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>
102 #define PAGEORDER_SIZE (PFBAK+PFFOR)
104 static int prefault_pageorder[] = {
105 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
106 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
107 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
108 -4 * PAGE_SIZE, 4 * PAGE_SIZE
111 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
112 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
114 #define VM_FAULT_READ_AHEAD 8
115 #define VM_FAULT_READ_BEHIND 7
116 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
123 vm_object_t first_object;
124 vm_pindex_t first_pindex;
126 vm_map_entry_t entry;
127 int lookup_still_valid;
132 release_page(struct faultstate *fs)
134 vm_page_lock_queues();
135 vm_page_wakeup(fs->m);
136 vm_page_deactivate(fs->m);
137 vm_page_unlock_queues();
142 unlock_map(struct faultstate *fs)
144 if (fs->lookup_still_valid) {
145 vm_map_lookup_done(fs->map, fs->entry);
146 fs->lookup_still_valid = FALSE;
151 unlock_and_deallocate(struct faultstate *fs)
154 vm_object_pip_wakeup(fs->object);
155 VM_OBJECT_UNLOCK(fs->object);
156 if (fs->object != fs->first_object) {
157 VM_OBJECT_LOCK(fs->first_object);
158 vm_page_lock_queues();
159 vm_page_free(fs->first_m);
160 vm_page_unlock_queues();
161 vm_object_pip_wakeup(fs->first_object);
162 VM_OBJECT_UNLOCK(fs->first_object);
165 vm_object_deallocate(fs->first_object);
167 if (fs->vp != NULL) {
173 if (!fs->map->system_map)
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_WIRE_MASK) == 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.
201 * The map in question must be referenced, and remains so.
202 * Caller may hold no locks.
205 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
209 int is_first_object_locked, result;
210 boolean_t growstack, wired;
212 vm_object_t next_object;
213 vm_page_t marray[VM_FAULT_READ];
216 struct faultstate fs;
220 atomic_add_int(&cnt.v_vm_faults, 1);
225 * Find the backing store object and offset into it to begin the
229 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
230 &fs.first_object, &fs.first_pindex, &prot, &wired);
231 if (result != KERN_SUCCESS) {
232 if (result != KERN_PROTECTION_FAILURE ||
233 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
234 if (growstack && result == KERN_INVALID_ADDRESS &&
235 map != kernel_map && curproc != NULL) {
236 result = vm_map_growstack(curproc, vaddr);
237 if (result != KERN_SUCCESS)
238 return (KERN_FAILURE);
246 * If we are user-wiring a r/w segment, and it is COW, then
247 * we need to do the COW operation. Note that we don't COW
248 * currently RO sections now, because it is NOT desirable
249 * to COW .text. We simply keep .text from ever being COW'ed
250 * and take the heat that one cannot debug wired .text sections.
252 result = vm_map_lookup(&fs.map, vaddr,
253 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
254 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
255 if (result != KERN_SUCCESS)
259 * If we don't COW now, on a user wire, the user will never
260 * be able to write to the mapping. If we don't make this
261 * restriction, the bookkeeping would be nearly impossible.
263 * XXX The following assignment modifies the map without
264 * holding a write lock on it.
266 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
267 fs.entry->max_protection &= ~VM_PROT_WRITE;
270 map_generation = fs.map->timestamp;
272 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
273 panic("vm_fault: fault on nofault entry, addr: %lx",
278 * Make a reference to this object to prevent its disposal while we
279 * are messing with it. Once we have the reference, the map is free
280 * to be diddled. Since objects reference their shadows (and copies),
281 * they will stay around as well.
283 * Bump the paging-in-progress count to prevent size changes (e.g.
284 * truncation operations) during I/O. This must be done after
285 * obtaining the vnode lock in order to avoid possible deadlocks.
287 * XXX vnode_pager_lock() can block without releasing the map lock.
