2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 * Permission to use, copy, modify and distribute this software and
50 * its documentation is hereby granted, provided that both the copyright
51 * notice and this permission notice appear in all copies of the
52 * software, derivative works or modified versions, and any portions
53 * thereof, and that both notices appear in supporting documentation.
55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 * Carnegie Mellon requests users of this software to return to
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
71 * Page fault handling module.
74 #include <sys/cdefs.h>
75 __FBSDID("$FreeBSD$");
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/mutex.h>
85 #include <sys/resourcevar.h>
86 #include <sys/sysctl.h>
87 #include <sys/vmmeter.h>
88 #include <sys/vnode.h>
91 #include <vm/vm_param.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_extern.h>
101 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
105 #define PAGEORDER_SIZE (PFBAK+PFFOR)
107 static int prefault_pageorder[] = {
108 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
109 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
110 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
111 -4 * PAGE_SIZE, 4 * PAGE_SIZE
114 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
115 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
117 #define VM_FAULT_READ_AHEAD 8
118 #define VM_FAULT_READ_BEHIND 7
119 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
126 vm_object_t first_object;
127 vm_pindex_t first_pindex;
129 vm_map_entry_t entry;
130 int lookup_still_valid;
136 release_page(struct faultstate *fs)
139 vm_page_wakeup(fs->m);
141 vm_page_deactivate(fs->m);
142 vm_page_unlock(fs->m);
147 unlock_map(struct faultstate *fs)
150 if (fs->lookup_still_valid) {
151 vm_map_lookup_done(fs->map, fs->entry);
152 fs->lookup_still_valid = FALSE;
157 unlock_and_deallocate(struct faultstate *fs)
160 vm_object_pip_wakeup(fs->object);
161 VM_OBJECT_UNLOCK(fs->object);
162 if (fs->object != fs->first_object) {
163 VM_OBJECT_LOCK(fs->first_object);
164 vm_page_lock(fs->first_m);
165 vm_page_free(fs->first_m);
166 vm_page_unlock(fs->first_m);
167 vm_object_pip_wakeup(fs->first_object);
168 VM_OBJECT_UNLOCK(fs->first_object);
171 vm_object_deallocate(fs->first_object);
173 if (fs->vp != NULL) {
177 VFS_UNLOCK_GIANT(fs->vfslocked);
182 * TRYPAGER - used by vm_fault to calculate whether the pager for the
183 * current object *might* contain the page.
185 * default objects are zero-fill, there is no real pager.
187 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
188 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
193 * Handle a page fault occurring at the given address,
194 * requiring the given permissions, in the map specified.
195 * If successful, the page is inserted into the
196 * associated physical map.
198 * NOTE: the given address should be truncated to the
199 * proper page address.
201 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
202 * a standard error specifying why the fault is fatal is returned.
204 * The map in question must be referenced, and remains so.
205 * Caller may hold no locks.
208 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
212 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
213 return (KERN_PROTECTION_FAILURE);
214 return (vm_fault_hold(map, vaddr, fault_type, fault_flags, NULL));
218 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
219 int fault_flags, vm_page_t *m_hold)
222 int is_first_object_locked, result;
223 boolean_t growstack, wired;
225 vm_object_t next_object;
226 vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
228 int faultcount, ahead, behind, alloc_req;
229 struct faultstate fs;
235 PCPU_INC(cnt.v_vm_faults);
238 faultcount = behind = 0;
243 * Find the backing store object and offset into it to begin the
247 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
248 &fs.first_object, &fs.first_pindex, &prot, &wired);
249 if (result != KERN_SUCCESS) {
250 if (growstack && result == KERN_INVALID_ADDRESS &&
252 result = vm_map_growstack(curproc, vaddr);
253 if (result != KERN_SUCCESS)
254 return (KERN_FAILURE);
261 map_generation = fs.map->timestamp;
263 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
264 panic("vm_fault: fault on nofault entry, addr: %lx",
269 * Make a reference to this object to prevent its disposal while we
270 * are messing with it. Once we have the reference, the map is free
271 * to be diddled. Since objects reference their shadows (and copies),
272 * they will stay around as well.
