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 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
179 vm_prot_t fault_type, int fault_flags, boolean_t set_wd)
181 boolean_t need_dirty;
183 if (((prot & VM_PROT_WRITE) == 0 &&
184 (fault_flags & VM_FAULT_DIRTY) == 0) ||
185 (m->oflags & VPO_UNMANAGED) != 0)
188 VM_OBJECT_ASSERT_LOCKED(m->object);
190 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
191 (fault_flags & VM_FAULT_CHANGE_WIRING) == 0) ||
192 (fault_flags & VM_FAULT_DIRTY) != 0;
195 vm_object_set_writeable_dirty(m->object);
198 * If two callers of vm_fault_dirty() with set_wd ==
199 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
200 * flag set, other with flag clear, race, it is
201 * possible for the no-NOSYNC thread to see m->dirty
202 * != 0 and not clear VPO_NOSYNC. Take vm_page lock
203 * around manipulation of VPO_NOSYNC and
204 * vm_page_dirty() call, to avoid the race and keep
205 * m->oflags consistent.
210 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
211 * if the page is already dirty to prevent data written with
212 * the expectation of being synced from not being synced.
213 * Likewise if this entry does not request NOSYNC then make
214 * sure the page isn't marked NOSYNC. Applications sharing
215 * data should use the same flags to avoid ping ponging.
217 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
219 m->oflags |= VPO_NOSYNC;
222 m->oflags &= ~VPO_NOSYNC;
226 * If the fault is a write, we know that this page is being
227 * written NOW so dirty it explicitly to save on
228 * pmap_is_modified() calls later.
230 * Also tell the backing pager, if any, that it should remove
231 * any swap backing since the page is now dirty.
238 vm_pager_page_unswapped(m);
242 * TRYPAGER - used by vm_fault to calculate whether the pager for the
243 * current object *might* contain the page.
245 * default objects are zero-fill, there is no real pager.
247 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
248 ((fault_flags & VM_FAULT_CHANGE_WIRING) == 0 || wired))
253 * Handle a page fault occurring at the given address,
254 * requiring the given permissions, in the map specified.
255 * If successful, the page is inserted into the
256 * associated physical map.
258 * NOTE: the given address should be truncated to the
259 * proper page address.
261 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
262 * a standard error specifying why the fault is fatal is returned.
264 * The map in question must be referenced, and remains so.
265 * Caller may hold no locks.
268 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
275 if ((td->td_pflags & TDP_NOFAULTING) != 0)
276 return (KERN_PROTECTION_FAILURE);
278 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
279 ktrfault(vaddr, fault_type);
281 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
284 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
291 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
292 int fault_flags, vm_page_t *m_hold)
296 int alloc_req, era, faultcount, nera, reqpage, result;
297 boolean_t growstack, is_first_object_locked, wired;
299 vm_object_t next_object;
300 vm_page_t marray[VM_FAULT_READ_MAX];
302 struct faultstate fs;
309 PCPU_INC(cnt.v_vm_faults);
311 faultcount = reqpage = 0;
316 * Find the backing store object and offset into it to begin the
320 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
321 &fs.first_object, &fs.first_pindex, &prot, &wired);
322 if (result != KERN_SUCCESS) {
323 if (growstack && result == KERN_INVALID_ADDRESS &&
325 result = vm_map_growstack(curproc, vaddr);
326 if (result != KERN_SUCCESS)
327 return (KERN_FAILURE);
334 map_generation = fs.map->timestamp;
336 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
337 if ((curthread->td_pflags & TDP_DEVMEMIO) != 0) {
338 vm_map_unlock_read(fs.map);
339 return (KERN_FAILURE);
341 panic("vm_fault: fault on nofault entry, addr: %lx",
345 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
346 fs.entry->wiring_thread != curthread) {
347 vm_map_unlock_read(fs.map);
349 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
350 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
351 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
352 vm_map_unlock_and_wait(fs.map, 0);
354 vm_map_unlock(fs.map);
