2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 * Permission to use, copy, modify and distribute this software and
50 * its documentation is hereby granted, provided that both the copyright
51 * notice and this permission notice appear in all copies of the
52 * software, derivative works or modified versions, and any portions
53 * thereof, and that both notices appear in supporting documentation.
55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 * Carnegie Mellon requests users of this software to return to
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
71 * Page fault handling module.
74 #include <sys/cdefs.h>
75 __FBSDID("$FreeBSD$");
77 #include "opt_ktrace.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
84 #include <sys/mutex.h>
86 #include <sys/resourcevar.h>
87 #include <sys/sysctl.h>
88 #include <sys/vmmeter.h>
89 #include <sys/vnode.h>
91 #include <sys/ktrace.h>
95 #include <vm/vm_param.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_kern.h>
102 #include <vm/vm_pager.h>
103 #include <vm/vm_extern.h>
105 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
109 #define PAGEORDER_SIZE (PFBAK+PFFOR)
111 static int prefault_pageorder[] = {
112 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
113 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
114 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
115 -4 * PAGE_SIZE, 4 * PAGE_SIZE
118 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
119 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
121 #define VM_FAULT_READ_BEHIND 8
122 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
123 #define VM_FAULT_NINCR (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
124 #define VM_FAULT_SUM (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
125 #define VM_FAULT_CACHE_BEHIND (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
132 vm_object_t first_object;
133 vm_pindex_t first_pindex;
135 vm_map_entry_t entry;
136 int lookup_still_valid;
141 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
144 release_page(struct faultstate *fs)
147 vm_page_wakeup(fs->m);
149 vm_page_deactivate(fs->m);
150 vm_page_unlock(fs->m);
155 unlock_map(struct faultstate *fs)
158 if (fs->lookup_still_valid) {
159 vm_map_lookup_done(fs->map, fs->entry);
160 fs->lookup_still_valid = FALSE;
165 unlock_and_deallocate(struct faultstate *fs)
168 vm_object_pip_wakeup(fs->object);
169 VM_OBJECT_UNLOCK(fs->object);
170 if (fs->object != fs->first_object) {
171 VM_OBJECT_LOCK(fs->first_object);
172 vm_page_lock(fs->first_m);
173 vm_page_free(fs->first_m);
174 vm_page_unlock(fs->first_m);
175 vm_object_pip_wakeup(fs->first_object);
176 VM_OBJECT_UNLOCK(fs->first_object);
179 vm_object_deallocate(fs->first_object);
181 if (fs->vp != NULL) {
185 VFS_UNLOCK_GIANT(fs->vfslocked);
190 * TRYPAGER - used by vm_fault to calculate whether the pager for the
191 * current object *might* contain the page.
193 * default objects are zero-fill, there is no real pager.
195 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
196 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
201 * Handle a page fault occurring at the given address,
202 * requiring the given permissions, in the map specified.
203 * If successful, the page is inserted into the
204 * associated physical map.
206 * NOTE: the given address should be truncated to the
207 * proper page address.
209 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
210 * a standard error specifying why the fault is fatal is returned.
212 * The map in question must be referenced, and remains so.
213 * Caller may hold no locks.
216 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
223 if ((td->td_pflags & TDP_NOFAULTING) != 0)
224 return (KERN_PROTECTION_FAILURE);
226 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
227 ktrfault(vaddr, fault_type);
229 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
232 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
239 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
240 int fault_flags, vm_page_t *m_hold)
244 int alloc_req, era, faultcount, nera, reqpage, result;
245 boolean_t growstack, is_first_object_locked, wired;
247 vm_object_t next_object;
248 vm_page_t marray[VM_FAULT_READ_MAX];
250 struct faultstate fs;
256 PCPU_INC(cnt.v_vm_faults);
259 faultcount = reqpage = 0;
264 * Find the backing store object and offset into it to begin the
268 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
269 &fs.first_object, &fs.first_pindex, &prot, &wired);
270 if (result != KERN_SUCCESS) {
271 if (growstack && result == KERN_INVALID_ADDRESS &&
273 result = vm_map_growstack(curproc, vaddr);
274 if (result != KERN_SUCCESS)
275 return (KERN_FAILURE);
282 map_generation = fs.map->timestamp;
284 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
285 panic("vm_fault: fault on nofault entry, addr: %lx",
289 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
290 fs.entry->wiring_thread != curthread) {
291 vm_map_unlock_read(fs.map);
293 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
294 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
295 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
296 vm_map_unlock_and_wait(fs.map, 0);
298 vm_map_unlock(fs.map);
303 * Make a reference to this object to prevent its disposal while we
304 * are messing with it. Once we have the reference, the map is free
305 * to be diddled. Since objects reference their shadows (and copies),
306 * they will stay around as well.
