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 <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
81 #include <sys/mutex.h>
83 #include <sys/resourcevar.h>
84 #include <sys/sysctl.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vnode_pager.h>
98 #include <vm/vm_extern.h>
100 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
104 #define PAGEORDER_SIZE (PFBAK+PFFOR)
106 static int prefault_pageorder[] = {
107 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
108 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
109 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
110 -4 * PAGE_SIZE, 4 * PAGE_SIZE
113 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
114 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
116 #define VM_FAULT_READ_AHEAD 8
117 #define VM_FAULT_READ_BEHIND 7
118 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
125 vm_object_t first_object;
126 vm_pindex_t first_pindex;
128 vm_map_entry_t entry;
129 int lookup_still_valid;
134 release_page(struct faultstate *fs)
136 vm_page_lock_queues();
137 vm_page_wakeup(fs->m);
138 vm_page_deactivate(fs->m);
139 vm_page_unlock_queues();
144 unlock_map(struct faultstate *fs)
146 if (fs->lookup_still_valid) {
147 vm_map_lookup_done(fs->map, fs->entry);
148 fs->lookup_still_valid = FALSE;
153 unlock_and_deallocate(struct faultstate *fs)
156 vm_object_pip_wakeup(fs->object);
157 VM_OBJECT_UNLOCK(fs->object);
158 if (fs->object != fs->first_object) {
159 VM_OBJECT_LOCK(fs->first_object);
160 vm_page_lock_queues();
161 vm_page_free(fs->first_m);
162 vm_page_unlock_queues();
163 vm_object_pip_wakeup(fs->first_object);
164 VM_OBJECT_UNLOCK(fs->first_object);
167 vm_object_deallocate(fs->first_object);
169 if (fs->vp != NULL) {
172 vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount);
175 VFS_UNLOCK_GIANT(vfslocked);
177 if (!fs->map->system_map)
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_WIRE_MASK) == 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 growstack, wired;
216 vm_object_t next_object;
217 vm_page_t marray[VM_FAULT_READ];
220 struct faultstate fs;
224 atomic_add_int(&cnt.v_vm_faults, 1);
229 * Find the backing store object and offset into it to begin the
233 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
234 &fs.first_object, &fs.first_pindex, &prot, &wired);
235 if (result != KERN_SUCCESS) {
236 if (result != KERN_PROTECTION_FAILURE ||
237 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
238 if (growstack && result == KERN_INVALID_ADDRESS &&
239 map != kernel_map && curproc != NULL) {
240 result = vm_map_growstack(curproc, vaddr);
241 if (result != KERN_SUCCESS)
242 return (KERN_FAILURE);
250 * If we are user-wiring a r/w segment, and it is COW, then
251 * we need to do the COW operation. Note that we don't COW
252 * currently RO sections now, because it is NOT desirable
253 * to COW .text. We simply keep .text from ever being COW'ed
254 * and take the heat that one cannot debug wired .text sections.
256 result = vm_map_lookup(&fs.map, vaddr,
257 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
258 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
259 if (result != KERN_SUCCESS)
263 * If we don't COW now, on a user wire, the user will never
264 * be able to write to the mapping. If we don't make this
265 * restriction, the bookkeeping would be nearly impossible.
267 * XXX The following assignment modifies the map without
268 * holding a write lock on it.
270 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
271 fs.entry->max_protection &= ~VM_PROT_WRITE;
274 map_generation = fs.map->timestamp;
276 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
277 panic("vm_fault: fault on nofault entry, addr: %lx",
282 * Make a reference to this object to prevent its disposal while we
283 * are messing with it. Once we have the reference, the map is free
284 * to be diddled. Since objects reference their shadows (and copies),
285 * they will stay around as well.
287 * Bump the paging-in-progress count to prevent size changes (e.g.
288 * truncation operations) during I/O. This must be done after
289 * obtaining the vnode lock in order to avoid possible deadlocks.
