2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
12 * This code is derived from software contributed to Berkeley by
13 * The Mach Operating System project at Carnegie-Mellon University.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
46 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47 * All rights reserved.
49 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
51 * Permission to use, copy, modify and distribute this software and
52 * its documentation is hereby granted, provided that both the copyright
53 * notice and this permission notice appear in all copies of the
54 * software, derivative works or modified versions, and any portions
55 * thereof, and that both notices appear in supporting documentation.
57 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
58 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
59 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
61 * Carnegie Mellon requests users of this software to return to
63 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
64 * School of Computer Science
65 * Carnegie Mellon University
66 * Pittsburgh PA 15213-3890
68 * any improvements or extensions that they make and grant Carnegie the
69 * rights to redistribute these changes.
73 * Page fault handling module.
76 #include <sys/cdefs.h>
77 __FBSDID("$FreeBSD$");
79 #include "opt_ktrace.h"
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
88 #include <sys/racct.h>
89 #include <sys/resourcevar.h>
90 #include <sys/rwlock.h>
91 #include <sys/sysctl.h>
92 #include <sys/vmmeter.h>
93 #include <sys/vnode.h>
95 #include <sys/ktrace.h>
99 #include <vm/vm_param.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_pageout.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_pager.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_reserv.h>
113 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
114 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
116 #define VM_FAULT_DONTNEED_MIN 1048576
123 vm_object_t first_object;
124 vm_pindex_t first_pindex;
126 vm_map_entry_t entry;
128 bool lookup_still_valid;
132 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
134 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
135 int backward, int forward, bool obj_locked);
138 release_page(struct faultstate *fs)
141 vm_page_xunbusy(fs->m);
143 vm_page_deactivate(fs->m);
144 vm_page_unlock(fs->m);
149 unlock_map(struct faultstate *fs)
152 if (fs->lookup_still_valid) {
153 vm_map_lookup_done(fs->map, fs->entry);
154 fs->lookup_still_valid = false;
159 unlock_vp(struct faultstate *fs)
162 if (fs->vp != NULL) {
169 unlock_and_deallocate(struct faultstate *fs)
172 vm_object_pip_wakeup(fs->object);
173 VM_OBJECT_WUNLOCK(fs->object);
174 if (fs->object != fs->first_object) {
175 VM_OBJECT_WLOCK(fs->first_object);
176 vm_page_lock(fs->first_m);
177 vm_page_free(fs->first_m);
178 vm_page_unlock(fs->first_m);
179 vm_object_pip_wakeup(fs->first_object);
180 VM_OBJECT_WUNLOCK(fs->first_object);
183 vm_object_deallocate(fs->first_object);
189 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
190 vm_prot_t fault_type, int fault_flags, bool set_wd)
194 if (((prot & VM_PROT_WRITE) == 0 &&
195 (fault_flags & VM_FAULT_DIRTY) == 0) ||
196 (m->oflags & VPO_UNMANAGED) != 0)
199 VM_OBJECT_ASSERT_LOCKED(m->object);
201 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
202 (fault_flags & VM_FAULT_WIRE) == 0) ||
203 (fault_flags & VM_FAULT_DIRTY) != 0;
206 vm_object_set_writeable_dirty(m->object);
209 * If two callers of vm_fault_dirty() with set_wd ==
210 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
211 * flag set, other with flag clear, race, it is
212 * possible for the no-NOSYNC thread to see m->dirty
213 * != 0 and not clear VPO_NOSYNC. Take vm_page lock
214 * around manipulation of VPO_NOSYNC and
215 * vm_page_dirty() call, to avoid the race and keep
216 * m->oflags consistent.
221 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
222 * if the page is already dirty to prevent data written with
223 * the expectation of being synced from not being synced.
224 * Likewise if this entry does not request NOSYNC then make
225 * sure the page isn't marked NOSYNC. Applications sharing
226 * data should use the same flags to avoid ping ponging.
228 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
230 m->oflags |= VPO_NOSYNC;
233 m->oflags &= ~VPO_NOSYNC;
237 * If the fault is a write, we know that this page is being
238 * written NOW so dirty it explicitly to save on
239 * pmap_is_modified() calls later.
241 * Also, since the page is now dirty, we can possibly tell
242 * the pager to release any swap backing the page. Calling
243 * the pager requires a write lock on the object.
250 vm_pager_page_unswapped(m);
254 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m)
257 if (m_hold != NULL) {
266 * Unlocks fs.first_object and fs.map on success.
269 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
270 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
273 #if defined(__amd64__) && VM_NRESERVLEVEL > 0
279 MPASS(fs->vp == NULL);
280 m = vm_page_lookup(fs->first_object, fs->first_pindex);
281 /* A busy page can be mapped for read|execute access. */
282 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
283 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
284 return (KERN_FAILURE);
287 #if defined(__amd64__) && VM_NRESERVLEVEL > 0
288 if ((m->flags & PG_FICTITIOUS) == 0 &&
289 (m_super = vm_reserv_to_superpage(m)) != NULL &&
290 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
291 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
292 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
293 (pagesizes[m_super->psind] - 1)) &&
294 pmap_ps_enabled(fs->map->pmap)) {
295 flags = PS_ALL_VALID;
296 if ((prot & VM_PROT_WRITE) != 0) {
298 * Create a superpage mapping allowing write access
299 * only if none of the constituent pages are busy and
300 * all of them are already dirty (except possibly for
301 * the page that was faulted on).