274 * Bump the paging-in-progress count to prevent size changes (e.g.
275 * truncation operations) during I/O. This must be done after
276 * obtaining the vnode lock in order to avoid possible deadlocks.
278 VM_OBJECT_LOCK(fs.first_object);
279 vm_object_reference_locked(fs.first_object);
280 vm_object_pip_add(fs.first_object, 1);
282 fs.lookup_still_valid = TRUE;
285 fault_type = prot | (fault_type & VM_PROT_COPY);
290 * Search for the page at object/offset.
292 fs.object = fs.first_object;
293 fs.pindex = fs.first_pindex;
296 * If the object is dead, we stop here
298 if (fs.object->flags & OBJ_DEAD) {
299 unlock_and_deallocate(&fs);
300 return (KERN_PROTECTION_FAILURE);
304 * See if page is resident
306 fs.m = vm_page_lookup(fs.object, fs.pindex);
309 * check for page-based copy on write.
310 * We check fs.object == fs.first_object so
311 * as to ensure the legacy COW mechanism is
312 * used when the page in question is part of
313 * a shadow object. Otherwise, vm_page_cowfault()
314 * removes the page from the backing object,
315 * which is not what we want.
319 (fault_type & VM_PROT_WRITE) &&
320 (fs.object == fs.first_object)) {
321 vm_page_cowfault(fs.m);
322 unlock_and_deallocate(&fs);
327 * Wait/Retry if the page is busy. We have to do this
328 * if the page is busy via either VPO_BUSY or
329 * vm_page_t->busy because the vm_pager may be using
330 * vm_page_t->busy for pageouts ( and even pageins if
331 * it is the vnode pager ), and we could end up trying
332 * to pagein and pageout the same page simultaneously.
334 * We can theoretically allow the busy case on a read
335 * fault if the page is marked valid, but since such
336 * pages are typically already pmap'd, putting that
337 * special case in might be more effort then it is
338 * worth. We cannot under any circumstances mess
339 * around with a vm_page_t->busy page except, perhaps,
342 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
344 * Reference the page before unlocking and
345 * sleeping so that the page daemon is less
346 * likely to reclaim it.
348 vm_page_aflag_set(fs.m, PGA_REFERENCED);
349 vm_page_unlock(fs.m);
350 if (fs.object != fs.first_object) {
351 if (!VM_OBJECT_TRYLOCK(
353 VM_OBJECT_UNLOCK(fs.object);
354 VM_OBJECT_LOCK(fs.first_object);
355 VM_OBJECT_LOCK(fs.object);
357 vm_page_lock(fs.first_m);
358 vm_page_free(fs.first_m);
359 vm_page_unlock(fs.first_m);
360 vm_object_pip_wakeup(fs.first_object);
361 VM_OBJECT_UNLOCK(fs.first_object);
365 if (fs.m == vm_page_lookup(fs.object,
367 vm_page_sleep_if_busy(fs.m, TRUE,
370 vm_object_pip_wakeup(fs.object);
371 VM_OBJECT_UNLOCK(fs.object);
372 PCPU_INC(cnt.v_intrans);
373 vm_object_deallocate(fs.first_object);
376 vm_pageq_remove(fs.m);
377 vm_page_unlock(fs.m);
380 * Mark page busy for other processes, and the
381 * pagedaemon. If it still isn't completely valid
382 * (readable), jump to readrest, else break-out ( we
386 if (fs.m->valid != VM_PAGE_BITS_ALL)
392 * Page is not resident, If this is the search termination
393 * or the pager might contain the page, allocate a new page.
395 if (TRYPAGER || fs.object == fs.first_object) {
396 if (fs.pindex >= fs.object->size) {
397 unlock_and_deallocate(&fs);
398 return (KERN_PROTECTION_FAILURE);
402 * Allocate a new page for this object/offset pair.
404 * Unlocked read of the p_flag is harmless. At
405 * worst, the P_KILLED might be not observed
406 * there, and allocation can fail, causing
407 * restart and new reading of the p_flag.