291 * XXX vnode_pager_lock() can block without releasing the map lock.
293 if (!fs.map->system_map)
295 VM_OBJECT_LOCK(fs.first_object);
296 vm_object_reference_locked(fs.first_object);
297 fs.vp = vnode_pager_lock(fs.first_object);
298 KASSERT(fs.vp == NULL || !fs.map->system_map,
299 ("vm_fault: vnode-backed object mapped by system map"));
300 if (debug_mpsafevm && !fs.map->system_map)
302 vm_object_pip_add(fs.first_object, 1);
304 fs.lookup_still_valid = TRUE;
312 * Search for the page at object/offset.
314 fs.object = fs.first_object;
315 fs.pindex = fs.first_pindex;
318 * If the object is dead, we stop here
320 if (fs.object->flags & OBJ_DEAD) {
321 unlock_and_deallocate(&fs);
322 return (KERN_PROTECTION_FAILURE);
326 * See if page is resident
328 fs.m = vm_page_lookup(fs.object, fs.pindex);
333 * check for page-based copy on write.
334 * We check fs.object == fs.first_object so
335 * as to ensure the legacy COW mechanism is
336 * used when the page in question is part of
337 * a shadow object. Otherwise, vm_page_cowfault()
338 * removes the page from the backing object,
339 * which is not what we want.
341 vm_page_lock_queues();
343 (fault_type & VM_PROT_WRITE) &&
344 (fs.object == fs.first_object)) {
345 vm_page_cowfault(fs.m);
346 vm_page_unlock_queues();
347 unlock_and_deallocate(&fs);
352 * Wait/Retry if the page is busy. We have to do this
353 * if the page is busy via either PG_BUSY or
354 * vm_page_t->busy because the vm_pager may be using
355 * vm_page_t->busy for pageouts ( and even pageins if
356 * it is the vnode pager ), and we could end up trying
357 * to pagein and pageout the same page simultaneously.
359 * We can theoretically allow the busy case on a read
360 * fault if the page is marked valid, but since such
361 * pages are typically already pmap'd, putting that
362 * special case in might be more effort then it is
363 * worth. We cannot under any circumstances mess
364 * around with a vm_page_t->busy page except, perhaps,
367 if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
368 vm_page_unlock_queues();
369 VM_OBJECT_UNLOCK(fs.object);
370 if (fs.object != fs.first_object) {
371 VM_OBJECT_LOCK(fs.first_object);
372 vm_page_lock_queues();
373 vm_page_free(fs.first_m);
374 vm_page_unlock_queues();
375 vm_object_pip_wakeup(fs.first_object);
376 VM_OBJECT_UNLOCK(fs.first_object);
386 VM_OBJECT_LOCK(fs.object);
387 if (fs.m == vm_page_lookup(fs.object,
389 vm_page_lock_queues();
390 if (!vm_page_sleep_if_busy(fs.m, TRUE,
392 vm_page_unlock_queues();
394 vm_object_pip_wakeup(fs.object);
395 VM_OBJECT_UNLOCK(fs.object);
396 atomic_add_int(&cnt.v_intrans, 1);
397 if (!fs.map->system_map)
399 vm_object_deallocate(fs.first_object);
404 vm_pageq_remove_nowakeup(fs.m);
406 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
407 vm_page_activate(fs.m);
408 vm_page_unlock_queues();
409 unlock_and_deallocate(&fs);
415 * Mark page busy for other processes, and the
416 * pagedaemon. If it still isn't completely valid
417 * (readable), jump to readrest, else break-out ( we
421 vm_page_unlock_queues();
422 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
423 fs.m->object != kernel_object && fs.m->object != kmem_object) {
431 * Page is not resident, If this is the search termination
432 * or the pager might contain the page, allocate a new page.
434 if (TRYPAGER || fs.object == fs.first_object) {
435 if (fs.pindex >= fs.object->size) {
436 unlock_and_deallocate(&fs);
437 return (KERN_PROTECTION_FAILURE);
441 * Allocate a new page for this object/offset pair.
