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, alloc_req;
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 if (fs.object != fs.first_object) {
365 if (!VM_OBJECT_TRYLOCK(
367 VM_OBJECT_UNLOCK(fs.object);
368 VM_OBJECT_LOCK(fs.first_object);
369 VM_OBJECT_LOCK(fs.object);
371 vm_page_lock_queues();
372 vm_page_free(fs.first_m);
373 vm_page_unlock_queues();
374 vm_object_pip_wakeup(fs.first_object);
375 VM_OBJECT_UNLOCK(fs.first_object);
379 if (fs.m == vm_page_lookup(fs.object,
381 vm_page_sleep_if_busy(fs.m, TRUE,
384 vm_object_pip_wakeup(fs.object);
385 VM_OBJECT_UNLOCK(fs.object);
386 PCPU_INC(cnt.v_intrans);
387 vm_object_deallocate(fs.first_object);
390 vm_pageq_remove(fs.m);
391 vm_page_unlock_queues();
394 * Mark page busy for other processes, and the
395 * pagedaemon. If it still isn't completely valid
396 * (readable), jump to readrest, else break-out ( we
400 if (fs.m->valid != VM_PAGE_BITS_ALL &&
401 fs.m->object != kernel_object && fs.m->object != kmem_object) {
409 * Page is not resident, If this is the search termination
410 * or the pager might contain the page, allocate a new page.
412 if (TRYPAGER || fs.object == fs.first_object) {
413 if (fs.pindex >= fs.object->size) {
414 unlock_and_deallocate(&fs);
415 return (KERN_PROTECTION_FAILURE);
419 * Allocate a new page for this object/offset pair.
421 * Unlocked read of the p_flag is harmless. At
422 * worst, the P_KILLED might be not observed
423 * there, and allocation can fail, causing
424 * restart and new reading of the p_flag.
427 if (!vm_page_count_severe() || P_KILLED(curproc)) {
428 #if VM_NRESERVLEVEL > 0
429 if ((fs.object->flags & OBJ_COLORED) == 0) {
430 fs.object->flags |= OBJ_COLORED;
431 fs.object->pg_color = atop(vaddr) -
435 alloc_req = P_KILLED(curproc) ?
436 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
437 if (fs.object->type != OBJT_VNODE &&
438 fs.object->backing_object == NULL)
439 alloc_req |= VM_ALLOC_ZERO;
440 fs.m = vm_page_alloc(fs.object, fs.pindex,
444 unlock_and_deallocate(&fs);
447 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
453 * We have found a valid page or we have allocated a new page.
454 * The page thus may not be valid or may not be entirely
457 * Attempt to fault-in the page if there is a chance that the
458 * pager has it, and potentially fault in additional pages
464 u_char behavior = vm_map_entry_behavior(fs.entry);
466 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
471 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
472 if (behind > VM_FAULT_READ_BEHIND)
473 behind = VM_FAULT_READ_BEHIND;
475 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
476 if (ahead > VM_FAULT_READ_AHEAD)
477 ahead = VM_FAULT_READ_AHEAD;
479 is_first_object_locked = FALSE;
480 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
481 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
482 fs.pindex >= fs.entry->lastr &&
483 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
484 (fs.first_object == fs.object ||
485 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
486 fs.first_object->type != OBJT_DEVICE &&
487 fs.first_object->type != OBJT_PHYS &&
488 fs.first_object->type != OBJT_SG) {
489 vm_pindex_t firstpindex, tmppindex;
491 if (fs.first_pindex < 2 * VM_FAULT_READ)
494 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
496 are_queues_locked = FALSE;
498 * note: partially valid pages cannot be
499 * included in the lookahead - NFS piecemeal
500 * writes will barf on it badly.
502 for (tmppindex = fs.first_pindex - 1;
503 tmppindex >= firstpindex;
507 mt = vm_page_lookup(fs.first_object, tmppindex);
508 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
511 (mt->oflags & VPO_BUSY))
513 if (!are_queues_locked) {
514 are_queues_locked = TRUE;
515 vm_page_lock_queues();
517 if (mt->hold_count ||
522 vm_page_deactivate(mt);
527 if (are_queues_locked)
528 vm_page_unlock_queues();
532 if (is_first_object_locked)
533 VM_OBJECT_UNLOCK(fs.first_object);
536 * Call the pager to retrieve the data, if any, after
537 * releasing the lock on the map. We hold a ref on
538 * fs.object and the pages are VPO_BUSY'd.