289 if (!fs.map->system_map)
291 VM_OBJECT_LOCK(fs.first_object);
292 vm_object_reference_locked(fs.first_object);
293 fs.vp = vnode_pager_lock(fs.first_object);
294 KASSERT(fs.vp == NULL || !fs.map->system_map,
295 ("vm_fault: vnode-backed object mapped by system map"));
296 if (debug_mpsafevm && !fs.map->system_map)
298 vm_object_pip_add(fs.first_object, 1);
300 fs.lookup_still_valid = TRUE;
308 * Search for the page at object/offset.
310 fs.object = fs.first_object;
311 fs.pindex = fs.first_pindex;
314 * If the object is dead, we stop here
316 if (fs.object->flags & OBJ_DEAD) {
317 unlock_and_deallocate(&fs);
318 return (KERN_PROTECTION_FAILURE);
322 * See if page is resident
324 fs.m = vm_page_lookup(fs.object, fs.pindex);
329 * check for page-based copy on write.
330 * We check fs.object == fs.first_object so
331 * as to ensure the legacy COW mechanism is
332 * used when the page in question is part of
333 * a shadow object. Otherwise, vm_page_cowfault()
334 * removes the page from the backing object,
335 * which is not what we want.
337 vm_page_lock_queues();
339 (fault_type & VM_PROT_WRITE) &&
340 (fs.object == fs.first_object)) {
341 vm_page_cowfault(fs.m);
342 vm_page_unlock_queues();
343 unlock_and_deallocate(&fs);
348 * Wait/Retry if the page is busy. We have to do this
349 * if the page is busy via either PG_BUSY or
350 * vm_page_t->busy because the vm_pager may be using
351 * vm_page_t->busy for pageouts ( and even pageins if
352 * it is the vnode pager ), and we could end up trying
353 * to pagein and pageout the same page simultaneously.
355 * We can theoretically allow the busy case on a read
356 * fault if the page is marked valid, but since such
357 * pages are typically already pmap'd, putting that
358 * special case in might be more effort then it is
359 * worth. We cannot under any circumstances mess
360 * around with a vm_page_t->busy page except, perhaps,
363 if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
364 vm_page_unlock_queues();
365 VM_OBJECT_UNLOCK(fs.object);
366 if (fs.object != fs.first_object) {
367 VM_OBJECT_LOCK(fs.first_object);
368 vm_page_lock_queues();
369 vm_page_free(fs.first_m);
370 vm_page_unlock_queues();
371 vm_object_pip_wakeup(fs.first_object);
372 VM_OBJECT_UNLOCK(fs.first_object);
382 VM_OBJECT_LOCK(fs.object);
383 if (fs.m == vm_page_lookup(fs.object,
385 vm_page_lock_queues();
386 if (!vm_page_sleep_if_busy(fs.m, TRUE,
388 vm_page_unlock_queues();
390 vm_object_pip_wakeup(fs.object);
391 VM_OBJECT_UNLOCK(fs.object);
392 atomic_add_int(&cnt.v_intrans, 1);
393 if (!fs.map->system_map)
395 vm_object_deallocate(fs.first_object);
400 vm_pageq_remove_nowakeup(fs.m);
402 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
403 vm_page_activate(fs.m);
404 vm_page_unlock_queues();
405 unlock_and_deallocate(&fs);
411 * Mark page busy for other processes, and the
412 * pagedaemon. If it still isn't completely valid
413 * (readable), jump to readrest, else break-out ( we
417 vm_page_unlock_queues();
418 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
419 fs.m->object != kernel_object && fs.m->object != kmem_object) {
427 * Page is not resident, If this is the search termination
428 * or the pager might contain the page, allocate a new page.
430 if (TRYPAGER || fs.object == fs.first_object) {
431 if (fs.pindex >= fs.object->size) {
432 unlock_and_deallocate(&fs);
433 return (KERN_PROTECTION_FAILURE);
437 * Allocate a new page for this object/offset pair.
