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->first_object->flags & OBJ_NEEDGIANT)
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.first_object->flags & OBJ_NEEDGIANT)
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 KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 ||
302 ("vm_fault: Object requiring giant mapped by system map"));
303 if (fs.first_object->flags & OBJ_NEEDGIANT && debug_mpsafevm)
305 vm_object_pip_add(fs.first_object, 1);
307 fs.lookup_still_valid = TRUE;
315 * Search for the page at object/offset.
317 fs.object = fs.first_object;
318 fs.pindex = fs.first_pindex;
321 * If the object is dead, we stop here
323 if (fs.object->flags & OBJ_DEAD) {
324 unlock_and_deallocate(&fs);
325 return (KERN_PROTECTION_FAILURE);
329 * See if page is resident
331 fs.m = vm_page_lookup(fs.object, fs.pindex);
336 * check for page-based copy on write.
337 * We check fs.object == fs.first_object so
338 * as to ensure the legacy COW mechanism is
339 * used when the page in question is part of
340 * a shadow object. Otherwise, vm_page_cowfault()
341 * removes the page from the backing object,
342 * which is not what we want.
344 vm_page_lock_queues();
346 (fault_type & VM_PROT_WRITE) &&
347 (fs.object == fs.first_object)) {
348 vm_page_cowfault(fs.m);
349 vm_page_unlock_queues();
350 unlock_and_deallocate(&fs);
355 * Wait/Retry if the page is busy. We have to do this
356 * if the page is busy via either PG_BUSY or
357 * vm_page_t->busy because the vm_pager may be using
358 * vm_page_t->busy for pageouts ( and even pageins if
359 * it is the vnode pager ), and we could end up trying
360 * to pagein and pageout the same page simultaneously.
362 * We can theoretically allow the busy case on a read
363 * fault if the page is marked valid, but since such
364 * pages are typically already pmap'd, putting that
365 * special case in might be more effort then it is
366 * worth. We cannot under any circumstances mess
367 * around with a vm_page_t->busy page except, perhaps,
370 if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
371 vm_page_unlock_queues();
372 VM_OBJECT_UNLOCK(fs.object);
373 if (fs.object != fs.first_object) {
374 VM_OBJECT_LOCK(fs.first_object);
375 vm_page_lock_queues();
376 vm_page_free(fs.first_m);
377 vm_page_unlock_queues();
378 vm_object_pip_wakeup(fs.first_object);
379 VM_OBJECT_UNLOCK(fs.first_object);
386 vfslck = VFS_LOCK_GIANT(fs.vp->v_mount);
389 VFS_UNLOCK_GIANT(vfslck);
391 VM_OBJECT_LOCK(fs.object);
392 if (fs.m == vm_page_lookup(fs.object,
394 vm_page_lock_queues();
395 if (!vm_page_sleep_if_busy(fs.m, TRUE,
397 vm_page_unlock_queues();
399 vm_object_pip_wakeup(fs.object);
400 VM_OBJECT_UNLOCK(fs.object);
401 atomic_add_int(&cnt.v_intrans, 1);
402 if (!fs.map->system_map)
404 vm_object_deallocate(fs.first_object);
409 vm_pageq_remove_nowakeup(fs.m);
411 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
412 vm_page_activate(fs.m);
413 vm_page_unlock_queues();
414 unlock_and_deallocate(&fs);
420 * Mark page busy for other processes, and the
421 * pagedaemon. If it still isn't completely valid
422 * (readable), jump to readrest, else break-out ( we
426 vm_page_unlock_queues();
427 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
428 fs.m->object != kernel_object && fs.m->object != kmem_object) {
436 * Page is not resident, If this is the search termination
437 * or the pager might contain the page, allocate a new page.
439 if (TRYPAGER || fs.object == fs.first_object) {
440 if (fs.pindex >= fs.object->size) {
441 unlock_and_deallocate(&fs);
442 return (KERN_PROTECTION_FAILURE);
446 * Allocate a new page for this object/offset pair.
449 if (!vm_page_count_severe()) {
450 fs.m = vm_page_alloc(fs.object, fs.pindex,
451 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
454 unlock_and_deallocate(&fs);
462 * We have found a valid page or we have allocated a new page.
