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/vnode_pager.h>
100 #include <vm/vm_extern.h>
102 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
106 #define PAGEORDER_SIZE (PFBAK+PFFOR)
108 static int prefault_pageorder[] = {
109 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
110 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
111 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
112 -4 * PAGE_SIZE, 4 * PAGE_SIZE
115 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
116 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
118 #define VM_FAULT_READ_AHEAD 8
119 #define VM_FAULT_READ_BEHIND 7
120 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
127 vm_object_t first_object;
128 vm_pindex_t first_pindex;
130 vm_map_entry_t entry;
131 int lookup_still_valid;
137 release_page(struct faultstate *fs)
140 vm_page_wakeup(fs->m);
141 vm_page_lock_queues();
142 vm_page_deactivate(fs->m);
143 vm_page_unlock_queues();
148 unlock_map(struct faultstate *fs)
151 if (fs->lookup_still_valid) {
152 vm_map_lookup_done(fs->map, fs->entry);
153 fs->lookup_still_valid = FALSE;
158 unlock_and_deallocate(struct faultstate *fs)
161 vm_object_pip_wakeup(fs->object);
162 VM_OBJECT_UNLOCK(fs->object);
163 if (fs->object != fs->first_object) {
164 VM_OBJECT_LOCK(fs->first_object);
165 vm_page_lock_queues();
166 vm_page_free(fs->first_m);
167 vm_page_unlock_queues();
168 vm_object_pip_wakeup(fs->first_object);
169 VM_OBJECT_UNLOCK(fs->first_object);
172 vm_object_deallocate(fs->first_object);
174 if (fs->vp != NULL) {
178 VFS_UNLOCK_GIANT(fs->vfslocked);
183 * TRYPAGER - used by vm_fault to calculate whether the pager for the
184 * current object *might* contain the page.
186 * default objects are zero-fill, there is no real pager.
188 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
189 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
194 * Handle a page fault occurring at the given address,
195 * requiring the given permissions, in the map specified.
196 * If successful, the page is inserted into the
197 * associated physical map.
199 * NOTE: the given address should be truncated to the
200 * proper page address.
202 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
203 * a standard error specifying why the fault is fatal is returned.
206 * The map in question must be referenced, and remains so.
207 * Caller may hold no locks.
210 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
214 int is_first_object_locked, result;
215 boolean_t are_queues_locked, growstack, wired;
217 vm_object_t next_object;
218 vm_page_t marray[VM_FAULT_READ];
220 int faultcount, ahead, behind;
221 struct faultstate fs;
227 PCPU_INC(cnt.v_vm_faults);
230 faultcount = behind = 0;
235 * Find the backing store object and offset into it to begin the
239 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
240 &fs.first_object, &fs.first_pindex, &prot, &wired);
241 if (result != KERN_SUCCESS) {
242 if (result != KERN_PROTECTION_FAILURE ||
243 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
244 if (growstack && result == KERN_INVALID_ADDRESS &&
245 map != kernel_map && curproc != NULL) {
246 result = vm_map_growstack(curproc, vaddr);
247 if (result != KERN_SUCCESS)
248 return (KERN_FAILURE);
256 * If we are user-wiring a r/w segment, and it is COW, then
257 * we need to do the COW operation. Note that we don't COW
258 * currently RO sections now, because it is NOT desirable
259 * to COW .text. We simply keep .text from ever being COW'ed
260 * and take the heat that one cannot debug wired .text sections.
262 result = vm_map_lookup(&fs.map, vaddr,
263 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
264 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
265 if (result != KERN_SUCCESS)
269 * If we don't COW now, on a user wire, the user will never
270 * be able to write to the mapping. If we don't make this
271 * restriction, the bookkeeping would be nearly impossible.
273 * XXX The following assignment modifies the map without
274 * holding a write lock on it.
276 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
277 fs.entry->max_protection &= ~VM_PROT_WRITE;
280 map_generation = fs.map->timestamp;
282 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
283 panic("vm_fault: fault on nofault entry, addr: %lx",
288 * Make a reference to this object to prevent its disposal while we
289 * are messing with it. Once we have the reference, the map is free
290 * to be diddled. Since objects reference their shadows (and copies),
291 * they will stay around as well.