543 if (fs.object->type == OBJT_VNODE) {
544 vp = fs.object->handle;
547 else if (fs.vp != NULL) {
551 locked = VOP_ISLOCKED(vp);
553 if (VFS_NEEDSGIANT(vp->v_mount) && !fs.vfslocked) {
555 if (!mtx_trylock(&Giant)) {
556 VM_OBJECT_UNLOCK(fs.object);
558 VM_OBJECT_LOCK(fs.object);
562 if (locked != LK_EXCLUSIVE)
564 /* Do not sleep for vnode lock while fs.m is busy */
565 error = vget(vp, locked | LK_CANRECURSE |
566 LK_NOWAIT, curthread);
570 vfslocked = fs.vfslocked;
571 fs.vfslocked = 0; /* Keep Giant */
574 unlock_and_deallocate(&fs);
575 error = vget(vp, locked | LK_RETRY |
576 LK_CANRECURSE, curthread);
579 fs.vfslocked = vfslocked;
581 ("vm_fault: vget failed"));
587 KASSERT(fs.vp == NULL || !fs.map->system_map,
588 ("vm_fault: vnode-backed object mapped by system map"));
591 * now we find out if any other pages should be paged
592 * in at this time this routine checks to see if the
593 * pages surrounding this fault reside in the same
594 * object as the page for this fault. If they do,
595 * then they are faulted in also into the object. The
596 * array "marray" returned contains an array of
597 * vm_page_t structs where one of them is the
598 * vm_page_t passed to the routine. The reqpage
599 * return value is the index into the marray for the
600 * vm_page_t passed to the routine.
602 * fs.m plus the additional pages are VPO_BUSY'd.
604 faultcount = vm_fault_additional_pages(
605 fs.m, behind, ahead, marray, &reqpage);
608 vm_pager_get_pages(fs.object, marray, faultcount,
609 reqpage) : VM_PAGER_FAIL;
611 if (rv == VM_PAGER_OK) {
613 * Found the page. Leave it busy while we play
618 * Relookup in case pager changed page. Pager
619 * is responsible for disposition of old page
622 fs.m = vm_page_lookup(fs.object, fs.pindex);
624 unlock_and_deallocate(&fs);
629 break; /* break to PAGE HAS BEEN FOUND */
632 * Remove the bogus page (which does not exist at this
633 * object/offset); before doing so, we must get back
634 * our object lock to preserve our invariant.
636 * Also wake up any other process that may want to bring
639 * If this is the top-level object, we must leave the
640 * busy page to prevent another process from rushing
641 * past us, and inserting the page in that object at
642 * the same time that we are.
644 if (rv == VM_PAGER_ERROR)
645 printf("vm_fault: pager read error, pid %d (%s)\n",
646 curproc->p_pid, curproc->p_comm);
648 * Data outside the range of the pager or an I/O error
651 * XXX - the check for kernel_map is a kludge to work
652 * around having the machine panic on a kernel space
653 * fault w/ I/O error.
655 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
656 (rv == VM_PAGER_BAD)) {
657 vm_page_lock_queues();
659 vm_page_unlock_queues();
661 unlock_and_deallocate(&fs);
662 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
664 if (fs.object != fs.first_object) {
665 vm_page_lock_queues();
667 vm_page_unlock_queues();
670 * XXX - we cannot just fall out at this
671 * point, m has been freed and is invalid!
677 * We get here if the object has default pager (or unwiring)
678 * or the pager doesn't have the page.
680 if (fs.object == fs.first_object)
684 * Move on to the next object. Lock the next object before
685 * unlocking the current one.