359 fault_type = prot | (fault_type & VM_PROT_COPY);
361 if (fs.vp == NULL /* avoid locked vnode leak */ &&
362 (fault_flags & (VM_FAULT_CHANGE_WIRING | VM_FAULT_DIRTY)) == 0 &&
363 /* avoid calling vm_object_set_writeable_dirty() */
364 ((prot & VM_PROT_WRITE) == 0 ||
365 fs.first_object->type != OBJT_VNODE ||
366 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
367 VM_OBJECT_RLOCK(fs.first_object);
368 if ((prot & VM_PROT_WRITE) != 0 &&
369 fs.first_object->type == OBJT_VNODE &&
370 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) == 0)
372 m = vm_page_lookup(fs.first_object, fs.first_pindex);
373 /* A busy page can be mapped for read|execute access. */
374 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
375 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
377 result = pmap_enter(fs.map->pmap, vaddr, m, prot,
378 fault_type | PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED :
380 if (result != KERN_SUCCESS)
382 if (m_hold != NULL) {
388 vm_fault_dirty(fs.entry, m, prot, fault_type, fault_flags,
390 VM_OBJECT_RUNLOCK(fs.first_object);
392 vm_fault_prefault(&fs, vaddr, 0, 0);
393 vm_map_lookup_done(fs.map, fs.entry);
394 curthread->td_ru.ru_minflt++;
395 return (KERN_SUCCESS);
397 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
398 VM_OBJECT_RUNLOCK(fs.first_object);
399 VM_OBJECT_WLOCK(fs.first_object);
402 VM_OBJECT_WLOCK(fs.first_object);
406 * Make a reference to this object to prevent its disposal while we
407 * are messing with it. Once we have the reference, the map is free
408 * to be diddled. Since objects reference their shadows (and copies),
409 * they will stay around as well.
411 * Bump the paging-in-progress count to prevent size changes (e.g.
412 * truncation operations) during I/O. This must be done after
413 * obtaining the vnode lock in order to avoid possible deadlocks.
415 vm_object_reference_locked(fs.first_object);
416 vm_object_pip_add(fs.first_object, 1);
418 fs.lookup_still_valid = TRUE;
423 * Search for the page at object/offset.
425 fs.object = fs.first_object;
426 fs.pindex = fs.first_pindex;
429 * If the object is dead, we stop here
431 if (fs.object->flags & OBJ_DEAD) {
432 unlock_and_deallocate(&fs);
433 return (KERN_PROTECTION_FAILURE);
437 * See if page is resident
439 fs.m = vm_page_lookup(fs.object, fs.pindex);
442 * Wait/Retry if the page is busy. We have to do this
443 * if the page is either exclusive or shared busy
444 * because the vm_pager may be using read busy for
445 * pageouts (and even pageins if it is the vnode
446 * pager), and we could end up trying to pagein and
447 * pageout the same page simultaneously.
449 * We can theoretically allow the busy case on a read
450 * fault if the page is marked valid, but since such
451 * pages are typically already pmap'd, putting that
452 * special case in might be more effort then it is
453 * worth. We cannot under any circumstances mess
454 * around with a shared busied page except, perhaps,
457 if (vm_page_busied(fs.m)) {
459 * Reference the page before unlocking and
460 * sleeping so that the page daemon is less
461 * likely to reclaim it.
463 vm_page_aflag_set(fs.m, PGA_REFERENCED);
464 if (fs.object != fs.first_object) {
465 if (!VM_OBJECT_TRYWLOCK(
467 VM_OBJECT_WUNLOCK(fs.object);
468 VM_OBJECT_WLOCK(fs.first_object);
469 VM_OBJECT_WLOCK(fs.object);
471 vm_page_lock(fs.first_m);
472 vm_page_free(fs.first_m);
473 vm_page_unlock(fs.first_m);
474 vm_object_pip_wakeup(fs.first_object);
475 VM_OBJECT_WUNLOCK(fs.first_object);
479 if (fs.m == vm_page_lookup(fs.object,
481 vm_page_sleep_if_busy(fs.m, "vmpfw");
483 vm_object_pip_wakeup(fs.object);
484 VM_OBJECT_WUNLOCK(fs.object);
485 PCPU_INC(cnt.v_intrans);
486 vm_object_deallocate(fs.first_object);
490 vm_page_remque(fs.m);
491 vm_page_unlock(fs.m);
494 * Mark page busy for other processes, and the
495 * pagedaemon. If it still isn't completely valid
496 * (readable), jump to readrest, else break-out ( we
500 if (fs.m->valid != VM_PAGE_BITS_ALL)
506 * Page is not resident, If this is the search termination
507 * or the pager might contain the page, allocate a new page.
509 if (TRYPAGER || fs.object == fs.first_object) {
510 if (fs.pindex >= fs.object->size) {
511 unlock_and_deallocate(&fs);
512 return (KERN_PROTECTION_FAILURE);
516 * Allocate a new page for this object/offset pair.
