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>
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>
91 #include <sys/ktrace.h>
95 #include <vm/vm_param.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>
108 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
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)
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_cache_behind(const struct faultstate *fs, int distance);
130 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
131 int faultcount, int reqpage);
134 release_page(struct faultstate *fs)
137 vm_page_xunbusy(fs->m);
139 vm_page_deactivate(fs->m);
140 vm_page_unlock(fs->m);
145 unlock_map(struct faultstate *fs)
148 if (fs->lookup_still_valid) {
149 vm_map_lookup_done(fs->map, fs->entry);
150 fs->lookup_still_valid = FALSE;
155 unlock_and_deallocate(struct faultstate *fs)
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);
169 vm_object_deallocate(fs->first_object);
171 if (fs->vp != NULL) {
178 * TRYPAGER - used by vm_fault to calculate whether the pager for the
179 * current object *might* contain the page.
181 * default objects are zero-fill, there is no real pager.
183 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
184 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
189 * Handle a page fault occurring at the given address,
190 * requiring the given permissions, in the map specified.
191 * If successful, the page is inserted into the
192 * associated physical map.
194 * NOTE: the given address should be truncated to the
195 * proper page address.
197 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
198 * a standard error specifying why the fault is fatal is returned.
200 * The map in question must be referenced, and remains so.
201 * Caller may hold no locks.
204 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
211 if ((td->td_pflags & TDP_NOFAULTING) != 0)
212 return (KERN_PROTECTION_FAILURE);
214 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
215 ktrfault(vaddr, fault_type);
217 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
220 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
227 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
228 int fault_flags, vm_page_t *m_hold)
232 int alloc_req, era, faultcount, nera, reqpage, result;
233 boolean_t growstack, is_first_object_locked, wired;
235 vm_object_t next_object;
236 vm_page_t marray[VM_FAULT_READ_MAX];
238 struct faultstate fs;
244 PCPU_INC(cnt.v_vm_faults);
246 faultcount = reqpage = 0;
251 * Find the backing store object and offset into it to begin the
255 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
256 &fs.first_object, &fs.first_pindex, &prot, &wired);
257 if (result != KERN_SUCCESS) {
258 if (growstack && result == KERN_INVALID_ADDRESS &&
260 result = vm_map_growstack(curproc, vaddr);
261 if (result != KERN_SUCCESS)
262 return (KERN_FAILURE);
269 map_generation = fs.map->timestamp;
271 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
272 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
273 vm_map_unlock_read(fs.map);
274 return (KERN_FAILURE);
276 panic("vm_fault: fault on nofault entry, addr: %lx",
280 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
281 fs.entry->wiring_thread != curthread) {
282 vm_map_unlock_read(fs.map);
284 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
285 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
286 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
287 vm_map_unlock_and_wait(fs.map, 0);
289 vm_map_unlock(fs.map);
294 * Make a reference to this object to prevent its disposal while we
295 * are messing with it. Once we have the reference, the map is free
296 * to be diddled. Since objects reference their shadows (and copies),
297 * they will stay around as well.
299 * Bump the paging-in-progress count to prevent size changes (e.g.
300 * truncation operations) during I/O. This must be done after
301 * obtaining the vnode lock in order to avoid possible deadlocks.
303 VM_OBJECT_WLOCK(fs.first_object);
304 vm_object_reference_locked(fs.first_object);
305 vm_object_pip_add(fs.first_object, 1);
307 fs.lookup_still_valid = TRUE;
310 fault_type = prot | (fault_type & VM_PROT_COPY);
315 * Search for the page at object/offset.
317 fs.object = fs.first_object;
318 fs.pindex = fs.first_pindex;
321 * If the object is dead, we stop here
323 if (fs.object->flags & OBJ_DEAD) {
324 unlock_and_deallocate(&fs);
325 return (KERN_PROTECTION_FAILURE);
329 * See if page is resident
331 fs.m = vm_page_lookup(fs.object, fs.pindex);
334 * Wait/Retry if the page is busy. We have to do this
335 * if the page is either exclusive or shared busy
336 * because the vm_pager may be using read busy for
337 * pageouts (and even pageins if it is the vnode
338 * pager), and we could end up trying to pagein and
339 * pageout the same page simultaneously.
