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("$FreeBSD$");
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/mutex.h>
85 #include <sys/resourcevar.h>
86 #include <sys/sysctl.h>
87 #include <sys/vmmeter.h>
88 #include <sys/vnode.h>
91 #include <vm/vm_param.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_extern.h>
101 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
105 #define PAGEORDER_SIZE (PFBAK+PFFOR)
107 static int prefault_pageorder[] = {
108 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
109 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
110 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
111 -4 * PAGE_SIZE, 4 * PAGE_SIZE
114 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
115 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
117 #define VM_FAULT_READ_AHEAD 8
118 #define VM_FAULT_READ_BEHIND 7
119 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
126 vm_object_t first_object;
127 vm_pindex_t first_pindex;
129 vm_map_entry_t entry;
130 int lookup_still_valid;
136 release_page(struct faultstate *fs)
139 vm_page_wakeup(fs->m);
140 vm_page_lock_queues();
141 vm_page_deactivate(fs->m);
142 vm_page_unlock_queues();
147 unlock_map(struct faultstate *fs)
150 if (fs->lookup_still_valid) {
151 vm_map_lookup_done(fs->map, fs->entry);
152 fs->lookup_still_valid = FALSE;
157 unlock_and_deallocate(struct faultstate *fs)
160 vm_object_pip_wakeup(fs->object);
161 VM_OBJECT_UNLOCK(fs->object);
162 if (fs->object != fs->first_object) {
163 VM_OBJECT_LOCK(fs->first_object);
164 vm_page_lock_queues();
165 vm_page_free(fs->first_m);
166 vm_page_unlock_queues();
167 vm_object_pip_wakeup(fs->first_object);
168 VM_OBJECT_UNLOCK(fs->first_object);
171 vm_object_deallocate(fs->first_object);
173 if (fs->vp != NULL) {
177 VFS_UNLOCK_GIANT(fs->vfslocked);
182 * TRYPAGER - used by vm_fault to calculate whether the pager for the
183 * current object *might* contain the page.
185 * default objects are zero-fill, there is no real pager.
187 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
188 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
193 * Handle a page fault occurring at the given address,
194 * requiring the given permissions, in the map specified.
195 * If successful, the page is inserted into the
196 * associated physical map.
198 * NOTE: the given address should be truncated to the
199 * proper page address.
201 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
202 * a standard error specifying why the fault is fatal is returned.
205 * The map in question must be referenced, and remains so.
206 * Caller may hold no locks.
209 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
213 int is_first_object_locked, result;
214 boolean_t are_queues_locked, growstack, wired;
216 vm_object_t next_object;
217 vm_page_t marray[VM_FAULT_READ];
219 int faultcount, ahead, behind;
220 struct faultstate fs;
226 PCPU_INC(cnt.v_vm_faults);
229 faultcount = behind = 0;
234 * Find the backing store object and offset into it to begin the
238 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
239 &fs.first_object, &fs.first_pindex, &prot, &wired);
240 if (result != KERN_SUCCESS) {
241 if (growstack && result == KERN_INVALID_ADDRESS &&
243 result = vm_map_growstack(curproc, vaddr);
244 if (result != KERN_SUCCESS)
245 return (KERN_FAILURE);
252 map_generation = fs.map->timestamp;
254 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
255 panic("vm_fault: fault on nofault entry, addr: %lx",
260 * Make a reference to this object to prevent its disposal while we
261 * are messing with it. Once we have the reference, the map is free
262 * to be diddled. Since objects reference their shadows (and copies),
263 * they will stay around as well.
265 * Bump the paging-in-progress count to prevent size changes (e.g.
266 * truncation operations) during I/O. This must be done after
267 * obtaining the vnode lock in order to avoid possible deadlocks.
269 VM_OBJECT_LOCK(fs.first_object);
270 vm_object_reference_locked(fs.first_object);
271 vm_object_pip_add(fs.first_object, 1);
273 fs.lookup_still_valid = TRUE;
276 fault_type = prot | (fault_type & VM_PROT_COPY);
281 * Search for the page at object/offset.
283 fs.object = fs.first_object;
284 fs.pindex = fs.first_pindex;
287 * If the object is dead, we stop here
289 if (fs.object->flags & OBJ_DEAD) {
290 unlock_and_deallocate(&fs);
291 return (KERN_PROTECTION_FAILURE);
295 * See if page is resident
297 fs.m = vm_page_lookup(fs.object, fs.pindex);
300 * check for page-based copy on write.