410 if (!vm_page_count_severe() || P_KILLED(curproc)) {
411 #if VM_NRESERVLEVEL > 0
412 if ((fs.object->flags & OBJ_COLORED) == 0) {
413 fs.object->flags |= OBJ_COLORED;
414 fs.object->pg_color = atop(vaddr) -
418 alloc_req = P_KILLED(curproc) ?
419 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
420 if (fs.object->type != OBJT_VNODE &&
421 fs.object->backing_object == NULL)
422 alloc_req |= VM_ALLOC_ZERO;
423 fs.m = vm_page_alloc(fs.object, fs.pindex,
427 unlock_and_deallocate(&fs);
430 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
436 * We have found a valid page or we have allocated a new page.
437 * The page thus may not be valid or may not be entirely
440 * Attempt to fault-in the page if there is a chance that the
441 * pager has it, and potentially fault in additional pages
447 u_char behavior = vm_map_entry_behavior(fs.entry);
449 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
454 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
455 if (behind > VM_FAULT_READ_BEHIND)
456 behind = VM_FAULT_READ_BEHIND;
458 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
459 if (ahead > VM_FAULT_READ_AHEAD)
460 ahead = VM_FAULT_READ_AHEAD;
462 is_first_object_locked = FALSE;
463 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
464 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
465 fs.pindex >= fs.entry->lastr &&
466 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
467 (fs.first_object == fs.object ||
468 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
469 fs.first_object->type != OBJT_DEVICE &&
470 fs.first_object->type != OBJT_PHYS &&
471 fs.first_object->type != OBJT_SG) {
472 vm_pindex_t firstpindex;
474 if (fs.first_pindex < 2 * VM_FAULT_READ)
477 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
478 mt = fs.first_object != fs.object ?
480 KASSERT(mt != NULL, ("vm_fault: missing mt"));
481 KASSERT((mt->oflags & VPO_BUSY) != 0,
482 ("vm_fault: mt %p not busy", mt));
483 mt_prev = vm_page_prev(mt);
486 * note: partially valid pages cannot be
487 * included in the lookahead - NFS piecemeal
488 * writes will barf on it badly.
490 while ((mt = mt_prev) != NULL &&
491 mt->pindex >= firstpindex &&
492 mt->valid == VM_PAGE_BITS_ALL) {
493 mt_prev = vm_page_prev(mt);
495 (mt->oflags & VPO_BUSY))
498 if (mt->hold_count ||
505 vm_page_deactivate(mt);
513 if (is_first_object_locked)
514 VM_OBJECT_UNLOCK(fs.first_object);
517 * Call the pager to retrieve the data, if any, after
518 * releasing the lock on the map. We hold a ref on
519 * fs.object and the pages are VPO_BUSY'd.
524 if (fs.object->type == OBJT_VNODE) {
525 vp = fs.object->handle;
528 else if (fs.vp != NULL) {
532 locked = VOP_ISLOCKED(vp);
534 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
536 if (!mtx_trylock(&Giant)) {
537 VM_OBJECT_UNLOCK(fs.object);
539 VM_OBJECT_LOCK(fs.object);
543 if (locked != LK_EXCLUSIVE)
545 /* Do not sleep for vnode lock while fs.m is busy */
546 error = vget(vp, locked | LK_CANRECURSE |
547 LK_NOWAIT, curthread);
551 vfslocked = fs.vfslocked;
552 fs.vfslocked = 0; /* Keep Giant */
555 unlock_and_deallocate(&fs);
556 error = vget(vp, locked | LK_RETRY |
557 LK_CANRECURSE, curthread);
560 fs.vfslocked = vfslocked;
562 ("vm_fault: vget failed"));
568 KASSERT(fs.vp == NULL || !fs.map->system_map,
569 ("vm_fault: vnode-backed object mapped by system map"));
572 * now we find out if any other pages should be paged
573 * in at this time this routine checks to see if the
574 * pages surrounding this fault reside in the same
575 * object as the page for this fault. If they do,
576 * then they are faulted in also into the object. The
577 * array "marray" returned contains an array of
578 * vm_page_t structs where one of them is the
579 * vm_page_t passed to the routine. The reqpage
580 * return value is the index into the marray for the
581 * vm_page_t passed to the routine.
583 * fs.m plus the additional pages are VPO_BUSY'd.
