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>
85 #include <sys/resourcevar.h>
86 #include <sys/rwlock.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>
108 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
110 #define VM_FAULT_READ_BEHIND 8
111 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
112 #define VM_FAULT_NINCR (VM_FAULT_READ_MAX / VM_FAULT_READ_BEHIND)
113 #define VM_FAULT_SUM (VM_FAULT_NINCR * (VM_FAULT_NINCR + 1) / 2)
114 #define VM_FAULT_CACHE_BEHIND (VM_FAULT_READ_BEHIND * VM_FAULT_SUM)
121 vm_object_t first_object;
122 vm_pindex_t first_pindex;
124 vm_map_entry_t entry;
125 int lookup_still_valid;
129 static void vm_fault_cache_behind(const struct faultstate *fs, int distance);
130 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
131 int faultcount, int reqpage);
134 release_page(struct faultstate *fs)
137 vm_page_xunbusy(fs->m);
139 vm_page_deactivate(fs->m);
140 vm_page_unlock(fs->m);
145 unlock_map(struct faultstate *fs)
148 if (fs->lookup_still_valid) {
149 vm_map_lookup_done(fs->map, fs->entry);
150 fs->lookup_still_valid = FALSE;
155 unlock_and_deallocate(struct faultstate *fs)
158 vm_object_pip_wakeup(fs->object);
159 VM_OBJECT_WUNLOCK(fs->object);
160 if (fs->object != fs->first_object) {
161 VM_OBJECT_WLOCK(fs->first_object);
162 vm_page_lock(fs->first_m);
163 vm_page_free(fs->first_m);
164 vm_page_unlock(fs->first_m);
165 vm_object_pip_wakeup(fs->first_object);
166 VM_OBJECT_WUNLOCK(fs->first_object);
169 vm_object_deallocate(fs->first_object);
171 if (fs->vp != NULL) {
178 * TRYPAGER - used by vm_fault to calculate whether the pager for the
179 * current object *might* contain the page.
181 * default objects are zero-fill, there is no real pager.
183 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
184 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
189 * Handle a page fault occurring at the given address,
190 * requiring the given permissions, in the map specified.
191 * If successful, the page is inserted into the
192 * associated physical map.
194 * NOTE: the given address should be truncated to the
195 * proper page address.
197 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
198 * a standard error specifying why the fault is fatal is returned.
200 * The map in question must be referenced, and remains so.
201 * Caller may hold no locks.
204 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
211 if ((td->td_pflags & TDP_NOFAULTING) != 0)
212 return (KERN_PROTECTION_FAILURE);
214 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
215 ktrfault(vaddr, fault_type);
217 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
220 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
227 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
228 int fault_flags, vm_page_t *m_hold)
232 int alloc_req, era, faultcount, nera, reqpage, result;
233 boolean_t growstack, is_first_object_locked, wired;
235 vm_object_t next_object;
236 vm_page_t marray[VM_FAULT_READ_MAX];
238 struct faultstate fs;
245 PCPU_INC(cnt.v_vm_faults);
247 faultcount = reqpage = 0;
252 * Find the backing store object and offset into it to begin the
256 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
257 &fs.first_object, &fs.first_pindex, &prot, &wired);
258 if (result != KERN_SUCCESS) {
259 if (growstack && result == KERN_INVALID_ADDRESS &&
261 result = vm_map_growstack(curproc, vaddr);
262 if (result != KERN_SUCCESS)
263 return (KERN_FAILURE);
270 map_generation = fs.map->timestamp;
272 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
273 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
274 vm_map_unlock_read(fs.map);
275 return (KERN_FAILURE);
277 panic("vm_fault: fault on nofault entry, addr: %lx",
281 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
282 fs.entry->wiring_thread != curthread) {
283 vm_map_unlock_read(fs.map);
285 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
286 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
287 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
288 vm_map_unlock_and_wait(fs.map, 0);
290 vm_map_unlock(fs.map);
295 fault_type = prot | (fault_type & VM_PROT_COPY);
297 if (fs.vp == NULL /* avoid locked vnode leak */ &&
298 (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
299 /* avoid calling vm_object_set_writeable_dirty() */
300 ((prot & VM_PROT_WRITE) == 0 ||
301 fs.first_object->type != OBJT_VNODE ||
302 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
303 VM_OBJECT_RLOCK(fs.first_object);
304 if ((prot & VM_PROT_WRITE) != 0 &&
305 fs.first_object->type == OBJT_VNODE &&
306 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
308 m = vm_page_lookup(fs.first_object, fs.first_pindex);
309 /* A busy page can be mapped for read|execute access. */
310 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
311 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
313 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
314 fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
316 if (result != KERN_SUCCESS)
318 if (m_hold != NULL) {
324 if ((fault_type & VM_PROT_WRITE) != 0 &&
325 (m->oflags & VPO_UNMANAGED) == 0) {
327 vm_pager_page_unswapped(m);
329 VM_OBJECT_RUNLOCK(fs.first_object);
331 vm_fault_prefault(&fs, vaddr, 0, 0);
332 vm_map_lookup_done(fs.map, fs.entry);
333 curthread->td_ru.ru_minflt++;
334 return (KERN_SUCCESS);
336 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
337 VM_OBJECT_RUNLOCK(fs.first_object);
338 VM_OBJECT_WLOCK(fs.first_object);
341 VM_OBJECT_WLOCK(fs.first_object);
345 * Make a reference to this object to prevent its disposal while we
346 * are messing with it. Once we have the reference, the map is free
347 * to be diddled. Since objects reference their shadows (and copies),
348 * they will stay around as well.