303 flags |= PS_NONE_BUSY;
304 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
305 flags |= PS_ALL_DIRTY;
307 if (vm_page_ps_test(m_super, flags, m)) {
309 psind = m_super->psind;
310 vaddr = rounddown2(vaddr, pagesizes[psind]);
311 /* Preset the modified bit for dirty superpages. */
312 if ((flags & PS_ALL_DIRTY) != 0)
313 fault_type |= VM_PROT_WRITE;
317 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
318 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
319 if (rv != KERN_SUCCESS)
321 vm_fault_fill_hold(m_hold, m);
322 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
323 if (psind == 0 && !wired)
324 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
325 VM_OBJECT_RUNLOCK(fs->first_object);
326 vm_map_lookup_done(fs->map, fs->entry);
327 curthread->td_ru.ru_minflt++;
328 return (KERN_SUCCESS);
332 vm_fault_restore_map_lock(struct faultstate *fs)
335 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
336 MPASS(fs->first_object->paging_in_progress > 0);
338 if (!vm_map_trylock_read(fs->map)) {
339 VM_OBJECT_WUNLOCK(fs->first_object);
340 vm_map_lock_read(fs->map);
341 VM_OBJECT_WLOCK(fs->first_object);
343 fs->lookup_still_valid = true;
347 vm_fault_populate_check_page(vm_page_t m)
351 * Check each page to ensure that the pager is obeying the
352 * interface: the page must be installed in the object, fully
353 * valid, and exclusively busied.
356 MPASS(m->valid == VM_PAGE_BITS_ALL);
357 MPASS(vm_page_xbusied(m));
361 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
367 VM_OBJECT_ASSERT_WLOCKED(object);
368 MPASS(first <= last);
369 for (pidx = first, m = vm_page_lookup(object, pidx);
370 pidx <= last; pidx++, m = vm_page_next(m)) {
371 vm_fault_populate_check_page(m);
373 vm_page_deactivate(m);
380 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
381 int fault_flags, boolean_t wired, vm_page_t *m_hold)
386 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
387 int i, npages, psind, rv;
389 MPASS(fs->object == fs->first_object);
390 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
391 MPASS(fs->first_object->paging_in_progress > 0);
392 MPASS(fs->first_object->backing_object == NULL);
393 MPASS(fs->lookup_still_valid);
395 pager_first = OFF_TO_IDX(fs->entry->offset);
396 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
401 * Call the pager (driver) populate() method.
403 * There is no guarantee that the method will be called again
404 * if the current fault is for read, and a future fault is
405 * for write. Report the entry's maximum allowed protection
408 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
409 fault_type, fs->entry->max_protection, &pager_first, &pager_last);
411 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
412 if (rv == VM_PAGER_BAD) {
414 * VM_PAGER_BAD is the backdoor for a pager to request
415 * normal fault handling.
417 vm_fault_restore_map_lock(fs);
418 if (fs->map->timestamp != fs->map_generation)
419 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
420 return (KERN_NOT_RECEIVER);
422 if (rv != VM_PAGER_OK)
423 return (KERN_FAILURE); /* AKA SIGSEGV */
425 /* Ensure that the driver is obeying the interface. */
426 MPASS(pager_first <= pager_last);
427 MPASS(fs->first_pindex <= pager_last);
428 MPASS(fs->first_pindex >= pager_first);
429 MPASS(pager_last < fs->first_object->size);
431 vm_fault_restore_map_lock(fs);
432 if (fs->map->timestamp != fs->map_generation) {
433 vm_fault_populate_cleanup(fs->first_object, pager_first,
435 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
439 * The map is unchanged after our last unlock. Process the fault.
441 * The range [pager_first, pager_last] that is given to the
442 * pager is only a hint. The pager may populate any range
443 * within the object that includes the requested page index.
444 * In case the pager expanded the range, clip it to fit into
447 map_first = OFF_TO_IDX(fs->entry->offset);
448 if (map_first > pager_first) {
449 vm_fault_populate_cleanup(fs->first_object, pager_first,
451 pager_first = map_first;
453 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
454 if (map_last < pager_last) {
455 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
457 pager_last = map_last;
459 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
461 pidx += npages, m = vm_page_next(&m[npages - 1])) {
462 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
463 #if defined(__amd64__)
465 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
466 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
467 !pmap_ps_enabled(fs->map->pmap)))
472 npages = atop(pagesizes[psind]);
473 for (i = 0; i < npages; i++) {
474 vm_fault_populate_check_page(&m[i]);
475 vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
478 VM_OBJECT_WUNLOCK(fs->first_object);
479 pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ?
480 PMAP_ENTER_WIRED : 0), psind);
481 VM_OBJECT_WLOCK(fs->first_object);
483 for (i = 0; i < npages; i++) {
484 vm_page_change_lock(&m[i], &m_mtx);
485 if ((fault_flags & VM_FAULT_WIRE) != 0)
488 vm_page_activate(&m[i]);
489 if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
493 vm_page_xunbusy_maybelocked(&m[i]);
498 curthread->td_ru.ru_majflt++;
499 return (KERN_SUCCESS);
505 * Handle a page fault occurring at the given address,
506 * requiring the given permissions, in the map specified.