341 * We can theoretically allow the busy case on a read
342 * fault if the page is marked valid, but since such
343 * pages are typically already pmap'd, putting that
344 * special case in might be more effort then it is
345 * worth. We cannot under any circumstances mess
346 * around with a shared busied page except, perhaps,
349 if (vm_page_busied(fs.m)) {
351 * Reference the page before unlocking and
352 * sleeping so that the page daemon is less
353 * likely to reclaim it.
355 vm_page_aflag_set(fs.m, PGA_REFERENCED);
356 if (fs.object != fs.first_object) {
357 if (!VM_OBJECT_TRYWLOCK(
359 VM_OBJECT_WUNLOCK(fs.object);
360 VM_OBJECT_WLOCK(fs.first_object);
361 VM_OBJECT_WLOCK(fs.object);
363 vm_page_lock(fs.first_m);
364 vm_page_free(fs.first_m);
365 vm_page_unlock(fs.first_m);
366 vm_object_pip_wakeup(fs.first_object);
367 VM_OBJECT_WUNLOCK(fs.first_object);
371 if (fs.m == vm_page_lookup(fs.object,
373 vm_page_sleep_if_busy(fs.m, "vmpfw");
375 vm_object_pip_wakeup(fs.object);
376 VM_OBJECT_WUNLOCK(fs.object);
377 PCPU_INC(cnt.v_intrans);
378 vm_object_deallocate(fs.first_object);
382 vm_page_remque(fs.m);
383 vm_page_unlock(fs.m);
386 * Mark page busy for other processes, and the
387 * pagedaemon. If it still isn't completely valid
388 * (readable), jump to readrest, else break-out ( we
392 if (fs.m->valid != VM_PAGE_BITS_ALL)
398 * Page is not resident, If this is the search termination
399 * or the pager might contain the page, allocate a new page.
401 if (TRYPAGER || fs.object == fs.first_object) {
402 if (fs.pindex >= fs.object->size) {
403 unlock_and_deallocate(&fs);
404 return (KERN_PROTECTION_FAILURE);
408 * Allocate a new page for this object/offset pair.
410 * Unlocked read of the p_flag is harmless. At
411 * worst, the P_KILLED might be not observed
412 * there, and allocation can fail, causing
413 * restart and new reading of the p_flag.
416 if (!vm_page_count_severe() || P_KILLED(curproc)) {
417 #if VM_NRESERVLEVEL > 0
418 if ((fs.object->flags & OBJ_COLORED) == 0) {
419 fs.object->flags |= OBJ_COLORED;
420 fs.object->pg_color = atop(vaddr) -
424 alloc_req = P_KILLED(curproc) ?
425 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
426 if (fs.object->type != OBJT_VNODE &&
427 fs.object->backing_object == NULL)
428 alloc_req |= VM_ALLOC_ZERO;
429 fs.m = vm_page_alloc(fs.object, fs.pindex,
433 unlock_and_deallocate(&fs);
436 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
442 * We have found a valid page or we have allocated a new page.
443 * The page thus may not be valid or may not be entirely
446 * Attempt to fault-in the page if there is a chance that the
447 * pager has it, and potentially fault in additional pages
452 u_char behavior = vm_map_entry_behavior(fs.entry);
454 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
458 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
460 ahead = atop(fs.entry->end - vaddr) - 1;
461 if (ahead > VM_FAULT_READ_AHEAD_MAX)
462 ahead = VM_FAULT_READ_AHEAD_MAX;
463 if (fs.pindex == fs.entry->next_read)
464 vm_fault_cache_behind(&fs,
468 * If this is a sequential page fault, then
469 * arithmetically increase the number of pages
470 * in the read-ahead window. Otherwise, reset
471 * the read-ahead window to its smallest size.
473 behind = atop(vaddr - fs.entry->start);
474 if (behind > VM_FAULT_READ_BEHIND)
475 behind = VM_FAULT_READ_BEHIND;
476 ahead = atop(fs.entry->end - vaddr) - 1;
477 era = fs.entry->read_ahead;
478 if (fs.pindex == fs.entry->next_read) {
480 if (nera > VM_FAULT_READ_AHEAD_MAX)
481 nera = VM_FAULT_READ_AHEAD_MAX;
485 if (era == VM_FAULT_READ_AHEAD_MAX)
486 vm_fault_cache_behind(&fs,
487 VM_FAULT_CACHE_BEHIND);
488 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
489 ahead = VM_FAULT_READ_AHEAD_MIN;
491 fs.entry->read_ahead = ahead;
495 * Call the pager to retrieve the data, if any, after
496 * releasing the lock on the map. We hold a ref on
497 * fs.object and the pages are exclusive busied.