350 * Bump the paging-in-progress count to prevent size changes (e.g.
351 * truncation operations) during I/O. This must be done after
352 * obtaining the vnode lock in order to avoid possible deadlocks.
354 vm_object_reference_locked(fs.first_object);
355 vm_object_pip_add(fs.first_object, 1);
357 fs.lookup_still_valid = TRUE;
362 * Search for the page at object/offset.
364 fs.object = fs.first_object;
365 fs.pindex = fs.first_pindex;
368 * If the object is dead, we stop here
370 if (fs.object->flags & OBJ_DEAD) {
371 unlock_and_deallocate(&fs);
372 return (KERN_PROTECTION_FAILURE);
376 * See if page is resident
378 fs.m = vm_page_lookup(fs.object, fs.pindex);
381 * Wait/Retry if the page is busy. We have to do this
382 * if the page is either exclusive or shared busy
383 * because the vm_pager may be using read busy for
384 * pageouts (and even pageins if it is the vnode
385 * pager), and we could end up trying to pagein and
386 * pageout the same page simultaneously.
388 * We can theoretically allow the busy case on a read
389 * fault if the page is marked valid, but since such
390 * pages are typically already pmap'd, putting that
391 * special case in might be more effort then it is
392 * worth. We cannot under any circumstances mess
393 * around with a shared busied page except, perhaps,
396 if (vm_page_busied(fs.m)) {
398 * Reference the page before unlocking and
399 * sleeping so that the page daemon is less
400 * likely to reclaim it.
402 vm_page_aflag_set(fs.m, PGA_REFERENCED);
403 if (fs.object != fs.first_object) {
404 if (!VM_OBJECT_TRYWLOCK(
406 VM_OBJECT_WUNLOCK(fs.object);
407 VM_OBJECT_WLOCK(fs.first_object);
408 VM_OBJECT_WLOCK(fs.object);
410 vm_page_lock(fs.first_m);
411 vm_page_free(fs.first_m);
412 vm_page_unlock(fs.first_m);
413 vm_object_pip_wakeup(fs.first_object);
414 VM_OBJECT_WUNLOCK(fs.first_object);
418 if (fs.m == vm_page_lookup(fs.object,
420 vm_page_sleep_if_busy(fs.m, "vmpfw");
422 vm_object_pip_wakeup(fs.object);
423 VM_OBJECT_WUNLOCK(fs.object);
424 PCPU_INC(cnt.v_intrans);
425 vm_object_deallocate(fs.first_object);
429 vm_page_remque(fs.m);
430 vm_page_unlock(fs.m);
433 * Mark page busy for other processes, and the
434 * pagedaemon. If it still isn't completely valid
435 * (readable), jump to readrest, else break-out ( we
439 if (fs.m->valid != VM_PAGE_BITS_ALL)
445 * Page is not resident, If this is the search termination
446 * or the pager might contain the page, allocate a new page.
448 if (TRYPAGER || fs.object == fs.first_object) {
449 if (fs.pindex >= fs.object->size) {
450 unlock_and_deallocate(&fs);
451 return (KERN_PROTECTION_FAILURE);
455 * Allocate a new page for this object/offset pair.
457 * Unlocked read of the p_flag is harmless. At
458 * worst, the P_KILLED might be not observed
459 * there, and allocation can fail, causing
460 * restart and new reading of the p_flag.