518 * Unlocked read of the p_flag is harmless. At
519 * worst, the P_KILLED might be not observed
520 * there, and allocation can fail, causing
521 * restart and new reading of the p_flag.
524 if (!vm_page_count_severe() || P_KILLED(curproc)) {
525 #if VM_NRESERVLEVEL > 0
526 if ((fs.object->flags & OBJ_COLORED) == 0) {
527 fs.object->flags |= OBJ_COLORED;
528 fs.object->pg_color = atop(vaddr) -
532 alloc_req = P_KILLED(curproc) ?
533 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
534 if (fs.object->type != OBJT_VNODE &&
535 fs.object->backing_object == NULL)
536 alloc_req |= VM_ALLOC_ZERO;
537 fs.m = vm_page_alloc(fs.object, fs.pindex,
541 unlock_and_deallocate(&fs);
544 } else if (fs.m->valid == VM_PAGE_BITS_ALL)
550 * We have found a valid page or we have allocated a new page.
551 * The page thus may not be valid or may not be entirely
554 * Attempt to fault-in the page if there is a chance that the
555 * pager has it, and potentially fault in additional pages
560 u_char behavior = vm_map_entry_behavior(fs.entry);
562 if (behavior == MAP_ENTRY_BEHAV_RANDOM ||
566 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
568 ahead = atop(fs.entry->end - vaddr) - 1;
569 if (ahead > VM_FAULT_READ_AHEAD_MAX)
570 ahead = VM_FAULT_READ_AHEAD_MAX;
571 if (fs.pindex == fs.entry->next_read)
572 vm_fault_cache_behind(&fs,
576 * If this is a sequential page fault, then
577 * arithmetically increase the number of pages
578 * in the read-ahead window. Otherwise, reset
579 * the read-ahead window to its smallest size.
581 behind = atop(vaddr - fs.entry->start);
582 if (behind > VM_FAULT_READ_BEHIND)
583 behind = VM_FAULT_READ_BEHIND;
584 ahead = atop(fs.entry->end - vaddr) - 1;
585 era = fs.entry->read_ahead;
586 if (fs.pindex == fs.entry->next_read) {
588 if (nera > VM_FAULT_READ_AHEAD_MAX)
589 nera = VM_FAULT_READ_AHEAD_MAX;
593 if (era == VM_FAULT_READ_AHEAD_MAX)
594 vm_fault_cache_behind(&fs,
595 VM_FAULT_CACHE_BEHIND);
596 } else if (ahead > VM_FAULT_READ_AHEAD_MIN)
597 ahead = VM_FAULT_READ_AHEAD_MIN;
599 fs.entry->read_ahead = ahead;
603 * Call the pager to retrieve the data, if any, after
604 * releasing the lock on the map. We hold a ref on
605 * fs.object and the pages are exclusive busied.
609 if (fs.object->type == OBJT_VNODE) {
610 vp = fs.object->handle;
613 else if (fs.vp != NULL) {
617 locked = VOP_ISLOCKED(vp);
619 if (locked != LK_EXCLUSIVE)
621 /* Do not sleep for vnode lock while fs.m is busy */
622 error = vget(vp, locked | LK_CANRECURSE |
623 LK_NOWAIT, curthread);
627 unlock_and_deallocate(&fs);
628 error = vget(vp, locked | LK_RETRY |
629 LK_CANRECURSE, curthread);
633 ("vm_fault: vget failed"));
639 KASSERT(fs.vp == NULL || !fs.map->system_map,
640 ("vm_fault: vnode-backed object mapped by system map"));
643 * now we find out if any other pages should be paged
644 * in at this time this routine checks to see if the
645 * pages surrounding this fault reside in the same
646 * object as the page for this fault. If they do,
647 * then they are faulted in also into the object. The
648 * array "marray" returned contains an array of
649 * vm_page_t structs where one of them is the
650 * vm_page_t passed to the routine. The reqpage
651 * return value is the index into the marray for the
652 * vm_page_t passed to the routine.
654 * fs.m plus the additional pages are exclusive busied.