501 if (fs.object->type == OBJT_VNODE) {
502 vp = fs.object->handle;
505 else if (fs.vp != NULL) {
509 locked = VOP_ISLOCKED(vp);
511 if (locked != LK_EXCLUSIVE)
513 /* Do not sleep for vnode lock while fs.m is busy */
514 error = vget(vp, locked | LK_CANRECURSE |
515 LK_NOWAIT, curthread);
519 unlock_and_deallocate(&fs);
520 error = vget(vp, locked | LK_RETRY |
521 LK_CANRECURSE, curthread);
525 ("vm_fault: vget failed"));
531 KASSERT(fs.vp == NULL || !fs.map->system_map,
532 ("vm_fault: vnode-backed object mapped by system map"));
535 * now we find out if any other pages should be paged
536 * in at this time this routine checks to see if the
537 * pages surrounding this fault reside in the same
538 * object as the page for this fault. If they do,
539 * then they are faulted in also into the object. The
540 * array "marray" returned contains an array of
541 * vm_page_t structs where one of them is the
542 * vm_page_t passed to the routine. The reqpage
543 * return value is the index into the marray for the
544 * vm_page_t passed to the routine.
546 * fs.m plus the additional pages are exclusive busied.
548 faultcount = vm_fault_additional_pages(
549 fs.m, behind, ahead, marray, &reqpage);
552 vm_pager_get_pages(fs.object, marray, faultcount,
553 reqpage) : VM_PAGER_FAIL;
555 if (rv == VM_PAGER_OK) {
557 * Found the page. Leave it busy while we play
562 * Relookup in case pager changed page. Pager
563 * is responsible for disposition of old page
566 fs.m = vm_page_lookup(fs.object, fs.pindex);
568 unlock_and_deallocate(&fs);
573 break; /* break to PAGE HAS BEEN FOUND */
576 * Remove the bogus page (which does not exist at this
577 * object/offset); before doing so, we must get back
578 * our object lock to preserve our invariant.
580 * Also wake up any other process that may want to bring
583 * If this is the top-level object, we must leave the
584 * busy page to prevent another process from rushing
585 * past us, and inserting the page in that object at
586 * the same time that we are.
588 if (rv == VM_PAGER_ERROR)
589 printf("vm_fault: pager read error, pid %d (%s)\n",
590 curproc->p_pid, curproc->p_comm);
592 * Data outside the range of the pager or an I/O error
595 * XXX - the check for kernel_map is a kludge to work
596 * around having the machine panic on a kernel space
597 * fault w/ I/O error.
599 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
600 (rv == VM_PAGER_BAD)) {
603 vm_page_unlock(fs.m);
605 unlock_and_deallocate(&fs);
606 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
608 if (fs.object != fs.first_object) {
611 vm_page_unlock(fs.m);
614 * XXX - we cannot just fall out at this
615 * point, m has been freed and is invalid!
621 * We get here if the object has default pager (or unwiring)
622 * or the pager doesn't have the page.
624 if (fs.object == fs.first_object)
628 * Move on to the next object. Lock the next object before
629 * unlocking the current one.
631 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
632 next_object = fs.object->backing_object;
633 if (next_object == NULL) {
635 * If there's no object left, fill the page in the top
638 if (fs.object != fs.first_object) {
639 vm_object_pip_wakeup(fs.object);
640 VM_OBJECT_WUNLOCK(fs.object);
642 fs.object = fs.first_object;
643 fs.pindex = fs.first_pindex;
645 VM_OBJECT_WLOCK(fs.object);
650 * Zero the page if necessary and mark it valid.