687 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
688 next_object = fs.object->backing_object;
689 if (next_object == NULL) {
691 * If there's no object left, fill the page in the top
694 if (fs.object != fs.first_object) {
695 vm_object_pip_wakeup(fs.object);
696 VM_OBJECT_UNLOCK(fs.object);
698 fs.object = fs.first_object;
699 fs.pindex = fs.first_pindex;
701 VM_OBJECT_LOCK(fs.object);
706 * Zero the page if necessary and mark it valid.
708 if ((fs.m->flags & PG_ZERO) == 0) {
709 pmap_zero_page(fs.m);
711 PCPU_INC(cnt.v_ozfod);
713 PCPU_INC(cnt.v_zfod);
714 fs.m->valid = VM_PAGE_BITS_ALL;
715 break; /* break to PAGE HAS BEEN FOUND */
717 KASSERT(fs.object != next_object,
718 ("object loop %p", next_object));
719 VM_OBJECT_LOCK(next_object);
720 vm_object_pip_add(next_object, 1);
721 if (fs.object != fs.first_object)
722 vm_object_pip_wakeup(fs.object);
723 VM_OBJECT_UNLOCK(fs.object);
724 fs.object = next_object;
728 KASSERT((fs.m->oflags & VPO_BUSY) != 0,
729 ("vm_fault: not busy after main loop"));
732 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
737 * If the page is being written, but isn't already owned by the
738 * top-level object, we have to copy it into a new page owned by the
741 if (fs.object != fs.first_object) {
743 * We only really need to copy if we want to write it.
745 if (fault_type & VM_PROT_WRITE) {
747 * This allows pages to be virtually copied from a
748 * backing_object into the first_object, where the
749 * backing object has no other refs to it, and cannot
750 * gain any more refs. Instead of a bcopy, we just
751 * move the page from the backing object to the
752 * first object. Note that we must mark the page
753 * dirty in the first object so that it will go out
754 * to swap when needed.
756 is_first_object_locked = FALSE;
759 * Only one shadow object
761 (fs.object->shadow_count == 1) &&
763 * No COW refs, except us
765 (fs.object->ref_count == 1) &&
767 * No one else can look this object up
769 (fs.object->handle == NULL) &&
771 * No other ways to look the object up
773 ((fs.object->type == OBJT_DEFAULT) ||
774 (fs.object->type == OBJT_SWAP)) &&
775 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
777 * We don't chase down the shadow chain
779 fs.object == fs.first_object->backing_object) {
780 vm_page_lock_queues();
782 * get rid of the unnecessary page
784 vm_page_free(fs.first_m);
786 * grab the page and put it into the
787 * process'es object. The page is
788 * automatically made dirty.
790 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
791 vm_page_unlock_queues();
795 PCPU_INC(cnt.v_cow_optim);
798 * Oh, well, lets copy it.
800 pmap_copy_page(fs.m, fs.first_m);
801 fs.first_m->valid = VM_PAGE_BITS_ALL;
805 * We no longer need the old page or object.
810 * fs.object != fs.first_object due to above
813 vm_object_pip_wakeup(fs.object);
814 VM_OBJECT_UNLOCK(fs.object);
816 * Only use the new page below...
818 fs.object = fs.first_object;
819 fs.pindex = fs.first_pindex;
821 if (!is_first_object_locked)
822 VM_OBJECT_LOCK(fs.object);
823 PCPU_INC(cnt.v_cow_faults);
825 prot &= ~VM_PROT_WRITE;
830 * We must verify that the maps have not changed since our last
833 if (!fs.lookup_still_valid) {
834 vm_object_t retry_object;
835 vm_pindex_t retry_pindex;
836 vm_prot_t retry_prot;
838 if (!vm_map_trylock_read(fs.map)) {
840 unlock_and_deallocate(&fs);
843 fs.lookup_still_valid = TRUE;
844 if (fs.map->timestamp != map_generation) {
845 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
846 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
849 * If we don't need the page any longer, put it on the inactive
850 * list (the easiest thing to do here). If no one needs it,
851 * pageout will grab it eventually.
853 if (result != KERN_SUCCESS) {
855 unlock_and_deallocate(&fs);
858 * If retry of map lookup would have blocked then
859 * retry fault from start.