293 * Bump the paging-in-progress count to prevent size changes (e.g.
294 * truncation operations) during I/O. This must be done after
295 * obtaining the vnode lock in order to avoid possible deadlocks.
297 VM_OBJECT_LOCK(fs.first_object);
298 vm_object_reference_locked(fs.first_object);
299 vm_object_pip_add(fs.first_object, 1);
301 fs.lookup_still_valid = TRUE;
309 * Search for the page at object/offset.
311 fs.object = fs.first_object;
312 fs.pindex = fs.first_pindex;
315 * If the object is dead, we stop here
317 if (fs.object->flags & OBJ_DEAD) {
318 unlock_and_deallocate(&fs);
319 return (KERN_PROTECTION_FAILURE);
323 * See if page is resident
325 fs.m = vm_page_lookup(fs.object, fs.pindex);
328 * check for page-based copy on write.
329 * We check fs.object == fs.first_object so
330 * as to ensure the legacy COW mechanism is
331 * used when the page in question is part of
332 * a shadow object. Otherwise, vm_page_cowfault()
333 * removes the page from the backing object,
334 * which is not what we want.
336 vm_page_lock_queues();
338 (fault_type & VM_PROT_WRITE) &&
339 (fs.object == fs.first_object)) {
340 vm_page_cowfault(fs.m);
341 vm_page_unlock_queues();
342 unlock_and_deallocate(&fs);
347 * Wait/Retry if the page is busy. We have to do this
348 * if the page is busy via either VPO_BUSY or
349 * vm_page_t->busy because the vm_pager may be using
350 * vm_page_t->busy for pageouts ( and even pageins if
351 * it is the vnode pager ), and we could end up trying
352 * to pagein and pageout the same page simultaneously.
354 * We can theoretically allow the busy case on a read
355 * fault if the page is marked valid, but since such
356 * pages are typically already pmap'd, putting that
357 * special case in might be more effort then it is
358 * worth. We cannot under any circumstances mess
359 * around with a vm_page_t->busy page except, perhaps,
362 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
363 vm_page_unlock_queues();
364 VM_OBJECT_UNLOCK(fs.object);
365 if (fs.object != fs.first_object) {
366 VM_OBJECT_LOCK(fs.first_object);
367 vm_page_lock_queues();
368 vm_page_free(fs.first_m);
369 vm_page_unlock_queues();
370 vm_object_pip_wakeup(fs.first_object);
371 VM_OBJECT_UNLOCK(fs.first_object);
375 VM_OBJECT_LOCK(fs.object);
376 if (fs.m == vm_page_lookup(fs.object,
378 vm_page_sleep_if_busy(fs.m, TRUE,
381 vm_object_pip_wakeup(fs.object);
382 VM_OBJECT_UNLOCK(fs.object);
383 PCPU_INC(cnt.v_intrans);
384 vm_object_deallocate(fs.first_object);
387 vm_pageq_remove(fs.m);
388 vm_page_unlock_queues();
391 * Mark page busy for other processes, and the
392 * pagedaemon. If it still isn't completely valid
393 * (readable), jump to readrest, else break-out ( we
397 if (fs.m->valid != VM_PAGE_BITS_ALL &&
398 fs.m->object != kernel_object && fs.m->object != kmem_object) {
406 * Page is not resident, If this is the search termination
407 * or the pager might contain the page, allocate a new page.
409 if (TRYPAGER || fs.object == fs.first_object) {
410 if (fs.pindex >= fs.object->size) {
411 unlock_and_deallocate(&fs);
412 return (KERN_PROTECTION_FAILURE);
416 * Allocate a new page for this object/offset pair.
419 if (!vm_page_count_severe()) {
420 #if VM_NRESERVLEVEL > 0
421 if ((fs.object->flags & OBJ_COLORED) == 0) {
422 fs.object->flags |= OBJ_COLORED;
423 fs.object->pg_color = atop(vaddr) -
427 fs.m = vm_page_alloc(fs.object, fs.pindex,
428 (fs.object->type == OBJT_VNODE ||
429 fs.object->backing_object != NULL) ?
430 VM_ALLOC_NORMAL : VM_ALLOC_ZERO);
433 unlock_and_deallocate(&fs);
436 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
442 * We have found a valid page or we have allocated a new page.