301 * We check fs.object == fs.first_object so
302 * as to ensure the legacy COW mechanism is
303 * used when the page in question is part of
304 * a shadow object. Otherwise, vm_page_cowfault()
305 * removes the page from the backing object,
306 * which is not what we want.
308 vm_page_lock_queues();
310 (fault_type & VM_PROT_WRITE) &&
311 (fs.object == fs.first_object)) {
312 vm_page_cowfault(fs.m);
313 vm_page_unlock_queues();
314 unlock_and_deallocate(&fs);
319 * Wait/Retry if the page is busy. We have to do this
320 * if the page is busy via either VPO_BUSY or
321 * vm_page_t->busy because the vm_pager may be using
322 * vm_page_t->busy for pageouts ( and even pageins if
323 * it is the vnode pager ), and we could end up trying
324 * to pagein and pageout the same page simultaneously.
326 * We can theoretically allow the busy case on a read
327 * fault if the page is marked valid, but since such
328 * pages are typically already pmap'd, putting that
329 * special case in might be more effort then it is
330 * worth. We cannot under any circumstances mess
331 * around with a vm_page_t->busy page except, perhaps,
334 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
335 vm_page_unlock_queues();
336 VM_OBJECT_UNLOCK(fs.object);
337 if (fs.object != fs.first_object) {
338 VM_OBJECT_LOCK(fs.first_object);
339 vm_page_lock_queues();
340 vm_page_free(fs.first_m);
341 vm_page_unlock_queues();
342 vm_object_pip_wakeup(fs.first_object);
343 VM_OBJECT_UNLOCK(fs.first_object);
347 VM_OBJECT_LOCK(fs.object);
348 if (fs.m == vm_page_lookup(fs.object,
350 vm_page_sleep_if_busy(fs.m, TRUE,
353 vm_object_pip_wakeup(fs.object);
354 VM_OBJECT_UNLOCK(fs.object);
355 PCPU_INC(cnt.v_intrans);
356 vm_object_deallocate(fs.first_object);
359 vm_pageq_remove(fs.m);
360 vm_page_unlock_queues();
363 * Mark page busy for other processes, and the
364 * pagedaemon. If it still isn't completely valid
365 * (readable), jump to readrest, else break-out ( we
369 if (fs.m->valid != VM_PAGE_BITS_ALL &&
370 fs.m->object != kernel_object && fs.m->object != kmem_object) {
378 * Page is not resident, If this is the search termination
379 * or the pager might contain the page, allocate a new page.
381 if (TRYPAGER || fs.object == fs.first_object) {
382 if (fs.pindex >= fs.object->size) {
383 unlock_and_deallocate(&fs);
384 return (KERN_PROTECTION_FAILURE);
388 * Allocate a new page for this object/offset pair.
391 if (!vm_page_count_severe()) {
392 #if VM_NRESERVLEVEL > 0
393 if ((fs.object->flags & OBJ_COLORED) == 0) {
394 fs.object->flags |= OBJ_COLORED;
395 fs.object->pg_color = atop(vaddr) -
399 fs.m = vm_page_alloc(fs.object, fs.pindex,
400 (fs.object->type == OBJT_VNODE ||
401 fs.object->backing_object != NULL) ?
402 VM_ALLOC_NORMAL : VM_ALLOC_ZERO);
405 unlock_and_deallocate(&fs);
408 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
414 * We have found a valid page or we have allocated a new page.
415 * The page thus may not be valid or may not be entirely
418 * Attempt to fault-in the page if there is a chance that the
419 * pager has it, and potentially fault in additional pages
425 u_char behavior = vm_map_entry_behavior(fs.entry);
427 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
431 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
432 if (behind > VM_FAULT_READ_BEHIND)
433 behind = VM_FAULT_READ_BEHIND;
435 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
436 if (ahead > VM_FAULT_READ_AHEAD)
437 ahead = VM_FAULT_READ_AHEAD;
439 is_first_object_locked = FALSE;
440 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
441 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
442 fs.pindex >= fs.entry->lastr &&
443 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
444 (fs.first_object == fs.object ||
445 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
446 fs.first_object->type != OBJT_DEVICE &&
447 fs.first_object->type != OBJT_PHYS &&
448 fs.first_object->type != OBJT_SG) {
449 vm_pindex_t firstpindex, tmppindex;
451 if (fs.first_pindex < 2 * VM_FAULT_READ)
454 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
456 are_queues_locked = FALSE;
458 * note: partially valid pages cannot be
459 * included in the lookahead - NFS piecemeal
460 * writes will barf on it badly.