585 faultcount = vm_fault_additional_pages(
586 fs.m, behind, ahead, marray, &reqpage);
589 vm_pager_get_pages(fs.object, marray, faultcount,
590 reqpage) : VM_PAGER_FAIL;
592 if (rv == VM_PAGER_OK) {
594 * Found the page. Leave it busy while we play
599 * Relookup in case pager changed page. Pager
600 * is responsible for disposition of old page
603 fs.m = vm_page_lookup(fs.object, fs.pindex);
605 unlock_and_deallocate(&fs);
610 break; /* break to PAGE HAS BEEN FOUND */
613 * Remove the bogus page (which does not exist at this
614 * object/offset); before doing so, we must get back
615 * our object lock to preserve our invariant.
617 * Also wake up any other process that may want to bring
620 * If this is the top-level object, we must leave the
621 * busy page to prevent another process from rushing
622 * past us, and inserting the page in that object at
623 * the same time that we are.
625 if (rv == VM_PAGER_ERROR)
626 printf("vm_fault: pager read error, pid %d (%s)\n",
627 curproc->p_pid, curproc->p_comm);
629 * Data outside the range of the pager or an I/O error
632 * XXX - the check for kernel_map is a kludge to work
633 * around having the machine panic on a kernel space
634 * fault w/ I/O error.
636 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
637 (rv == VM_PAGER_BAD)) {
640 vm_page_unlock(fs.m);
642 unlock_and_deallocate(&fs);
643 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
645 if (fs.object != fs.first_object) {
648 vm_page_unlock(fs.m);
651 * XXX - we cannot just fall out at this
652 * point, m has been freed and is invalid!
658 * We get here if the object has default pager (or unwiring)
659 * or the pager doesn't have the page.
661 if (fs.object == fs.first_object)
665 * Move on to the next object. Lock the next object before
666 * unlocking the current one.
668 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
669 next_object = fs.object->backing_object;
670 if (next_object == NULL) {
672 * If there's no object left, fill the page in the top
675 if (fs.object != fs.first_object) {
676 vm_object_pip_wakeup(fs.object);
677 VM_OBJECT_UNLOCK(fs.object);
679 fs.object = fs.first_object;
680 fs.pindex = fs.first_pindex;
682 VM_OBJECT_LOCK(fs.object);
687 * Zero the page if necessary and mark it valid.
689 if ((fs.m->flags & PG_ZERO) == 0) {
690 pmap_zero_page(fs.m);
692 PCPU_INC(cnt.v_ozfod);
694 PCPU_INC(cnt.v_zfod);
695 fs.m->valid = VM_PAGE_BITS_ALL;
696 break; /* break to PAGE HAS BEEN FOUND */
698 KASSERT(fs.object != next_object,
699 ("object loop %p", next_object));
700 VM_OBJECT_LOCK(next_object);
701 vm_object_pip_add(next_object, 1);
702 if (fs.object != fs.first_object)
703 vm_object_pip_wakeup(fs.object);
704 VM_OBJECT_UNLOCK(fs.object);
705 fs.object = next_object;
709 KASSERT((fs.m->oflags & VPO_BUSY) != 0,
710 ("vm_fault: not busy after main loop"));
713 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
718 * If the page is being written, but isn't already owned by the
719 * top-level object, we have to copy it into a new page owned by the
722 if (fs.object != fs.first_object) {
724 * We only really need to copy if we want to write it.
726 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
728 * This allows pages to be virtually copied from a
729 * backing_object into the first_object, where the
730 * backing object has no other refs to it, and cannot
731 * gain any more refs. Instead of a bcopy, we just
732 * move the page from the backing object to the
733 * first object. Note that we must mark the page
734 * dirty in the first object so that it will go out
735 * to swap when needed.