308 * Bump the paging-in-progress count to prevent size changes (e.g.
309 * truncation operations) during I/O. This must be done after
310 * obtaining the vnode lock in order to avoid possible deadlocks.
312 VM_OBJECT_LOCK(fs.first_object);
313 vm_object_reference_locked(fs.first_object);
314 vm_object_pip_add(fs.first_object, 1);
316 fs.lookup_still_valid = TRUE;
319 fault_type = prot | (fault_type & VM_PROT_COPY);
324 * Search for the page at object/offset.
326 fs.object = fs.first_object;
327 fs.pindex = fs.first_pindex;
330 * If the object is dead, we stop here
332 if (fs.object->flags & OBJ_DEAD) {
333 unlock_and_deallocate(&fs);
334 return (KERN_PROTECTION_FAILURE);
338 * See if page is resident
340 fs.m = vm_page_lookup(fs.object, fs.pindex);
343 * check for page-based copy on write.
344 * We check fs.object == fs.first_object so
345 * as to ensure the legacy COW mechanism is
346 * used when the page in question is part of
347 * a shadow object. Otherwise, vm_page_cowfault()
348 * removes the page from the backing object,
349 * which is not what we want.
353 (fault_type & VM_PROT_WRITE) &&
354 (fs.object == fs.first_object)) {
355 vm_page_cowfault(fs.m);
356 unlock_and_deallocate(&fs);
361 * Wait/Retry if the page is busy. We have to do this
362 * if the page is busy via either VPO_BUSY or
363 * vm_page_t->busy because the vm_pager may be using
364 * vm_page_t->busy for pageouts ( and even pageins if
365 * it is the vnode pager ), and we could end up trying
366 * to pagein and pageout the same page simultaneously.
368 * We can theoretically allow the busy case on a read
369 * fault if the page is marked valid, but since such
370 * pages are typically already pmap'd, putting that
371 * special case in might be more effort then it is
372 * worth. We cannot under any circumstances mess
373 * around with a vm_page_t->busy page except, perhaps,
376 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
378 * Reference the page before unlocking and
379 * sleeping so that the page daemon is less
380 * likely to reclaim it.
382 vm_page_aflag_set(fs.m, PGA_REFERENCED);
383 vm_page_unlock(fs.m);
384 if (fs.object != fs.first_object) {
385 if (!VM_OBJECT_TRYLOCK(
387 VM_OBJECT_UNLOCK(fs.object);
388 VM_OBJECT_LOCK(fs.first_object);
389 VM_OBJECT_LOCK(fs.object);
391 vm_page_lock(fs.first_m);
392 vm_page_free(fs.first_m);
393 vm_page_unlock(fs.first_m);
394 vm_object_pip_wakeup(fs.first_object);
395 VM_OBJECT_UNLOCK(fs.first_object);
399 if (fs.m == vm_page_lookup(fs.object,
401 vm_page_sleep_if_busy(fs.m, TRUE,
404 vm_object_pip_wakeup(fs.object);
405 VM_OBJECT_UNLOCK(fs.object);
406 PCPU_INC(cnt.v_intrans);
407 vm_object_deallocate(fs.first_object);
410 vm_pageq_remove(fs.m);
411 vm_page_unlock(fs.m);
414 * Mark page busy for other processes, and the
415 * pagedaemon. If it still isn't completely valid
416 * (readable), jump to readrest, else break-out ( we
420 if (fs.m->valid != VM_PAGE_BITS_ALL)
426 * Page is not resident, If this is the search termination
427 * or the pager might contain the page, allocate a new page.
429 if (TRYPAGER || fs.object == fs.first_object) {
430 if (fs.pindex >= fs.object->size) {
431 unlock_and_deallocate(&fs);
432 return (KERN_PROTECTION_FAILURE);
436 * Allocate a new page for this object/offset pair.