463 if (!vm_page_count_severe() || P_KILLED(curproc)) {
464 #if VM_NRESERVLEVEL > 0
465 if ((fs.object->flags & OBJ_COLORED) == 0) {
466 fs.object->flags |= OBJ_COLORED;
467 fs.object->pg_color = atop(vaddr) -
471 alloc_req = P_KILLED(curproc) ?
472 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
473 if (fs.object->type != OBJT_VNODE &&
474 fs.object->backing_object == NULL)
475 alloc_req |= VM_ALLOC_ZERO;
476 fs.m = vm_page_alloc(fs.object, fs.pindex,
480 unlock_and_deallocate(&fs);
483 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
489 * We have found a valid page or we have allocated a new page.
490 * The page thus may not be valid or may not be entirely
493 * Attempt to fault-in the page if there is a chance that the
494 * pager has it, and potentially fault in additional pages
499 u_char behavior = vm_map_entry_behavior(fs.entry);
501 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
505 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
507 ahead = atop(fs.entry->end - vaddr) - 1;
508 if (ahead > VM_FAULT_READ_AHEAD_MAX)
509 ahead = VM_FAULT_READ_AHEAD_MAX;
510 if (fs.pindex == fs.entry->next_read)
511 vm_fault_cache_behind(&fs,
515 * If this is a sequential page fault, then
516 * arithmetically increase the number of pages
517 * in the read-ahead window. Otherwise, reset
518 * the read-ahead window to its smallest size.
520 behind = atop(vaddr - fs.entry->start);
521 if (behind > VM_FAULT_READ_BEHIND)
522 behind = VM_FAULT_READ_BEHIND;
523 ahead = atop(fs.entry->end - vaddr) - 1;
524 era = fs.entry->read_ahead;
525 if (fs.pindex == fs.entry->next_read) {
527 if (nera > VM_FAULT_READ_AHEAD_MAX)
528 nera = VM_FAULT_READ_AHEAD_MAX;
532 if (era == VM_FAULT_READ_AHEAD_MAX)
533 vm_fault_cache_behind(&fs,
534 VM_FAULT_CACHE_BEHIND);
535 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
536 ahead = VM_FAULT_READ_AHEAD_MIN;
538 fs.entry->read_ahead = ahead;
542 * Call the pager to retrieve the data, if any, after
543 * releasing the lock on the map. We hold a ref on
544 * fs.object and the pages are exclusive busied.
548 if (fs.object->type == OBJT_VNODE) {
549 vp = fs.object->handle;
552 else if (fs.vp != NULL) {
556 locked = VOP_ISLOCKED(vp);
558 if (locked != LK_EXCLUSIVE)
560 /* Do not sleep for vnode lock while fs.m is busy */
561 error = vget(vp, locked | LK_CANRECURSE |
562 LK_NOWAIT, curthread);
566 unlock_and_deallocate(&fs);
567 error = vget(vp, locked | LK_RETRY |
568 LK_CANRECURSE, curthread);
572 ("vm_fault: vget failed"));
578 KASSERT(fs.vp == NULL || !fs.map->system_map,
579 ("vm_fault: vnode-backed object mapped by system map"));
582 * now we find out if any other pages should be paged
583 * in at this time this routine checks to see if the
584 * pages surrounding this fault reside in the same
585 * object as the page for this fault. If they do,
586 * then they are faulted in also into the object. The
587 * array "marray" returned contains an array of
588 * vm_page_t structs where one of them is the
589 * vm_page_t passed to the routine. The reqpage
590 * return value is the index into the marray for the
591 * vm_page_t passed to the routine.
593 * fs.m plus the additional pages are exclusive busied.
595 faultcount = vm_fault_additional_pages(
596 fs.m, behind, ahead, marray, &reqpage);
599 vm_pager_get_pages(fs.object, marray, faultcount,
600 reqpage) : VM_PAGER_FAIL;
602 if (rv == VM_PAGER_OK) {
604 * Found the page. Leave it busy while we play
609 * Relookup in case pager changed page. Pager
610 * is responsible for disposition of old page
613 fs.m = vm_page_lookup(fs.object, fs.pindex);
615 unlock_and_deallocate(&fs);
620 break; /* break to PAGE HAS BEEN FOUND */
623 * Remove the bogus page (which does not exist at this
624 * object/offset); before doing so, we must get back
625 * our object lock to preserve our invariant.