656 faultcount = vm_fault_additional_pages(
657 fs.m, behind, ahead, marray, &reqpage);
660 vm_pager_get_pages(fs.object, marray, faultcount,
661 reqpage) : VM_PAGER_FAIL;
663 if (rv == VM_PAGER_OK) {
665 * Found the page. Leave it busy while we play
670 * Relookup in case pager changed page. Pager
671 * is responsible for disposition of old page
674 fs.m = vm_page_lookup(fs.object, fs.pindex);
676 unlock_and_deallocate(&fs);
681 break; /* break to PAGE HAS BEEN FOUND */
684 * Remove the bogus page (which does not exist at this
685 * object/offset); before doing so, we must get back
686 * our object lock to preserve our invariant.
688 * Also wake up any other process that may want to bring
691 * If this is the top-level object, we must leave the
692 * busy page to prevent another process from rushing
693 * past us, and inserting the page in that object at
694 * the same time that we are.
696 if (rv == VM_PAGER_ERROR)
697 printf("vm_fault: pager read error, pid %d (%s)\n",
698 curproc->p_pid, curproc->p_comm);
700 * Data outside the range of the pager or an I/O error
703 * XXX - the check for kernel_map is a kludge to work
704 * around having the machine panic on a kernel space
705 * fault w/ I/O error.
707 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
708 (rv == VM_PAGER_BAD)) {
711 vm_page_unlock(fs.m);
713 unlock_and_deallocate(&fs);
714 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
716 if (fs.object != fs.first_object) {
719 vm_page_unlock(fs.m);
722 * XXX - we cannot just fall out at this
723 * point, m has been freed and is invalid!
729 * We get here if the object has default pager (or unwiring)
730 * or the pager doesn't have the page.
732 if (fs.object == fs.first_object)
736 * Move on to the next object. Lock the next object before
737 * unlocking the current one.
739 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
740 next_object = fs.object->backing_object;
741 if (next_object == NULL) {
743 * If there's no object left, fill the page in the top
746 if (fs.object != fs.first_object) {
747 vm_object_pip_wakeup(fs.object);
748 VM_OBJECT_WUNLOCK(fs.object);
750 fs.object = fs.first_object;
751 fs.pindex = fs.first_pindex;
753 VM_OBJECT_WLOCK(fs.object);
758 * Zero the page if necessary and mark it valid.
760 if ((fs.m->flags & PG_ZERO) == 0) {
761 pmap_zero_page(fs.m);
763 PCPU_INC(cnt.v_ozfod);
765 PCPU_INC(cnt.v_zfod);
766 fs.m->valid = VM_PAGE_BITS_ALL;
767 /* Don't try to prefault neighboring pages. */
769 break; /* break to PAGE HAS BEEN FOUND */
771 KASSERT(fs.object != next_object,
772 ("object loop %p", next_object));
773 VM_OBJECT_WLOCK(next_object);
774 vm_object_pip_add(next_object, 1);
775 if (fs.object != fs.first_object)
776 vm_object_pip_wakeup(fs.object);
777 VM_OBJECT_WUNLOCK(fs.object);
778 fs.object = next_object;
782 vm_page_assert_xbusied(fs.m);
785 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
790 * If the page is being written, but isn't already owned by the
791 * top-level object, we have to copy it into a new page owned by the
794 if (fs.object != fs.first_object) {
796 * We only really need to copy if we want to write it.
798 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
800 * This allows pages to be virtually copied from a
801 * backing_object into the first_object, where the
802 * backing object has no other refs to it, and cannot
803 * gain any more refs. Instead of a bcopy, we just
804 * move the page from the backing object to the
805 * first object. Note that we must mark the page
806 * dirty in the first object so that it will go out
807 * to swap when needed.
809 is_first_object_locked = FALSE;
812 * Only one shadow object
814 (fs.object->shadow_count == 1) &&
816 * No COW refs, except us
818 (fs.object->ref_count == 1) &&
820 * No one else can look this object up
822 (fs.object->handle == NULL) &&
824 * No other ways to look the object up
826 ((fs.object->type == OBJT_DEFAULT) ||
827 (fs.object->type == OBJT_SWAP)) &&
828 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
830 * We don't chase down the shadow chain
832 fs.object == fs.first_object->backing_object) {
834 * get rid of the unnecessary page
836 vm_page_lock(fs.first_m);
837 vm_page_free(fs.first_m);
838 vm_page_unlock(fs.first_m);
840 * grab the page and put it into the
841 * process'es object. The page is
842 * automatically made dirty.
844 if (vm_page_rename(fs.m, fs.first_object,
846 unlock_and_deallocate(&fs);
852 PCPU_INC(cnt.v_cow_optim);
855 * Oh, well, lets copy it.