462 for (tmppindex = fs.first_pindex - 1;
463 tmppindex >= firstpindex;
467 mt = vm_page_lookup(fs.first_object, tmppindex);
468 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
471 (mt->oflags & VPO_BUSY))
473 if (!are_queues_locked) {
474 are_queues_locked = TRUE;
475 vm_page_lock_queues();
477 if (mt->hold_count ||
482 vm_page_deactivate(mt);
487 if (are_queues_locked)
488 vm_page_unlock_queues();
492 if (is_first_object_locked)
493 VM_OBJECT_UNLOCK(fs.first_object);
496 * Call the pager to retrieve the data, if any, after
497 * releasing the lock on the map. We hold a ref on
498 * fs.object and the pages are VPO_BUSY'd.
503 if (fs.object->type == OBJT_VNODE) {
504 vp = fs.object->handle;
507 else if (fs.vp != NULL) {
511 locked = VOP_ISLOCKED(vp);
513 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
515 if (!mtx_trylock(&Giant)) {
516 VM_OBJECT_UNLOCK(fs.object);
518 VM_OBJECT_LOCK(fs.object);
522 if (locked != LK_EXCLUSIVE)
524 /* Do not sleep for vnode lock while fs.m is busy */
525 error = vget(vp, locked | LK_CANRECURSE |
526 LK_NOWAIT, curthread);
530 vfslocked = fs.vfslocked;
531 fs.vfslocked = 0; /* Keep Giant */
534 unlock_and_deallocate(&fs);
535 error = vget(vp, locked | LK_RETRY |
536 LK_CANRECURSE, curthread);
539 fs.vfslocked = vfslocked;
541 ("vm_fault: vget failed"));
547 KASSERT(fs.vp == NULL || !fs.map->system_map,
548 ("vm_fault: vnode-backed object mapped by system map"));
551 * now we find out if any other pages should be paged
552 * in at this time this routine checks to see if the
553 * pages surrounding this fault reside in the same
554 * object as the page for this fault. If they do,
555 * then they are faulted in also into the object. The
556 * array "marray" returned contains an array of
557 * vm_page_t structs where one of them is the
558 * vm_page_t passed to the routine. The reqpage
559 * return value is the index into the marray for the
560 * vm_page_t passed to the routine.
562 * fs.m plus the additional pages are VPO_BUSY'd.
564 faultcount = vm_fault_additional_pages(
565 fs.m, behind, ahead, marray, &reqpage);
568 vm_pager_get_pages(fs.object, marray, faultcount,
569 reqpage) : VM_PAGER_FAIL;
571 if (rv == VM_PAGER_OK) {
573 * Found the page. Leave it busy while we play
578 * Relookup in case pager changed page. Pager
579 * is responsible for disposition of old page
582 fs.m = vm_page_lookup(fs.object, fs.pindex);
584 unlock_and_deallocate(&fs);
589 break; /* break to PAGE HAS BEEN FOUND */
592 * Remove the bogus page (which does not exist at this
593 * object/offset); before doing so, we must get back
594 * our object lock to preserve our invariant.
596 * Also wake up any other process that may want to bring
599 * If this is the top-level object, we must leave the
600 * busy page to prevent another process from rushing
601 * past us, and inserting the page in that object at
602 * the same time that we are.
604 if (rv == VM_PAGER_ERROR)
605 printf("vm_fault: pager read error, pid %d (%s)\n",
606 curproc->p_pid, curproc->p_comm);
608 * Data outside the range of the pager or an I/O error
611 * XXX - the check for kernel_map is a kludge to work
612 * around having the machine panic on a kernel space
613 * fault w/ I/O error.
615 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
616 (rv == VM_PAGER_BAD)) {
617 vm_page_lock_queues();
619 vm_page_unlock_queues();
621 unlock_and_deallocate(&fs);
622 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
624 if (fs.object != fs.first_object) {
625 vm_page_lock_queues();
627 vm_page_unlock_queues();
630 * XXX - we cannot just fall out at this
631 * point, m has been freed and is invalid!