440 if (!vm_page_count_severe()) {
441 fs.m = vm_page_alloc(fs.object, fs.pindex,
442 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
445 unlock_and_deallocate(&fs);
453 * We have found a valid page or we have allocated a new page.
454 * The page thus may not be valid or may not be entirely
457 * Attempt to fault-in the page if there is a chance that the
458 * pager has it, and potentially fault in additional pages
465 u_char behavior = vm_map_entry_behavior(fs.entry);
467 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
471 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
472 if (behind > VM_FAULT_READ_BEHIND)
473 behind = VM_FAULT_READ_BEHIND;
475 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
476 if (ahead > VM_FAULT_READ_AHEAD)
477 ahead = VM_FAULT_READ_AHEAD;
479 is_first_object_locked = FALSE;
480 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
481 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
482 fs.pindex >= fs.entry->lastr &&
483 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
484 (fs.first_object == fs.object ||
485 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
486 fs.first_object->type != OBJT_DEVICE) {
487 vm_pindex_t firstpindex, tmppindex;
489 if (fs.first_pindex < 2 * VM_FAULT_READ)
492 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
494 vm_page_lock_queues();
496 * note: partially valid pages cannot be
497 * included in the lookahead - NFS piecemeal
498 * writes will barf on it badly.
500 for (tmppindex = fs.first_pindex - 1;
501 tmppindex >= firstpindex;
505 mt = vm_page_lookup(fs.first_object, tmppindex);
506 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
509 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
515 vm_page_deactivate(mt);
520 vm_page_unlock_queues();
524 if (is_first_object_locked)
525 VM_OBJECT_UNLOCK(fs.first_object);
527 * now we find out if any other pages should be paged
528 * in at this time this routine checks to see if the
529 * pages surrounding this fault reside in the same
530 * object as the page for this fault. If they do,
531 * then they are faulted in also into the object. The
532 * array "marray" returned contains an array of
533 * vm_page_t structs where one of them is the
534 * vm_page_t passed to the routine. The reqpage
535 * return value is the index into the marray for the
536 * vm_page_t passed to the routine.
538 * fs.m plus the additional pages are PG_BUSY'd.
540 * XXX vm_fault_additional_pages() can block
541 * without releasing the map lock.
543 faultcount = vm_fault_additional_pages(
544 fs.m, behind, ahead, marray, &reqpage);
547 * update lastr imperfectly (we do not know how much
548 * getpages will actually read), but good enough.
550 * XXX The following assignment modifies the map
551 * without holding a write lock on it.
553 fs.entry->lastr = fs.pindex + faultcount - behind;
556 * Call the pager to retrieve the data, if any, after
557 * releasing the lock on the map. We hold a ref on
558 * fs.object and the pages are PG_BUSY'd.
563 vm_pager_get_pages(fs.object, marray, faultcount,
564 reqpage) : VM_PAGER_FAIL;
566 if (rv == VM_PAGER_OK) {
568 * Found the page. Leave it busy while we play
573 * Relookup in case pager changed page. Pager
574 * is responsible for disposition of old page
577 fs.m = vm_page_lookup(fs.object, fs.pindex);
579 unlock_and_deallocate(&fs);
584 break; /* break to PAGE HAS BEEN FOUND */
587 * Remove the bogus page (which does not exist at this
588 * object/offset); before doing so, we must get back
589 * our object lock to preserve our invariant.
591 * Also wake up any other process that may want to bring
594 * If this is the top-level object, we must leave the
595 * busy page to prevent another process from rushing
596 * past us, and inserting the page in that object at
597 * the same time that we are.
599 if (rv == VM_PAGER_ERROR)
600 printf("vm_fault: pager read error, pid %d (%s)\n",
601 curproc->p_pid, curproc->p_comm);
603 * Data outside the range of the pager or an I/O error
606 * XXX - the check for kernel_map is a kludge to work
607 * around having the machine panic on a kernel space
608 * fault w/ I/O error.