861 if (result == KERN_FAILURE)
865 if ((retry_object != fs.first_object) ||
866 (retry_pindex != fs.first_pindex)) {
868 unlock_and_deallocate(&fs);
873 * Check whether the protection has changed or the object has
874 * been copied while we left the map unlocked. Changing from
875 * read to write permission is OK - we leave the page
876 * write-protected, and catch the write fault. Changing from
877 * write to read permission means that we can't mark the page
878 * write-enabled after all.
884 * If the page was filled by a pager, update the map entry's
885 * last read offset. Since the pager does not return the
886 * actual set of pages that it read, this update is based on
887 * the requested set. Typically, the requested and actual
890 * XXX The following assignment modifies the map
891 * without holding a write lock on it.
894 fs.entry->lastr = fs.pindex + faultcount - behind;
896 if (prot & VM_PROT_WRITE) {
897 vm_object_set_writeable_dirty(fs.object);
900 * If the fault is a write, we know that this page is being
901 * written NOW so dirty it explicitly to save on
902 * pmap_is_modified() calls later.
904 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
905 * if the page is already dirty to prevent data written with
906 * the expectation of being synced from not being synced.
907 * Likewise if this entry does not request NOSYNC then make
908 * sure the page isn't marked NOSYNC. Applications sharing
909 * data should use the same flags to avoid ping ponging.
911 * Also tell the backing pager, if any, that it should remove
912 * any swap backing since the page is now dirty.
914 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
915 if (fs.m->dirty == 0)
916 fs.m->oflags |= VPO_NOSYNC;
918 fs.m->oflags &= ~VPO_NOSYNC;
920 if (fault_flags & VM_FAULT_DIRTY) {
922 vm_pager_page_unswapped(fs.m);
927 * Page had better still be busy
929 KASSERT(fs.m->oflags & VPO_BUSY,
930 ("vm_fault: page %p not busy!", fs.m));
932 * Page must be completely valid or it is not fit to
933 * map into user space. vm_pager_get_pages() ensures this.
935 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
936 ("vm_fault: page %p partially invalid", fs.m));
937 VM_OBJECT_UNLOCK(fs.object);
940 * Put this page into the physical map. We had to do the unlock above
941 * because pmap_enter() may sleep. We don't put the page
942 * back on the active queue until later so that the pageout daemon
943 * won't find it (yet).
945 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
946 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
947 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
949 VM_OBJECT_LOCK(fs.object);
950 vm_page_lock_queues();
951 vm_page_flag_set(fs.m, PG_REFERENCED);
954 * If the page is not wired down, then put it where the pageout daemon
957 if (fault_flags & VM_FAULT_WIRE_MASK) {
961 vm_page_unwire(fs.m, 1);
963 vm_page_activate(fs.m);
965 vm_page_unlock_queues();
966 vm_page_wakeup(fs.m);
969 * Unlock everything, and return
971 unlock_and_deallocate(&fs);
973 curthread->td_ru.ru_majflt++;
975 curthread->td_ru.ru_minflt++;
977 return (KERN_SUCCESS);
981 * vm_fault_prefault provides a quick way of clustering
982 * pagefaults into a processes address space. It is a "cousin"
983 * of vm_map_pmap_enter, except it runs at page fault time instead
987 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
990 vm_offset_t addr, starta;
995 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
998 object = entry->object.vm_object;
1000 starta = addra - PFBAK * PAGE_SIZE;
1001 if (starta < entry->start) {
1002 starta = entry->start;
1003 } else if (starta > addra) {
1007 for (i = 0; i < PAGEORDER_SIZE; i++) {
1008 vm_object_t backing_object, lobject;
1010 addr = addra + prefault_pageorder[i];
1011 if (addr > addra + (PFFOR * PAGE_SIZE))
1014 if (addr < starta || addr >= entry->end)
1017 if (!pmap_is_prefaultable(pmap, addr))
1020 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1022 VM_OBJECT_LOCK(lobject);
1023 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1024 lobject->type == OBJT_DEFAULT &&
1025 (backing_object = lobject->backing_object) != NULL) {
1026 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1027 0, ("vm_fault_prefault: unaligned object offset"));
1028 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1029 VM_OBJECT_LOCK(backing_object);
1030 VM_OBJECT_UNLOCK(lobject);
1031 lobject = backing_object;
1034 * give-up when a page is not in memory
1037 VM_OBJECT_UNLOCK(lobject);
1040 if (m->valid == VM_PAGE_BITS_ALL &&
1041 (m->flags & PG_FICTITIOUS) == 0) {
1042 vm_page_lock_queues();
1043 pmap_enter_quick(pmap, addr, m, entry->protection);
1044 vm_page_unlock_queues();
1046 VM_OBJECT_UNLOCK(lobject);
1053 * Ensure that the requested virtual address, which may be in userland,
1054 * is valid. Fault-in the page if necessary. Return -1 on failure.