637 * We get here if the object has default pager (or unwiring)
638 * or the pager doesn't have the page.
640 if (fs.object == fs.first_object)
644 * Move on to the next object. Lock the next object before
645 * unlocking the current one.
647 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
648 next_object = fs.object->backing_object;
649 if (next_object == NULL) {
651 * If there's no object left, fill the page in the top
654 if (fs.object != fs.first_object) {
655 vm_object_pip_wakeup(fs.object);
656 VM_OBJECT_UNLOCK(fs.object);
658 fs.object = fs.first_object;
659 fs.pindex = fs.first_pindex;
661 VM_OBJECT_LOCK(fs.object);
666 * Zero the page if necessary and mark it valid.
668 if ((fs.m->flags & PG_ZERO) == 0) {
669 pmap_zero_page(fs.m);
671 PCPU_INC(cnt.v_ozfod);
673 PCPU_INC(cnt.v_zfod);
674 fs.m->valid = VM_PAGE_BITS_ALL;
675 break; /* break to PAGE HAS BEEN FOUND */
677 KASSERT(fs.object != next_object,
678 ("object loop %p", next_object));
679 VM_OBJECT_LOCK(next_object);
680 vm_object_pip_add(next_object, 1);
681 if (fs.object != fs.first_object)
682 vm_object_pip_wakeup(fs.object);
683 VM_OBJECT_UNLOCK(fs.object);
684 fs.object = next_object;
688 KASSERT((fs.m->oflags & VPO_BUSY) != 0,
689 ("vm_fault: not busy after main loop"));
692 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
697 * If the page is being written, but isn't already owned by the
698 * top-level object, we have to copy it into a new page owned by the
701 if (fs.object != fs.first_object) {
703 * We only really need to copy if we want to write it.
705 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
707 * This allows pages to be virtually copied from a
708 * backing_object into the first_object, where the
709 * backing object has no other refs to it, and cannot
710 * gain any more refs. Instead of a bcopy, we just
711 * move the page from the backing object to the
712 * first object. Note that we must mark the page
713 * dirty in the first object so that it will go out
714 * to swap when needed.
716 is_first_object_locked = FALSE;
719 * Only one shadow object
721 (fs.object->shadow_count == 1) &&
723 * No COW refs, except us
725 (fs.object->ref_count == 1) &&
727 * No one else can look this object up
729 (fs.object->handle == NULL) &&
731 * No other ways to look the object up
733 ((fs.object->type == OBJT_DEFAULT) ||
734 (fs.object->type == OBJT_SWAP)) &&
735 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
737 * We don't chase down the shadow chain
739 fs.object == fs.first_object->backing_object) {
740 vm_page_lock_queues();
742 * get rid of the unnecessary page
744 vm_page_free(fs.first_m);
746 * grab the page and put it into the
747 * process'es object. The page is
748 * automatically made dirty.
750 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
751 vm_page_unlock_queues();
755 PCPU_INC(cnt.v_cow_optim);
758 * Oh, well, lets copy it.
760 pmap_copy_page(fs.m, fs.first_m);
761 fs.first_m->valid = VM_PAGE_BITS_ALL;
762 if (wired && (fault_flags &
763 VM_FAULT_CHANGE_WIRING) == 0) {
764 vm_page_lock_queues();
765 vm_page_wire(fs.first_m);
766 vm_page_unwire(fs.m, FALSE);
767 vm_page_unlock_queues();
770 * We no longer need the old page or object.
775 * fs.object != fs.first_object due to above
778 vm_object_pip_wakeup(fs.object);
779 VM_OBJECT_UNLOCK(fs.object);
781 * Only use the new page below...
783 fs.object = fs.first_object;
784 fs.pindex = fs.first_pindex;
786 if (!is_first_object_locked)
787 VM_OBJECT_LOCK(fs.object);
788 PCPU_INC(cnt.v_cow_faults);
790 prot &= ~VM_PROT_WRITE;
795 * We must verify that the maps have not changed since our last
798 if (!fs.lookup_still_valid) {
799 vm_object_t retry_object;
800 vm_pindex_t retry_pindex;
801 vm_prot_t retry_prot;
803 if (!vm_map_trylock_read(fs.map)) {
805 unlock_and_deallocate(&fs);
808 fs.lookup_still_valid = TRUE;
809 if (fs.map->timestamp != map_generation) {
810 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
811 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
814 * If we don't need the page any longer, put it on the inactive
815 * list (the easiest thing to do here). If no one needs it,
816 * pageout will grab it eventually.