627 * Also wake up any other process that may want to bring
630 * If this is the top-level object, we must leave the
631 * busy page to prevent another process from rushing
632 * past us, and inserting the page in that object at
633 * the same time that we are.
635 if (rv == VM_PAGER_ERROR)
636 printf("vm_fault: pager read error, pid %d (%s)\n",
637 curproc->p_pid, curproc->p_comm);
639 * Data outside the range of the pager or an I/O error
642 * XXX - the check for kernel_map is a kludge to work
643 * around having the machine panic on a kernel space
644 * fault w/ I/O error.
646 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
647 (rv == VM_PAGER_BAD)) {
650 vm_page_unlock(fs.m);
652 unlock_and_deallocate(&fs);
653 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
655 if (fs.object != fs.first_object) {
658 vm_page_unlock(fs.m);
661 * XXX - we cannot just fall out at this
662 * point, m has been freed and is invalid!
668 * We get here if the object has default pager (or unwiring)
669 * or the pager doesn't have the page.
671 if (fs.object == fs.first_object)
675 * Move on to the next object. Lock the next object before
676 * unlocking the current one.
678 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
679 next_object = fs.object->backing_object;
680 if (next_object == NULL) {
682 * If there's no object left, fill the page in the top
685 if (fs.object != fs.first_object) {
686 vm_object_pip_wakeup(fs.object);
687 VM_OBJECT_WUNLOCK(fs.object);
689 fs.object = fs.first_object;
690 fs.pindex = fs.first_pindex;
692 VM_OBJECT_WLOCK(fs.object);
697 * Zero the page if necessary and mark it valid.
699 if ((fs.m->flags & PG_ZERO) == 0) {
700 pmap_zero_page(fs.m);
702 PCPU_INC(cnt.v_ozfod);
704 PCPU_INC(cnt.v_zfod);
705 fs.m->valid = VM_PAGE_BITS_ALL;
706 /* Don't try to prefault neighboring pages. */
708 break; /* break to PAGE HAS BEEN FOUND */
710 KASSERT(fs.object != next_object,
711 ("object loop %p", next_object));
712 VM_OBJECT_WLOCK(next_object);
713 vm_object_pip_add(next_object, 1);
714 if (fs.object != fs.first_object)
715 vm_object_pip_wakeup(fs.object);
716 VM_OBJECT_WUNLOCK(fs.object);
717 fs.object = next_object;
721 vm_page_assert_xbusied(fs.m);
724 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
729 * If the page is being written, but isn't already owned by the
730 * top-level object, we have to copy it into a new page owned by the
733 if (fs.object != fs.first_object) {
735 * We only really need to copy if we want to write it.
737 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
739 * This allows pages to be virtually copied from a
740 * backing_object into the first_object, where the
741 * backing object has no other refs to it, and cannot
742 * gain any more refs. Instead of a bcopy, we just
743 * move the page from the backing object to the
744 * first object. Note that we must mark the page
745 * dirty in the first object so that it will go out
746 * to swap when needed.
748 is_first_object_locked = FALSE;
751 * Only one shadow object
753 (fs.object->shadow_count == 1) &&
755 * No COW refs, except us
757 (fs.object->ref_count == 1) &&
759 * No one else can look this object up
761 (fs.object->handle == NULL) &&
763 * No other ways to look the object up
765 ((fs.object->type == OBJT_DEFAULT) ||
766 (fs.object->type == OBJT_SWAP)) &&
767 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
769 * We don't chase down the shadow chain
771 fs.object == fs.first_object->backing_object) {
773 * get rid of the unnecessary page
775 vm_page_lock(fs.first_m);
776 vm_page_free(fs.first_m);
777 vm_page_unlock(fs.first_m);
779 * grab the page and put it into the
780 * process'es object. The page is
781 * automatically made dirty.
783 if (vm_page_rename(fs.m, fs.first_object,
785 unlock_and_deallocate(&fs);
791 PCPU_INC(cnt.v_cow_optim);
794 * Oh, well, lets copy it.
796 pmap_copy_page(fs.m, fs.first_m);
797 fs.first_m->valid = VM_PAGE_BITS_ALL;
798 if (wired && (fault_flags &
799 VM_FAULT_CHANGE_WIRING) == 0) {
800 vm_page_lock(fs.first_m);
801 vm_page_wire(fs.first_m);
802 vm_page_unlock(fs.first_m);
805 vm_page_unwire(fs.m, FALSE);
806 vm_page_unlock(fs.m);
809 * We no longer need the old page or object.