438 * Unlocked read of the p_flag is harmless. At
439 * worst, the P_KILLED might be not observed
440 * there, and allocation can fail, causing
441 * restart and new reading of the p_flag.
444 if (!vm_page_count_severe() || P_KILLED(curproc)) {
445 #if VM_NRESERVLEVEL > 0
446 if ((fs.object->flags & OBJ_COLORED) == 0) {
447 fs.object->flags |= OBJ_COLORED;
448 fs.object->pg_color = atop(vaddr) -
452 alloc_req = P_KILLED(curproc) ?
453 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
454 if (fs.object->type != OBJT_VNODE &&
455 fs.object->backing_object == NULL)
456 alloc_req |= VM_ALLOC_ZERO;
457 fs.m = vm_page_alloc(fs.object, fs.pindex,
461 unlock_and_deallocate(&fs);
464 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
470 * We have found a valid page or we have allocated a new page.
471 * The page thus may not be valid or may not be entirely
474 * Attempt to fault-in the page if there is a chance that the
475 * pager has it, and potentially fault in additional pages
480 u_char behavior = vm_map_entry_behavior(fs.entry);
482 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
486 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
488 ahead = atop(fs.entry->end - vaddr) - 1;
489 if (ahead > VM_FAULT_READ_AHEAD_MAX)
490 ahead = VM_FAULT_READ_AHEAD_MAX;
491 if (fs.pindex == fs.entry->next_read)
492 vm_fault_cache_behind(&fs,
496 * If this is a sequential page fault, then
497 * arithmetically increase the number of pages
498 * in the read-ahead window. Otherwise, reset
499 * the read-ahead window to its smallest size.
501 behind = atop(vaddr - fs.entry->start);
502 if (behind > VM_FAULT_READ_BEHIND)
503 behind = VM_FAULT_READ_BEHIND;
504 ahead = atop(fs.entry->end - vaddr) - 1;
505 era = fs.entry->read_ahead;
506 if (fs.pindex == fs.entry->next_read) {
508 if (nera > VM_FAULT_READ_AHEAD_MAX)
509 nera = VM_FAULT_READ_AHEAD_MAX;
513 if (era == VM_FAULT_READ_AHEAD_MAX)
514 vm_fault_cache_behind(&fs,
515 VM_FAULT_CACHE_BEHIND);
516 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
517 ahead = VM_FAULT_READ_AHEAD_MIN;
519 fs.entry->read_ahead = ahead;
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)) {
646 vm_page_unlock(fs.m);
648 unlock_and_deallocate(&fs);
649 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
651 if (fs.object != fs.first_object) {
654 vm_page_unlock(fs.m);
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_COPY | VM_PROT_WRITE)) != 0) {
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) {
768 * get rid of the unnecessary page
770 vm_page_lock(fs.first_m);
771 vm_page_free(fs.first_m);
772 vm_page_unlock(fs.first_m);
774 * grab the page and put it into the
775 * process'es object. The page is
776 * automatically made dirty.
779 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
780 vm_page_unlock(fs.m);
784 PCPU_INC(cnt.v_cow_optim);
787 * Oh, well, lets copy it.
789 pmap_copy_page(fs.m, fs.first_m);
790 fs.first_m->valid = VM_PAGE_BITS_ALL;
791 if (wired && (fault_flags &
792 VM_FAULT_CHANGE_WIRING) == 0) {
793 vm_page_lock(fs.first_m);
794 vm_page_wire(fs.first_m);
795 vm_page_unlock(fs.first_m);
798 vm_page_unwire(fs.m, FALSE);
799 vm_page_unlock(fs.m);
802 * We no longer need the old page or object.
807 * fs.object != fs.first_object due to above
810 vm_object_pip_wakeup(fs.object);
811 VM_OBJECT_UNLOCK(fs.object);
813 * Only use the new page below...
815 fs.object = fs.first_object;
816 fs.pindex = fs.first_pindex;
818 if (!is_first_object_locked)
819 VM_OBJECT_LOCK(fs.object);
820 PCPU_INC(cnt.v_cow_faults);
823 prot &= ~VM_PROT_WRITE;
828 * We must verify that the maps have not changed since our last
831 if (!fs.lookup_still_valid) {
832 vm_object_t retry_object;
833 vm_pindex_t retry_pindex;
834 vm_prot_t retry_prot;
836 if (!vm_map_trylock_read(fs.map)) {
838 unlock_and_deallocate(&fs);
841 fs.lookup_still_valid = TRUE;
842 if (fs.map->timestamp != map_generation) {
843 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
844 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
847 * If we don't need the page any longer, put it on the inactive
848 * list (the easiest thing to do here). If no one needs it,
849 * pageout will grab it eventually.