463 * The page thus may not be valid or may not be entirely
466 * Attempt to fault-in the page if there is a chance that the
467 * pager has it, and potentially fault in additional pages
474 u_char behavior = vm_map_entry_behavior(fs.entry);
476 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
480 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
481 if (behind > VM_FAULT_READ_BEHIND)
482 behind = VM_FAULT_READ_BEHIND;
484 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
485 if (ahead > VM_FAULT_READ_AHEAD)
486 ahead = VM_FAULT_READ_AHEAD;
488 is_first_object_locked = FALSE;
489 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
490 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
491 fs.pindex >= fs.entry->lastr &&
492 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
493 (fs.first_object == fs.object ||
494 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
495 fs.first_object->type != OBJT_DEVICE) {
496 vm_pindex_t firstpindex, tmppindex;
498 if (fs.first_pindex < 2 * VM_FAULT_READ)
501 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
503 vm_page_lock_queues();
505 * note: partially valid pages cannot be
506 * included in the lookahead - NFS piecemeal
507 * writes will barf on it badly.
509 for (tmppindex = fs.first_pindex - 1;
510 tmppindex >= firstpindex;
514 mt = vm_page_lookup(fs.first_object, tmppindex);
515 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
518 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
524 vm_page_deactivate(mt);
529 vm_page_unlock_queues();
533 if (is_first_object_locked)
534 VM_OBJECT_UNLOCK(fs.first_object);
536 * now we find out if any other pages should be paged
537 * in at this time this routine checks to see if the
538 * pages surrounding this fault reside in the same
539 * object as the page for this fault. If they do,
540 * then they are faulted in also into the object. The
541 * array "marray" returned contains an array of
542 * vm_page_t structs where one of them is the
543 * vm_page_t passed to the routine. The reqpage
544 * return value is the index into the marray for the
545 * vm_page_t passed to the routine.
547 * fs.m plus the additional pages are PG_BUSY'd.
549 * XXX vm_fault_additional_pages() can block
550 * without releasing the map lock.
552 faultcount = vm_fault_additional_pages(
553 fs.m, behind, ahead, marray, &reqpage);
556 * update lastr imperfectly (we do not know how much
557 * getpages will actually read), but good enough.
559 * XXX The following assignment modifies the map
560 * without holding a write lock on it.
562 fs.entry->lastr = fs.pindex + faultcount - behind;
565 * Call the pager to retrieve the data, if any, after
566 * releasing the lock on the map. We hold a ref on
567 * fs.object and the pages are PG_BUSY'd.
572 vm_pager_get_pages(fs.object, marray, faultcount,
573 reqpage) : VM_PAGER_FAIL;
575 if (rv == VM_PAGER_OK) {
577 * Found the page. Leave it busy while we play
582 * Relookup in case pager changed page. Pager
583 * is responsible for disposition of old page
586 fs.m = vm_page_lookup(fs.object, fs.pindex);
588 unlock_and_deallocate(&fs);
593 break; /* break to PAGE HAS BEEN FOUND */
596 * Remove the bogus page (which does not exist at this
597 * object/offset); before doing so, we must get back
598 * our object lock to preserve our invariant.
600 * Also wake up any other process that may want to bring
603 * If this is the top-level object, we must leave the
604 * busy page to prevent another process from rushing
605 * past us, and inserting the page in that object at
606 * the same time that we are.
608 if (rv == VM_PAGER_ERROR)
609 printf("vm_fault: pager read error, pid %d (%s)\n",
610 curproc->p_pid, curproc->p_comm);
612 * Data outside the range of the pager or an I/O error
615 * XXX - the check for kernel_map is a kludge to work
616 * around having the machine panic on a kernel space
617 * fault w/ I/O error.
619 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
620 (rv == VM_PAGER_BAD)) {
621 vm_page_lock_queues();
623 vm_page_unlock_queues();
625 unlock_and_deallocate(&fs);
626 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
628 if (fs.object != fs.first_object) {
629 vm_page_lock_queues();
631 vm_page_unlock_queues();
634 * XXX - we cannot just fall out at this
635 * point, m has been freed and is invalid!
641 * We get here if the object has default pager (or unwiring)
642 * or the pager doesn't have the page.