444 if (!vm_page_count_severe()) {
445 fs.m = vm_page_alloc(fs.object, fs.pindex,
446 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
449 unlock_and_deallocate(&fs);
457 * We have found a valid page or we have allocated a new page.
458 * The page thus may not be valid or may not be entirely
461 * Attempt to fault-in the page if there is a chance that the
462 * pager has it, and potentially fault in additional pages
469 u_char behavior = vm_map_entry_behavior(fs.entry);
471 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
475 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
476 if (behind > VM_FAULT_READ_BEHIND)
477 behind = VM_FAULT_READ_BEHIND;
479 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
480 if (ahead > VM_FAULT_READ_AHEAD)
481 ahead = VM_FAULT_READ_AHEAD;
483 is_first_object_locked = FALSE;
484 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
485 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
486 fs.pindex >= fs.entry->lastr &&
487 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
488 (fs.first_object == fs.object ||
489 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
490 fs.first_object->type != OBJT_DEVICE) {
491 vm_pindex_t firstpindex, tmppindex;
493 if (fs.first_pindex < 2 * VM_FAULT_READ)
496 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
498 vm_page_lock_queues();
500 * note: partially valid pages cannot be
501 * included in the lookahead - NFS piecemeal
502 * writes will barf on it badly.
504 for (tmppindex = fs.first_pindex - 1;
505 tmppindex >= firstpindex;
509 mt = vm_page_lookup(fs.first_object, tmppindex);
510 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
513 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
519 vm_page_deactivate(mt);
524 vm_page_unlock_queues();
528 if (is_first_object_locked)
529 VM_OBJECT_UNLOCK(fs.first_object);
531 * now we find out if any other pages should be paged
532 * in at this time this routine checks to see if the
533 * pages surrounding this fault reside in the same
534 * object as the page for this fault. If they do,
535 * then they are faulted in also into the object. The
536 * array "marray" returned contains an array of
537 * vm_page_t structs where one of them is the
538 * vm_page_t passed to the routine. The reqpage
539 * return value is the index into the marray for the
540 * vm_page_t passed to the routine.
542 * fs.m plus the additional pages are PG_BUSY'd.
544 * XXX vm_fault_additional_pages() can block
545 * without releasing the map lock.
547 faultcount = vm_fault_additional_pages(
548 fs.m, behind, ahead, marray, &reqpage);
551 * update lastr imperfectly (we do not know how much
552 * getpages will actually read), but good enough.
554 * XXX The following assignment modifies the map
555 * without holding a write lock on it.
557 fs.entry->lastr = fs.pindex + faultcount - behind;
560 * Call the pager to retrieve the data, if any, after
561 * releasing the lock on the map. We hold a ref on
562 * fs.object and the pages are PG_BUSY'd.
567 vm_pager_get_pages(fs.object, marray, faultcount,
568 reqpage) : VM_PAGER_FAIL;
570 if (rv == VM_PAGER_OK) {
572 * Found the page. Leave it busy while we play
577 * Relookup in case pager changed page. Pager
578 * is responsible for disposition of old page
581 fs.m = vm_page_lookup(fs.object, fs.pindex);
583 unlock_and_deallocate(&fs);
588 break; /* break to PAGE HAS BEEN FOUND */
591 * Remove the bogus page (which does not exist at this
592 * object/offset); before doing so, we must get back
593 * our object lock to preserve our invariant.
595 * Also wake up any other process that may want to bring
598 * If this is the top-level object, we must leave the
599 * busy page to prevent another process from rushing
600 * past us, and inserting the page in that object at
601 * the same time that we are.
603 if (rv == VM_PAGER_ERROR)
604 printf("vm_fault: pager read error, pid %d (%s)\n",
605 curproc->p_pid, curproc->p_comm);
607 * Data outside the range of the pager or an I/O error
610 * XXX - the check for kernel_map is a kludge to work
611 * around having the machine panic on a kernel space
612 * fault w/ I/O error.