652 if ((fs.m->flags & PG_ZERO) == 0) {
653 pmap_zero_page(fs.m);
655 PCPU_INC(cnt.v_ozfod);
657 PCPU_INC(cnt.v_zfod);
658 fs.m->valid = VM_PAGE_BITS_ALL;
659 /* Don't try to prefault neighboring pages. */
661 break; /* break to PAGE HAS BEEN FOUND */
663 KASSERT(fs.object != next_object,
664 ("object loop %p", next_object));
665 VM_OBJECT_WLOCK(next_object);
666 vm_object_pip_add(next_object, 1);
667 if (fs.object != fs.first_object)
668 vm_object_pip_wakeup(fs.object);
669 VM_OBJECT_WUNLOCK(fs.object);
670 fs.object = next_object;
674 vm_page_assert_xbusied(fs.m);
677 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
682 * If the page is being written, but isn't already owned by the
683 * top-level object, we have to copy it into a new page owned by the
686 if (fs.object != fs.first_object) {
688 * We only really need to copy if we want to write it.
690 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
692 * This allows pages to be virtually copied from a
693 * backing_object into the first_object, where the
694 * backing object has no other refs to it, and cannot
695 * gain any more refs. Instead of a bcopy, we just
696 * move the page from the backing object to the
697 * first object. Note that we must mark the page
698 * dirty in the first object so that it will go out
699 * to swap when needed.
701 is_first_object_locked = FALSE;
704 * Only one shadow object
706 (fs.object->shadow_count == 1) &&
708 * No COW refs, except us
710 (fs.object->ref_count == 1) &&
712 * No one else can look this object up
714 (fs.object->handle == NULL) &&
716 * No other ways to look the object up
718 ((fs.object->type == OBJT_DEFAULT) ||
719 (fs.object->type == OBJT_SWAP)) &&
720 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
722 * We don't chase down the shadow chain
724 fs.object == fs.first_object->backing_object) {
726 * get rid of the unnecessary page
728 vm_page_lock(fs.first_m);
729 vm_page_free(fs.first_m);
730 vm_page_unlock(fs.first_m);
732 * grab the page and put it into the
733 * process'es object. The page is
734 * automatically made dirty.
736 if (vm_page_rename(fs.m, fs.first_object,
738 unlock_and_deallocate(&fs);
744 PCPU_INC(cnt.v_cow_optim);
747 * Oh, well, lets copy it.
749 pmap_copy_page(fs.m, fs.first_m);
750 fs.first_m->valid = VM_PAGE_BITS_ALL;
751 if (wired && (fault_flags &
752 VM_FAULT_CHANGE_WIRING) == 0) {
753 vm_page_lock(fs.first_m);
754 vm_page_wire(fs.first_m);
755 vm_page_unlock(fs.first_m);
758 vm_page_unwire(fs.m, FALSE);
759 vm_page_unlock(fs.m);
762 * We no longer need the old page or object.
767 * fs.object != fs.first_object due to above
770 vm_object_pip_wakeup(fs.object);
771 VM_OBJECT_WUNLOCK(fs.object);
773 * Only use the new page below...
775 fs.object = fs.first_object;
776 fs.pindex = fs.first_pindex;
778 if (!is_first_object_locked)
779 VM_OBJECT_WLOCK(fs.object);
780 PCPU_INC(cnt.v_cow_faults);
783 prot &= ~VM_PROT_WRITE;
788 * We must verify that the maps have not changed since our last
791 if (!fs.lookup_still_valid) {
792 vm_object_t retry_object;
793 vm_pindex_t retry_pindex;
794 vm_prot_t retry_prot;
796 if (!vm_map_trylock_read(fs.map)) {
798 unlock_and_deallocate(&fs);
801 fs.lookup_still_valid = TRUE;
802 if (fs.map->timestamp != map_generation) {
803 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
804 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
807 * If we don't need the page any longer, put it on the inactive
808 * list (the easiest thing to do here). If no one needs it,
809 * pageout will grab it eventually.
811 if (result != KERN_SUCCESS) {
813 unlock_and_deallocate(&fs);
816 * If retry of map lookup would have blocked then
817 * retry fault from start.
819 if (result == KERN_FAILURE)
823 if ((retry_object != fs.first_object) ||
824 (retry_pindex != fs.first_pindex)) {
826 unlock_and_deallocate(&fs);
831 * Check whether the protection has changed or the object has
832 * been copied while we left the map unlocked. Changing from
833 * read to write permission is OK - we leave the page
834 * write-protected, and catch the write fault. Changing from
835 * write to read permission means that we can't mark the page
836 * write-enabled after all.