814 * fs.object != fs.first_object due to above
817 vm_object_pip_wakeup(fs.object);
818 VM_OBJECT_WUNLOCK(fs.object);
820 * Only use the new page below...
822 fs.object = fs.first_object;
823 fs.pindex = fs.first_pindex;
825 if (!is_first_object_locked)
826 VM_OBJECT_WLOCK(fs.object);
827 PCPU_INC(cnt.v_cow_faults);
830 prot &= ~VM_PROT_WRITE;
835 * We must verify that the maps have not changed since our last
838 if (!fs.lookup_still_valid) {
839 vm_object_t retry_object;
840 vm_pindex_t retry_pindex;
841 vm_prot_t retry_prot;
843 if (!vm_map_trylock_read(fs.map)) {
845 unlock_and_deallocate(&fs);
848 fs.lookup_still_valid = TRUE;
849 if (fs.map->timestamp != map_generation) {
850 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
851 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
854 * If we don't need the page any longer, put it on the inactive
855 * list (the easiest thing to do here). If no one needs it,
856 * pageout will grab it eventually.
858 if (result != KERN_SUCCESS) {
860 unlock_and_deallocate(&fs);
863 * If retry of map lookup would have blocked then
864 * retry fault from start.
866 if (result == KERN_FAILURE)
870 if ((retry_object != fs.first_object) ||
871 (retry_pindex != fs.first_pindex)) {
873 unlock_and_deallocate(&fs);
878 * Check whether the protection has changed or the object has
879 * been copied while we left the map unlocked. Changing from
880 * read to write permission is OK - we leave the page
881 * write-protected, and catch the write fault. Changing from
882 * write to read permission means that we can't mark the page
883 * write-enabled after all.
889 * If the page was filled by a pager, update the map entry's
890 * last read offset. Since the pager does not return the
891 * actual set of pages that it read, this update is based on
892 * the requested set. Typically, the requested and actual
895 * XXX The following assignment modifies the map
896 * without holding a write lock on it.
899 fs.entry->next_read = fs.pindex + faultcount - reqpage;
901 if (((prot & VM_PROT_WRITE) != 0 ||
902 (fault_flags & VM_FAULT_DIRTY) != 0) &&
903 (fs.m->oflags & VPO_UNMANAGED) == 0) {
904 vm_object_set_writeable_dirty(fs.object);
907 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
908 * if the page is already dirty to prevent data written with
909 * the expectation of being synced from not being synced.
910 * Likewise if this entry does not request NOSYNC then make
911 * sure the page isn't marked NOSYNC. Applications sharing
912 * data should use the same flags to avoid ping ponging.
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;
922 * If the fault is a write, we know that this page is being
923 * written NOW so dirty it explicitly to save on
924 * pmap_is_modified() calls later.
926 * Also tell the backing pager, if any, that it should remove
927 * any swap backing since the page is now dirty.
929 if (((fault_type & VM_PROT_WRITE) != 0 &&
930 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
931 (fault_flags & VM_FAULT_DIRTY) != 0) {
933 vm_pager_page_unswapped(fs.m);
937 vm_page_assert_xbusied(fs.m);
940 * Page must be completely valid or it is not fit to
941 * map into user space. vm_pager_get_pages() ensures this.
943 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
944 ("vm_fault: page %p partially invalid", fs.m));
945 VM_OBJECT_WUNLOCK(fs.object);
948 * Put this page into the physical map. We had to do the unlock above
949 * because pmap_enter() may sleep. We don't put the page
950 * back on the active queue until later so that the pageout daemon
951 * won't find it (yet).
953 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
954 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
955 if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
957 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
958 VM_OBJECT_WLOCK(fs.object);
962 * If the page is not wired down, then put it where the pageout daemon
965 if (fault_flags & VM_FAULT_CHANGE_WIRING) {
969 vm_page_unwire(fs.m, 1);
971 vm_page_activate(fs.m);
972 if (m_hold != NULL) {
976 vm_page_unlock(fs.m);
977 vm_page_xunbusy(fs.m);
980 * Unlock everything, and return
982 unlock_and_deallocate(&fs);
984 PCPU_INC(cnt.v_io_faults);
985 curthread->td_ru.ru_majflt++;
987 curthread->td_ru.ru_minflt++;
989 return (KERN_SUCCESS);
993 * Speed up the reclamation of up to "distance" pages that precede the
994 * faulting pindex within the first object of the shadow chain.