857 pmap_copy_page(fs.m, fs.first_m);
858 fs.first_m->valid = VM_PAGE_BITS_ALL;
859 if (wired && (fault_flags &
860 VM_FAULT_CHANGE_WIRING) == 0) {
861 vm_page_lock(fs.first_m);
862 vm_page_wire(fs.first_m);
863 vm_page_unlock(fs.first_m);
866 vm_page_unwire(fs.m, PQ_INACTIVE);
867 vm_page_unlock(fs.m);
870 * We no longer need the old page or object.
875 * fs.object != fs.first_object due to above
878 vm_object_pip_wakeup(fs.object);
879 VM_OBJECT_WUNLOCK(fs.object);
881 * Only use the new page below...
883 fs.object = fs.first_object;
884 fs.pindex = fs.first_pindex;
886 if (!is_first_object_locked)
887 VM_OBJECT_WLOCK(fs.object);
888 PCPU_INC(cnt.v_cow_faults);
891 prot &= ~VM_PROT_WRITE;
896 * We must verify that the maps have not changed since our last
899 if (!fs.lookup_still_valid) {
900 vm_object_t retry_object;
901 vm_pindex_t retry_pindex;
902 vm_prot_t retry_prot;
904 if (!vm_map_trylock_read(fs.map)) {
906 unlock_and_deallocate(&fs);
909 fs.lookup_still_valid = TRUE;
910 if (fs.map->timestamp != map_generation) {
911 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
912 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
915 * If we don't need the page any longer, put it on the inactive
916 * list (the easiest thing to do here). If no one needs it,
917 * pageout will grab it eventually.
919 if (result != KERN_SUCCESS) {
921 unlock_and_deallocate(&fs);
924 * If retry of map lookup would have blocked then
925 * retry fault from start.
927 if (result == KERN_FAILURE)
931 if ((retry_object != fs.first_object) ||
932 (retry_pindex != fs.first_pindex)) {
934 unlock_and_deallocate(&fs);
939 * Check whether the protection has changed or the object has
940 * been copied while we left the map unlocked. Changing from
941 * read to write permission is OK - we leave the page
942 * write-protected, and catch the write fault. Changing from
943 * write to read permission means that we can't mark the page
944 * write-enabled after all.
950 * If the page was filled by a pager, update the map entry's
951 * last read offset. Since the pager does not return the
952 * actual set of pages that it read, this update is based on
953 * the requested set. Typically, the requested and actual
956 * XXX The following assignment modifies the map
957 * without holding a write lock on it.
960 fs.entry->next_read = fs.pindex + faultcount - reqpage;
962 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, TRUE);
963 vm_page_assert_xbusied(fs.m);
966 * Page must be completely valid or it is not fit to
967 * map into user space. vm_pager_get_pages() ensures this.
969 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
970 ("vm_fault: page %p partially invalid", fs.m));
971 VM_OBJECT_WUNLOCK(fs.object);
974 * Put this page into the physical map. We had to do the unlock above
975 * because pmap_enter() may sleep. We don't put the page
976 * back on the active queue until later so that the pageout daemon
977 * won't find it (yet).
979 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
980 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
981 if (faultcount != 1 && (fault_flags & VM_FAULT_CHANGE_WIRING) == 0 &&
983 vm_fault_prefault(&fs, vaddr, faultcount, reqpage);
984 VM_OBJECT_WLOCK(fs.object);
988 * If the page is not wired down, then put it where the pageout daemon
991 if (fault_flags & VM_FAULT_CHANGE_WIRING) {
995 vm_page_unwire(fs.m, PQ_ACTIVE);
997 vm_page_activate(fs.m);
998 if (m_hold != NULL) {
1002 vm_page_unlock(fs.m);
1003 vm_page_xunbusy(fs.m);
1006 * Unlock everything, and return
1008 unlock_and_deallocate(&fs);
1010 PCPU_INC(cnt.v_io_faults);
1011 curthread->td_ru.ru_majflt++;
1013 curthread->td_ru.ru_minflt++;
1015 return (KERN_SUCCESS);
1019 * Speed up the reclamation of up to "distance" pages that precede the
1020 * faulting pindex within the first object of the shadow chain.