842 * If the page was filled by a pager, update the map entry's
843 * last read offset. Since the pager does not return the
844 * actual set of pages that it read, this update is based on
845 * the requested set. Typically, the requested and actual
848 * XXX The following assignment modifies the map
849 * without holding a write lock on it.
852 fs.entry->next_read = fs.pindex + faultcount - reqpage;
854 if ((prot & VM_PROT_WRITE) != 0 ||
855 (fault_flags & VM_FAULT_DIRTY) != 0) {
856 vm_object_set_writeable_dirty(fs.object);
859 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
860 * if the page is already dirty to prevent data written with
861 * the expectation of being synced from not being synced.
862 * Likewise if this entry does not request NOSYNC then make
863 * sure the page isn't marked NOSYNC. Applications sharing
864 * data should use the same flags to avoid ping ponging.
866 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
867 if (fs.m->dirty == 0)
868 fs.m->oflags |= VPO_NOSYNC;
870 fs.m->oflags &= ~VPO_NOSYNC;
874 * If the fault is a write, we know that this page is being
875 * written NOW so dirty it explicitly to save on
876 * pmap_is_modified() calls later.
878 * Also tell the backing pager, if any, that it should remove
879 * any swap backing since the page is now dirty.
881 if (((fault_type & VM_PROT_WRITE) != 0 &&
882 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
883 (fault_flags & VM_FAULT_DIRTY) != 0) {
885 vm_pager_page_unswapped(fs.m);
889 vm_page_assert_xbusied(fs.m);
892 * Page must be completely valid or it is not fit to
893 * map into user space. vm_pager_get_pages() ensures this.
895 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
896 ("vm_fault: page %p partially invalid", fs.m));
897 VM_OBJECT_WUNLOCK(fs.object);
900 * Put this page into the physical map. We had to do the unlock above
901 * because pmap_enter() may sleep. We don't put the page
902 * back on the active queue until later so that the pageout daemon
903 * won't find it (yet).
905 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
906 if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
908 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
909 VM_OBJECT_WLOCK(fs.object);
913 * If the page is not wired down, then put it where the pageout daemon
916 if (fault_flags & VM_FAULT_CHANGE_WIRING) {
920 vm_page_unwire(fs.m, 1);
922 vm_page_activate(fs.m);
923 if (m_hold != NULL) {
927 vm_page_unlock(fs.m);
928 vm_page_xunbusy(fs.m);
931 * Unlock everything, and return
933 unlock_and_deallocate(&fs);
935 PCPU_INC(cnt.v_io_faults);
936 curthread->td_ru.ru_majflt++;
938 curthread->td_ru.ru_minflt++;
940 return (KERN_SUCCESS);
944 * Speed up the reclamation of up to "distance" pages that precede the
945 * faulting pindex within the first object of the shadow chain.
948 vm_fault_cache_behind(const struct faultstate *fs, int distance)
950 vm_object_t first_object, object;
955 VM_OBJECT_ASSERT_WLOCKED(object);
956 first_object = fs->first_object;
957 if (first_object != object) {
958 if (!VM_OBJECT_TRYWLOCK(first_object)) {
959 VM_OBJECT_WUNLOCK(object);
960 VM_OBJECT_WLOCK(first_object);
961 VM_OBJECT_WLOCK(object);
964 /* Neither fictitious nor unmanaged pages can be cached. */
965 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
966 if (fs->first_pindex < distance)
969 pindex = fs->first_pindex - distance;
970 if (pindex < OFF_TO_IDX(fs->entry->offset))
971 pindex = OFF_TO_IDX(fs->entry->offset);
972 m = first_object != object ? fs->first_m : fs->m;
973 vm_page_assert_xbusied(m);
974 m_prev = vm_page_prev(m);
975 while ((m = m_prev) != NULL && m->pindex >= pindex &&
976 m->valid == VM_PAGE_BITS_ALL) {
977 m_prev = vm_page_prev(m);
978 if (vm_page_busied(m))
981 if (m->hold_count == 0 && m->wire_count == 0) {
983 vm_page_aflag_clear(m, PGA_REFERENCED);
985 vm_page_deactivate(m);
992 if (first_object != object)
993 VM_OBJECT_WUNLOCK(first_object);
997 * vm_fault_prefault provides a quick way of clustering
998 * pagefaults into a processes address space. It is a "cousin"
999 * of vm_map_pmap_enter, except it runs at page fault time instead
1003 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1004 int faultcount, int reqpage)
1007 vm_map_entry_t entry;
1008 vm_object_t backing_object, lobject;
1009 vm_offset_t addr, starta;
1012 int backward, forward, i;
1014 pmap = fs->map->pmap;
1015 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1018 if (faultcount > 0) {
1020 forward = faultcount - reqpage - 1;
1027 starta = addra - backward * PAGE_SIZE;
1028 if (starta < entry->start) {
1029 starta = entry->start;
1030 } else if (starta > addra) {
1035 * Generate the sequence of virtual addresses that are candidates for
1036 * prefaulting in an outward spiral from the faulting virtual address,
1037 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1038 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1039 * If the candidate address doesn't have a backing physical page, then
1040 * the loop immediately terminates.