737 is_first_object_locked = FALSE;
740 * Only one shadow object
742 (fs.object->shadow_count == 1) &&
744 * No COW refs, except us
746 (fs.object->ref_count == 1) &&
748 * No one else can look this object up
750 (fs.object->handle == NULL) &&
752 * No other ways to look the object up
754 ((fs.object->type == OBJT_DEFAULT) ||
755 (fs.object->type == OBJT_SWAP)) &&
756 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
758 * We don't chase down the shadow chain
760 fs.object == fs.first_object->backing_object) {
762 * get rid of the unnecessary page
764 vm_page_lock(fs.first_m);
765 vm_page_free(fs.first_m);
766 vm_page_unlock(fs.first_m);
768 * grab the page and put it into the
769 * process'es object. The page is
770 * automatically made dirty.
773 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
774 vm_page_unlock(fs.m);
778 PCPU_INC(cnt.v_cow_optim);
781 * Oh, well, lets copy it.
783 pmap_copy_page(fs.m, fs.first_m);
784 fs.first_m->valid = VM_PAGE_BITS_ALL;
785 if (wired && (fault_flags &
786 VM_FAULT_CHANGE_WIRING) == 0) {
787 vm_page_lock(fs.first_m);
788 vm_page_wire(fs.first_m);
789 vm_page_unlock(fs.first_m);
792 vm_page_unwire(fs.m, FALSE);
793 vm_page_unlock(fs.m);
796 * We no longer need the old page or object.
801 * fs.object != fs.first_object due to above
804 vm_object_pip_wakeup(fs.object);
805 VM_OBJECT_UNLOCK(fs.object);
807 * Only use the new page below...
809 fs.object = fs.first_object;
810 fs.pindex = fs.first_pindex;
812 if (!is_first_object_locked)
813 VM_OBJECT_LOCK(fs.object);
814 PCPU_INC(cnt.v_cow_faults);
816 prot &= ~VM_PROT_WRITE;
821 * We must verify that the maps have not changed since our last
824 if (!fs.lookup_still_valid) {
825 vm_object_t retry_object;
826 vm_pindex_t retry_pindex;
827 vm_prot_t retry_prot;
829 if (!vm_map_trylock_read(fs.map)) {
831 unlock_and_deallocate(&fs);
834 fs.lookup_still_valid = TRUE;
835 if (fs.map->timestamp != map_generation) {
836 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
837 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
840 * If we don't need the page any longer, put it on the inactive
841 * list (the easiest thing to do here). If no one needs it,
842 * pageout will grab it eventually.
844 if (result != KERN_SUCCESS) {
846 unlock_and_deallocate(&fs);
849 * If retry of map lookup would have blocked then
850 * retry fault from start.
852 if (result == KERN_FAILURE)
856 if ((retry_object != fs.first_object) ||
857 (retry_pindex != fs.first_pindex)) {
859 unlock_and_deallocate(&fs);
864 * Check whether the protection has changed or the object has
865 * been copied while we left the map unlocked. Changing from
866 * read to write permission is OK - we leave the page
867 * write-protected, and catch the write fault. Changing from
868 * write to read permission means that we can't mark the page
869 * write-enabled after all.
875 * If the page was filled by a pager, update the map entry's
876 * last read offset. Since the pager does not return the
877 * actual set of pages that it read, this update is based on
878 * the requested set. Typically, the requested and actual
881 * XXX The following assignment modifies the map
882 * without holding a write lock on it.
885 fs.entry->lastr = fs.pindex + faultcount - behind;
887 if ((prot & VM_PROT_WRITE) != 0 ||
888 (fault_flags & VM_FAULT_DIRTY) != 0) {
889 vm_object_set_writeable_dirty(fs.object);
892 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
893 * if the page is already dirty to prevent data written with
894 * the expectation of being synced from not being synced.
895 * Likewise if this entry does not request NOSYNC then make
896 * sure the page isn't marked NOSYNC. Applications sharing
897 * data should use the same flags to avoid ping ponging.
899 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
900 if (fs.m->dirty == 0)
901 fs.m->oflags |= VPO_NOSYNC;
903 fs.m->oflags &= ~VPO_NOSYNC;
907 * If the fault is a write, we know that this page is being
908 * written NOW so dirty it explicitly to save on
909 * pmap_is_modified() calls later.
911 * Also tell the backing pager, if any, that it should remove
912 * any swap backing since the page is now dirty.