1023 vm_fault_cache_behind(const struct faultstate *fs, int distance)
1025 vm_object_t first_object, object;
1026 vm_page_t m, m_prev;
1029 object = fs->object;
1030 VM_OBJECT_ASSERT_WLOCKED(object);
1031 first_object = fs->first_object;
1032 if (first_object != object) {
1033 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1034 VM_OBJECT_WUNLOCK(object);
1035 VM_OBJECT_WLOCK(first_object);
1036 VM_OBJECT_WLOCK(object);
1039 /* Neither fictitious nor unmanaged pages can be cached. */
1040 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1041 if (fs->first_pindex < distance)
1044 pindex = fs->first_pindex - distance;
1045 if (pindex < OFF_TO_IDX(fs->entry->offset))
1046 pindex = OFF_TO_IDX(fs->entry->offset);
1047 m = first_object != object ? fs->first_m : fs->m;
1048 vm_page_assert_xbusied(m);
1049 m_prev = vm_page_prev(m);
1050 while ((m = m_prev) != NULL && m->pindex >= pindex &&
1051 m->valid == VM_PAGE_BITS_ALL) {
1052 m_prev = vm_page_prev(m);
1053 if (vm_page_busied(m))
1056 if (m->hold_count == 0 && m->wire_count == 0) {
1058 vm_page_aflag_clear(m, PGA_REFERENCED);
1060 vm_page_deactivate(m);
1067 if (first_object != object)
1068 VM_OBJECT_WUNLOCK(first_object);
1072 * vm_fault_prefault provides a quick way of clustering
1073 * pagefaults into a processes address space. It is a "cousin"
1074 * of vm_map_pmap_enter, except it runs at page fault time instead
1078 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1079 int faultcount, int reqpage)
1082 vm_map_entry_t entry;
1083 vm_object_t backing_object, lobject;
1084 vm_offset_t addr, starta;
1087 int backward, forward, i;
1089 pmap = fs->map->pmap;
1090 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1093 if (faultcount > 0) {
1095 forward = faultcount - reqpage - 1;
1102 starta = addra - backward * PAGE_SIZE;
1103 if (starta < entry->start) {
1104 starta = entry->start;
1105 } else if (starta > addra) {
1110 * Generate the sequence of virtual addresses that are candidates for
1111 * prefaulting in an outward spiral from the faulting virtual address,
1112 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1113 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1114 * If the candidate address doesn't have a backing physical page, then
1115 * the loop immediately terminates.
1117 for (i = 0; i < 2 * imax(backward, forward); i++) {
1118 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1120 if (addr > addra + forward * PAGE_SIZE)
1123 if (addr < starta || addr >= entry->end)
1126 if (!pmap_is_prefaultable(pmap, addr))
1129 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1130 lobject = entry->object.vm_object;
1131 VM_OBJECT_RLOCK(lobject);
1132 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1133 lobject->type == OBJT_DEFAULT &&
1134 (backing_object = lobject->backing_object) != NULL) {
1135 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1136 0, ("vm_fault_prefault: unaligned object offset"));
1137 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1138 VM_OBJECT_RLOCK(backing_object);
1139 VM_OBJECT_RUNLOCK(lobject);
1140 lobject = backing_object;
1143 VM_OBJECT_RUNLOCK(lobject);
1146 if (m->valid == VM_PAGE_BITS_ALL &&
1147 (m->flags & PG_FICTITIOUS) == 0)
1148 pmap_enter_quick(pmap, addr, m, entry->protection);
1149 VM_OBJECT_RUNLOCK(lobject);
1154 * Hold each of the physical pages that are mapped by the specified range of
1155 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1156 * and allow the specified types of access, "prot". If all of the implied
1157 * pages are successfully held, then the number of held pages is returned
1158 * together with pointers to those pages in the array "ma". However, if any
1159 * of the pages cannot be held, -1 is returned.
1162 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1163 vm_prot_t prot, vm_page_t *ma, int max_count)
1165 vm_offset_t end, va;
1168 boolean_t pmap_failed;
1172 end = round_page(addr + len);
1173 addr = trunc_page(addr);
1176 * Check for illegal addresses.
1178 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1181 if (atop(end - addr) > max_count)
1182 panic("vm_fault_quick_hold_pages: count > max_count");
1183 count = atop(end - addr);
1186 * Most likely, the physical pages are resident in the pmap, so it is
1187 * faster to try pmap_extract_and_hold() first.