818 if (result != KERN_SUCCESS) {
820 unlock_and_deallocate(&fs);
823 * If retry of map lookup would have blocked then
824 * retry fault from start.
826 if (result == KERN_FAILURE)
830 if ((retry_object != fs.first_object) ||
831 (retry_pindex != fs.first_pindex)) {
833 unlock_and_deallocate(&fs);
838 * Check whether the protection has changed or the object has
839 * been copied while we left the map unlocked. Changing from
840 * read to write permission is OK - we leave the page
841 * write-protected, and catch the write fault. Changing from
842 * write to read permission means that we can't mark the page
843 * write-enabled after all.
849 * If the page was filled by a pager, update the map entry's
850 * last read offset. Since the pager does not return the
851 * actual set of pages that it read, this update is based on
852 * the requested set. Typically, the requested and actual
855 * XXX The following assignment modifies the map
856 * without holding a write lock on it.
859 fs.entry->lastr = fs.pindex + faultcount - behind;
861 if (prot & VM_PROT_WRITE) {
862 vm_object_set_writeable_dirty(fs.object);
865 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
866 * if the page is already dirty to prevent data written with
867 * the expectation of being synced from not being synced.
868 * Likewise if this entry does not request NOSYNC then make
869 * sure the page isn't marked NOSYNC. Applications sharing
870 * data should use the same flags to avoid ping ponging.
872 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
873 if (fs.m->dirty == 0)
874 fs.m->oflags |= VPO_NOSYNC;
876 fs.m->oflags &= ~VPO_NOSYNC;
880 * If the fault is a write, we know that this page is being
881 * written NOW so dirty it explicitly to save on
882 * pmap_is_modified() calls later.
884 * Also tell the backing pager, if any, that it should remove
885 * any swap backing since the page is now dirty.
887 if ((fault_type & VM_PROT_WRITE) != 0 &&
888 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) {
890 vm_pager_page_unswapped(fs.m);
895 * Page had better still be busy
897 KASSERT(fs.m->oflags & VPO_BUSY,
898 ("vm_fault: page %p not busy!", fs.m));
900 * Page must be completely valid or it is not fit to
901 * map into user space. vm_pager_get_pages() ensures this.
903 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
904 ("vm_fault: page %p partially invalid", fs.m));
905 VM_OBJECT_UNLOCK(fs.object);
908 * Put this page into the physical map. We had to do the unlock above
909 * because pmap_enter() may sleep. We don't put the page
910 * back on the active queue until later so that the pageout daemon
911 * won't find it (yet).
913 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
914 if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
915 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
916 VM_OBJECT_LOCK(fs.object);
917 vm_page_lock_queues();
918 vm_page_flag_set(fs.m, PG_REFERENCED);
921 * If the page is not wired down, then put it where the pageout daemon
924 if (fault_flags & VM_FAULT_CHANGE_WIRING) {
928 vm_page_unwire(fs.m, 1);
930 vm_page_activate(fs.m);
932 vm_page_unlock_queues();
933 vm_page_wakeup(fs.m);
936 * Unlock everything, and return
938 unlock_and_deallocate(&fs);
940 curthread->td_ru.ru_majflt++;
942 curthread->td_ru.ru_minflt++;
944 return (KERN_SUCCESS);
948 * vm_fault_prefault provides a quick way of clustering
949 * pagefaults into a processes address space. It is a "cousin"
950 * of vm_map_pmap_enter, except it runs at page fault time instead
954 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
957 vm_offset_t addr, starta;
962 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
965 object = entry->object.vm_object;
967 starta = addra - PFBAK * PAGE_SIZE;
968 if (starta < entry->start) {
969 starta = entry->start;
970 } else if (starta > addra) {
974 for (i = 0; i < PAGEORDER_SIZE; i++) {
975 vm_object_t backing_object, lobject;
977 addr = addra + prefault_pageorder[i];
978 if (addr > addra + (PFFOR * PAGE_SIZE))
981 if (addr < starta || addr >= entry->end)
984 if (!pmap_is_prefaultable(pmap, addr))
987 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
989 VM_OBJECT_LOCK(lobject);
990 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
991 lobject->type == OBJT_DEFAULT &&
992 (backing_object = lobject->backing_object) != NULL) {
993 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
994 0, ("vm_fault_prefault: unaligned object offset"));
995 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
996 VM_OBJECT_LOCK(backing_object);
997 VM_OBJECT_UNLOCK(lobject);
998 lobject = backing_object;
1001 * give-up when a page is not in memory
1004 VM_OBJECT_UNLOCK(lobject);
1007 if (m->valid == VM_PAGE_BITS_ALL &&
1008 (m->flags & PG_FICTITIOUS) == 0) {
1009 vm_page_lock_queues();
1010 pmap_enter_quick(pmap, addr, m, entry->protection);
1011 vm_page_unlock_queues();
1013 VM_OBJECT_UNLOCK(lobject);
1020 * Ensure that the requested virtual address, which may be in userland,
1021 * is valid. Fault-in the page if necessary. Return -1 on failure.