1042 for (i = 0; i < 2 * imax(backward, forward); i++) {
1043 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1045 if (addr > addra + forward * PAGE_SIZE)
1048 if (addr < starta || addr >= entry->end)
1051 if (!pmap_is_prefaultable(pmap, addr))
1054 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1055 lobject = entry->object.vm_object;
1056 VM_OBJECT_RLOCK(lobject);
1057 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1058 lobject->type == OBJT_DEFAULT &&
1059 (backing_object = lobject->backing_object) != NULL) {
1060 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1061 0, ("vm_fault_prefault: unaligned object offset"));
1062 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1063 VM_OBJECT_RLOCK(backing_object);
1064 VM_OBJECT_RUNLOCK(lobject);
1065 lobject = backing_object;
1068 VM_OBJECT_RUNLOCK(lobject);
1071 if (m->valid == VM_PAGE_BITS_ALL &&
1072 (m->flags & PG_FICTITIOUS) == 0)
1073 pmap_enter_quick(pmap, addr, m, entry->protection);
1074 VM_OBJECT_RUNLOCK(lobject);
1079 * Hold each of the physical pages that are mapped by the specified range of
1080 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1081 * and allow the specified types of access, "prot". If all of the implied
1082 * pages are successfully held, then the number of held pages is returned
1083 * together with pointers to those pages in the array "ma". However, if any
1084 * of the pages cannot be held, -1 is returned.
1087 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1088 vm_prot_t prot, vm_page_t *ma, int max_count)
1090 vm_offset_t end, va;
1093 boolean_t pmap_failed;
1097 end = round_page(addr + len);
1098 addr = trunc_page(addr);
1101 * Check for illegal addresses.
1103 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1106 if (atop(end - addr) > max_count)
1107 panic("vm_fault_quick_hold_pages: count > max_count");
1108 count = atop(end - addr);
1111 * Most likely, the physical pages are resident in the pmap, so it is
1112 * faster to try pmap_extract_and_hold() first.
1114 pmap_failed = FALSE;
1115 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1116 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1119 else if ((prot & VM_PROT_WRITE) != 0 &&
1120 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1122 * Explicitly dirty the physical page. Otherwise, the
1123 * caller's changes may go unnoticed because they are
1124 * performed through an unmanaged mapping or by a DMA
1127 * The object lock is not held here.
1128 * See vm_page_clear_dirty_mask().
1135 * One or more pages could not be held by the pmap. Either no
1136 * page was mapped at the specified virtual address or that
1137 * mapping had insufficient permissions. Attempt to fault in
1138 * and hold these pages.
1140 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1141 if (*mp == NULL && vm_fault_hold(map, va, prot,
1142 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1147 for (mp = ma; mp < ma + count; mp++)
1150 vm_page_unhold(*mp);
1151 vm_page_unlock(*mp);
1159 * Wire down a range of virtual addresses in a map.
1162 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1163 boolean_t fictitious)
1169 * We simulate a fault to get the page and enter it in the physical
1170 * map. For user wiring, we only ask for read access on currently
1171 * read-only sections.
1173 for (va = start; va < end; va += PAGE_SIZE) {
1174 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1177 vm_fault_unwire(map, start, va, fictitious);
1181 return (KERN_SUCCESS);
1187 * Unwire a range of virtual addresses in a map.