997 vm_fault_cache_behind(const struct faultstate *fs, int distance)
999 vm_object_t first_object, object;
1000 vm_page_t m, m_prev;
1003 object = fs->object;
1004 VM_OBJECT_ASSERT_WLOCKED(object);
1005 first_object = fs->first_object;
1006 if (first_object != object) {
1007 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1008 VM_OBJECT_WUNLOCK(object);
1009 VM_OBJECT_WLOCK(first_object);
1010 VM_OBJECT_WLOCK(object);
1013 /* Neither fictitious nor unmanaged pages can be cached. */
1014 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1015 if (fs->first_pindex < distance)
1018 pindex = fs->first_pindex - distance;
1019 if (pindex < OFF_TO_IDX(fs->entry->offset))
1020 pindex = OFF_TO_IDX(fs->entry->offset);
1021 m = first_object != object ? fs->first_m : fs->m;
1022 vm_page_assert_xbusied(m);
1023 m_prev = vm_page_prev(m);
1024 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1025 m->valid == VM_PAGE_BITS_ALL) {
1026 m_prev = vm_page_prev(m);
1027 if (vm_page_busied(m))
1030 if (m->hold_count == 0 && m->wire_count == 0) {
1032 vm_page_aflag_clear(m, PGA_REFERENCED);
1034 vm_page_deactivate(m);
1041 if (first_object != object)
1042 VM_OBJECT_WUNLOCK(first_object);
1046 * vm_fault_prefault provides a quick way of clustering
1047 * pagefaults into a processes address space. It is a "cousin"
1048 * of vm_map_pmap_enter, except it runs at page fault time instead
1052 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1053 int faultcount, int reqpage)
1056 vm_map_entry_t entry;
1057 vm_object_t backing_object, lobject;
1058 vm_offset_t addr, starta;
1061 int backward, forward, i;
1063 pmap = fs->map->pmap;
1064 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1067 if (faultcount > 0) {
1069 forward = faultcount - reqpage - 1;
1076 starta = addra - backward * PAGE_SIZE;
1077 if (starta < entry->start) {
1078 starta = entry->start;
1079 } else if (starta > addra) {
1084 * Generate the sequence of virtual addresses that are candidates for
1085 * prefaulting in an outward spiral from the faulting virtual address,
1086 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1087 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1088 * If the candidate address doesn't have a backing physical page, then
1089 * the loop immediately terminates.
1091 for (i = 0; i < 2 * imax(backward, forward); i++) {
1092 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1094 if (addr > addra + forward * PAGE_SIZE)
1097 if (addr < starta || addr >= entry->end)
1100 if (!pmap_is_prefaultable(pmap, addr))
1103 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1104 lobject = entry->object.vm_object;
1105 VM_OBJECT_RLOCK(lobject);
1106 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1107 lobject->type == OBJT_DEFAULT &&
1108 (backing_object = lobject->backing_object) != NULL) {
1109 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1110 0, ("vm_fault_prefault: unaligned object offset"));
1111 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1112 VM_OBJECT_RLOCK(backing_object);
1113 VM_OBJECT_RUNLOCK(lobject);
1114 lobject = backing_object;
1117 VM_OBJECT_RUNLOCK(lobject);
1120 if (m->valid == VM_PAGE_BITS_ALL &&
1121 (m->flags & PG_FICTITIOUS) == 0)
1122 pmap_enter_quick(pmap, addr, m, entry->protection);
1123 VM_OBJECT_RUNLOCK(lobject);
1128 * Hold each of the physical pages that are mapped by the specified range of
1129 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1130 * and allow the specified types of access, "prot". If all of the implied
1131 * pages are successfully held, then the number of held pages is returned
1132 * together with pointers to those pages in the array "ma". However, if any
1133 * of the pages cannot be held, -1 is returned.
1136 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1137 vm_prot_t prot, vm_page_t *ma, int max_count)
1139 vm_offset_t end, va;
1142 boolean_t pmap_failed;
1146 end = round_page(addr + len);
1147 addr = trunc_page(addr);
1150 * Check for illegal addresses.
1152 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1155 if (atop(end - addr) > max_count)
1156 panic("vm_fault_quick_hold_pages: count > max_count");
1157 count = atop(end - addr);
1160 * Most likely, the physical pages are resident in the pmap, so it is
1161 * faster to try pmap_extract_and_hold() first.