851 if (result != KERN_SUCCESS) {
853 unlock_and_deallocate(&fs);
856 * If retry of map lookup would have blocked then
857 * retry fault from start.
859 if (result == KERN_FAILURE)
863 if ((retry_object != fs.first_object) ||
864 (retry_pindex != fs.first_pindex)) {
866 unlock_and_deallocate(&fs);
871 * Check whether the protection has changed or the object has
872 * been copied while we left the map unlocked. Changing from
873 * read to write permission is OK - we leave the page
874 * write-protected, and catch the write fault. Changing from
875 * write to read permission means that we can't mark the page
876 * write-enabled after all.
882 * If the page was filled by a pager, update the map entry's
883 * last read offset. Since the pager does not return the
884 * actual set of pages that it read, this update is based on
885 * the requested set. Typically, the requested and actual
888 * XXX The following assignment modifies the map
889 * without holding a write lock on it.
892 fs.entry->next_read = fs.pindex + faultcount - reqpage;
894 if ((prot & VM_PROT_WRITE) != 0 ||
895 (fault_flags & VM_FAULT_DIRTY) != 0) {
896 vm_object_set_writeable_dirty(fs.object);
899 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
900 * if the page is already dirty to prevent data written with
901 * the expectation of being synced from not being synced.
902 * Likewise if this entry does not request NOSYNC then make
903 * sure the page isn't marked NOSYNC. Applications sharing
904 * data should use the same flags to avoid ping ponging.
906 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
907 if (fs.m->dirty == 0)
908 fs.m->oflags |= VPO_NOSYNC;
910 fs.m->oflags &= ~VPO_NOSYNC;
914 * If the fault is a write, we know that this page is being
915 * written NOW so dirty it explicitly to save on
916 * pmap_is_modified() calls later.
918 * Also tell the backing pager, if any, that it should remove
919 * any swap backing since the page is now dirty.
921 if (((fault_type & VM_PROT_WRITE) != 0 &&
922 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
923 (fault_flags & VM_FAULT_DIRTY) != 0) {
925 vm_pager_page_unswapped(fs.m);
930 * Page had better still be busy
932 KASSERT(fs.m->oflags & VPO_BUSY,
933 ("vm_fault: page %p not busy!", fs.m));
935 * Page must be completely valid or it is not fit to
936 * map into user space. vm_pager_get_pages() ensures this.
938 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
939 ("vm_fault: page %p partially invalid", fs.m));
940 VM_OBJECT_UNLOCK(fs.object);
943 * Put this page into the physical map. We had to do the unlock above
944 * because pmap_enter() may sleep. We don't put the page
945 * back on the active queue until later so that the pageout daemon
946 * won't find it (yet).
948 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
949 if ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 && wired == 0)
950 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
951 VM_OBJECT_LOCK(fs.object);
955 * If the page is not wired down, then put it where the pageout daemon
958 if (fault_flags & VM_FAULT_CHANGE_WIRING) {
962 vm_page_unwire(fs.m, 1);
964 vm_page_activate(fs.m);
965 if (m_hold != NULL) {
969 vm_page_unlock(fs.m);
970 vm_page_wakeup(fs.m);
973 * Unlock everything, and return
975 unlock_and_deallocate(&fs);
977 curthread->td_ru.ru_majflt++;
979 curthread->td_ru.ru_minflt++;
981 return (KERN_SUCCESS);
985 * Speed up the reclamation of up to "distance" pages that precede the
986 * faulting pindex within the first object of the shadow chain.
