRUE;
884 if (fs.map->timestamp != map_generation) {
885 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
886 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
889 * If we don't need the page any longer, put it on the inactive
890 * list (the easiest thing to do here). If no one needs it,
891 * pageout will grab it eventually.
893 if (result != KERN_SUCCESS) {
895 unlock_and_deallocate(&fs);
898 * If retry of map lookup would have blocked then
899 * retry fault from start.
901 if (result == KERN_FAILURE)
905 if ((retry_object != fs.first_object) ||
906 (retry_pindex != fs.first_pindex)) {
908 unlock_and_deallocate(&fs);
913 * Check whether the protection has changed or the object has
914 * been copied while we left the map unlocked. Changing from
915 * read to write permission is OK - we leave the page
916 * write-protected, and catch the write fault. Changing from
917 * write to read permission means that we can't mark the page
918 * write-enabled after all.
924 * If the page was filled by a pager, update the map entry's
927 * XXX The following assignment modifies the map
928 * without holding a write lock on it.
931 fs.entry->next_read = fs.pindex + ahead + 1;
933 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
934 vm_page_assert_xbusied(fs.m);
937 * Page must be completely valid or it is not fit to
938 * map into user space. vm_pager_get_pages() ensures this.
940 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
941 ("vm_fault: page %p partially invalid", fs.m));
942 VM_OBJECT_WUNLOCK(fs.object);
945 * Put this page into the physical map. We had to do the unlock above
946 * because pmap_enter() may sleep. We don't put the page
947 * back on the active queue until later so that the pageout daemon
948 * won't find it (yet).
950 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
951 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
952 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
954 vm_fault_prefault(&fs, vaddr,
955 faultcount > 0 ? behind : PFBAK,
956 faultcount > 0 ? ahead : PFFOR);
957 VM_OBJECT_WLOCK(fs.object);
961 * If the page is not wired down, then put it where the pageout daemon
964 if ((fault_flags & VM_FAULT_WIRE) != 0) {
965 KASSERT(wired, ("VM_FAULT_WIRE && !wired"));
968 vm_page_activate(fs.m);
969 if (m_hold != NULL) {
973 vm_page_unlock(fs.m);
974 vm_page_xunbusy(fs.m);
977 * Unlock everything, and return
979 unlock_and_deallocate(&fs);
981 PCPU_INC(cnt.v_io_faults);
982 curthread->td_ru.ru_majflt++;
984 if (racct_enable && fs.object->type == OBJT_VNODE) {
986 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
987 racct_add_force(curproc, RACCT_WRITEBPS,
988 PAGE_SIZE + behind * PAGE_SIZE);
989 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
991 racct_add_force(curproc, RACCT_READBPS,
992 PAGE_SIZE + ahead * PAGE_SIZE);
993 racct_add_force(curproc, RACCT_READIOPS, 1);
995 PROC_UNLOCK(curproc);
999 curthread->td_ru.ru_minflt++;
1001 return (KERN_SUCCESS);
1005 * Speed up the reclamation of pages that precede the faulting pindex within
1006 * the first object of the shadow chain. Essentially, perform the equivalent
1007 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1008 * the faulting pindex by the cluster size when the pages read by vm_fault()
1009 * cross a cluster-size boundary. The cluster size is the greater of the
1010 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1012 * When "fs->first_object" is a shadow object, the pages in the backing object
1013 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1014 * function must only be concerned with pages in the first object.
1017 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1019 vm_map_entry_t entry;
1020 vm_object_t first_object, object;
1021 vm_offset_t end, start;
1022 vm_page_t m, m_next;
1023 vm_pindex_t pend, pstart;
1026 object = fs->object;
1027 VM_OBJECT_ASSERT_WLOCKED(object);
1028 first_object = fs->first_object;
1029 if (first_object != object) {
1030 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1031 VM_OBJECT_WUNLOCK(object);
1032 VM_OBJECT_WLOCK(first_object);
1033 VM_OBJECT_WLOCK(object);
1036 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1037 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1038 size = VM_FAULT_DONTNEED_MIN;
1039 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1040 size = pagesizes[1];
1041 end = rounddown2(vaddr, size);
1042 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1043 (entry = fs->entry)->start < end) {
1044 if (end - entry->start < size)
1045 start = entry->start;
1048 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1049 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1051 m_next = vm_page_find_least(first_object, pstart);
1052 pend = OFF_TO_IDX(entry->offset) + atop(end -
1054 while ((m = m_next) != NULL && m->pindex < pend) {
1055 m_next = TAILQ_NEXT(m, listq);
1056 if (m->valid != VM_PAGE_BITS_ALL ||
1061 * Don't clear PGA_REFERENCED, since it would
1062 * likely represent a reference by a different
1065 * Typically, at this point, prefetched pages
1066 * are still in the inactive queue. Only
1067 * pages that triggered page faults are in the
1071 vm_page_deactivate(m);
1076 if (first_object != object)
1077 VM_OBJECT_WUNLOCK(first_object);
1081 * vm_fault_prefault provides a quick way of clustering
1082 * pagefaults into a processes address space. It is a "cousin"
1083 * of vm_map_pmap_enter, except it runs at page fault time instead
1087 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1088 int backward, int forward)
1091 vm_map_entry_t entry;
1092 vm_object_t backing_object, lobject;
1093 vm_offset_t addr, starta;
1098 pmap = fs->map->pmap;
1099 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1104 starta = addra - backward * PAGE_SIZE;
1105 if (starta < entry->start) {
1106 starta = entry->start;
1107 } else if (starta > addra) {
1112 * Generate the sequence of virtual addresses that are candidates for
1113 * prefaulting in an outward spiral from the faulting virtual address,
1114 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1115 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1116 * If the candidate address doesn't have a backing physical page, then
1117 * the loop immediately terminates.