644 if (fs.object == fs.first_object)
648 * Move on to the next object. Lock the next object before
649 * unlocking the current one.
651 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
652 next_object = fs.object->backing_object;
653 if (next_object == NULL) {
655 * If there's no object left, fill the page in the top
658 if (fs.object != fs.first_object) {
659 vm_object_pip_wakeup(fs.object);
660 VM_OBJECT_UNLOCK(fs.object);
662 fs.object = fs.first_object;
663 fs.pindex = fs.first_pindex;
665 VM_OBJECT_LOCK(fs.object);
670 * Zero the page if necessary and mark it valid.
672 if ((fs.m->flags & PG_ZERO) == 0) {
673 pmap_zero_page(fs.m);
675 atomic_add_int(&cnt.v_ozfod, 1);
677 atomic_add_int(&cnt.v_zfod, 1);
678 fs.m->valid = VM_PAGE_BITS_ALL;
679 break; /* break to PAGE HAS BEEN FOUND */
681 KASSERT(fs.object != next_object,
682 ("object loop %p", next_object));
683 VM_OBJECT_LOCK(next_object);
684 vm_object_pip_add(next_object, 1);
685 if (fs.object != fs.first_object)
686 vm_object_pip_wakeup(fs.object);
687 VM_OBJECT_UNLOCK(fs.object);
688 fs.object = next_object;
692 KASSERT((fs.m->flags & PG_BUSY) != 0,
693 ("vm_fault: not busy after main loop"));
696 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
701 * If the page is being written, but isn't already owned by the
702 * top-level object, we have to copy it into a new page owned by the
705 if (fs.object != fs.first_object) {
707 * We only really need to copy if we want to write it.
709 if (fault_type & VM_PROT_WRITE) {
711 * This allows pages to be virtually copied from a
712 * backing_object into the first_object, where the
713 * backing object has no other refs to it, and cannot
714 * gain any more refs. Instead of a bcopy, we just
715 * move the page from the backing object to the
716 * first object. Note that we must mark the page
717 * dirty in the first object so that it will go out
718 * to swap when needed.
720 is_first_object_locked = FALSE;
723 * Only one shadow object
725 (fs.object->shadow_count == 1) &&
727 * No COW refs, except us
729 (fs.object->ref_count == 1) &&
731 * No one else can look this object up
733 (fs.object->handle == NULL) &&
735 * No other ways to look the object up
737 ((fs.object->type == OBJT_DEFAULT) ||
738 (fs.object->type == OBJT_SWAP)) &&
739 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
741 * We don't chase down the shadow chain
743 fs.object == fs.first_object->backing_object) {
744 vm_page_lock_queues();
746 * get rid of the unnecessary page
748 pmap_remove_all(fs.first_m);
749 vm_page_free(fs.first_m);
751 * grab the page and put it into the
752 * process'es object. The page is
753 * automatically made dirty.
755 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
757 vm_page_unlock_queues();
760 atomic_add_int(&cnt.v_cow_optim, 1);
763 * Oh, well, lets copy it.
765 pmap_copy_page(fs.m, fs.first_m);
766 fs.first_m->valid = VM_PAGE_BITS_ALL;
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 atomic_add_int(&cnt.v_cow_faults, 1);
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 active
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.
848 if (prot & VM_PROT_WRITE) {
849 vm_page_lock_queues();
850 vm_page_flag_set(fs.m, PG_WRITEABLE);
851 vm_object_set_writeable_dirty(fs.m->object);
854 * If the fault is a write, we know that this page is being
855 * written NOW so dirty it explicitly to save on
856 * pmap_is_modified() calls later.
858 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
859 * if the page is already dirty to prevent data written with
860 * the expectation of being synced from not being synced.
861 * Likewise if this entry does not request NOSYNC then make
862 * sure the page isn't marked NOSYNC. Applications sharing
863 * data should use the same flags to avoid ping ponging.
865 * Also tell the backing pager, if any, that it should remove
866 * any swap backing since the page is now dirty.