610 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
611 (rv == VM_PAGER_BAD)) {
612 vm_page_lock_queues();
614 vm_page_unlock_queues();
616 unlock_and_deallocate(&fs);
617 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
619 if (fs.object != fs.first_object) {
620 vm_page_lock_queues();
622 vm_page_unlock_queues();
625 * XXX - we cannot just fall out at this
626 * point, m has been freed and is invalid!
632 * We get here if the object has default pager (or unwiring)
633 * or the pager doesn't have the page.
635 if (fs.object == fs.first_object)
639 * Move on to the next object. Lock the next object before
640 * unlocking the current one.
642 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
643 next_object = fs.object->backing_object;
644 if (next_object == NULL) {
646 * If there's no object left, fill the page in the top
649 if (fs.object != fs.first_object) {
650 vm_object_pip_wakeup(fs.object);
651 VM_OBJECT_UNLOCK(fs.object);
653 fs.object = fs.first_object;
654 fs.pindex = fs.first_pindex;
656 VM_OBJECT_LOCK(fs.object);
661 * Zero the page if necessary and mark it valid.
663 if ((fs.m->flags & PG_ZERO) == 0) {
664 pmap_zero_page(fs.m);
666 atomic_add_int(&cnt.v_ozfod, 1);
668 atomic_add_int(&cnt.v_zfod, 1);
669 fs.m->valid = VM_PAGE_BITS_ALL;
670 break; /* break to PAGE HAS BEEN FOUND */
672 KASSERT(fs.object != next_object,
673 ("object loop %p", next_object));
674 VM_OBJECT_LOCK(next_object);
675 vm_object_pip_add(next_object, 1);
676 if (fs.object != fs.first_object)
677 vm_object_pip_wakeup(fs.object);
678 VM_OBJECT_UNLOCK(fs.object);
679 fs.object = next_object;
683 KASSERT((fs.m->flags & PG_BUSY) != 0,
684 ("vm_fault: not busy after main loop"));
687 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
692 * If the page is being written, but isn't already owned by the
693 * top-level object, we have to copy it into a new page owned by the
696 if (fs.object != fs.first_object) {
698 * We only really need to copy if we want to write it.
700 if (fault_type & VM_PROT_WRITE) {
702 * This allows pages to be virtually copied from a
703 * backing_object into the first_object, where the
704 * backing object has no other refs to it, and cannot
705 * gain any more refs. Instead of a bcopy, we just
706 * move the page from the backing object to the
707 * first object. Note that we must mark the page
708 * dirty in the first object so that it will go out
709 * to swap when needed.
711 is_first_object_locked = FALSE;
714 * Only one shadow object
716 (fs.object->shadow_count == 1) &&
718 * No COW refs, except us
720 (fs.object->ref_count == 1) &&
722 * No one else can look this object up
724 (fs.object->handle == NULL) &&
726 * No other ways to look the object up
728 ((fs.object->type == OBJT_DEFAULT) ||
729 (fs.object->type == OBJT_SWAP)) &&
730 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
732 * We don't chase down the shadow chain
734 fs.object == fs.first_object->backing_object) {
735 vm_page_lock_queues();
737 * get rid of the unnecessary page
739 pmap_remove_all(fs.first_m);
740 vm_page_free(fs.first_m);
742 * grab the page and put it into the
743 * process'es object. The page is
744 * automatically made dirty.
746 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
748 vm_page_unlock_queues();
751 atomic_add_int(&cnt.v_cow_optim, 1);
754 * Oh, well, lets copy it.
756 pmap_copy_page(fs.m, fs.first_m);
757 fs.first_m->valid = VM_PAGE_BITS_ALL;
761 * We no longer need the old page or object.
766 * fs.object != fs.first_object due to above
769 vm_object_pip_wakeup(fs.object);
770 VM_OBJECT_UNLOCK(fs.object);
772 * Only use the new page below...