507 * If successful, the page is inserted into the
508 * associated physical map.
510 * NOTE: the given address should be truncated to the
511 * proper page address.
513 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
514 * a standard error specifying why the fault is fatal is returned.
516 * The map in question must be referenced, and remains so.
517 * Caller may hold no locks.
520 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
527 if ((td->td_pflags & TDP_NOFAULTING) != 0)
528 return (KERN_PROTECTION_FAILURE);
530 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
531 ktrfault(vaddr, fault_type);
533 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
536 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
543 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
544 int fault_flags, vm_page_t *m_hold)
546 struct faultstate fs;
548 vm_object_t next_object, retry_object;
549 vm_offset_t e_end, e_start;
550 vm_pindex_t retry_pindex;
551 vm_prot_t prot, retry_prot;
552 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
553 int locked, nera, result, rv;
555 boolean_t wired; /* Passed by reference. */
556 bool dead, hardfault, is_first_object_locked;
558 VM_CNT_INC(v_vm_faults);
567 * Find the backing store object and offset into it to begin the
571 result = vm_map_lookup(&fs.map, vaddr, fault_type |
572 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
573 &fs.first_pindex, &prot, &wired);
574 if (result != KERN_SUCCESS) {
579 fs.map_generation = fs.map->timestamp;
581 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
582 panic("%s: fault on nofault entry, addr: %#lx",
583 __func__, (u_long)vaddr);
586 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
587 fs.entry->wiring_thread != curthread) {
588 vm_map_unlock_read(fs.map);
590 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
591 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
593 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
594 vm_map_unlock_and_wait(fs.map, 0);
596 vm_map_unlock(fs.map);
600 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
603 fault_type = prot | (fault_type & VM_PROT_COPY);
605 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
606 ("!wired && VM_FAULT_WIRE"));
609 * Try to avoid lock contention on the top-level object through
610 * special-case handling of some types of page faults, specifically,
611 * those that are both (1) mapping an existing page from the top-
612 * level object and (2) not having to mark that object as containing
613 * dirty pages. Under these conditions, a read lock on the top-level
614 * object suffices, allowing multiple page faults of a similar type to
615 * run in parallel on the same top-level object.
617 if (fs.vp == NULL /* avoid locked vnode leak */ &&
618 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
619 /* avoid calling vm_object_set_writeable_dirty() */
620 ((prot & VM_PROT_WRITE) == 0 ||
621 (fs.first_object->type != OBJT_VNODE &&
622 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
623 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
624 VM_OBJECT_RLOCK(fs.first_object);
625 if ((prot & VM_PROT_WRITE) == 0 ||
626 (fs.first_object->type != OBJT_VNODE &&
627 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
628 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
629 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
630 fault_flags, wired, m_hold);
631 if (rv == KERN_SUCCESS)
634 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
635 VM_OBJECT_RUNLOCK(fs.first_object);
636 VM_OBJECT_WLOCK(fs.first_object);
639 VM_OBJECT_WLOCK(fs.first_object);
643 * Make a reference to this object to prevent its disposal while we
644 * are messing with it. Once we have the reference, the map is free
645 * to be diddled. Since objects reference their shadows (and copies),
646 * they will stay around as well.
648 * Bump the paging-in-progress count to prevent size changes (e.g.
649 * truncation operations) during I/O.
651 vm_object_reference_locked(fs.first_object);
652 vm_object_pip_add(fs.first_object, 1);
654 fs.lookup_still_valid = true;
659 * Search for the page at object/offset.
661 fs.object = fs.first_object;
662 fs.pindex = fs.first_pindex;
665 * If the object is marked for imminent termination,
666 * we retry here, since the collapse pass has raced
667 * with us. Otherwise, if we see terminally dead
668 * object, return fail.
670 if ((fs.object->flags & OBJ_DEAD) != 0) {
671 dead = fs.object->type == OBJT_DEAD;
672 unlock_and_deallocate(&fs);
674 return (KERN_PROTECTION_FAILURE);
680 * See if page is resident
682 fs.m = vm_page_lookup(fs.object, fs.pindex);
685 * Wait/Retry if the page is busy. We have to do this
686 * if the page is either exclusive or shared busy
687 * because the vm_pager may be using read busy for
688 * pageouts (and even pageins if it is the vnode
689 * pager), and we could end up trying to pagein and
690 * pageout the same page simultaneously.
692 * We can theoretically allow the busy case on a read
693 * fault if the page is marked valid, but since such
694 * pages are typically already pmap'd, putting that
695 * special case in might be more effort then it is
696 * worth. We cannot under any circumstances mess
697 * around with a shared busied page except, perhaps,
700 if (vm_page_busied(fs.m)) {
702 * Reference the page before unlocking and
703 * sleeping so that the page daemon is less
704 * likely to reclaim it.