614 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
615 (rv == VM_PAGER_BAD)) {
616 vm_page_lock_queues();
618 vm_page_unlock_queues();
620 unlock_and_deallocate(&fs);
621 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
623 if (fs.object != fs.first_object) {
624 vm_page_lock_queues();
626 vm_page_unlock_queues();
629 * XXX - we cannot just fall out at this
630 * point, m has been freed and is invalid!
636 * We get here if the object has default pager (or unwiring)
637 * or the pager doesn't have the page.
639 if (fs.object == fs.first_object)
643 * Move on to the next object. Lock the next object before
644 * unlocking the current one.
646 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
647 next_object = fs.object->backing_object;
648 if (next_object == NULL) {
650 * If there's no object left, fill the page in the top
653 if (fs.object != fs.first_object) {
654 vm_object_pip_wakeup(fs.object);
655 VM_OBJECT_UNLOCK(fs.object);
657 fs.object = fs.first_object;
658 fs.pindex = fs.first_pindex;
660 VM_OBJECT_LOCK(fs.object);
665 * Zero the page if necessary and mark it valid.
667 if ((fs.m->flags & PG_ZERO) == 0) {
668 pmap_zero_page(fs.m);
670 atomic_add_int(&cnt.v_ozfod, 1);
672 atomic_add_int(&cnt.v_zfod, 1);
673 fs.m->valid = VM_PAGE_BITS_ALL;
674 break; /* break to PAGE HAS BEEN FOUND */
676 KASSERT(fs.object != next_object,
677 ("object loop %p", next_object));
678 VM_OBJECT_LOCK(next_object);
679 vm_object_pip_add(next_object, 1);
680 if (fs.object != fs.first_object)
681 vm_object_pip_wakeup(fs.object);
682 VM_OBJECT_UNLOCK(fs.object);
683 fs.object = next_object;
687 KASSERT((fs.m->flags & PG_BUSY) != 0,
688 ("vm_fault: not busy after main loop"));
691 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
696 * If the page is being written, but isn't already owned by the
697 * top-level object, we have to copy it into a new page owned by the
700 if (fs.object != fs.first_object) {
702 * We only really need to copy if we want to write it.
704 if (fault_type & VM_PROT_WRITE) {
706 * This allows pages to be virtually copied from a
707 * backing_object into the first_object, where the
708 * backing object has no other refs to it, and cannot
709 * gain any more refs. Instead of a bcopy, we just
710 * move the page from the backing object to the
711 * first object. Note that we must mark the page
712 * dirty in the first object so that it will go out
713 * to swap when needed.
715 is_first_object_locked = FALSE;
718 * Only one shadow object
720 (fs.object->shadow_count == 1) &&
722 * No COW refs, except us
724 (fs.object->ref_count == 1) &&
726 * No one else can look this object up
728 (fs.object->handle == NULL) &&
730 * No other ways to look the object up
732 ((fs.object->type == OBJT_DEFAULT) ||
733 (fs.object->type == OBJT_SWAP)) &&
734 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
736 * We don't chase down the shadow chain
738 fs.object == fs.first_object->backing_object) {
739 vm_page_lock_queues();
741 * get rid of the unnecessary page
743 pmap_remove_all(fs.first_m);
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);
752 vm_page_unlock_queues();
755 atomic_add_int(&cnt.v_cow_optim, 1);
758 * Oh, well, lets copy it.
760 pmap_copy_page(fs.m, fs.first_m);
761 fs.first_m->valid = VM_PAGE_BITS_ALL;
765 * We no longer need the old page or object.
770 * fs.object != fs.first_object due to above
773 vm_object_pip_wakeup(fs.object);
774 VM_OBJECT_UNLOCK(fs.object);
776 * Only use the new page below...