443 * The page thus may not be valid or may not be entirely
446 * Attempt to fault-in the page if there is a chance that the
447 * pager has it, and potentially fault in additional pages
453 u_char behavior = vm_map_entry_behavior(fs.entry);
455 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
459 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
460 if (behind > VM_FAULT_READ_BEHIND)
461 behind = VM_FAULT_READ_BEHIND;
463 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
464 if (ahead > VM_FAULT_READ_AHEAD)
465 ahead = VM_FAULT_READ_AHEAD;
467 is_first_object_locked = FALSE;
468 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
469 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
470 fs.pindex >= fs.entry->lastr &&
471 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
472 (fs.first_object == fs.object ||
473 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
474 fs.first_object->type != OBJT_DEVICE &&
475 fs.first_object->type != OBJT_PHYS) {
476 vm_pindex_t firstpindex, tmppindex;
478 if (fs.first_pindex < 2 * VM_FAULT_READ)
481 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
483 are_queues_locked = FALSE;
485 * note: partially valid pages cannot be
486 * included in the lookahead - NFS piecemeal
487 * writes will barf on it badly.
489 for (tmppindex = fs.first_pindex - 1;
490 tmppindex >= firstpindex;
494 mt = vm_page_lookup(fs.first_object, tmppindex);
495 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
498 (mt->oflags & VPO_BUSY))
500 if (!are_queues_locked) {
501 are_queues_locked = TRUE;
502 vm_page_lock_queues();
504 if (mt->hold_count ||
509 vm_page_deactivate(mt);
514 if (are_queues_locked)
515 vm_page_unlock_queues();
519 if (is_first_object_locked)
520 VM_OBJECT_UNLOCK(fs.first_object);
523 * Call the pager to retrieve the data, if any, after
524 * releasing the lock on the map. We hold a ref on
525 * fs.object and the pages are VPO_BUSY'd.
530 if (fs.object->type == OBJT_VNODE) {
531 vp = fs.object->handle;
534 else if (fs.vp != NULL) {
538 locked = VOP_ISLOCKED(vp);
540 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
542 if (!mtx_trylock(&Giant)) {
543 VM_OBJECT_UNLOCK(fs.object);
545 VM_OBJECT_LOCK(fs.object);
549 if (locked != LK_EXCLUSIVE)
551 /* Do not sleep for vnode lock while fs.m is busy */
552 error = vget(vp, locked | LK_CANRECURSE |
553 LK_NOWAIT, curthread);
557 vfslocked = fs.vfslocked;
558 fs.vfslocked = 0; /* Keep Giant */
561 unlock_and_deallocate(&fs);
562 error = vget(vp, locked | LK_RETRY |
563 LK_CANRECURSE, curthread);
566 fs.vfslocked = vfslocked;
568 ("vm_fault: vget failed"));
574 KASSERT(fs.vp == NULL || !fs.map->system_map,
575 ("vm_fault: vnode-backed object mapped by system map"));
578 * now we find out if any other pages should be paged
579 * in at this time this routine checks to see if the
580 * pages surrounding this fault reside in the same
581 * object as the page for this fault. If they do,
582 * then they are faulted in also into the object. The
583 * array "marray" returned contains an array of
584 * vm_page_t structs where one of them is the
585 * vm_page_t passed to the routine. The reqpage
586 * return value is the index into the marray for the
587 * vm_page_t passed to the routine.
589 * fs.m plus the additional pages are VPO_BUSY'd.
591 faultcount = vm_fault_additional_pages(
592 fs.m, behind, ahead, marray, &reqpage);
595 vm_pager_get_pages(fs.object, marray, faultcount,
596 reqpage) : VM_PAGER_FAIL;
598 if (rv == VM_PAGER_OK) {
600 * Found the page. Leave it busy while we play
605 * Relookup in case pager changed page. Pager
606 * is responsible for disposition of old page
609 fs.m = vm_page_lookup(fs.object, fs.pindex);
611 unlock_and_deallocate(&fs);
616 break; /* break to PAGE HAS BEEN FOUND */
619 * Remove the bogus page (which does not exist at this
620 * object/offset); before doing so, we must get back
621 * our object lock to preserve our invariant.