1189 pmap_failed = FALSE;
1190 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1191 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1194 else if ((prot & VM_PROT_WRITE) != 0 &&
1195 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1197 * Explicitly dirty the physical page. Otherwise, the
1198 * caller's changes may go unnoticed because they are
1199 * performed through an unmanaged mapping or by a DMA
1202 * The object lock is not held here.
1203 * See vm_page_clear_dirty_mask().
1210 * One or more pages could not be held by the pmap. Either no
1211 * page was mapped at the specified virtual address or that
1212 * mapping had insufficient permissions. Attempt to fault in
1213 * and hold these pages.
1215 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1216 if (*mp == NULL && vm_fault_hold(map, va, prot,
1217 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1222 for (mp = ma; mp < ma + count; mp++)
1225 vm_page_unhold(*mp);
1226 vm_page_unlock(*mp);
1233 * vm_fault_copy_entry
1235 * Create new shadow object backing dst_entry with private copy of
1236 * all underlying pages. When src_entry is equal to dst_entry,
1237 * function implements COW for wired-down map entry. Otherwise,
1238 * it forks wired entry into dst_map.
1240 * In/out conditions:
1241 * The source and destination maps must be locked for write.
1242 * The source map entry must be wired down (or be a sharing map
1243 * entry corresponding to a main map entry that is wired down).
1246 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1247 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1248 vm_ooffset_t *fork_charge)
1250 vm_object_t backing_object, dst_object, object, src_object;
1251 vm_pindex_t dst_pindex, pindex, src_pindex;
1252 vm_prot_t access, prot;
1262 upgrade = src_entry == dst_entry;
1263 access = prot = dst_entry->protection;
1265 src_object = src_entry->object.vm_object;
1266 src_pindex = OFF_TO_IDX(src_entry->offset);
1268 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1269 dst_object = src_object;
1270 vm_object_reference(dst_object);
1273 * Create the top-level object for the destination entry. (Doesn't
1274 * actually shadow anything - we copy the pages directly.)
1276 dst_object = vm_object_allocate(OBJT_DEFAULT,
1277 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1278 #if VM_NRESERVLEVEL > 0
1279 dst_object->flags |= OBJ_COLORED;
1280 dst_object->pg_color = atop(dst_entry->start);
1284 VM_OBJECT_WLOCK(dst_object);
1285 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1286 ("vm_fault_copy_entry: vm_object not NULL"));
1287 if (src_object != dst_object) {
1288 dst_entry->object.vm_object = dst_object;
1289 dst_entry->offset = 0;
1290 dst_object->charge = dst_entry->end - dst_entry->start;
1292 if (fork_charge != NULL) {
1293 KASSERT(dst_entry->cred == NULL,
1294 ("vm_fault_copy_entry: leaked swp charge"));
1295 dst_object->cred = curthread->td_ucred;
1296 crhold(dst_object->cred);
1297 *fork_charge += dst_object->charge;
1298 } else if (dst_object->cred == NULL) {
1299 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1301 dst_object->cred = dst_entry->cred;
1302 dst_entry->cred = NULL;
1306 * If not an upgrade, then enter the mappings in the pmap as
1307 * read and/or execute accesses. Otherwise, enter them as
1310 * A writeable large page mapping is only created if all of
1311 * the constituent small page mappings are modified. Marking
1312 * PTEs as modified on inception allows promotion to happen
1313 * without taking potentially large number of soft faults.
1316 access &= ~VM_PROT_WRITE;
1319 * Loop through all of the virtual pages within the entry's
1320 * range, copying each page from the source object to the
1321 * destination object. Since the source is wired, those pages
1322 * must exist. In contrast, the destination is pageable.
1323 * Since the destination object does share any backing storage
1324 * with the source object, all of its pages must be dirtied,
1325 * regardless of whether they can be written.
1327 for (vaddr = dst_entry->start, dst_pindex = 0;
1328 vaddr < dst_entry->end;
1329 vaddr += PAGE_SIZE, dst_pindex++) {
1332 * Find the page in the source object, and copy it in.
1333 * Because the source is wired down, the page will be
1336 if (src_object != dst_object)
1337 VM_OBJECT_RLOCK(src_object);
1338 object = src_object;
1339 pindex = src_pindex + dst_pindex;
1340 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1341 (backing_object = object->backing_object) != NULL) {
1343 * Unless the source mapping is read-only or
1344 * it is presently being upgraded from
1345 * read-only, the first object in the shadow
1346 * chain should provide all of the pages. In
1347 * other words, this loop body should never be
1348 * executed when the source mapping is already
1351 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1353 ("vm_fault_copy_entry: main object missing page"));
1355 VM_OBJECT_RLOCK(backing_object);
1356 pindex += OFF_TO_IDX(object->backing_object_offset);
1357 if (object != dst_object)
1358 VM_OBJECT_RUNLOCK(object);
1359 object = backing_object;
1361 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1363 if (object != dst_object) {
1365 * Allocate a page in the destination object.