989 vm_fault_cache_behind(const struct faultstate *fs, int distance)
991 vm_object_t first_object, object;
996 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
997 first_object = fs->first_object;
998 if (first_object != object) {
999 if (!VM_OBJECT_TRYLOCK(first_object)) {
1000 VM_OBJECT_UNLOCK(object);
1001 VM_OBJECT_LOCK(first_object);
1002 VM_OBJECT_LOCK(object);
1005 if (first_object->type != OBJT_DEVICE &&
1006 first_object->type != OBJT_PHYS && first_object->type != OBJT_SG) {
1007 if (fs->first_pindex < distance)
1010 pindex = fs->first_pindex - distance;
1011 if (pindex < OFF_TO_IDX(fs->entry->offset))
1012 pindex = OFF_TO_IDX(fs->entry->offset);
1013 m = first_object != object ? fs->first_m : fs->m;
1014 KASSERT((m->oflags & VPO_BUSY) != 0,
1015 ("vm_fault_cache_behind: page %p is not busy", m));
1016 m_prev = vm_page_prev(m);
1017 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1018 m->valid == VM_PAGE_BITS_ALL) {
1019 m_prev = vm_page_prev(m);
1020 if (m->busy != 0 || (m->oflags & VPO_BUSY) != 0)
1023 if (m->hold_count == 0 && m->wire_count == 0) {
1025 vm_page_aflag_clear(m, PGA_REFERENCED);
1027 vm_page_deactivate(m);
1034 if (first_object != object)
1035 VM_OBJECT_UNLOCK(first_object);
1039 * vm_fault_prefault provides a quick way of clustering
1040 * pagefaults into a processes address space. It is a "cousin"
1041 * of vm_map_pmap_enter, except it runs at page fault time instead
1045 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
1048 vm_offset_t addr, starta;
1053 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1056 object = entry->object.vm_object;
1058 starta = addra - PFBAK * PAGE_SIZE;
1059 if (starta < entry->start) {
1060 starta = entry->start;
1061 } else if (starta > addra) {
1065 for (i = 0; i < PAGEORDER_SIZE; i++) {
1066 vm_object_t backing_object, lobject;
1068 addr = addra + prefault_pageorder[i];
1069 if (addr > addra + (PFFOR * PAGE_SIZE))
1072 if (addr < starta || addr >= entry->end)
1075 if (!pmap_is_prefaultable(pmap, addr))
1078 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1080 VM_OBJECT_LOCK(lobject);
1081 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1082 lobject->type == OBJT_DEFAULT &&
1083 (backing_object = lobject->backing_object) != NULL) {
1084 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1085 0, ("vm_fault_prefault: unaligned object offset"));
1086 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1087 VM_OBJECT_LOCK(backing_object);
1088 VM_OBJECT_UNLOCK(lobject);
1089 lobject = backing_object;
1092 * give-up when a page is not in memory
1095 VM_OBJECT_UNLOCK(lobject);
1098 if (m->valid == VM_PAGE_BITS_ALL &&
1099 (m->flags & PG_FICTITIOUS) == 0)
1100 pmap_enter_quick(pmap, addr, m, entry->protection);
1101 VM_OBJECT_UNLOCK(lobject);
1106 * Hold each of the physical pages that are mapped by the specified range of
1107 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1108 * and allow the specified types of access, "prot". If all of the implied
1109 * pages are successfully held, then the number of held pages is returned
1110 * together with pointers to those pages in the array "ma". However, if any
1111 * of the pages cannot be held, -1 is returned.
1114 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1115 vm_prot_t prot, vm_page_t *ma, int max_count)
1117 vm_offset_t end, va;
1120 boolean_t pmap_failed;
1124 end = round_page(addr + len);
1125 addr = trunc_page(addr);
1128 * Check for illegal addresses.
1130 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1133 if (atop(end - addr) > max_count)
1134 panic("vm_fault_quick_hold_pages: count > max_count");
1135 count = atop(end - addr);
1138 * Most likely, the physical pages are resident in the pmap, so it is
1139 * faster to try pmap_extract_and_hold() first.
1141 pmap_failed = FALSE;
1142 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1143 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1146 else if ((prot & VM_PROT_WRITE) != 0 &&
1147 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1149 * Explicitly dirty the physical page. Otherwise, the
1150 * caller's changes may go unnoticed because they are
1151 * performed through an unmanaged mapping or by a DMA
1154 * The object lock is not held here.
1155 * See vm_page_clear_dirty_mask().
1162 * One or more pages could not be held by the pmap. Either no
1163 * page was mapped at the specified virtual address or that
1164 * mapping had insufficient permissions. Attempt to fault in
1165 * and hold these pages.