1163 pmap_failed = FALSE;
1164 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1165 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1168 else if ((prot & VM_PROT_WRITE) != 0 &&
1169 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1171 * Explicitly dirty the physical page. Otherwise, the
1172 * caller's changes may go unnoticed because they are
1173 * performed through an unmanaged mapping or by a DMA
1176 * The object lock is not held here.
1177 * See vm_page_clear_dirty_mask().
1184 * One or more pages could not be held by the pmap. Either no
1185 * page was mapped at the specified virtual address or that
1186 * mapping had insufficient permissions. Attempt to fault in
1187 * and hold these pages.
1189 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1190 if (*mp == NULL && vm_fault_hold(map, va, prot,
1191 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1196 for (mp = ma; mp < ma + count; mp++)
1199 vm_page_unhold(*mp);
1200 vm_page_unlock(*mp);
1207 * vm_fault_copy_entry
1209 * Create new shadow object backing dst_entry with private copy of
1210 * all underlying pages. When src_entry is equal to dst_entry,
1211 * function implements COW for wired-down map entry. Otherwise,
1212 * it forks wired entry into dst_map.
1214 * In/out conditions:
1215 * The source and destination maps must be locked for write.
1216 * The source map entry must be wired down (or be a sharing map
1217 * entry corresponding to a main map entry that is wired down).
1220 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1221 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1222 vm_ooffset_t *fork_charge)
1224 vm_object_t backing_object, dst_object, object, src_object;
1225 vm_pindex_t dst_pindex, pindex, src_pindex;
1226 vm_prot_t access, prot;
1236 upgrade = src_entry == dst_entry;
1237 access = prot = dst_entry->protection;
1239 src_object = src_entry->object.vm_object;
1240 src_pindex = OFF_TO_IDX(src_entry->offset);
1242 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1243 dst_object = src_object;
1244 vm_object_reference(dst_object);
1247 * Create the top-level object for the destination entry. (Doesn't
1248 * actually shadow anything - we copy the pages directly.)
1250 dst_object = vm_object_allocate(OBJT_DEFAULT,
1251 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1252 #if VM_NRESERVLEVEL > 0
1253 dst_object->flags |= OBJ_COLORED;
1254 dst_object->pg_color = atop(dst_entry->start);
1258 VM_OBJECT_WLOCK(dst_object);
1259 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1260 ("vm_fault_copy_entry: vm_object not NULL"));
1261 if (src_object != dst_object) {
1262 dst_entry->object.vm_object = dst_object;
1263 dst_entry->offset = 0;
1264 dst_object->charge = dst_entry->end - dst_entry->start;
1266 if (fork_charge != NULL) {
1267 KASSERT(dst_entry->cred == NULL,
1268 ("vm_fault_copy_entry: leaked swp charge"));
1269 dst_object->cred = curthread->td_ucred;
1270 crhold(dst_object->cred);
1271 *fork_charge += dst_object->charge;
1272 } else if (dst_object->cred == NULL) {
1273 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1275 dst_object->cred = dst_entry->cred;
1276 dst_entry->cred = NULL;
1280 * If not an upgrade, then enter the mappings in the pmap as
1281 * read and/or execute accesses. Otherwise, enter them as
1284 * A writeable large page mapping is only created if all of
1285 * the constituent small page mappings are modified. Marking
1286 * PTEs as modified on inception allows promotion to happen
1287 * without taking potentially large number of soft faults.
1290 access &= ~VM_PROT_WRITE;
1293 * Loop through all of the virtual pages within the entry's
1294 * range, copying each page from the source object to the
1295 * destination object. Since the source is wired, those pages
1296 * must exist. In contrast, the destination is pageable.
1297 * Since the destination object does share any backing storage
1298 * with the source object, all of its pages must be dirtied,
1299 * regardless of whether they can be written.
1301 for (vaddr = dst_entry->start, dst_pindex = 0;
1302 vaddr < dst_entry->end;
1303 vaddr += PAGE_SIZE, dst_pindex++) {
1306 * Find the page in the source object, and copy it in.