1057 vm_fault_quick(caddr_t v, int prot)
1061 if (prot & VM_PROT_WRITE)
1062 r = subyte(v, fubyte(v));
1071 * Wire down a range of virtual addresses in a map.
1074 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1075 boolean_t user_wire, boolean_t fictitious)
1081 * We simulate a fault to get the page and enter it in the physical
1082 * map. For user wiring, we only ask for read access on currently
1083 * read-only sections.
1085 for (va = start; va < end; va += PAGE_SIZE) {
1086 rv = vm_fault(map, va,
1087 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1088 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1091 vm_fault_unwire(map, start, va, fictitious);
1095 return (KERN_SUCCESS);
1101 * Unwire a range of virtual addresses in a map.
1104 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1105 boolean_t fictitious)
1111 pmap = vm_map_pmap(map);
1114 * Since the pages are wired down, we must be able to get their
1115 * mappings from the physical map system.
1117 for (va = start; va < end; va += PAGE_SIZE) {
1118 pa = pmap_extract(pmap, va);
1120 pmap_change_wiring(pmap, va, FALSE);
1122 vm_page_lock_queues();
1123 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1124 vm_page_unlock_queues();
1132 * vm_fault_copy_entry
1134 * Create new shadow object backing dst_entry with private copy of
1135 * all underlying pages. When src_entry is equal to dst_entry,
1136 * function implements COW for wired-down map entry. Otherwise,
1137 * it forks wired entry into dst_map.
1139 * In/out conditions:
1140 * The source and destination maps must be locked for write.
1141 * The source map entry must be wired down (or be a sharing map
1142 * entry corresponding to a main map entry that is wired down).
1145 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1146 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1147 vm_ooffset_t *fork_charge)
1149 vm_object_t backing_object, dst_object, object, src_object;
1150 vm_pindex_t dst_pindex, pindex, src_pindex;
1151 vm_prot_t access, prot;
1155 boolean_t src_readonly, upgrade;
1161 upgrade = src_entry == dst_entry;
1163 src_object = src_entry->object.vm_object;
1164 src_pindex = OFF_TO_IDX(src_entry->offset);
1165 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1168 * Create the top-level object for the destination entry. (Doesn't
1169 * actually shadow anything - we copy the pages directly.)
1171 dst_object = vm_object_allocate(OBJT_DEFAULT,
1172 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1173 #if VM_NRESERVLEVEL > 0
1174 dst_object->flags |= OBJ_COLORED;
1175 dst_object->pg_color = atop(dst_entry->start);
1178 VM_OBJECT_LOCK(dst_object);
1179 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1180 ("vm_fault_copy_entry: vm_object not NULL"));
1181 dst_entry->object.vm_object = dst_object;
1182 dst_entry->offset = 0;
1183 dst_object->charge = dst_entry->end - dst_entry->start;
1184 if (fork_charge != NULL) {
1185 KASSERT(dst_entry->uip == NULL,
1186 ("vm_fault_copy_entry: leaked swp charge"));
1187 dst_object->uip = curthread->td_ucred->cr_ruidinfo;
1188 uihold(dst_object->uip);
1189 *fork_charge += dst_object->charge;
1191 dst_object->uip = dst_entry->uip;
1192 dst_entry->uip = NULL;
1194 access = prot = dst_entry->max_protection;
1196 * If not an upgrade, then enter the mappings in the pmap as
1197 * read and/or execute accesses. Otherwise, enter them as
1200 * A writeable large page mapping is only created if all of
1201 * the constituent small page mappings are modified. Marking
1202 * PTEs as modified on inception allows promotion to happen
1203 * without taking potentially large number of soft faults.