1167 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1168 if (*mp == NULL && vm_fault_hold(map, va, prot,
1169 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1174 for (mp = ma; mp < ma + count; mp++)
1177 vm_page_unhold(*mp);
1178 vm_page_unlock(*mp);
1186 * Wire down a range of virtual addresses in a map.
1189 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1190 boolean_t fictitious)
1196 * We simulate a fault to get the page and enter it in the physical
1197 * map. For user wiring, we only ask for read access on currently
1198 * read-only sections.
1200 for (va = start; va < end; va += PAGE_SIZE) {
1201 rv = vm_fault(map, va, VM_PROT_NONE, VM_FAULT_CHANGE_WIRING);
1204 vm_fault_unwire(map, start, va, fictitious);
1208 return (KERN_SUCCESS);
1214 * Unwire a range of virtual addresses in a map.
1217 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1218 boolean_t fictitious)
1225 pmap = vm_map_pmap(map);
1228 * Since the pages are wired down, we must be able to get their
1229 * mappings from the physical map system.
1231 for (va = start; va < end; va += PAGE_SIZE) {
1232 pa = pmap_extract(pmap, va);
1234 pmap_change_wiring(pmap, va, FALSE);
1236 m = PHYS_TO_VM_PAGE(pa);
1238 vm_page_unwire(m, TRUE);
1247 * vm_fault_copy_entry
1249 * Create new shadow object backing dst_entry with private copy of
1250 * all underlying pages. When src_entry is equal to dst_entry,
1251 * function implements COW for wired-down map entry. Otherwise,
1252 * it forks wired entry into dst_map.
1254 * In/out conditions:
1255 * The source and destination maps must be locked for write.
1256 * The source map entry must be wired down (or be a sharing map
1257 * entry corresponding to a main map entry that is wired down).
1260 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1261 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1262 vm_ooffset_t *fork_charge)
1264 vm_object_t backing_object, dst_object, object, src_object;
1265 vm_pindex_t dst_pindex, pindex, src_pindex;
1266 vm_prot_t access, prot;
1270 boolean_t src_readonly, upgrade;
1276 upgrade = src_entry == dst_entry;
1278 src_object = src_entry->object.vm_object;
1279 src_pindex = OFF_TO_IDX(src_entry->offset);
1280 src_readonly = (src_entry->protection & VM_PROT_WRITE) == 0;
1283 * Create the top-level object for the destination entry. (Doesn't
1284 * actually shadow anything - we copy the pages directly.)
1286 dst_object = vm_object_allocate(OBJT_DEFAULT,
1287 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1288 #if VM_NRESERVLEVEL > 0
1289 dst_object->flags |= OBJ_COLORED;
1290 dst_object->pg_color = atop(dst_entry->start);
1293 VM_OBJECT_LOCK(dst_object);
1294 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1295 ("vm_fault_copy_entry: vm_object not NULL"));
1296 dst_entry->object.vm_object = dst_object;
1297 dst_entry->offset = 0;
1298 dst_object->charge = dst_entry->end - dst_entry->start;
1299 if (fork_charge != NULL) {
1300 KASSERT(dst_entry->cred == NULL,
1301 ("vm_fault_copy_entry: leaked swp charge"));
1302 dst_object->cred = curthread->td_ucred;
1303 crhold(dst_object->cred);
1304 *fork_charge += dst_object->charge;
1306 dst_object->cred = dst_entry->cred;
1307 dst_entry->cred = NULL;
1309 access = prot = dst_entry->protection;
1311 * If not an upgrade, then enter the mappings in the pmap as
1312 * read and/or execute accesses. Otherwise, enter them as
1315 * A writeable large page mapping is only created if all of
1316 * the constituent small page mappings are modified. Marking
1317 * PTEs as modified on inception allows promotion to happen
1318 * without taking potentially large number of soft faults.
1321 access &= ~VM_PROT_WRITE;
1324 * Loop through all of the virtual pages within the entry's
1325 * range, copying each page from the source object to the
1326 * destination object. Since the source is wired, those pages
1327 * must exist. In contrast, the destination is pageable.
1328 * Since the destination object does share any backing storage
1329 * with the source object, all of its pages must be dirtied,
1330 * regardless of whether they can be written.
1332 for (vaddr = dst_entry->start, dst_pindex = 0;
1333 vaddr < dst_entry->end;
1334 vaddr += PAGE_SIZE, dst_pindex++) {
1337 * Allocate a page in the destination object.