1119 for (i = 0; i < 2 * imax(backward, forward); i++) {
1120 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1122 if (addr > addra + forward * PAGE_SIZE)
1125 if (addr < starta || addr >= entry->end)
1128 if (!pmap_is_prefaultable(pmap, addr))
1131 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1132 lobject = entry->object.vm_object;
1133 VM_OBJECT_RLOCK(lobject);
1134 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1135 lobject->type == OBJT_DEFAULT &&
1136 (backing_object = lobject->backing_object) != NULL) {
1137 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1138 0, ("vm_fault_prefault: unaligned object offset"));
1139 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1140 VM_OBJECT_RLOCK(backing_object);
1141 VM_OBJECT_RUNLOCK(lobject);
1142 lobject = backing_object;
1145 VM_OBJECT_RUNLOCK(lobject);
1148 if (m->valid == VM_PAGE_BITS_ALL &&
1149 (m->flags & PG_FICTITIOUS) == 0)
1150 pmap_enter_quick(pmap, addr, m, entry->protection);
1151 VM_OBJECT_RUNLOCK(lobject);
1156 * Hold each of the physical pages that are mapped by the specified range of
1157 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1158 * and allow the specified types of access, "prot". If all of the implied
1159 * pages are successfully held, then the number of held pages is returned
1160 * together with pointers to those pages in the array "ma". However, if any
1161 * of the pages cannot be held, -1 is returned.
1164 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1165 vm_prot_t prot, vm_page_t *ma, int max_count)
1167 vm_offset_t end, va;
1170 boolean_t pmap_failed;
1174 end = round_page(addr + len);
1175 addr = trunc_page(addr);
1178 * Check for illegal addresses.
1180 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1183 if (atop(end - addr) > max_count)
1184 panic("vm_fault_quick_hold_pages: count > max_count");
1185 count = atop(end - addr);
1188 * Most likely, the physical pages are resident in the pmap, so it is
1189 * faster to try pmap_extract_and_hold() first.
1191 pmap_failed = FALSE;
1192 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1193 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1196 else if ((prot & VM_PROT_WRITE) != 0 &&
1197 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1199 * Explicitly dirty the physical page. Otherwise, the
1200 * caller's changes may go unnoticed because they are
1201 * performed through an unmanaged mapping or by a DMA
1204 * The object lock is not held here.
1205 * See vm_page_clear_dirty_mask().
1212 * One or more pages could not be held by the pmap. Either no
1213 * page was mapped at the specified virtual address or that
1214 * mapping had insufficient permissions. Attempt to fault in
1215 * and hold these pages.
1217 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1218 if (*mp == NULL && vm_fault_hold(map, va, prot,
1219 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1224 for (mp = ma; mp < ma + count; mp++)
1227 vm_page_unhold(*mp);
1228 vm_page_unlock(*mp);
1235 * vm_fault_copy_entry
1237 * Create new shadow object backing dst_entry with private copy of
1238 * all underlying pages. When src_entry is equal to dst_entry,
1239 * function implements COW for wired-down map entry. Otherwise,
1240 * it forks wired entry into dst_map.
1242 * In/out conditions:
1243 * The source and destination maps must be locked for write.
1244 * The source map entry must be wired down (or be a sharing map
1245 * entry corresponding to a main map entry that is wired down).
1248 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1249 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1250 vm_ooffset_t *fork_charge)
1252 vm_object_t backing_object, dst_object, object, src_object;
1253 vm_pindex_t dst_pindex, pindex, src_pindex;
1254 vm_prot_t access, prot;
1264 upgrade = src_entry == dst_entry;
1265 access = prot = dst_entry->protection;
1267 src_object = src_entry->object.vm_object;
1268 src_pindex = OFF_TO_IDX(src_entry->offset);
1270 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1271 dst_object = src_object;
1272 vm_object_reference(dst_object);
1275 * Create the top-level object for the destination entry. (Doesn't
1276 * actually shadow anything - we copy the pages directly.)