914 if (((fault_type & VM_PROT_WRITE) != 0 &&
915 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
916 (fault_flags & VM_FAULT_DIRTY) != 0) {
918 vm_pager_page_unswapped(fs.m);
923 * Page had better still be busy
925 KASSERT(fs.m->oflags & VPO_BUSY,
926 ("vm_fault: page %p not busy!", fs.m));
928 * Page must be completely valid or it is not fit to
929 * map into user space. vm_pager_get_pages() ensures this.
931 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
932 ("vm_fault: page %p partially invalid", fs.m));
933 VM_OBJECT_UNLOCK(fs.object);
936 * Put this page into the physical map. We had to do the unlock above
937 * because pmap_enter() may sleep. We don't put the page
938 * back on the active queue until later so that the pageout daemon
939 * won't find it (yet).
941 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
942 if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
943 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
944 VM_OBJECT_LOCK(fs.object);
948 * If the page is not wired down, then put it where the pageout daemon
951 if (fault_flags & VM_FAULT_CHANGE_WIRING) {
955 vm_page_unwire(fs.m, 1);
957 vm_page_activate(fs.m);
958 if (m_hold != NULL) {
962 vm_page_unlock(fs.m);
963 vm_page_wakeup(fs.m);
966 * Unlock everything, and return
968 unlock_and_deallocate(&fs);
970 curthread->td_ru.ru_majflt++;
972 curthread->td_ru.ru_minflt++;
974 return (KERN_SUCCESS);
978 * vm_fault_prefault provides a quick way of clustering
979 * pagefaults into a processes address space. It is a "cousin"
980 * of vm_map_pmap_enter, except it runs at page fault time instead
984 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
987 vm_offset_t addr, starta;
992 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
995 object = entry->object.vm_object;
997 starta = addra - PFBAK * PAGE_SIZE;
998 if (starta < entry->start) {
999 starta = entry->start;
1000 } else if (starta > addra) {
1004 for (i = 0; i < PAGEORDER_SIZE; i++) {
1005 vm_object_t backing_object, lobject;
1007 addr = addra + prefault_pageorder[i];
1008 if (addr > addra + (PFFOR * PAGE_SIZE))
1011 if (addr < starta || addr >= entry->end)
1014 if (!pmap_is_prefaultable(pmap, addr))
1017 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1019 VM_OBJECT_LOCK(lobject);
1020 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1021 lobject->type == OBJT_DEFAULT &&
1022 (backing_object = lobject->backing_object) != NULL) {
1023 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1024 0, ("vm_fault_prefault: unaligned object offset"));
1025 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1026 VM_OBJECT_LOCK(backing_object);
1027 VM_OBJECT_UNLOCK(lobject);
1028 lobject = backing_object;
1031 * give-up when a page is not in memory
1034 VM_OBJECT_UNLOCK(lobject);
1037 if (m->valid == VM_PAGE_BITS_ALL &&
1038 (m->flags & PG_FICTITIOUS) == 0)
1039 pmap_enter_quick(pmap, addr, m, entry->protection);
1040 VM_OBJECT_UNLOCK(lobject);
1045 * Hold each of the physical pages that are mapped by the specified range of
1046 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1047 * and allow the specified types of access, "prot". If all of the implied
1048 * pages are successfully held, then the number of held pages is returned
1049 * together with pointers to those pages in the array "ma". However, if any
1050 * of the pages cannot be held, -1 is returned.
1053 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1054 vm_prot_t prot, vm_page_t *ma, int max_count)
1056 vm_offset_t end, va;
1059 boolean_t pmap_failed;
1063 end = round_page(addr + len);
1064 addr = trunc_page(addr);
1067 * Check for illegal addresses.
1069 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1072 count = howmany(end - addr, PAGE_SIZE);
1073 if (count > max_count)
1074 panic("vm_fault_quick_hold_pages: count > max_count");
1077 * Most likely, the physical pages are resident in the pmap, so it is
1078 * faster to try pmap_extract_and_hold() first.
1080 pmap_failed = FALSE;
1081 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1082 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1085 else if ((prot & VM_PROT_WRITE) != 0 &&
1086 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1088 * Explicitly dirty the physical page. Otherwise, the
1089 * caller's changes may go unnoticed because they are
1090 * performed through an unmanaged mapping or by a DMA
1093 * The object lock is not held here.