1206 access &= ~VM_PROT_WRITE;
1209 * Loop through all of the pages in the entry's range, copying each
1210 * one from the source object (it should be there) to the destination
1213 for (vaddr = dst_entry->start, dst_pindex = 0;
1214 vaddr < dst_entry->end;
1215 vaddr += PAGE_SIZE, dst_pindex++) {
1218 * Allocate a page in the destination object.
1221 dst_m = vm_page_alloc(dst_object, dst_pindex,
1223 if (dst_m == NULL) {
1224 VM_OBJECT_UNLOCK(dst_object);
1226 VM_OBJECT_LOCK(dst_object);
1228 } while (dst_m == NULL);
1231 * Find the page in the source object, and copy it in.
1232 * (Because the source is wired down, the page will be in
1235 VM_OBJECT_LOCK(src_object);
1236 object = src_object;
1237 pindex = src_pindex + dst_pindex;
1238 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1240 (backing_object = object->backing_object) != NULL) {
1242 * Allow fallback to backing objects if we are reading.
1244 VM_OBJECT_LOCK(backing_object);
1245 pindex += OFF_TO_IDX(object->backing_object_offset);
1246 VM_OBJECT_UNLOCK(object);
1247 object = backing_object;
1250 panic("vm_fault_copy_wired: page missing");
1251 pmap_copy_page(src_m, dst_m);
1252 VM_OBJECT_UNLOCK(object);
1253 dst_m->valid = VM_PAGE_BITS_ALL;
1254 VM_OBJECT_UNLOCK(dst_object);
1257 * Enter it in the pmap. If a wired, copy-on-write
1258 * mapping is being replaced by a write-enabled
1259 * mapping, then wire that new mapping.
1261 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1264 * Mark it no longer busy, and put it on the active list.
1266 VM_OBJECT_LOCK(dst_object);
1267 vm_page_lock_queues();
1269 vm_page_unwire(src_m, 0);
1270 vm_page_wire(dst_m);
1272 vm_page_activate(dst_m);
1273 vm_page_unlock_queues();
1274 vm_page_wakeup(dst_m);
1276 VM_OBJECT_UNLOCK(dst_object);
1278 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1279 vm_object_deallocate(src_object);
1285 * This routine checks around the requested page for other pages that
1286 * might be able to be faulted in. This routine brackets the viable
1287 * pages for the pages to be paged in.
1290 * m, rbehind, rahead
1293 * marray (array of vm_page_t), reqpage (index of requested page)
1296 * number of pages in marray
1299 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1308 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1310 int cbehind, cahead;
1312 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1316 cbehind = cahead = 0;
1319 * if the requested page is not available, then give up now
1321 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1325 if ((cbehind == 0) && (cahead == 0)) {
1331 if (rahead > cahead) {
1335 if (rbehind > cbehind) {
1340 * scan backward for the read behind pages -- in memory
1343 if (rbehind > pindex) {
1347 startpindex = pindex - rbehind;
1350 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1351 rtm->pindex >= startpindex)
1352 startpindex = rtm->pindex + 1;
1354 /* tpindex is unsigned; beware of numeric underflow. */
1355 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1356 tpindex < pindex; i++, tpindex--) {
1358 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1359 VM_ALLOC_IFNOTCACHED);
1362 * Shift the allocated pages to the
1363 * beginning of the array.
1365 for (j = 0; j < i; j++) {
1366 marray[j] = marray[j + tpindex + 1 -
1372 marray[tpindex - startpindex] = rtm;
1380 /* page offset of the required page */
1383 tpindex = pindex + 1;
1387 * scan forward for the read ahead pages
1389 endpindex = tpindex + rahead;
1390 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1391 endpindex = rtm->pindex;
1392 if (endpindex > object->size)
1393 endpindex = object->size;
1395 for (; tpindex < endpindex; i++, tpindex++) {
1397 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1398 VM_ALLOC_IFNOTCACHED);
1406 /* return number of pages */