774 fs.object = fs.first_object;
775 fs.pindex = fs.first_pindex;
777 if (!is_first_object_locked)
778 VM_OBJECT_LOCK(fs.object);
779 atomic_add_int(&cnt.v_cow_faults, 1);
781 prot &= ~VM_PROT_WRITE;
786 * We must verify that the maps have not changed since our last
789 if (!fs.lookup_still_valid) {
790 vm_object_t retry_object;
791 vm_pindex_t retry_pindex;
792 vm_prot_t retry_prot;
794 if (!vm_map_trylock_read(fs.map)) {
796 unlock_and_deallocate(&fs);
799 fs.lookup_still_valid = TRUE;
800 if (fs.map->timestamp != map_generation) {
801 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
802 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
805 * If we don't need the page any longer, put it on the active
806 * list (the easiest thing to do here). If no one needs it,
807 * pageout will grab it eventually.
809 if (result != KERN_SUCCESS) {
811 unlock_and_deallocate(&fs);
814 * If retry of map lookup would have blocked then
815 * retry fault from start.
817 if (result == KERN_FAILURE)
821 if ((retry_object != fs.first_object) ||
822 (retry_pindex != fs.first_pindex)) {
824 unlock_and_deallocate(&fs);
829 * Check whether the protection has changed or the object has
830 * been copied while we left the map unlocked. Changing from
831 * read to write permission is OK - we leave the page
832 * write-protected, and catch the write fault. Changing from
833 * write to read permission means that we can't mark the page
834 * write-enabled after all.
839 if (prot & VM_PROT_WRITE) {
840 vm_page_lock_queues();
841 vm_page_flag_set(fs.m, PG_WRITEABLE);
842 vm_object_set_writeable_dirty(fs.m->object);
845 * If the fault is a write, we know that this page is being
846 * written NOW so dirty it explicitly to save on
847 * pmap_is_modified() calls later.
849 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
850 * if the page is already dirty to prevent data written with
851 * the expectation of being synced from not being synced.
852 * Likewise if this entry does not request NOSYNC then make
853 * sure the page isn't marked NOSYNC. Applications sharing
854 * data should use the same flags to avoid ping ponging.
856 * Also tell the backing pager, if any, that it should remove
857 * any swap backing since the page is now dirty.
859 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
860 if (fs.m->dirty == 0)
861 vm_page_flag_set(fs.m, PG_NOSYNC);
863 vm_page_flag_clear(fs.m, PG_NOSYNC);
865 vm_page_unlock_queues();
866 if (fault_flags & VM_FAULT_DIRTY) {
868 vm_pager_page_unswapped(fs.m);
873 * Page had better still be busy
875 KASSERT(fs.m->flags & PG_BUSY,
876 ("vm_fault: page %p not busy!", fs.m));
878 * Sanity check: page must be completely valid or it is not fit to
879 * map into user space. vm_pager_get_pages() ensures this.
881 if (fs.m->valid != VM_PAGE_BITS_ALL) {
882 vm_page_zero_invalid(fs.m, TRUE);
883 printf("Warning: page %p partially invalid on fault\n", fs.m);
885 VM_OBJECT_UNLOCK(fs.object);
888 * Put this page into the physical map. We had to do the unlock above
889 * because pmap_enter() may sleep. We don't put the page
890 * back on the active queue until later so that the pageout daemon
891 * won't find it (yet).