1024 vm_fault_quick(caddr_t v, int prot)
1028 if (prot & VM_PROT_WRITE)
1029 r = subyte(v, fubyte(v));
1038 * Wire down a range of virtual addresses in a map.
1041 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1042 boolean_t fictitious)
1048 * We simulate a fault to get the page and enter it in the physical
1049 * map. For user wiring, we only ask for read access on currently
1050 * read-only sections.
1052 for (va = start; va < end; va += PAGE_SIZE) {
1053 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1056 vm_fault_unwire(map, start, va, fictitious);
1060 return (KERN_SUCCESS);
1066 * Unwire a range of virtual addresses in a map.
1069 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1070 boolean_t fictitious)
1076 pmap = vm_map_pmap(map);
1079 * Since the pages are wired down, we must be able to get their
1080 * mappings from the physical map system.
1082 for (va = start; va < end; va += PAGE_SIZE) {
1083 pa = pmap_extract(pmap, va);
1085 pmap_change_wiring(pmap, va, FALSE);
1087 vm_page_lock_queues();
1088 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1089 vm_page_unlock_queues();
1097 * vm_fault_copy_entry
1099 * Create new shadow object backing dst_entry with private copy of
1100 * all underlying pages. When src_entry is equal to dst_entry,
1101 * function implements COW for wired-down map entry. Otherwise,
1102 * it forks wired entry into dst_map.
1104 * In/out conditions:
1105 * The source and destination maps must be locked for write.
1106 * The source map entry must be wired down (or be a sharing map
1107 * entry corresponding to a main map entry that is wired down).
1110 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1111 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1112 vm_ooffset_t *fork_charge)
1114 vm_object_t backing_object, dst_object, object, src_object;
1115 vm_pindex_t dst_pindex, pindex, src_pindex;
1116 vm_prot_t access, prot;
1120 boolean_t src_readonly, upgrade;
1126 upgrade = src_entry == dst_entry;
1128 src_object = src_entry->object.vm_object;
1129 src_pindex = OFF_TO_IDX(src_entry->offset);
1130 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1133 * Create the top-level object for the destination entry. (Doesn't
1134 * actually shadow anything - we copy the pages directly.)
1136 dst_object = vm_object_allocate(OBJT_DEFAULT,
1137 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1138 #if VM_NRESERVLEVEL > 0
1139 dst_object->flags |= OBJ_COLORED;
1140 dst_object->pg_color = atop(dst_entry->start);
1143 VM_OBJECT_LOCK(dst_object);
1144 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1145 ("vm_fault_copy_entry: vm_object not NULL"));
1146 dst_entry->object.vm_object = dst_object;
1147 dst_entry->offset = 0;
1148 dst_object->charge = dst_entry->end - dst_entry->start;
1149 if (fork_charge != NULL) {
1150 KASSERT(dst_entry->uip == NULL,
1151 ("vm_fault_copy_entry: leaked swp charge"));
1152 dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1153 uihold(dst_object->uip);
1154 *fork_charge += dst_object->charge;
1156 dst_object->uip = dst_entry->uip;
1157 dst_entry->uip = NULL;
1159 access = prot = dst_entry->protection;
1161 * If not an upgrade, then enter the mappings in the pmap as
1162 * read and/or execute accesses. Otherwise, enter them as
1165 * A writeable large page mapping is only created if all of
1166 * the constituent small page mappings are modified. Marking
1167 * PTEs as modified on inception allows promotion to happen
1168 * without taking potentially large number of soft faults.