706 vm_page_aflag_set(fs.m, PGA_REFERENCED);
707 if (fs.object != fs.first_object) {
708 if (!VM_OBJECT_TRYWLOCK(
710 VM_OBJECT_WUNLOCK(fs.object);
711 VM_OBJECT_WLOCK(fs.first_object);
712 VM_OBJECT_WLOCK(fs.object);
714 vm_page_lock(fs.first_m);
715 vm_page_free(fs.first_m);
716 vm_page_unlock(fs.first_m);
717 vm_object_pip_wakeup(fs.first_object);
718 VM_OBJECT_WUNLOCK(fs.first_object);
722 if (fs.m == vm_page_lookup(fs.object,
724 vm_page_sleep_if_busy(fs.m, "vmpfw");
726 vm_object_pip_wakeup(fs.object);
727 VM_OBJECT_WUNLOCK(fs.object);
728 VM_CNT_INC(v_intrans);
729 vm_object_deallocate(fs.first_object);
734 * Mark page busy for other processes, and the
735 * pagedaemon. If it still isn't completely valid
736 * (readable), jump to readrest, else break-out ( we
740 if (fs.m->valid != VM_PAGE_BITS_ALL)
742 break; /* break to PAGE HAS BEEN FOUND */
744 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
747 * Page is not resident. If the pager might contain the page
748 * or this is the beginning of the search, allocate a new
749 * page. (Default objects are zero-fill, so there is no real
752 if (fs.object->type != OBJT_DEFAULT ||
753 fs.object == fs.first_object) {
754 if (fs.pindex >= fs.object->size) {
755 unlock_and_deallocate(&fs);
756 return (KERN_PROTECTION_FAILURE);
759 if (fs.object == fs.first_object &&
760 (fs.first_object->flags & OBJ_POPULATE) != 0 &&
761 fs.first_object->shadow_count == 0) {
762 rv = vm_fault_populate(&fs, prot, fault_type,
763 fault_flags, wired, m_hold);
767 unlock_and_deallocate(&fs);
769 case KERN_RESOURCE_SHORTAGE:
770 unlock_and_deallocate(&fs);
772 case KERN_NOT_RECEIVER:
774 * Pager's populate() method
775 * returned VM_PAGER_BAD.
779 panic("inconsistent return codes");
784 * Allocate a new page for this object/offset pair.
786 * Unlocked read of the p_flag is harmless. At
787 * worst, the P_KILLED might be not observed
788 * there, and allocation can fail, causing
789 * restart and new reading of the p_flag.
791 if (!vm_page_count_severe() || P_KILLED(curproc)) {
792 #if VM_NRESERVLEVEL > 0
793 vm_object_color(fs.object, atop(vaddr) -
796 alloc_req = P_KILLED(curproc) ?
797 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
798 if (fs.object->type != OBJT_VNODE &&
799 fs.object->backing_object == NULL)
800 alloc_req |= VM_ALLOC_ZERO;
801 fs.m = vm_page_alloc(fs.object, fs.pindex,
805 unlock_and_deallocate(&fs);
813 * At this point, we have either allocated a new page or found
814 * an existing page that is only partially valid.
816 * We hold a reference on the current object and the page is
821 * If the pager for the current object might have the page,
822 * then determine the number of additional pages to read and
823 * potentially reprioritize previously read pages for earlier
824 * reclamation. These operations should only be performed
825 * once per page fault. Even if the current pager doesn't
826 * have the page, the number of additional pages to read will
827 * apply to subsequent objects in the shadow chain.
829 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
830 !P_KILLED(curproc)) {
831 KASSERT(fs.lookup_still_valid, ("map unlocked"));
832 era = fs.entry->read_ahead;
833 behavior = vm_map_entry_behavior(fs.entry);
834 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
836 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
837 nera = VM_FAULT_READ_AHEAD_MAX;
838 if (vaddr == fs.entry->next_read)
839 vm_fault_dontneed(&fs, vaddr, nera);
840 } else if (vaddr == fs.entry->next_read) {
842 * This is a sequential fault. Arithmetically
843 * increase the requested number of pages in
844 * the read-ahead window. The requested
845 * number of pages is "# of sequential faults
846 * x (read ahead min + 1) + read ahead min"
848 nera = VM_FAULT_READ_AHEAD_MIN;
851 if (nera > VM_FAULT_READ_AHEAD_MAX)
852 nera = VM_FAULT_READ_AHEAD_MAX;
854 if (era == VM_FAULT_READ_AHEAD_MAX)
855 vm_fault_dontneed(&fs, vaddr, nera);
858 * This is a non-sequential fault.
864 * A read lock on the map suffices to update
865 * the read ahead count safely.
867 fs.entry->read_ahead = nera;
871 * Prepare for unlocking the map. Save the map
872 * entry's start and end addresses, which are used to
873 * optimize the size of the pager operation below.
874 * Even if the map entry's addresses change after
875 * unlocking the map, using the saved addresses is
878 e_start = fs.entry->start;
879 e_end = fs.entry->end;
883 * Call the pager to retrieve the page if there is a chance
884 * that the pager has it, and potentially retrieve additional
885 * pages at the same time.
887 if (fs.object->type != OBJT_DEFAULT) {
889 * Release the map lock before locking the vnode or
890 * sleeping in the pager. (If the current object has
891 * a shadow, then an earlier iteration of this loop
892 * may have already unlocked the map.)
896 if (fs.object->type == OBJT_VNODE &&
897 (vp = fs.object->handle) != fs.vp) {
899 * Perform an unlock in case the desired vnode
900 * changed while the map was unlocked during a
905 locked = VOP_ISLOCKED(vp);
906 if (locked != LK_EXCLUSIVE)
910 * We must not sleep acquiring the vnode lock
911 * while we have the page exclusive busied or
912 * the object's paging-in-progress count
913 * incremented. Otherwise, we could deadlock.