893 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
894 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
895 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
897 VM_OBJECT_LOCK(fs.object);
898 vm_page_lock_queues();
899 vm_page_flag_set(fs.m, PG_REFERENCED);
902 * If the page is not wired down, then put it where the pageout daemon
905 if (fault_flags & VM_FAULT_WIRE_MASK) {
909 vm_page_unwire(fs.m, 1);
911 vm_page_activate(fs.m);
913 vm_page_wakeup(fs.m);
914 vm_page_unlock_queues();
917 * Unlock everything, and return
919 unlock_and_deallocate(&fs);
921 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
923 curproc->p_stats->p_ru.ru_majflt++;
925 curproc->p_stats->p_ru.ru_minflt++;
928 PROC_UNLOCK(curproc);
930 return (KERN_SUCCESS);
934 * vm_fault_prefault provides a quick way of clustering
935 * pagefaults into a processes address space. It is a "cousin"
936 * of vm_map_pmap_enter, except it runs at page fault time instead
940 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
943 vm_offset_t addr, starta;
948 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
951 object = entry->object.vm_object;
953 starta = addra - PFBAK * PAGE_SIZE;
954 if (starta < entry->start) {
955 starta = entry->start;
956 } else if (starta > addra) {
961 for (i = 0; i < PAGEORDER_SIZE; i++) {
962 vm_object_t backing_object, lobject;
964 addr = addra + prefault_pageorder[i];
965 if (addr > addra + (PFFOR * PAGE_SIZE))
968 if (addr < starta || addr >= entry->end)
971 if (!pmap_is_prefaultable(pmap, addr))
974 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
976 VM_OBJECT_LOCK(lobject);
977 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
978 lobject->type == OBJT_DEFAULT &&
979 (backing_object = lobject->backing_object) != NULL) {
980 if (lobject->backing_object_offset & PAGE_MASK)
982 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
983 VM_OBJECT_LOCK(backing_object);
984 VM_OBJECT_UNLOCK(lobject);
985 lobject = backing_object;
988 * give-up when a page is not in memory
991 VM_OBJECT_UNLOCK(lobject);
994 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
996 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
998 vm_page_lock_queues();
999 if ((m->queue - m->pc) == PQ_CACHE)
1000 vm_page_deactivate(m);
1001 mpte = pmap_enter_quick(pmap, addr, m, mpte);
1002 vm_page_unlock_queues();
1004 VM_OBJECT_UNLOCK(lobject);
1011 * Ensure that the requested virtual address, which may be in userland,
1012 * is valid. Fault-in the page if necessary. Return -1 on failure.
1015 vm_fault_quick(caddr_t v, int prot)
1019 if (prot & VM_PROT_WRITE)
1020 r = subyte(v, fubyte(v));
1029 * Wire down a range of virtual addresses in a map.
1032 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1033 boolean_t user_wire, boolean_t fictitious)
1039 * We simulate a fault to get the page and enter it in the physical
1040 * map. For user wiring, we only ask for read access on currently
1041 * read-only sections.
1043 for (va = start; va < end; va += PAGE_SIZE) {
1044 rv = vm_fault(map, va,
1045 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1046 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1049 vm_fault_unwire(map, start, va, fictitious);
1053 return (KERN_SUCCESS);
1059 * Unwire a range of virtual addresses in a map.
1062 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1063 boolean_t fictitious)
1069 pmap = vm_map_pmap(map);
1072 * Since the pages are wired down, we must be able to get their
1073 * mappings from the physical map system.
1075 for (va = start; va < end; va += PAGE_SIZE) {
1076 pa = pmap_extract(pmap, va);
1078 pmap_change_wiring(pmap, va, FALSE);
1080 vm_page_lock_queues();
1081 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1082 vm_page_unlock_queues();
1090 * vm_fault_copy_entry
1092 * Copy all of the pages from a wired-down map entry to another.
1094 * In/out conditions:
1095 * The source and destination maps must be locked for write.
1096 * The source map entry must be wired down (or be a sharing map
1097 * entry corresponding to a main map entry that is wired down).
1100 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1103 vm_map_entry_t dst_entry;
1104 vm_map_entry_t src_entry;
1106 vm_object_t backing_object, dst_object, object;
1107 vm_object_t src_object;
1108 vm_ooffset_t dst_offset;
1109 vm_ooffset_t src_offset;
1120 src_object = src_entry->object.vm_object;
1121 src_offset = src_entry->offset;
1124 * Create the top-level object for the destination entry. (Doesn't
1125 * actually shadow anything - we copy the pages directly.)