868 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
869 if (fs.m->dirty == 0)
870 vm_page_flag_set(fs.m, PG_NOSYNC);
872 vm_page_flag_clear(fs.m, PG_NOSYNC);
874 vm_page_unlock_queues();
875 if (fault_flags & VM_FAULT_DIRTY) {
877 vm_pager_page_unswapped(fs.m);
882 * Page had better still be busy
884 KASSERT(fs.m->flags & PG_BUSY,
885 ("vm_fault: page %p not busy!", fs.m));
887 * Sanity check: page must be completely valid or it is not fit to
888 * map into user space. vm_pager_get_pages() ensures this.
890 if (fs.m->valid != VM_PAGE_BITS_ALL) {
891 vm_page_zero_invalid(fs.m, TRUE);
892 printf("Warning: page %p partially invalid on fault\n", fs.m);
894 VM_OBJECT_UNLOCK(fs.object);
897 * Put this page into the physical map. We had to do the unlock above
898 * because pmap_enter() may sleep. We don't put the page
899 * back on the active queue until later so that the pageout daemon
900 * won't find it (yet).
902 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
903 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
904 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
906 VM_OBJECT_LOCK(fs.object);
907 vm_page_lock_queues();
908 vm_page_flag_set(fs.m, PG_REFERENCED);
911 * If the page is not wired down, then put it where the pageout daemon
914 if (fault_flags & VM_FAULT_WIRE_MASK) {
918 vm_page_unwire(fs.m, 1);
920 vm_page_activate(fs.m);
922 vm_page_wakeup(fs.m);
923 vm_page_unlock_queues();
926 * Unlock everything, and return
928 unlock_and_deallocate(&fs);
930 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
932 curproc->p_stats->p_ru.ru_majflt++;
934 curproc->p_stats->p_ru.ru_minflt++;
937 PROC_UNLOCK(curproc);
939 return (KERN_SUCCESS);
943 * vm_fault_prefault provides a quick way of clustering
944 * pagefaults into a processes address space. It is a "cousin"
945 * of vm_map_pmap_enter, except it runs at page fault time instead
949 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
952 vm_offset_t addr, starta;
957 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
960 object = entry->object.vm_object;
962 starta = addra - PFBAK * PAGE_SIZE;
963 if (starta < entry->start) {
964 starta = entry->start;
965 } else if (starta > addra) {
970 for (i = 0; i < PAGEORDER_SIZE; i++) {
971 vm_object_t backing_object, lobject;
973 addr = addra + prefault_pageorder[i];
974 if (addr > addra + (PFFOR * PAGE_SIZE))
977 if (addr < starta || addr >= entry->end)
980 if (!pmap_is_prefaultable(pmap, addr))
983 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
985 VM_OBJECT_LOCK(lobject);
986 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
987 lobject->type == OBJT_DEFAULT &&
988 (backing_object = lobject->backing_object) != NULL) {
989 if (lobject->backing_object_offset & PAGE_MASK)
991 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
992 VM_OBJECT_LOCK(backing_object);
993 VM_OBJECT_UNLOCK(lobject);
994 lobject = backing_object;
997 * give-up when a page is not in memory
1000 VM_OBJECT_UNLOCK(lobject);
1003 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1005 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1007 vm_page_lock_queues();
1008 if ((m->queue - m->pc) == PQ_CACHE)
1009 vm_page_deactivate(m);
1010 mpte = pmap_enter_quick(pmap, addr, m, mpte);
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 user_wire, 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,
1054 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1055 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1058 vm_fault_unwire(map, start, va, fictitious);
1062 return (KERN_SUCCESS);
1068 * Unwire a range of virtual addresses in a map.
1071 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1072 boolean_t fictitious)
1078 pmap = vm_map_pmap(map);
1081 * Since the pages are wired down, we must be able to get their
1082 * mappings from the physical map system.
1084 for (va = start; va < end; va += PAGE_SIZE) {
1085 pa = pmap_extract(pmap, va);
1087 pmap_change_wiring(pmap, va, FALSE);
1089 vm_page_lock_queues();
1090 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1091 vm_page_unlock_queues();
1099 * vm_fault_copy_entry
1101 * Copy all of the pages from a wired-down map entry to another.
1103 * In/out conditions:
1104 * The source and destination maps must be locked for write.
1105 * The source map entry must be wired down (or be a sharing map
1106 * entry corresponding to a main map entry that is wired down).