1367 dst_m = vm_page_alloc(dst_object, (src_object ==
1368 dst_object ? src_pindex : 0) + dst_pindex,
1370 if (dst_m == NULL) {
1371 VM_OBJECT_WUNLOCK(dst_object);
1372 VM_OBJECT_RUNLOCK(object);
1374 VM_OBJECT_WLOCK(dst_object);
1377 pmap_copy_page(src_m, dst_m);
1378 VM_OBJECT_RUNLOCK(object);
1379 dst_m->valid = VM_PAGE_BITS_ALL;
1380 dst_m->dirty = VM_PAGE_BITS_ALL;
1383 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1385 vm_page_xbusy(dst_m);
1386 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1387 ("invalid dst page %p", dst_m));
1389 VM_OBJECT_WUNLOCK(dst_object);
1392 * Enter it in the pmap. If a wired, copy-on-write
1393 * mapping is being replaced by a write-enabled
1394 * mapping, then wire that new mapping.
1396 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1397 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1400 * Mark it no longer busy, and put it on the active list.
1402 VM_OBJECT_WLOCK(dst_object);
1405 if (src_m != dst_m) {
1406 vm_page_lock(src_m);
1407 vm_page_unwire(src_m, PQ_INACTIVE);
1408 vm_page_unlock(src_m);
1409 vm_page_lock(dst_m);
1410 vm_page_wire(dst_m);
1411 vm_page_unlock(dst_m);
1413 KASSERT(dst_m->wire_count > 0,
1414 ("dst_m %p is not wired", dst_m));
1417 vm_page_lock(dst_m);
1418 vm_page_activate(dst_m);
1419 vm_page_unlock(dst_m);
1421 vm_page_xunbusy(dst_m);
1423 VM_OBJECT_WUNLOCK(dst_object);
1425 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1426 vm_object_deallocate(src_object);
1432 * This routine checks around the requested page for other pages that
1433 * might be able to be faulted in. This routine brackets the viable
1434 * pages for the pages to be paged in.
1437 * m, rbehind, rahead
1440 * marray (array of vm_page_t), reqpage (index of requested page)
1443 * number of pages in marray
1446 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1455 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1457 int cbehind, cahead;
1459 VM_OBJECT_ASSERT_WLOCKED(m->object);
1463 cbehind = cahead = 0;
1466 * if the requested page is not available, then give up now
1468 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1472 if ((cbehind == 0) && (cahead == 0)) {
1478 if (rahead > cahead) {
1482 if (rbehind > cbehind) {
1487 * scan backward for the read behind pages -- in memory
1490 if (rbehind > pindex) {
1494 startpindex = pindex - rbehind;
1497 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1498 rtm->pindex >= startpindex)
1499 startpindex = rtm->pindex + 1;
1501 /* tpindex is unsigned; beware of numeric underflow. */
1502 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1503 tpindex < pindex; i++, tpindex--) {
1505 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1506 VM_ALLOC_IFNOTCACHED);
1509 * Shift the allocated pages to the
1510 * beginning of the array.
1512 for (j = 0; j < i; j++) {
1513 marray[j] = marray[j + tpindex + 1 -
1519 marray[tpindex - startpindex] = rtm;
1527 /* page offset of the required page */
1530 tpindex = pindex + 1;
1534 * scan forward for the read ahead pages
1536 endpindex = tpindex + rahead;
1537 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1538 endpindex = rtm->pindex;
1539 if (endpindex > object->size)
1540 endpindex = object->size;
1542 for (; tpindex < endpindex; i++, tpindex++) {
1544 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1545 VM_ALLOC_IFNOTCACHED);
1553 /* return number of pages */
1558 * Block entry into the machine-independent layer's page fault handler by
1559 * the calling thread. Subsequent calls to vm_fault() by that thread will
1560 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1561 * spurious page faults.
1564 vm_fault_disable_pagefaults(void)
1567 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1571 vm_fault_enable_pagefaults(int save)
1574 curthread_pflags_restore(save);