1094 * See vm_page_clear_dirty_mask().
1101 * One or more pages could not be held by the pmap. Either no
1102 * page was mapped at the specified virtual address or that
1103 * mapping had insufficient permissions. Attempt to fault in
1104 * and hold these pages.
1106 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1107 if (*mp == NULL && vm_fault_hold(map, va, prot,
1108 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1113 for (mp = ma; mp < ma + count; mp++)
1116 vm_page_unhold(*mp);
1117 vm_page_unlock(*mp);
1125 * Wire down a range of virtual addresses in a map.
1128 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1129 boolean_t fictitious)
1135 * We simulate a fault to get the page and enter it in the physical
1136 * map. For user wiring, we only ask for read access on currently
1137 * read-only sections.
1139 for (va = start; va < end; va += PAGE_SIZE) {
1140 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1143 vm_fault_unwire(map, start, va, fictitious);
1147 return (KERN_SUCCESS);
1153 * Unwire a range of virtual addresses in a map.
1156 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1157 boolean_t fictitious)
1164 pmap = vm_map_pmap(map);
1167 * Since the pages are wired down, we must be able to get their
1168 * mappings from the physical map system.
1170 for (va = start; va < end; va += PAGE_SIZE) {
1171 pa = pmap_extract(pmap, va);
1173 pmap_change_wiring(pmap, va, FALSE);
1175 m = PHYS_TO_VM_PAGE(pa);
1177 vm_page_unwire(m, TRUE);
1186 * vm_fault_copy_entry
1188 * Create new shadow object backing dst_entry with private copy of
1189 * all underlying pages. When src_entry is equal to dst_entry,
1190 * function implements COW for wired-down map entry. Otherwise,
1191 * it forks wired entry into dst_map.
1193 * In/out conditions:
1194 * The source and destination maps must be locked for write.
1195 * The source map entry must be wired down (or be a sharing map
1196 * entry corresponding to a main map entry that is wired down).
1199 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1200 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1201 vm_ooffset_t *fork_charge)
1203 vm_object_t backing_object, dst_object, object, src_object;
1204 vm_pindex_t dst_pindex, pindex, src_pindex;
1205 vm_prot_t access, prot;
1209 boolean_t src_readonly, upgrade;
1215 upgrade = src_entry == dst_entry;
1217 src_object = src_entry->object.vm_object;
1218 src_pindex = OFF_TO_IDX(src_entry->offset);
1219 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1222 * Create the top-level object for the destination entry. (Doesn't
1223 * actually shadow anything - we copy the pages directly.)
1225 dst_object = vm_object_allocate(OBJT_DEFAULT,
1226 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1227 #if VM_NRESERVLEVEL > 0
1228 dst_object->flags |= OBJ_COLORED;
1229 dst_object->pg_color = atop(dst_entry->start);
1232 VM_OBJECT_LOCK(dst_object);
1233 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1234 ("vm_fault_copy_entry: vm_object not NULL"));
1235 dst_entry->object.vm_object = dst_object;
1236 dst_entry->offset = 0;
1237 dst_object->charge = dst_entry->end - dst_entry->start;
1238 if (fork_charge != NULL) {
1239 KASSERT(dst_entry->cred == NULL,
1240 ("vm_fault_copy_entry: leaked swp charge"));
1241 dst_object->cred = curthread->td_ucred;
1242 crhold(dst_object->cred);
1243 *fork_charge += dst_object->charge;
1245 dst_object->cred = dst_entry->cred;
1246 dst_entry->cred = NULL;
1248 access = prot = dst_entry->protection;
1250 * If not an upgrade, then enter the mappings in the pmap as
1251 * read and/or execute accesses. Otherwise, enter them as
1254 * A writeable large page mapping is only created if all of
1255 * the constituent small page mappings are modified. Marking
1256 * PTEs as modified on inception allows promotion to happen
1257 * without taking potentially large number of soft faults.
1260 access &= ~VM_PROT_WRITE;
1263 * Loop through all of the pages in the entry's range, copying each
1264 * one from the source object (it should be there) to the destination
1267 for (vaddr = dst_entry->start, dst_pindex = 0;
1268 vaddr < dst_entry->end;
1269 vaddr += PAGE_SIZE, dst_pindex++) {
1272 * Allocate a page in the destination object.