778 fs.object = fs.first_object;
779 fs.pindex = fs.first_pindex;
781 if (!is_first_object_locked)
782 VM_OBJECT_LOCK(fs.object);
783 atomic_add_int(&cnt.v_cow_faults, 1);
785 prot &= ~VM_PROT_WRITE;
790 * We must verify that the maps have not changed since our last
793 if (!fs.lookup_still_valid) {
794 vm_object_t retry_object;
795 vm_pindex_t retry_pindex;
796 vm_prot_t retry_prot;
798 if (!vm_map_trylock_read(fs.map)) {
800 unlock_and_deallocate(&fs);
803 fs.lookup_still_valid = TRUE;
804 if (fs.map->timestamp != map_generation) {
805 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
806 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
809 * If we don't need the page any longer, put it on the active
810 * list (the easiest thing to do here). If no one needs it,
811 * pageout will grab it eventually.
813 if (result != KERN_SUCCESS) {
815 unlock_and_deallocate(&fs);
818 * If retry of map lookup would have blocked then
819 * retry fault from start.
821 if (result == KERN_FAILURE)
825 if ((retry_object != fs.first_object) ||
826 (retry_pindex != fs.first_pindex)) {
828 unlock_and_deallocate(&fs);
833 * Check whether the protection has changed or the object has
834 * been copied while we left the map unlocked. Changing from
835 * read to write permission is OK - we leave the page
836 * write-protected, and catch the write fault. Changing from
837 * write to read permission means that we can't mark the page
838 * write-enabled after all.
843 if (prot & VM_PROT_WRITE) {
844 vm_page_lock_queues();
845 vm_page_flag_set(fs.m, PG_WRITEABLE);
846 vm_object_set_writeable_dirty(fs.m->object);
849 * If the fault is a write, we know that this page is being
850 * written NOW so dirty it explicitly to save on
851 * pmap_is_modified() calls later.
853 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
854 * if the page is already dirty to prevent data written with
855 * the expectation of being synced from not being synced.
856 * Likewise if this entry does not request NOSYNC then make
857 * sure the page isn't marked NOSYNC. Applications sharing
858 * data should use the same flags to avoid ping ponging.
860 * Also tell the backing pager, if any, that it should remove
861 * any swap backing since the page is now dirty.
863 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
864 if (fs.m->dirty == 0)
865 vm_page_flag_set(fs.m, PG_NOSYNC);
867 vm_page_flag_clear(fs.m, PG_NOSYNC);
869 vm_page_unlock_queues();
870 if (fault_flags & VM_FAULT_DIRTY) {
872 vm_pager_page_unswapped(fs.m);
877 * Page had better still be busy
879 KASSERT(fs.m->flags & PG_BUSY,
880 ("vm_fault: page %p not busy!", fs.m));
882 * Sanity check: page must be completely valid or it is not fit to
883 * map into user space. vm_pager_get_pages() ensures this.
885 if (fs.m->valid != VM_PAGE_BITS_ALL) {
886 vm_page_zero_invalid(fs.m, TRUE);
887 printf("Warning: page %p partially invalid on fault\n", fs.m);
889 VM_OBJECT_UNLOCK(fs.object);
892 * Put this page into the physical map. We had to do the unlock above
893 * because pmap_enter() may sleep. We don't put the page
894 * back on the active queue until later so that the pageout daemon
895 * won't find it (yet).