1307 * Because the source is wired down, the page will be
1310 if (src_object != dst_object)
1311 VM_OBJECT_RLOCK(src_object);
1312 object = src_object;
1313 pindex = src_pindex + dst_pindex;
1314 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1315 (backing_object = object->backing_object) != NULL) {
1317 * Unless the source mapping is read-only or
1318 * it is presently being upgraded from
1319 * read-only, the first object in the shadow
1320 * chain should provide all of the pages. In
1321 * other words, this loop body should never be
1322 * executed when the source mapping is already
1325 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1327 ("vm_fault_copy_entry: main object missing page"));
1329 VM_OBJECT_RLOCK(backing_object);
1330 pindex += OFF_TO_IDX(object->backing_object_offset);
1331 if (object != dst_object)
1332 VM_OBJECT_RUNLOCK(object);
1333 object = backing_object;
1335 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1337 if (object != dst_object) {
1339 * Allocate a page in the destination object.
1341 dst_m = vm_page_alloc(dst_object, (src_object ==
1342 dst_object ? src_pindex : 0) + dst_pindex,
1344 if (dst_m == NULL) {
1345 VM_OBJECT_WUNLOCK(dst_object);
1346 VM_OBJECT_RUNLOCK(object);
1348 VM_OBJECT_WLOCK(dst_object);
1351 pmap_copy_page(src_m, dst_m);
1352 VM_OBJECT_RUNLOCK(object);
1353 dst_m->valid = VM_PAGE_BITS_ALL;
1354 dst_m->dirty = VM_PAGE_BITS_ALL;
1357 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1359 vm_page_xbusy(dst_m);
1360 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1361 ("invalid dst page %p", dst_m));
1363 VM_OBJECT_WUNLOCK(dst_object);
1366 * Enter it in the pmap. If a wired, copy-on-write
1367 * mapping is being replaced by a write-enabled
1368 * mapping, then wire that new mapping.
1370 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1371 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1374 * Mark it no longer busy, and put it on the active list.
1376 VM_OBJECT_WLOCK(dst_object);
1379 if (src_m != dst_m) {
1380 vm_page_lock(src_m);
1381 vm_page_unwire(src_m, 0);
1382 vm_page_unlock(src_m);
1383 vm_page_lock(dst_m);
1384 vm_page_wire(dst_m);
1385 vm_page_unlock(dst_m);
1387 KASSERT(dst_m->wire_count > 0,
1388 ("dst_m %p is not wired", dst_m));
1391 vm_page_lock(dst_m);
1392 vm_page_activate(dst_m);
1393 vm_page_unlock(dst_m);
1395 vm_page_xunbusy(dst_m);
1397 VM_OBJECT_WUNLOCK(dst_object);
1399 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1400 vm_object_deallocate(src_object);
1406 * This routine checks around the requested page for other pages that
1407 * might be able to be faulted in. This routine brackets the viable
1408 * pages for the pages to be paged in.
1411 * m, rbehind, rahead
1414 * marray (array of vm_page_t), reqpage (index of requested page)
1417 * number of pages in marray
1420 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1429 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1431 int cbehind, cahead;
1433 VM_OBJECT_ASSERT_WLOCKED(m->object);
1437 cbehind = cahead = 0;
1440 * if the requested page is not available, then give up now
1442 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1446 if ((cbehind == 0) && (cahead == 0)) {
1452 if (rahead > cahead) {
1456 if (rbehind > cbehind) {
1461 * scan backward for the read behind pages -- in memory
1464 if (rbehind > pindex) {
1468 startpindex = pindex - rbehind;
1471 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1472 rtm->pindex >= startpindex)
1473 startpindex = rtm->pindex + 1;
1475 /* tpindex is unsigned; beware of numeric underflow. */
1476 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1477 tpindex < pindex; i++, tpindex--) {
1479 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1480 VM_ALLOC_IFNOTCACHED);
1483 * Shift the allocated pages to the
1484 * beginning of the array.
1486 for (j = 0; j < i; j++) {
1487 marray[j] = marray[j + tpindex + 1 -
1493 marray[tpindex - startpindex] = rtm;
1501 /* page offset of the required page */
1504 tpindex = pindex + 1;
1508 * scan forward for the read ahead pages
1510 endpindex = tpindex + rahead;
1511 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1512 endpindex = rtm->pindex;
1513 if (endpindex > object->size)
1514 endpindex = object->size;
1516 for (; tpindex < endpindex; i++, tpindex++) {
1518 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1519 VM_ALLOC_IFNOTCACHED);
1527 /* return number of pages */
1532 * Block entry into the machine-independent layer's page fault handler by
1533 * the calling thread. Subsequent calls to vm_fault() by that thread will
1534 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1535 * spurious page faults.
1538 vm_fault_disable_pagefaults(void)
1541 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1545 vm_fault_enable_pagefaults(int save)
1548 curthread_pflags_restore(save);