623 * Also wake up any other process that may want to bring
626 * If this is the top-level object, we must leave the
627 * busy page to prevent another process from rushing
628 * past us, and inserting the page in that object at
629 * the same time that we are.
631 if (rv == VM_PAGER_ERROR)
632 printf("vm_fault: pager read error, pid %d (%s)\n",
633 curproc->p_pid, curproc->p_comm);
635 * Data outside the range of the pager or an I/O error
638 * XXX - the check for kernel_map is a kludge to work
639 * around having the machine panic on a kernel space
640 * fault w/ I/O error.
642 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
643 (rv == VM_PAGER_BAD)) {
644 vm_page_lock_queues();
646 vm_page_unlock_queues();
648 unlock_and_deallocate(&fs);
649 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
651 if (fs.object != fs.first_object) {
652 vm_page_lock_queues();
654 vm_page_unlock_queues();
657 * XXX - we cannot just fall out at this
658 * point, m has been freed and is invalid!
664 * We get here if the object has default pager (or unwiring)
665 * or the pager doesn't have the page.
667 if (fs.object == fs.first_object)
671 * Move on to the next object. Lock the next object before
672 * unlocking the current one.
674 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
675 next_object = fs.object->backing_object;
676 if (next_object == NULL) {
678 * If there's no object left, fill the page in the top
681 if (fs.object != fs.first_object) {
682 vm_object_pip_wakeup(fs.object);
683 VM_OBJECT_UNLOCK(fs.object);
685 fs.object = fs.first_object;
686 fs.pindex = fs.first_pindex;
688 VM_OBJECT_LOCK(fs.object);
693 * Zero the page if necessary and mark it valid.
695 if ((fs.m->flags & PG_ZERO) == 0) {
696 pmap_zero_page(fs.m);
698 PCPU_INC(cnt.v_ozfod);
700 PCPU_INC(cnt.v_zfod);
701 fs.m->valid = VM_PAGE_BITS_ALL;
702 break; /* break to PAGE HAS BEEN FOUND */
704 KASSERT(fs.object != next_object,
705 ("object loop %p", next_object));
706 VM_OBJECT_LOCK(next_object);
707 vm_object_pip_add(next_object, 1);
708 if (fs.object != fs.first_object)
709 vm_object_pip_wakeup(fs.object);
710 VM_OBJECT_UNLOCK(fs.object);
711 fs.object = next_object;
715 KASSERT((fs.m->oflags & VPO_BUSY) != 0,
716 ("vm_fault: not busy after main loop"));
719 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
724 * If the page is being written, but isn't already owned by the
725 * top-level object, we have to copy it into a new page owned by the
728 if (fs.object != fs.first_object) {
730 * We only really need to copy if we want to write it.
732 if (fault_type & VM_PROT_WRITE) {
734 * This allows pages to be virtually copied from a
735 * backing_object into the first_object, where the
736 * backing object has no other refs to it, and cannot
737 * gain any more refs. Instead of a bcopy, we just
738 * move the page from the backing object to the
739 * first object. Note that we must mark the page
740 * dirty in the first object so that it will go out
741 * to swap when needed.
743 is_first_object_locked = FALSE;
746 * Only one shadow object
748 (fs.object->shadow_count == 1) &&
750 * No COW refs, except us
752 (fs.object->ref_count == 1) &&
754 * No one else can look this object up
756 (fs.object->handle == NULL) &&
758 * No other ways to look the object up
760 ((fs.object->type == OBJT_DEFAULT) ||
761 (fs.object->type == OBJT_SWAP)) &&
762 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
764 * We don't chase down the shadow chain
766 fs.object == fs.first_object->backing_object) {
767 vm_page_lock_queues();
769 * get rid of the unnecessary page
771 vm_page_free(fs.first_m);
773 * grab the page and put it into the
774 * process'es object. The page is
775 * automatically made dirty.
777 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
778 vm_page_unlock_queues();
782 PCPU_INC(cnt.v_cow_optim);
785 * Oh, well, lets copy it.
787 pmap_copy_page(fs.m, fs.first_m);
788 fs.first_m->valid = VM_PAGE_BITS_ALL;
792 * We no longer need the old page or object.