897 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
898 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
899 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
901 VM_OBJECT_LOCK(fs.object);
902 vm_page_lock_queues();
903 vm_page_flag_set(fs.m, PG_REFERENCED);
906 * If the page is not wired down, then put it where the pageout daemon
909 if (fault_flags & VM_FAULT_WIRE_MASK) {
913 vm_page_unwire(fs.m, 1);
915 vm_page_activate(fs.m);
917 vm_page_wakeup(fs.m);
918 vm_page_unlock_queues();
921 * Unlock everything, and return
923 unlock_and_deallocate(&fs);
925 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
927 curproc->p_stats->p_ru.ru_majflt++;
929 curproc->p_stats->p_ru.ru_minflt++;
932 PROC_UNLOCK(curproc);
934 return (KERN_SUCCESS);
938 * vm_fault_prefault provides a quick way of clustering
939 * pagefaults into a processes address space. It is a "cousin"
940 * of vm_map_pmap_enter, except it runs at page fault time instead
944 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
947 vm_offset_t addr, starta;
952 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
955 object = entry->object.vm_object;
957 starta = addra - PFBAK * PAGE_SIZE;
958 if (starta < entry->start) {
959 starta = entry->start;
960 } else if (starta > addra) {
965 for (i = 0; i < PAGEORDER_SIZE; i++) {
966 vm_object_t backing_object, lobject;
968 addr = addra + prefault_pageorder[i];
969 if (addr > addra + (PFFOR * PAGE_SIZE))
972 if (addr < starta || addr >= entry->end)
975 if (!pmap_is_prefaultable(pmap, addr))
978 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
980 VM_OBJECT_LOCK(lobject);
981 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
982 lobject->type == OBJT_DEFAULT &&
983 (backing_object = lobject->backing_object) != NULL) {
984 if (lobject->backing_object_offset & PAGE_MASK)
986 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
987 VM_OBJECT_LOCK(backing_object);
988 VM_OBJECT_UNLOCK(lobject);
989 lobject = backing_object;
992 * give-up when a page is not in memory
995 VM_OBJECT_UNLOCK(lobject);
998 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1000 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1002 vm_page_lock_queues();
1003 if ((m->queue - m->pc) == PQ_CACHE)
1004 vm_page_deactivate(m);
1005 mpte = pmap_enter_quick(pmap, addr, m, mpte);
1006 vm_page_unlock_queues();
1008 VM_OBJECT_UNLOCK(lobject);
1015 * Ensure that the requested virtual address, which may be in userland,
1016 * is valid. Fault-in the page if necessary. Return -1 on failure.
1019 vm_fault_quick(caddr_t v, int prot)
1023 if (prot & VM_PROT_WRITE)
1024 r = subyte(v, fubyte(v));
1033 * Wire down a range of virtual addresses in a map.
1036 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1037 boolean_t user_wire, boolean_t fictitious)
1043 * We simulate a fault to get the page and enter it in the physical
1044 * map. For user wiring, we only ask for read access on currently
1045 * read-only sections.
1047 for (va = start; va < end; va += PAGE_SIZE) {
1048 rv = vm_fault(map, va,
1049 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1050 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1053 vm_fault_unwire(map, start, va, fictitious);
1057 return (KERN_SUCCESS);
1063 * Unwire a range of virtual addresses in a map.
1066 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1067 boolean_t fictitious)
1073 pmap = vm_map_pmap(map);
1076 * Since the pages are wired down, we must be able to get their
1077 * mappings from the physical map system.
1079 for (va = start; va < end; va += PAGE_SIZE) {
1080 pa = pmap_extract(pmap, va);
1082 pmap_change_wiring(pmap, va, FALSE);
1084 vm_page_lock_queues();
1085 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1086 vm_page_unlock_queues();
1094 * vm_fault_copy_entry
1096 * Copy all of the pages from a wired-down map entry to another.
1098 * In/out conditions:
1099 * The source and destination maps must be locked for write.
1100 * The source map entry must be wired down (or be a sharing map
1101 * entry corresponding to a main map entry that is wired down).
1104 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1107 vm_map_entry_t dst_entry;
1108 vm_map_entry_t src_entry;
1110 vm_object_t backing_object, dst_object, object;
1111 vm_object_t src_object;
1112 vm_ooffset_t dst_offset;
1113 vm_ooffset_t src_offset;
1124 src_object = src_entry->object.vm_object;
1125 src_offset = src_entry->offset;
1128 * Create the top-level object for the destination entry. (Doesn't
1129 * actually shadow anything - we copy the pages directly.)