915 error = vget(vp, locked | LK_CANRECURSE |
916 LK_NOWAIT, curthread);
920 unlock_and_deallocate(&fs);
921 error = vget(vp, locked | LK_RETRY |
922 LK_CANRECURSE, curthread);
926 ("vm_fault: vget failed"));
931 KASSERT(fs.vp == NULL || !fs.map->system_map,
932 ("vm_fault: vnode-backed object mapped by system map"));
935 * Page in the requested page and hint the pager,
936 * that it may bring up surrounding pages.
938 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
943 /* Is this a sequential fault? */
949 * Request a cluster of pages that is
950 * aligned to a VM_FAULT_READ_DEFAULT
951 * page offset boundary within the
952 * object. Alignment to a page offset
953 * boundary is more likely to coincide
954 * with the underlying file system
955 * block than alignment to a virtual
958 cluster_offset = fs.pindex %
959 VM_FAULT_READ_DEFAULT;
960 behind = ulmin(cluster_offset,
961 atop(vaddr - e_start));
962 ahead = VM_FAULT_READ_DEFAULT - 1 -
965 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
967 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
969 if (rv == VM_PAGER_OK) {
970 faultcount = behind + 1 + ahead;
972 break; /* break to PAGE HAS BEEN FOUND */
974 if (rv == VM_PAGER_ERROR)
975 printf("vm_fault: pager read error, pid %d (%s)\n",
976 curproc->p_pid, curproc->p_comm);
979 * If an I/O error occurred or the requested page was
980 * outside the range of the pager, clean up and return
983 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
985 if (fs.m->wire_count == 0)
988 vm_page_xunbusy_maybelocked(fs.m);
989 vm_page_unlock(fs.m);
991 unlock_and_deallocate(&fs);
992 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
993 KERN_PROTECTION_FAILURE);
997 * The requested page does not exist at this object/
998 * offset. Remove the invalid page from the object,
999 * waking up anyone waiting for it, and continue on to
1000 * the next object. However, if this is the top-level
1001 * object, we must leave the busy page in place to
1002 * prevent another process from rushing past us, and
1003 * inserting the page in that object at the same time
1006 if (fs.object != fs.first_object) {
1008 if (fs.m->wire_count == 0)
1011 vm_page_xunbusy_maybelocked(fs.m);
1012 vm_page_unlock(fs.m);
1018 * We get here if the object has default pager (or unwiring)
1019 * or the pager doesn't have the page.
1021 if (fs.object == fs.first_object)
1025 * Move on to the next object. Lock the next object before
1026 * unlocking the current one.
1028 next_object = fs.object->backing_object;
1029 if (next_object == NULL) {
1031 * If there's no object left, fill the page in the top
1032 * object with zeros.
1034 if (fs.object != fs.first_object) {
1035 vm_object_pip_wakeup(fs.object);
1036 VM_OBJECT_WUNLOCK(fs.object);
1038 fs.object = fs.first_object;
1039 fs.pindex = fs.first_pindex;
1041 VM_OBJECT_WLOCK(fs.object);
1046 * Zero the page if necessary and mark it valid.
1048 if ((fs.m->flags & PG_ZERO) == 0) {
1049 pmap_zero_page(fs.m);
1051 VM_CNT_INC(v_ozfod);
1054 fs.m->valid = VM_PAGE_BITS_ALL;
1055 /* Don't try to prefault neighboring pages. */
1057 break; /* break to PAGE HAS BEEN FOUND */
1059 KASSERT(fs.object != next_object,
1060 ("object loop %p", next_object));
1061 VM_OBJECT_WLOCK(next_object);
1062 vm_object_pip_add(next_object, 1);
1063 if (fs.object != fs.first_object)
1064 vm_object_pip_wakeup(fs.object);
1066 OFF_TO_IDX(fs.object->backing_object_offset);
1067 VM_OBJECT_WUNLOCK(fs.object);
1068 fs.object = next_object;
1072 vm_page_assert_xbusied(fs.m);
1075 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1080 * If the page is being written, but isn't already owned by the
1081 * top-level object, we have to copy it into a new page owned by the
1084 if (fs.object != fs.first_object) {
1086 * We only really need to copy if we want to write it.
1088 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1090 * This allows pages to be virtually copied from a
1091 * backing_object into the first_object, where the
1092 * backing object has no other refs to it, and cannot
1093 * gain any more refs. Instead of a bcopy, we just
1094 * move the page from the backing object to the
1095 * first object. Note that we must mark the page
1096 * dirty in the first object so that it will go out
1097 * to swap when needed.
1099 is_first_object_locked = false;
1102 * Only one shadow object
1104 (fs.object->shadow_count == 1) &&
1106 * No COW refs, except us
1108 (fs.object->ref_count == 1) &&
1110 * No one else can look this object up
1112 (fs.object->handle == NULL) &&
1114 * No other ways to look the object up
1116 ((fs.object->type == OBJT_DEFAULT) ||
1117 (fs.object->type == OBJT_SWAP)) &&
1118 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1120 * We don't chase down the shadow chain
1122 fs.object == fs.first_object->backing_object) {
1124 vm_page_remque(fs.m);
1125 vm_page_remove(fs.m);
1126 vm_page_unlock(fs.m);
1127 vm_page_lock(fs.first_m);
1128 vm_page_replace_checked(fs.m, fs.first_object,
1129 fs.first_pindex, fs.first_m);
1130 vm_page_free(fs.first_m);
1131 vm_page_unlock(fs.first_m);
1132 vm_page_dirty(fs.m);
1133 #if VM_NRESERVLEVEL > 0
1135 * Rename the reservation.