797 * fs.object != fs.first_object due to above
800 vm_object_pip_wakeup(fs.object);
801 VM_OBJECT_UNLOCK(fs.object);
803 * Only use the new page below...
805 fs.object = fs.first_object;
806 fs.pindex = fs.first_pindex;
808 if (!is_first_object_locked)
809 VM_OBJECT_LOCK(fs.object);
810 PCPU_INC(cnt.v_cow_faults);
812 prot &= ~VM_PROT_WRITE;
817 * We must verify that the maps have not changed since our last
820 if (!fs.lookup_still_valid) {
821 vm_object_t retry_object;
822 vm_pindex_t retry_pindex;
823 vm_prot_t retry_prot;
825 if (!vm_map_trylock_read(fs.map)) {
827 unlock_and_deallocate(&fs);
830 fs.lookup_still_valid = TRUE;
831 if (fs.map->timestamp != map_generation) {
832 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
833 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
836 * If we don't need the page any longer, put it on the inactive
837 * list (the easiest thing to do here). If no one needs it,
838 * pageout will grab it eventually.
840 if (result != KERN_SUCCESS) {
842 unlock_and_deallocate(&fs);
845 * If retry of map lookup would have blocked then
846 * retry fault from start.
848 if (result == KERN_FAILURE)
852 if ((retry_object != fs.first_object) ||
853 (retry_pindex != fs.first_pindex)) {
855 unlock_and_deallocate(&fs);
860 * Check whether the protection has changed or the object has
861 * been copied while we left the map unlocked. Changing from
862 * read to write permission is OK - we leave the page
863 * write-protected, and catch the write fault. Changing from
864 * write to read permission means that we can't mark the page
865 * write-enabled after all.
871 * If the page was filled by a pager, update the map entry's
872 * last read offset. Since the pager does not return the
873 * actual set of pages that it read, this update is based on
874 * the requested set. Typically, the requested and actual
877 * XXX The following assignment modifies the map
878 * without holding a write lock on it.
881 fs.entry->lastr = fs.pindex + faultcount - behind;
883 if (prot & VM_PROT_WRITE) {
884 vm_object_set_writeable_dirty(fs.object);
887 * If the fault is a write, we know that this page is being
888 * written NOW so dirty it explicitly to save on
889 * pmap_is_modified() calls later.
891 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
892 * if the page is already dirty to prevent data written with
893 * the expectation of being synced from not being synced.
894 * Likewise if this entry does not request NOSYNC then make
895 * sure the page isn't marked NOSYNC. Applications sharing
896 * data should use the same flags to avoid ping ponging.
898 * Also tell the backing pager, if any, that it should remove
899 * any swap backing since the page is now dirty.
901 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
902 if (fs.m->dirty == 0)
903 fs.m->oflags |= VPO_NOSYNC;
905 fs.m->oflags &= ~VPO_NOSYNC;
907 if (fault_flags & VM_FAULT_DIRTY) {
909 vm_pager_page_unswapped(fs.m);
914 * Page had better still be busy
916 KASSERT(fs.m->oflags & VPO_BUSY,
917 ("vm_fault: page %p not busy!", fs.m));
919 * Page must be completely valid or it is not fit to
920 * map into user space. vm_pager_get_pages() ensures this.
922 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
923 ("vm_fault: page %p partially invalid", fs.m));
924 VM_OBJECT_UNLOCK(fs.object);
927 * Put this page into the physical map. We had to do the unlock above
928 * because pmap_enter() may sleep. We don't put the page
929 * back on the active queue until later so that the pageout daemon
930 * won't find it (yet).