1131 dst_object = vm_object_allocate(OBJT_DEFAULT,
1132 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1134 VM_OBJECT_LOCK(dst_object);
1135 dst_entry->object.vm_object = dst_object;
1136 dst_entry->offset = 0;
1138 prot = dst_entry->max_protection;
1141 * Loop through all of the pages in the entry's range, copying each
1142 * one from the source object (it should be there) to the destination
1145 for (vaddr = dst_entry->start, dst_offset = 0;
1146 vaddr < dst_entry->end;
1147 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1150 * Allocate a page in the destination object
1153 dst_m = vm_page_alloc(dst_object,
1154 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1155 if (dst_m == NULL) {
1156 VM_OBJECT_UNLOCK(dst_object);
1158 VM_OBJECT_LOCK(dst_object);
1160 } while (dst_m == NULL);
1163 * Find the page in the source object, and copy it in.
1164 * (Because the source is wired down, the page will be in
1167 VM_OBJECT_LOCK(src_object);
1168 object = src_object;
1170 while ((src_m = vm_page_lookup(object, pindex +
1171 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1172 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1173 (backing_object = object->backing_object) != NULL) {
1175 * Allow fallback to backing objects if we are reading.
1177 VM_OBJECT_LOCK(backing_object);
1178 pindex += OFF_TO_IDX(object->backing_object_offset);
1179 VM_OBJECT_UNLOCK(object);
1180 object = backing_object;
1183 panic("vm_fault_copy_wired: page missing");
1184 pmap_copy_page(src_m, dst_m);
1185 VM_OBJECT_UNLOCK(object);
1186 dst_m->valid = VM_PAGE_BITS_ALL;
1187 VM_OBJECT_UNLOCK(dst_object);
1190 * Enter it in the pmap...
1192 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1193 VM_OBJECT_LOCK(dst_object);
1194 vm_page_lock_queues();
1195 if ((prot & VM_PROT_WRITE) != 0)
1196 vm_page_flag_set(dst_m, PG_WRITEABLE);
1199 * Mark it no longer busy, and put it on the active list.
1201 vm_page_activate(dst_m);
1202 vm_page_wakeup(dst_m);
1203 vm_page_unlock_queues();
1205 VM_OBJECT_UNLOCK(dst_object);
1210 * This routine checks around the requested page for other pages that
1211 * might be able to be faulted in. This routine brackets the viable
1212 * pages for the pages to be paged in.
1215 * m, rbehind, rahead
1218 * marray (array of vm_page_t), reqpage (index of requested page)
1221 * number of pages in marray
1223 * This routine can't block.
1226 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1235 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1237 int cbehind, cahead;
1239 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1245 * we don't fault-ahead for device pager
1247 if (object->type == OBJT_DEVICE) {
1254 * if the requested page is not available, then give up now
1256 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1260 if ((cbehind == 0) && (cahead == 0)) {
1266 if (rahead > cahead) {
1270 if (rbehind > cbehind) {
1275 * try to do any readahead that we might have free pages for.
1277 if ((rahead + rbehind) >
1278 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1279 pagedaemon_wakeup();
1286 * scan backward for the read behind pages -- in memory
1289 if (rbehind > pindex) {
1293 startpindex = pindex - rbehind;
1296 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1297 if (vm_page_lookup(object, tpindex)) {
1298 startpindex = tpindex + 1;
1305 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1307 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1309 vm_page_lock_queues();
1310 for (j = 0; j < i; j++) {
1311 vm_page_free(marray[j]);
1313 vm_page_unlock_queues();
1327 /* page offset of the required page */
1330 tpindex = pindex + 1;
1334 * scan forward for the read ahead pages
1336 endpindex = tpindex + rahead;
1337 if (endpindex > object->size)
1338 endpindex = object->size;
1340 for (; tpindex < endpindex; i++, tpindex++) {
1342 if (vm_page_lookup(object, tpindex)) {
1346 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1354 /* return number of bytes of pages */