1278 dst_object = vm_object_allocate(OBJT_DEFAULT,
1279 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1280 #if VM_NRESERVLEVEL > 0
1281 dst_object->flags |= OBJ_COLORED;
1282 dst_object->pg_color = atop(dst_entry->start);
1286 VM_OBJECT_WLOCK(dst_object);
1287 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1288 ("vm_fault_copy_entry: vm_object not NULL"));
1289 if (src_object != dst_object) {
1290 dst_entry->object.vm_object = dst_object;
1291 dst_entry->offset = 0;
1292 dst_object->charge = dst_entry->end - dst_entry->start;
1294 if (fork_charge != NULL) {
1295 KASSERT(dst_entry->cred == NULL,
1296 ("vm_fault_copy_entry: leaked swp charge"));
1297 dst_object->cred = curthread->td_ucred;
1298 crhold(dst_object->cred);
1299 *fork_charge += dst_object->charge;
1300 } else if (dst_object->cred == NULL) {
1301 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1303 dst_object->cred = dst_entry->cred;
1304 dst_entry->cred = NULL;
1308 * If not an upgrade, then enter the mappings in the pmap as
1309 * read and/or execute accesses. Otherwise, enter them as
1312 * A writeable large page mapping is only created if all of
1313 * the constituent small page mappings are modified. Marking
1314 * PTEs as modified on inception allows promotion to happen
1315 * without taking potentially large number of soft faults.
1318 access &= ~VM_PROT_WRITE;
1321 * Loop through all of the virtual pages within the entry's
1322 * range, copying each page from the source object to the
1323 * destination object. Since the source is wired, those pages
1324 * must exist. In contrast, the destination is pageable.
1325 * Since the destination object does share any backing storage
1326 * with the source object, all of its pages must be dirtied,
1327 * regardless of whether they can be written.
1329 for (vaddr = dst_entry->start, dst_pindex = 0;
1330 vaddr < dst_entry->end;
1331 vaddr += PAGE_SIZE, dst_pindex++) {
1334 * Find the page in the source object, and copy it in.
1335 * Because the source is wired down, the page will be
1338 if (src_object != dst_object)
1339 VM_OBJECT_RLOCK(src_object);
1340 object = src_object;
1341 pindex = src_pindex + dst_pindex;
1342 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1343 (backing_object = object->backing_object) != NULL) {
1345 * Unless the source mapping is read-only or
1346 * it is presently being upgraded from
1347 * read-only, the first object in the shadow
1348 * chain should provide all of the pages. In
1349 * other words, this loop body should never be
1350 * executed when the source mapping is already
1353 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1355 ("vm_fault_copy_entry: main object missing page"));
1357 VM_OBJECT_RLOCK(backing_object);
1358 pindex += OFF_TO_IDX(object->backing_object_offset);
1359 if (object != dst_object)
1360 VM_OBJECT_RUNLOCK(object);
1361 object = backing_object;
1363 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1365 if (object != dst_object) {
1367 * Allocate a page in the destination object.
1369 dst_m = vm_page_alloc(dst_object, (src_object ==
1370 dst_object ? src_pindex : 0) + dst_pindex,
1372 if (dst_m == NULL) {
1373 VM_OBJECT_WUNLOCK(dst_object);
1374 VM_OBJECT_RUNLOCK(object);
1376 VM_OBJECT_WLOCK(dst_object);
1379 pmap_copy_page(src_m, dst_m);
1380 VM_OBJECT_RUNLOCK(object);
1381 dst_m->valid = VM_PAGE_BITS_ALL;
1382 dst_m->dirty = VM_PAGE_BITS_ALL;
1385 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1387 vm_page_xbusy(dst_m);
1388 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1389 ("invalid dst page %p", dst_m));
1391 VM_OBJECT_WUNLOCK(dst_object);
1394 * Enter it in the pmap. If a wired, copy-on-write
1395 * mapping is being replaced by a write-enabled
1396 * mapping, then wire that new mapping.
1398 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1399 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1402 * Mark it no longer busy, and put it on the active list.
1404 VM_OBJECT_WLOCK(dst_object);
1407 if (src_m != dst_m) {
1408 vm_page_lock(src_m);
1409 vm_page_unwire(src_m, PQ_INACTIVE);
1410 vm_page_unlock(src_m);
1411 vm_page_lock(dst_m);
1412 vm_page_wire(dst_m);
1413 vm_page_unlock(dst_m);
1415 KASSERT(dst_m->wire_count > 0,
1416 ("dst_m %p is not wired", dst_m));
1419 vm_page_lock(dst_m);
1420 vm_page_activate(dst_m);
1421 vm_page_unlock(dst_m);
1423 vm_page_xunbusy(dst_m);
1425 VM_OBJECT_WUNLOCK(dst_object);
1427 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1428 vm_object_deallocate(src_object);
1433 * Block entry into the machine-independent layer's page fault handler by
1434 * the calling thread. Subsequent calls to vm_fault() by that thread will
1435 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1436 * spurious page faults.
1439 vm_fault_disable_pagefaults(void)
1442 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1446 vm_fault_enable_pagefaults(int save)
1449 curthread_pflags_restore(save);