1340 dst_m = vm_page_alloc(dst_object, dst_pindex,
1342 if (dst_m == NULL) {
1343 VM_OBJECT_UNLOCK(dst_object);
1345 VM_OBJECT_LOCK(dst_object);
1347 } while (dst_m == NULL);
1350 * Find the page in the source object, and copy it in.
1351 * (Because the source is wired down, the page will be in
1354 VM_OBJECT_LOCK(src_object);
1355 object = src_object;
1356 pindex = src_pindex + dst_pindex;
1357 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1359 (backing_object = object->backing_object) != NULL) {
1361 * Allow fallback to backing objects if we are reading.
1363 VM_OBJECT_LOCK(backing_object);
1364 pindex += OFF_TO_IDX(object->backing_object_offset);
1365 VM_OBJECT_UNLOCK(object);
1366 object = backing_object;
1369 panic("vm_fault_copy_wired: page missing");
1370 pmap_copy_page(src_m, dst_m);
1371 VM_OBJECT_UNLOCK(object);
1372 dst_m->valid = VM_PAGE_BITS_ALL;
1373 dst_m->dirty = VM_PAGE_BITS_ALL;
1374 VM_OBJECT_UNLOCK(dst_object);
1377 * Enter it in the pmap. If a wired, copy-on-write
1378 * mapping is being replaced by a write-enabled
1379 * mapping, then wire that new mapping.
1381 pmap_enter(dst_map->pmap, vaddr, access, dst_m, prot, upgrade);
1384 * Mark it no longer busy, and put it on the active list.
1386 VM_OBJECT_LOCK(dst_object);
1389 vm_page_lock(src_m);
1390 vm_page_unwire(src_m, 0);
1391 vm_page_unlock(src_m);
1393 vm_page_lock(dst_m);
1394 vm_page_wire(dst_m);
1395 vm_page_unlock(dst_m);
1397 vm_page_lock(dst_m);
1398 vm_page_activate(dst_m);
1399 vm_page_unlock(dst_m);
1401 vm_page_wakeup(dst_m);
1403 VM_OBJECT_UNLOCK(dst_object);
1405 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1406 vm_object_deallocate(src_object);
1412 * This routine checks around the requested page for other pages that
1413 * might be able to be faulted in. This routine brackets the viable
1414 * pages for the pages to be paged in.
1417 * m, rbehind, rahead
1420 * marray (array of vm_page_t), reqpage (index of requested page)
1423 * number of pages in marray
1426 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1435 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1437 int cbehind, cahead;
1439 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1443 cbehind = cahead = 0;
1446 * if the requested page is not available, then give up now
1448 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1452 if ((cbehind == 0) && (cahead == 0)) {
1458 if (rahead > cahead) {
1462 if (rbehind > cbehind) {
1467 * scan backward for the read behind pages -- in memory
1470 if (rbehind > pindex) {
1474 startpindex = pindex - rbehind;
1477 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1478 rtm->pindex >= startpindex)
1479 startpindex = rtm->pindex + 1;
1481 /* tpindex is unsigned; beware of numeric underflow. */
1482 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1483 tpindex < pindex; i++, tpindex--) {
1485 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1486 VM_ALLOC_IFNOTCACHED);
1489 * Shift the allocated pages to the
1490 * beginning of the array.
1492 for (j = 0; j < i; j++) {
1493 marray[j] = marray[j + tpindex + 1 -
1499 marray[tpindex - startpindex] = rtm;
1507 /* page offset of the required page */
1510 tpindex = pindex + 1;
1514 * scan forward for the read ahead pages
1516 endpindex = tpindex + rahead;
1517 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1518 endpindex = rtm->pindex;
1519 if (endpindex > object->size)
1520 endpindex = object->size;
1522 for (; tpindex < endpindex; i++, tpindex++) {
1524 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1525 VM_ALLOC_IFNOTCACHED);
1533 /* return number of pages */
1538 * Block entry into the machine-independent layer's page fault handler by
1539 * the calling thread. Subsequent calls to vm_fault() by that thread will
1540 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1541 * spurious page faults.
1544 vm_fault_disable_pagefaults(void)
1547 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1551 vm_fault_enable_pagefaults(int save)
1554 curthread_pflags_restore(save);