1109 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1112 vm_map_entry_t dst_entry;
1113 vm_map_entry_t src_entry;
1115 vm_object_t backing_object, dst_object, object;
1116 vm_object_t src_object;
1117 vm_ooffset_t dst_offset;
1118 vm_ooffset_t src_offset;
1129 src_object = src_entry->object.vm_object;
1130 src_offset = src_entry->offset;
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 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1139 VM_OBJECT_LOCK(dst_object);
1140 dst_entry->object.vm_object = dst_object;
1141 dst_entry->offset = 0;
1143 prot = dst_entry->max_protection;
1146 * Loop through all of the pages in the entry's range, copying each
1147 * one from the source object (it should be there) to the destination
1150 for (vaddr = dst_entry->start, dst_offset = 0;
1151 vaddr < dst_entry->end;
1152 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1155 * Allocate a page in the destination object
1158 dst_m = vm_page_alloc(dst_object,
1159 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1160 if (dst_m == NULL) {
1161 VM_OBJECT_UNLOCK(dst_object);
1163 VM_OBJECT_LOCK(dst_object);
1165 } while (dst_m == NULL);
1168 * Find the page in the source object, and copy it in.
1169 * (Because the source is wired down, the page will be in
1172 VM_OBJECT_LOCK(src_object);
1173 object = src_object;
1175 while ((src_m = vm_page_lookup(object, pindex +
1176 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1177 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1178 (backing_object = object->backing_object) != NULL) {
1180 * Allow fallback to backing objects if we are reading.
1182 VM_OBJECT_LOCK(backing_object);
1183 pindex += OFF_TO_IDX(object->backing_object_offset);
1184 VM_OBJECT_UNLOCK(object);
1185 object = backing_object;
1188 panic("vm_fault_copy_wired: page missing");
1189 pmap_copy_page(src_m, dst_m);
1190 VM_OBJECT_UNLOCK(object);
1191 dst_m->valid = VM_PAGE_BITS_ALL;
1192 VM_OBJECT_UNLOCK(dst_object);
1195 * Enter it in the pmap...
1197 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1198 VM_OBJECT_LOCK(dst_object);
1199 vm_page_lock_queues();
1200 if ((prot & VM_PROT_WRITE) != 0)
1201 vm_page_flag_set(dst_m, PG_WRITEABLE);
1204 * Mark it no longer busy, and put it on the active list.
1206 vm_page_activate(dst_m);
1207 vm_page_wakeup(dst_m);
1208 vm_page_unlock_queues();
1210 VM_OBJECT_UNLOCK(dst_object);
1215 * This routine checks around the requested page for other pages that
1216 * might be able to be faulted in. This routine brackets the viable
1217 * pages for the pages to be paged in.
1220 * m, rbehind, rahead
1223 * marray (array of vm_page_t), reqpage (index of requested page)
1226 * number of pages in marray
1228 * This routine can't block.
1231 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1240 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1242 int cbehind, cahead;
1244 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1250 * we don't fault-ahead for device pager
1252 if (object->type == OBJT_DEVICE) {
1259 * if the requested page is not available, then give up now
1261 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1265 if ((cbehind == 0) && (cahead == 0)) {
1271 if (rahead > cahead) {
1275 if (rbehind > cbehind) {
1280 * try to do any readahead that we might have free pages for.
1282 if ((rahead + rbehind) >
1283 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1284 pagedaemon_wakeup();
1291 * scan backward for the read behind pages -- in memory
1294 if (rbehind > pindex) {
1298 startpindex = pindex - rbehind;
1301 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1302 if (vm_page_lookup(object, tpindex)) {
1303 startpindex = tpindex + 1;
1310 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1312 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1314 vm_page_lock_queues();
1315 for (j = 0; j < i; j++) {
1316 vm_page_free(marray[j]);
1318 vm_page_unlock_queues();
1332 /* page offset of the required page */
1335 tpindex = pindex + 1;
1339 * scan forward for the read ahead pages
1341 endpindex = tpindex + rahead;
1342 if (endpindex > object->size)
1343 endpindex = object->size;
1345 for (; tpindex < endpindex; i++, tpindex++) {
1347 if (vm_page_lookup(object, tpindex)) {
1351 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1359 /* return number of bytes of pages */