1137 vm_reserv_rename(fs.m, fs.first_object,
1138 fs.object, OFF_TO_IDX(
1139 fs.first_object->backing_object_offset));
1142 * Removing the page from the backing object
1145 vm_page_xbusy(fs.m);
1148 VM_CNT_INC(v_cow_optim);
1151 * Oh, well, lets copy it.
1153 pmap_copy_page(fs.m, fs.first_m);
1154 fs.first_m->valid = VM_PAGE_BITS_ALL;
1155 if ((fault_flags & VM_FAULT_WIRE) == 0) {
1156 prot &= ~VM_PROT_WRITE;
1157 fault_type &= ~VM_PROT_WRITE;
1159 if (wired && (fault_flags &
1160 VM_FAULT_WIRE) == 0) {
1161 vm_page_lock(fs.first_m);
1162 vm_page_wire(fs.first_m);
1163 vm_page_unlock(fs.first_m);
1166 vm_page_unwire(fs.m, PQ_INACTIVE);
1167 vm_page_unlock(fs.m);
1170 * We no longer need the old page or object.
1175 * fs.object != fs.first_object due to above
1178 vm_object_pip_wakeup(fs.object);
1179 VM_OBJECT_WUNLOCK(fs.object);
1181 * Only use the new page below...
1183 fs.object = fs.first_object;
1184 fs.pindex = fs.first_pindex;
1186 if (!is_first_object_locked)
1187 VM_OBJECT_WLOCK(fs.object);
1188 VM_CNT_INC(v_cow_faults);
1189 curthread->td_cow++;
1191 prot &= ~VM_PROT_WRITE;
1196 * We must verify that the maps have not changed since our last
1199 if (!fs.lookup_still_valid) {
1200 if (!vm_map_trylock_read(fs.map)) {
1202 unlock_and_deallocate(&fs);
1205 fs.lookup_still_valid = true;
1206 if (fs.map->timestamp != fs.map_generation) {
1207 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1208 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1211 * If we don't need the page any longer, put it on the inactive
1212 * list (the easiest thing to do here). If no one needs it,
1213 * pageout will grab it eventually.
1215 if (result != KERN_SUCCESS) {
1217 unlock_and_deallocate(&fs);
1220 * If retry of map lookup would have blocked then
1221 * retry fault from start.
1223 if (result == KERN_FAILURE)
1227 if ((retry_object != fs.first_object) ||
1228 (retry_pindex != fs.first_pindex)) {
1230 unlock_and_deallocate(&fs);
1235 * Check whether the protection has changed or the object has
1236 * been copied while we left the map unlocked. Changing from
1237 * read to write permission is OK - we leave the page
1238 * write-protected, and catch the write fault. Changing from
1239 * write to read permission means that we can't mark the page
1240 * write-enabled after all.
1243 fault_type &= retry_prot;
1246 unlock_and_deallocate(&fs);
1250 /* Reassert because wired may have changed. */
1251 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1252 ("!wired && VM_FAULT_WIRE"));
1257 * If the page was filled by a pager, save the virtual address that
1258 * should be faulted on next under a sequential access pattern to the
1259 * map entry. A read lock on the map suffices to update this address
1263 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1265 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1266 vm_page_assert_xbusied(fs.m);
1269 * Page must be completely valid or it is not fit to
1270 * map into user space. vm_pager_get_pages() ensures this.
1272 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1273 ("vm_fault: page %p partially invalid", fs.m));
1274 VM_OBJECT_WUNLOCK(fs.object);
1277 * Put this page into the physical map. We had to do the unlock above
1278 * because pmap_enter() may sleep. We don't put the page
1279 * back on the active queue until later so that the pageout daemon
1280 * won't find it (yet).
1282 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1283 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1284 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1286 vm_fault_prefault(&fs, vaddr,
1287 faultcount > 0 ? behind : PFBAK,
1288 faultcount > 0 ? ahead : PFFOR, false);
1289 VM_OBJECT_WLOCK(fs.object);
1293 * If the page is not wired down, then put it where the pageout daemon
1296 if ((fault_flags & VM_FAULT_WIRE) != 0)
1299 vm_page_activate(fs.m);
1300 if (m_hold != NULL) {
1304 vm_page_unlock(fs.m);
1305 vm_page_xunbusy(fs.m);
1308 * Unlock everything, and return
1310 unlock_and_deallocate(&fs);
1312 VM_CNT_INC(v_io_faults);
1313 curthread->td_ru.ru_majflt++;
1315 if (racct_enable && fs.object->type == OBJT_VNODE) {
1317 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1318 racct_add_force(curproc, RACCT_WRITEBPS,
1319 PAGE_SIZE + behind * PAGE_SIZE);
1320 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1322 racct_add_force(curproc, RACCT_READBPS,
1323 PAGE_SIZE + ahead * PAGE_SIZE);
1324 racct_add_force(curproc, RACCT_READIOPS, 1);
1326 PROC_UNLOCK(curproc);
1330 curthread->td_ru.ru_minflt++;
1332 return (KERN_SUCCESS);
1336 * Speed up the reclamation of pages that precede the faulting pindex within
1337 * the first object of the shadow chain. Essentially, perform the equivalent
1338 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1339 * the faulting pindex by the cluster size when the pages read by vm_fault()
1340 * cross a cluster-size boundary. The cluster size is the greater of the
1341 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1343 * When "fs->first_object" is a shadow object, the pages in the backing object
1344 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1345 * function must only be concerned with pages in the first object.