1190 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1191 boolean_t fictitious)
1198 pmap = vm_map_pmap(map);
1201 * Since the pages are wired down, we must be able to get their
1202 * mappings from the physical map system.
1204 for (va = start; va < end; va += PAGE_SIZE) {
1205 pa = pmap_extract(pmap, va);
1207 pmap_change_wiring(pmap, va, FALSE);
1209 m = PHYS_TO_VM_PAGE(pa);
1211 vm_page_unwire(m, TRUE);
1220 * vm_fault_copy_entry
1222 * Create new shadow object backing dst_entry with private copy of
1223 * all underlying pages. When src_entry is equal to dst_entry,
1224 * function implements COW for wired-down map entry. Otherwise,
1225 * it forks wired entry into dst_map.
1227 * In/out conditions:
1228 * The source and destination maps must be locked for write.
1229 * The source map entry must be wired down (or be a sharing map
1230 * entry corresponding to a main map entry that is wired down).
1233 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1234 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1235 vm_ooffset_t *fork_charge)
1237 vm_object_t backing_object, dst_object, object, src_object;
1238 vm_pindex_t dst_pindex, pindex, src_pindex;
1239 vm_prot_t access, prot;
1249 upgrade = src_entry == dst_entry;
1250 access = prot = dst_entry->protection;
1252 src_object = src_entry->object.vm_object;
1253 src_pindex = OFF_TO_IDX(src_entry->offset);
1255 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1256 dst_object = src_object;
1257 vm_object_reference(dst_object);
1260 * Create the top-level object for the destination entry. (Doesn't
1261 * actually shadow anything - we copy the pages directly.)
1263 dst_object = vm_object_allocate(OBJT_DEFAULT,
1264 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1265 #if VM_NRESERVLEVEL > 0
1266 dst_object->flags |= OBJ_COLORED;
1267 dst_object->pg_color = atop(dst_entry->start);
1271 VM_OBJECT_WLOCK(dst_object);
1272 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1273 ("vm_fault_copy_entry: vm_object not NULL"));
1274 if (src_object != dst_object) {
1275 dst_entry->object.vm_object = dst_object;
1276 dst_entry->offset = 0;
1277 dst_object->charge = dst_entry->end - dst_entry->start;
1279 if (fork_charge != NULL) {
1280 KASSERT(dst_entry->cred == NULL,
1281 ("vm_fault_copy_entry: leaked swp charge"));
1282 dst_object->cred = curthread->td_ucred;
1283 crhold(dst_object->cred);
1284 *fork_charge += dst_object->charge;
1285 } else if (dst_object->cred == NULL) {
1286 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1288 dst_object->cred = dst_entry->cred;
1289 dst_entry->cred = NULL;
1293 * If not an upgrade, then enter the mappings in the pmap as
1294 * read and/or execute accesses. Otherwise, enter them as
1297 * A writeable large page mapping is only created if all of
1298 * the constituent small page mappings are modified. Marking
1299 * PTEs as modified on inception allows promotion to happen
1300 * without taking potentially large number of soft faults.
1303 access &= ~VM_PROT_WRITE;
1306 * Loop through all of the virtual pages within the entry's
1307 * range, copying each page from the source object to the
1308 * destination object. Since the source is wired, those pages
1309 * must exist. In contrast, the destination is pageable.
1310 * Since the destination object does share any backing storage
1311 * with the source object, all of its pages must be dirtied,
1312 * regardless of whether they can be written.
1314 for (vaddr = dst_entry->start, dst_pindex = 0;
1315 vaddr < dst_entry->end;
1316 vaddr += PAGE_SIZE, dst_pindex++) {
1319 * Find the page in the source object, and copy it in.
1320 * Because the source is wired down, the page will be
1323 if (src_object != dst_object)
1324 VM_OBJECT_RLOCK(src_object);
1325 object = src_object;
1326 pindex = src_pindex + dst_pindex;
1327 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1328 (backing_object = object->backing_object) != NULL) {
1330 * Unless the source mapping is read-only or
1331 * it is presently being upgraded from
1332 * read-only, the first object in the shadow
1333 * chain should provide all of the pages. In
1334 * other words, this loop body should never be
1335 * executed when the source mapping is already
1338 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1340 ("vm_fault_copy_entry: main object missing page"));
1342 VM_OBJECT_RLOCK(backing_object);
1343 pindex += OFF_TO_IDX(object->backing_object_offset);
1344 if (object != dst_object)
1345 VM_OBJECT_RUNLOCK(object);
1346 object = backing_object;
1348 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1350 if (object != dst_object) {
1352 * Allocate a page in the destination object.