932 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
933 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
934 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
936 VM_OBJECT_LOCK(fs.object);
937 vm_page_lock_queues();
938 vm_page_flag_set(fs.m, PG_REFERENCED);
941 * If the page is not wired down, then put it where the pageout daemon
944 if (fault_flags & VM_FAULT_WIRE_MASK) {
948 vm_page_unwire(fs.m, 1);
950 vm_page_activate(fs.m);
952 vm_page_unlock_queues();
953 vm_page_wakeup(fs.m);
956 * Unlock everything, and return
958 unlock_and_deallocate(&fs);
960 curthread->td_ru.ru_majflt++;
962 curthread->td_ru.ru_minflt++;
964 return (KERN_SUCCESS);
968 * vm_fault_prefault provides a quick way of clustering
969 * pagefaults into a processes address space. It is a "cousin"
970 * of vm_map_pmap_enter, except it runs at page fault time instead
974 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
977 vm_offset_t addr, starta;
982 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
985 object = entry->object.vm_object;
987 starta = addra - PFBAK * PAGE_SIZE;
988 if (starta < entry->start) {
989 starta = entry->start;
990 } else if (starta > addra) {
994 for (i = 0; i < PAGEORDER_SIZE; i++) {
995 vm_object_t backing_object, lobject;
997 addr = addra + prefault_pageorder[i];
998 if (addr > addra + (PFFOR * PAGE_SIZE))
1001 if (addr < starta || addr >= entry->end)
1004 if (!pmap_is_prefaultable(pmap, addr))
1007 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1009 VM_OBJECT_LOCK(lobject);
1010 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1011 lobject->type == OBJT_DEFAULT &&
1012 (backing_object = lobject->backing_object) != NULL) {
1013 if (lobject->backing_object_offset & PAGE_MASK)
1015 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1016 VM_OBJECT_LOCK(backing_object);
1017 VM_OBJECT_UNLOCK(lobject);
1018 lobject = backing_object;
1021 * give-up when a page is not in memory
1024 VM_OBJECT_UNLOCK(lobject);
1027 if (m->valid == VM_PAGE_BITS_ALL &&
1028 (m->flags & PG_FICTITIOUS) == 0) {
1029 vm_page_lock_queues();
1030 pmap_enter_quick(pmap, addr, m, entry->protection);
1031 vm_page_unlock_queues();
1033 VM_OBJECT_UNLOCK(lobject);
1040 * Ensure that the requested virtual address, which may be in userland,
1041 * is valid. Fault-in the page if necessary. Return -1 on failure.
1044 vm_fault_quick(caddr_t v, int prot)
1048 if (prot & VM_PROT_WRITE)
1049 r = subyte(v, fubyte(v));
1058 * Wire down a range of virtual addresses in a map.
1061 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1062 boolean_t user_wire, boolean_t fictitious)
1068 * We simulate a fault to get the page and enter it in the physical
1069 * map. For user wiring, we only ask for read access on currently
1070 * read-only sections.
1072 for (va = start; va < end; va += PAGE_SIZE) {
1073 rv = vm_fault(map, va,
1074 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1075 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1078 vm_fault_unwire(map, start, va, fictitious);
1082 return (KERN_SUCCESS);
1088 * Unwire a range of virtual addresses in a map.
1091 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1092 boolean_t fictitious)
1098 pmap = vm_map_pmap(map);
1101 * Since the pages are wired down, we must be able to get their
1102 * mappings from the physical map system.
1104 for (va = start; va < end; va += PAGE_SIZE) {
1105 pa = pmap_extract(pmap, va);
1107 pmap_change_wiring(pmap, va, FALSE);
1109 vm_page_lock_queues();
1110 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1111 vm_page_unlock_queues();
1119 * vm_fault_copy_entry
1121 * Copy all of the pages from a wired-down map entry to another.
1123 * In/out conditions:
1124 * The source and destination maps must be locked for write.
1125 * The source map entry must be wired down (or be a sharing map
1126 * entry corresponding to a main map entry that is wired down).
1129 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1130 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1131 vm_ooffset_t *fork_charge)
1133 vm_object_t backing_object, dst_object, object;
1134 vm_object_t src_object;
1135 vm_ooffset_t dst_offset;
1136 vm_ooffset_t src_offset;
1147 src_object = src_entry->object.vm_object;
1148 src_offset = src_entry->offset;
1151 * Create the top-level object for the destination entry. (Doesn't
1152 * actually shadow anything - we copy the pages directly.)