1127 dst_object = vm_object_allocate(OBJT_DEFAULT,
1128 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1130 VM_OBJECT_LOCK(dst_object);
1131 dst_entry->object.vm_object = dst_object;
1132 dst_entry->offset = 0;
1134 prot = dst_entry->max_protection;
1137 * Loop through all of the pages in the entry's range, copying each
1138 * one from the source object (it should be there) to the destination
1141 for (vaddr = dst_entry->start, dst_offset = 0;
1142 vaddr < dst_entry->end;
1143 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1146 * Allocate a page in the destination object
1149 dst_m = vm_page_alloc(dst_object,
1150 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1151 if (dst_m == NULL) {
1152 VM_OBJECT_UNLOCK(dst_object);
1154 VM_OBJECT_LOCK(dst_object);
1156 } while (dst_m == NULL);
1159 * Find the page in the source object, and copy it in.
1160 * (Because the source is wired down, the page will be in
1163 VM_OBJECT_LOCK(src_object);
1164 object = src_object;
1166 while ((src_m = vm_page_lookup(object, pindex +
1167 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1168 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1169 (backing_object = object->backing_object) != NULL) {
1171 * Allow fallback to backing objects if we are reading.
1173 VM_OBJECT_LOCK(backing_object);
1174 pindex += OFF_TO_IDX(object->backing_object_offset);
1175 VM_OBJECT_UNLOCK(object);
1176 object = backing_object;
1179 panic("vm_fault_copy_wired: page missing");
1180 pmap_copy_page(src_m, dst_m);
1181 VM_OBJECT_UNLOCK(object);
1182 dst_m->valid = VM_PAGE_BITS_ALL;
1183 VM_OBJECT_UNLOCK(dst_object);
1186 * Enter it in the pmap...
1188 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1189 VM_OBJECT_LOCK(dst_object);
1190 vm_page_lock_queues();
1191 if ((prot & VM_PROT_WRITE) != 0)
1192 vm_page_flag_set(dst_m, PG_WRITEABLE);
1195 * Mark it no longer busy, and put it on the active list.
1197 vm_page_activate(dst_m);
1198 vm_page_wakeup(dst_m);
1199 vm_page_unlock_queues();
1201 VM_OBJECT_UNLOCK(dst_object);
1206 * This routine checks around the requested page for other pages that
1207 * might be able to be faulted in. This routine brackets the viable
1208 * pages for the pages to be paged in.
1211 * m, rbehind, rahead
1214 * marray (array of vm_page_t), reqpage (index of requested page)
1217 * number of pages in marray
1219 * This routine can't block.
1222 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1231 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1233 int cbehind, cahead;
1235 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1241 * we don't fault-ahead for device pager
1243 if (object->type == OBJT_DEVICE) {
1250 * if the requested page is not available, then give up now
1252 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1256 if ((cbehind == 0) && (cahead == 0)) {
1262 if (rahead > cahead) {
1266 if (rbehind > cbehind) {
1271 * try to do any readahead that we might have free pages for.
1273 if ((rahead + rbehind) >
1274 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1275 pagedaemon_wakeup();
1282 * scan backward for the read behind pages -- in memory
1285 if (rbehind > pindex) {
1289 startpindex = pindex - rbehind;
1292 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1293 if (vm_page_lookup(object, tpindex)) {
1294 startpindex = tpindex + 1;
1301 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1303 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1305 vm_page_lock_queues();
1306 for (j = 0; j < i; j++) {
1307 vm_page_free(marray[j]);
1309 vm_page_unlock_queues();
1323 /* page offset of the required page */
1326 tpindex = pindex + 1;
1330 * scan forward for the read ahead pages
1332 endpindex = tpindex + rahead;
1333 if (endpindex > object->size)
1334 endpindex = object->size;
1336 for (; tpindex < endpindex; i++, tpindex++) {
1338 if (vm_page_lookup(object, tpindex)) {
1342 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1350 /* return number of bytes of pages */