1348 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1350 vm_map_entry_t entry;
1351 vm_object_t first_object, object;
1352 vm_offset_t end, start;
1353 vm_page_t m, m_next;
1354 vm_pindex_t pend, pstart;
1357 object = fs->object;
1358 VM_OBJECT_ASSERT_WLOCKED(object);
1359 first_object = fs->first_object;
1360 if (first_object != object) {
1361 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1362 VM_OBJECT_WUNLOCK(object);
1363 VM_OBJECT_WLOCK(first_object);
1364 VM_OBJECT_WLOCK(object);
1367 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1368 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1369 size = VM_FAULT_DONTNEED_MIN;
1370 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1371 size = pagesizes[1];
1372 end = rounddown2(vaddr, size);
1373 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1374 (entry = fs->entry)->start < end) {
1375 if (end - entry->start < size)
1376 start = entry->start;
1379 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1380 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1382 m_next = vm_page_find_least(first_object, pstart);
1383 pend = OFF_TO_IDX(entry->offset) + atop(end -
1385 while ((m = m_next) != NULL && m->pindex < pend) {
1386 m_next = TAILQ_NEXT(m, listq);
1387 if (m->valid != VM_PAGE_BITS_ALL ||
1392 * Don't clear PGA_REFERENCED, since it would
1393 * likely represent a reference by a different
1396 * Typically, at this point, prefetched pages
1397 * are still in the inactive queue. Only
1398 * pages that triggered page faults are in the
1402 if (!vm_page_inactive(m))
1403 vm_page_deactivate(m);
1408 if (first_object != object)
1409 VM_OBJECT_WUNLOCK(first_object);
1413 * vm_fault_prefault provides a quick way of clustering
1414 * pagefaults into a processes address space. It is a "cousin"
1415 * of vm_map_pmap_enter, except it runs at page fault time instead
1419 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1420 int backward, int forward, bool obj_locked)
1423 vm_map_entry_t entry;
1424 vm_object_t backing_object, lobject;
1425 vm_offset_t addr, starta;
1430 pmap = fs->map->pmap;
1431 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1436 if (addra < backward * PAGE_SIZE) {
1437 starta = entry->start;
1439 starta = addra - backward * PAGE_SIZE;
1440 if (starta < entry->start)
1441 starta = entry->start;
1445 * Generate the sequence of virtual addresses that are candidates for
1446 * prefaulting in an outward spiral from the faulting virtual address,
1447 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1448 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1449 * If the candidate address doesn't have a backing physical page, then
1450 * the loop immediately terminates.
1452 for (i = 0; i < 2 * imax(backward, forward); i++) {
1453 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1455 if (addr > addra + forward * PAGE_SIZE)
1458 if (addr < starta || addr >= entry->end)
1461 if (!pmap_is_prefaultable(pmap, addr))
1464 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1465 lobject = entry->object.vm_object;
1467 VM_OBJECT_RLOCK(lobject);
1468 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1469 lobject->type == OBJT_DEFAULT &&
1470 (backing_object = lobject->backing_object) != NULL) {
1471 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1472 0, ("vm_fault_prefault: unaligned object offset"));
1473 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1474 VM_OBJECT_RLOCK(backing_object);
1475 if (!obj_locked || lobject != entry->object.vm_object)
1476 VM_OBJECT_RUNLOCK(lobject);
1477 lobject = backing_object;
1480 if (!obj_locked || lobject != entry->object.vm_object)
1481 VM_OBJECT_RUNLOCK(lobject);
1484 if (m->valid == VM_PAGE_BITS_ALL &&
1485 (m->flags & PG_FICTITIOUS) == 0)
1486 pmap_enter_quick(pmap, addr, m, entry->protection);
1487 if (!obj_locked || lobject != entry->object.vm_object)
1488 VM_OBJECT_RUNLOCK(lobject);
1493 * Hold each of the physical pages that are mapped by the specified range of
1494 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1495 * and allow the specified types of access, "prot". If all of the implied
1496 * pages are successfully held, then the number of held pages is returned
1497 * together with pointers to those pages in the array "ma". However, if any
1498 * of the pages cannot be held, -1 is returned.
1501 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1502 vm_prot_t prot, vm_page_t *ma, int max_count)
1504 vm_offset_t end, va;
1507 boolean_t pmap_failed;
1511 end = round_page(addr + len);
1512 addr = trunc_page(addr);
1515 * Check for illegal addresses.
1517 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1520 if (atop(end - addr) > max_count)
1521 panic("vm_fault_quick_hold_pages: count > max_count");
1522 count = atop(end - addr);
1525 * Most likely, the physical pages are resident in the pmap, so it is
1526 * faster to try pmap_extract_and_hold() first.
1528 pmap_failed = FALSE;
1529 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1530 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1533 else if ((prot & VM_PROT_WRITE) != 0 &&
1534 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1536 * Explicitly dirty the physical page. Otherwise, the
1537 * caller's changes may go unnoticed because they are
1538 * performed through an unmanaged mapping or by a DMA
1541 * The object lock is not held here.