1154 dst_object = vm_object_allocate(OBJT_DEFAULT,
1155 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1156 #if VM_NRESERVLEVEL > 0
1157 dst_object->flags |= OBJ_COLORED;
1158 dst_object->pg_color = atop(dst_entry->start);
1161 VM_OBJECT_LOCK(dst_object);
1162 KASSERT(dst_entry->object.vm_object == NULL,
1163 ("vm_fault_copy_entry: vm_object not NULL"));
1164 dst_entry->object.vm_object = dst_object;
1165 dst_entry->offset = 0;
1166 dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1167 uihold(dst_object->uip);
1168 dst_object->charge = dst_entry->end - dst_entry->start;
1169 KASSERT(dst_entry->uip == NULL,
1170 ("vm_fault_copy_entry: leaked swp charge"));
1171 *fork_charge += dst_object->charge;
1172 prot = dst_entry->max_protection;
1175 * Loop through all of the pages in the entry's range, copying each
1176 * one from the source object (it should be there) to the destination
1179 for (vaddr = dst_entry->start, dst_offset = 0;
1180 vaddr < dst_entry->end;
1181 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1184 * Allocate a page in the destination object
1187 dst_m = vm_page_alloc(dst_object,
1188 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1189 if (dst_m == NULL) {
1190 VM_OBJECT_UNLOCK(dst_object);
1192 VM_OBJECT_LOCK(dst_object);
1194 } while (dst_m == NULL);
1197 * Find the page in the source object, and copy it in.
1198 * (Because the source is wired down, the page will be in
1201 VM_OBJECT_LOCK(src_object);
1202 object = src_object;
1204 while ((src_m = vm_page_lookup(object, pindex +
1205 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1206 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1207 (backing_object = object->backing_object) != NULL) {
1209 * Allow fallback to backing objects if we are reading.
1211 VM_OBJECT_LOCK(backing_object);
1212 pindex += OFF_TO_IDX(object->backing_object_offset);
1213 VM_OBJECT_UNLOCK(object);
1214 object = backing_object;
1217 panic("vm_fault_copy_wired: page missing");
1218 pmap_copy_page(src_m, dst_m);
1219 VM_OBJECT_UNLOCK(object);
1220 dst_m->valid = VM_PAGE_BITS_ALL;
1221 VM_OBJECT_UNLOCK(dst_object);
1224 * Enter it in the pmap as a read and/or execute access.
1226 pmap_enter(dst_map->pmap, vaddr, prot & ~VM_PROT_WRITE, dst_m,
1230 * Mark it no longer busy, and put it on the active list.
1232 VM_OBJECT_LOCK(dst_object);
1233 vm_page_lock_queues();
1234 vm_page_activate(dst_m);
1235 vm_page_unlock_queues();
1236 vm_page_wakeup(dst_m);
1238 VM_OBJECT_UNLOCK(dst_object);
1243 * This routine checks around the requested page for other pages that
1244 * might be able to be faulted in. This routine brackets the viable
1245 * pages for the pages to be paged in.
1248 * m, rbehind, rahead
1251 * marray (array of vm_page_t), reqpage (index of requested page)
1254 * number of pages in marray
1257 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1266 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1268 int cbehind, cahead;
1270 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1274 cbehind = cahead = 0;
1277 * if the requested page is not available, then give up now
1279 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1283 if ((cbehind == 0) && (cahead == 0)) {
1289 if (rahead > cahead) {
1293 if (rbehind > cbehind) {
1298 * scan backward for the read behind pages -- in memory
1301 if (rbehind > pindex) {
1305 startpindex = pindex - rbehind;
1308 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1309 rtm->pindex >= startpindex)
1310 startpindex = rtm->pindex + 1;
1312 /* tpindex is unsigned; beware of numeric underflow. */
1313 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1314 tpindex < pindex; i++, tpindex--) {
1316 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1317 VM_ALLOC_IFNOTCACHED);
1320 * Shift the allocated pages to the
1321 * beginning of the array.
1323 for (j = 0; j < i; j++) {
1324 marray[j] = marray[j + tpindex + 1 -
1330 marray[tpindex - startpindex] = rtm;
1338 /* page offset of the required page */
1341 tpindex = pindex + 1;
1345 * scan forward for the read ahead pages
1347 endpindex = tpindex + rahead;
1348 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1349 endpindex = rtm->pindex;
1350 if (endpindex > object->size)
1351 endpindex = object->size;
1353 for (; tpindex < endpindex; i++, tpindex++) {
1355 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1356 VM_ALLOC_IFNOTCACHED);
1364 /* return number of pages */