1354 dst_m = vm_page_alloc(dst_object, (src_object ==
1355 dst_object ? src_pindex : 0) + dst_pindex,
1357 if (dst_m == NULL) {
1358 VM_OBJECT_WUNLOCK(dst_object);
1359 VM_OBJECT_RUNLOCK(object);
1361 VM_OBJECT_WLOCK(dst_object);
1364 pmap_copy_page(src_m, dst_m);
1365 VM_OBJECT_RUNLOCK(object);
1366 dst_m->valid = VM_PAGE_BITS_ALL;
1367 dst_m->dirty = VM_PAGE_BITS_ALL;
1370 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1372 vm_page_xbusy(dst_m);
1373 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1374 ("invalid dst page %p", dst_m));
1376 VM_OBJECT_WUNLOCK(dst_object);
1379 * Enter it in the pmap. If a wired, copy-on-write
1380 * mapping is being replaced by a write-enabled
1381 * mapping, then wire that new mapping.
1383 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1386 * Mark it no longer busy, and put it on the active list.
1388 VM_OBJECT_WLOCK(dst_object);
1391 if (src_m != dst_m) {
1392 vm_page_lock(src_m);
1393 vm_page_unwire(src_m, 0);
1394 vm_page_unlock(src_m);
1395 vm_page_lock(dst_m);
1396 vm_page_wire(dst_m);
1397 vm_page_unlock(dst_m);
1399 KASSERT(dst_m->wire_count > 0,
1400 ("dst_m %p is not wired", dst_m));
1403 vm_page_lock(dst_m);
1404 vm_page_activate(dst_m);
1405 vm_page_unlock(dst_m);
1407 vm_page_xunbusy(dst_m);
1409 VM_OBJECT_WUNLOCK(dst_object);
1411 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1412 vm_object_deallocate(src_object);
1418 * This routine checks around the requested page for other pages that
1419 * might be able to be faulted in. This routine brackets the viable
1420 * pages for the pages to be paged in.
1423 * m, rbehind, rahead
1426 * marray (array of vm_page_t), reqpage (index of requested page)
1429 * number of pages in marray
1432 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1441 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1443 int cbehind, cahead;
1445 VM_OBJECT_ASSERT_WLOCKED(m->object);
1449 cbehind = cahead = 0;
1452 * if the requested page is not available, then give up now
1454 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1458 if ((cbehind == 0) && (cahead == 0)) {
1464 if (rahead > cahead) {
1468 if (rbehind > cbehind) {
1473 * scan backward for the read behind pages -- in memory
1476 if (rbehind > pindex) {
1480 startpindex = pindex - rbehind;
1483 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1484 rtm->pindex >= startpindex)
1485 startpindex = rtm->pindex + 1;
1487 /* tpindex is unsigned; beware of numeric underflow. */
1488 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1489 tpindex < pindex; i++, tpindex--) {
1491 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1492 VM_ALLOC_IFNOTCACHED);
1495 * Shift the allocated pages to the
1496 * beginning of the array.
1498 for (j = 0; j < i; j++) {
1499 marray[j] = marray[j + tpindex + 1 -
1505 marray[tpindex - startpindex] = rtm;
1513 /* page offset of the required page */
1516 tpindex = pindex + 1;
1520 * scan forward for the read ahead pages
1522 endpindex = tpindex + rahead;
1523 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1524 endpindex = rtm->pindex;
1525 if (endpindex > object->size)
1526 endpindex = object->size;
1528 for (; tpindex < endpindex; i++, tpindex++) {
1530 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1531 VM_ALLOC_IFNOTCACHED);
1539 /* return number of pages */
1544 * Block entry into the machine-independent layer's page fault handler by
1545 * the calling thread. Subsequent calls to vm_fault() by that thread will
1546 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1547 * spurious page faults.
1550 vm_fault_disable_pagefaults(void)
1553 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1557 vm_fault_enable_pagefaults(int save)
1560 curthread_pflags_restore(save);