1275 dst_m = vm_page_alloc(dst_object, dst_pindex,
1277 if (dst_m == NULL) {
1278 VM_OBJECT_UNLOCK(dst_object);
1280 VM_OBJECT_LOCK(dst_object);
1282 } while (dst_m == NULL);
1285 * Find the page in the source object, and copy it in.
1286 * (Because the source is wired down, the page will be in
1289 VM_OBJECT_LOCK(src_object);
1290 object = src_object;
1291 pindex = src_pindex + dst_pindex;
1292 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1294 (backing_object = object->backing_object) != NULL) {
1296 * Allow fallback to backing objects if we are reading.
1298 VM_OBJECT_LOCK(backing_object);
1299 pindex += OFF_TO_IDX(object->backing_object_offset);
1300 VM_OBJECT_UNLOCK(object);
1301 object = backing_object;
1304 panic("vm_fault_copy_wired: page missing");
1305 pmap_copy_page(src_m, dst_m);
1306 VM_OBJECT_UNLOCK(object);
1307 dst_m->valid = VM_PAGE_BITS_ALL;
1308 VM_OBJECT_UNLOCK(dst_object);
1311 * Enter it in the pmap. If a wired, copy-on-write
1312 * mapping is being replaced by a write-enabled
1313 * mapping, then wire that new mapping.
1315 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1318 * Mark it no longer busy, and put it on the active list.
1320 VM_OBJECT_LOCK(dst_object);
1323 vm_page_lock(src_m);
1324 vm_page_unwire(src_m, 0);
1325 vm_page_unlock(src_m);
1327 vm_page_lock(dst_m);
1328 vm_page_wire(dst_m);
1329 vm_page_unlock(dst_m);
1331 vm_page_lock(dst_m);
1332 vm_page_activate(dst_m);
1333 vm_page_unlock(dst_m);
1335 vm_page_wakeup(dst_m);
1337 VM_OBJECT_UNLOCK(dst_object);
1339 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1340 vm_object_deallocate(src_object);
1346 * This routine checks around the requested page for other pages that
1347 * might be able to be faulted in. This routine brackets the viable
1348 * pages for the pages to be paged in.
1351 * m, rbehind, rahead
1354 * marray (array of vm_page_t), reqpage (index of requested page)
1357 * number of pages in marray
1360 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1369 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1371 int cbehind, cahead;
1373 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1377 cbehind = cahead = 0;
1380 * if the requested page is not available, then give up now
1382 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1386 if ((cbehind == 0) && (cahead == 0)) {
1392 if (rahead > cahead) {
1396 if (rbehind > cbehind) {
1401 * scan backward for the read behind pages -- in memory
1404 if (rbehind > pindex) {
1408 startpindex = pindex - rbehind;
1411 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1412 rtm->pindex >= startpindex)
1413 startpindex = rtm->pindex + 1;
1415 /* tpindex is unsigned; beware of numeric underflow. */
1416 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1417 tpindex < pindex; i++, tpindex--) {
1419 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1420 VM_ALLOC_IFNOTCACHED);
1423 * Shift the allocated pages to the
1424 * beginning of the array.
1426 for (j = 0; j < i; j++) {
1427 marray[j] = marray[j + tpindex + 1 -
1433 marray[tpindex - startpindex] = rtm;
1441 /* page offset of the required page */
1444 tpindex = pindex + 1;
1448 * scan forward for the read ahead pages
1450 endpindex = tpindex + rahead;
1451 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1452 endpindex = rtm->pindex;
1453 if (endpindex > object->size)
1454 endpindex = object->size;
1456 for (; tpindex < endpindex; i++, tpindex++) {
1458 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1459 VM_ALLOC_IFNOTCACHED);
1467 /* return number of pages */
1472 vm_fault_disable_pagefaults(void)
1475 return (curthread_pflags_set(TDP_NOFAULTING));
1479 vm_fault_enable_pagefaults(int save)
1482 curthread_pflags_restore(save);