1171 access &= ~VM_PROT_WRITE;
1174 * Loop through all of the pages in the entry's range, copying each
1175 * one from the source object (it should be there) to the destination
1178 for (vaddr = dst_entry->start, dst_pindex = 0;
1179 vaddr < dst_entry->end;
1180 vaddr += PAGE_SIZE, dst_pindex++) {
1183 * Allocate a page in the destination object.
1186 dst_m = vm_page_alloc(dst_object, dst_pindex,
1188 if (dst_m == NULL) {
1189 VM_OBJECT_UNLOCK(dst_object);
1191 VM_OBJECT_LOCK(dst_object);
1193 } while (dst_m == NULL);
1196 * Find the page in the source object, and copy it in.
1197 * (Because the source is wired down, the page will be in
1200 VM_OBJECT_LOCK(src_object);
1201 object = src_object;
1202 pindex = src_pindex + dst_pindex;
1203 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1205 (backing_object = object->backing_object) != NULL) {
1207 * Allow fallback to backing objects if we are reading.
1209 VM_OBJECT_LOCK(backing_object);
1210 pindex += OFF_TO_IDX(object->backing_object_offset);
1211 VM_OBJECT_UNLOCK(object);
1212 object = backing_object;
1215 panic("vm_fault_copy_wired: page missing");
1216 pmap_copy_page(src_m, dst_m);
1217 VM_OBJECT_UNLOCK(object);
1218 dst_m->valid = VM_PAGE_BITS_ALL;
1219 VM_OBJECT_UNLOCK(dst_object);
1222 * Enter it in the pmap. If a wired, copy-on-write
1223 * mapping is being replaced by a write-enabled
1224 * mapping, then wire that new mapping.
1226 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1229 * Mark it no longer busy, and put it on the active list.
1231 VM_OBJECT_LOCK(dst_object);
1232 vm_page_lock_queues();
1234 vm_page_unwire(src_m, 0);
1235 vm_page_wire(dst_m);
1237 vm_page_activate(dst_m);
1238 vm_page_unlock_queues();
1239 vm_page_wakeup(dst_m);
1241 VM_OBJECT_UNLOCK(dst_object);
1243 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1244 vm_object_deallocate(src_object);
1250 * This routine checks around the requested page for other pages that
1251 * might be able to be faulted in. This routine brackets the viable
1252 * pages for the pages to be paged in.
1255 * m, rbehind, rahead
1258 * marray (array of vm_page_t), reqpage (index of requested page)
1261 * number of pages in marray
1264 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1273 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1275 int cbehind, cahead;
1277 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1281 cbehind = cahead = 0;
1284 * if the requested page is not available, then give up now
1286 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1290 if ((cbehind == 0) && (cahead == 0)) {
1296 if (rahead > cahead) {
1300 if (rbehind > cbehind) {
1305 * scan backward for the read behind pages -- in memory
1308 if (rbehind > pindex) {
1312 startpindex = pindex - rbehind;
1315 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1316 rtm->pindex >= startpindex)
1317 startpindex = rtm->pindex + 1;
1319 /* tpindex is unsigned; beware of numeric underflow. */
1320 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1321 tpindex < pindex; i++, tpindex--) {
1323 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1324 VM_ALLOC_IFNOTCACHED);
1327 * Shift the allocated pages to the
1328 * beginning of the array.
1330 for (j = 0; j < i; j++) {
1331 marray[j] = marray[j + tpindex + 1 -
1337 marray[tpindex - startpindex] = rtm;
1345 /* page offset of the required page */
1348 tpindex = pindex + 1;
1352 * scan forward for the read ahead pages
1354 endpindex = tpindex + rahead;
1355 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1356 endpindex = rtm->pindex;
1357 if (endpindex > object->size)
1358 endpindex = object->size;
1360 for (; tpindex < endpindex; i++, tpindex++) {
1362 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1363 VM_ALLOC_IFNOTCACHED);
1371 /* return number of pages */