1542 * See vm_page_clear_dirty_mask().
1549 * One or more pages could not be held by the pmap. Either no
1550 * page was mapped at the specified virtual address or that
1551 * mapping had insufficient permissions. Attempt to fault in
1552 * and hold these pages.
1554 * If vm_fault_disable_pagefaults() was called,
1555 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1556 * acquire MD VM locks, which means we must not call
1557 * vm_fault_hold(). Some (out of tree) callers mark
1558 * too wide a code area with vm_fault_disable_pagefaults()
1559 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1560 * the proper behaviour explicitly.
1562 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1563 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1565 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1566 if (*mp == NULL && vm_fault_hold(map, va, prot,
1567 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1572 for (mp = ma; mp < ma + count; mp++)
1575 vm_page_unhold(*mp);
1576 vm_page_unlock(*mp);
1583 * vm_fault_copy_entry
1585 * Create new shadow object backing dst_entry with private copy of
1586 * all underlying pages. When src_entry is equal to dst_entry,
1587 * function implements COW for wired-down map entry. Otherwise,
1588 * it forks wired entry into dst_map.
1590 * In/out conditions:
1591 * The source and destination maps must be locked for write.
1592 * The source map entry must be wired down (or be a sharing map
1593 * entry corresponding to a main map entry that is wired down).
1596 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1597 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1598 vm_ooffset_t *fork_charge)
1600 vm_object_t backing_object, dst_object, object, src_object;
1601 vm_pindex_t dst_pindex, pindex, src_pindex;
1602 vm_prot_t access, prot;
1612 upgrade = src_entry == dst_entry;
1613 access = prot = dst_entry->protection;
1615 src_object = src_entry->object.vm_object;
1616 src_pindex = OFF_TO_IDX(src_entry->offset);
1618 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1619 dst_object = src_object;
1620 vm_object_reference(dst_object);
1623 * Create the top-level object for the destination entry. (Doesn't
1624 * actually shadow anything - we copy the pages directly.)
1626 dst_object = vm_object_allocate(OBJT_DEFAULT,
1627 atop(dst_entry->end - dst_entry->start));
1628 #if VM_NRESERVLEVEL > 0
1629 dst_object->flags |= OBJ_COLORED;
1630 dst_object->pg_color = atop(dst_entry->start);
1634 VM_OBJECT_WLOCK(dst_object);
1635 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1636 ("vm_fault_copy_entry: vm_object not NULL"));
1637 if (src_object != dst_object) {
1638 dst_object->domain = src_object->domain;
1639 dst_entry->object.vm_object = dst_object;
1640 dst_entry->offset = 0;
1641 dst_object->charge = dst_entry->end - dst_entry->start;
1643 if (fork_charge != NULL) {
1644 KASSERT(dst_entry->cred == NULL,
1645 ("vm_fault_copy_entry: leaked swp charge"));
1646 dst_object->cred = curthread->td_ucred;
1647 crhold(dst_object->cred);
1648 *fork_charge += dst_object->charge;
1649 } else if (dst_object->cred == NULL) {
1650 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1652 dst_object->cred = dst_entry->cred;
1653 dst_entry->cred = NULL;
1657 * If not an upgrade, then enter the mappings in the pmap as
1658 * read and/or execute accesses. Otherwise, enter them as
1661 * A writeable large page mapping is only created if all of
1662 * the constituent small page mappings are modified. Marking
1663 * PTEs as modified on inception allows promotion to happen
1664 * without taking potentially large number of soft faults.
1667 access &= ~VM_PROT_WRITE;
1670 * Loop through all of the virtual pages within the entry's
1671 * range, copying each page from the source object to the
1672 * destination object. Since the source is wired, those pages
1673 * must exist. In contrast, the destination is pageable.
1674 * Since the destination object doesn't share any backing storage
1675 * with the source object, all of its pages must be dirtied,
1676 * regardless of whether they can be written.
1678 for (vaddr = dst_entry->start, dst_pindex = 0;
1679 vaddr < dst_entry->end;
1680 vaddr += PAGE_SIZE, dst_pindex++) {
1683 * Find the page in the source object, and copy it in.
1684 * Because the source is wired down, the page will be
1687 if (src_object != dst_object)
1688 VM_OBJECT_RLOCK(src_object);
1689 object = src_object;
1690 pindex = src_pindex + dst_pindex;
1691 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1692 (backing_object = object->backing_object) != NULL) {
1694 * Unless the source mapping is read-only or
1695 * it is presently being upgraded from
1696 * read-only, the first object in the shadow
1697 * chain should provide all of the pages. In
1698 * other words, this loop body should never be
1699 * executed when the source mapping is already
1702 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1704 ("vm_fault_copy_entry: main object missing page"));
1706 VM_OBJECT_RLOCK(backing_object);
1707 pindex += OFF_TO_IDX(object->backing_object_offset);
1708 if (object != dst_object)
1709 VM_OBJECT_RUNLOCK(object);
1710 object = backing_object;
1712 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1714 if (object != dst_object) {
1716 * Allocate a page in the destination object.
1718 dst_m = vm_page_alloc(dst_object, (src_object ==
1719 dst_object ? src_pindex : 0) + dst_pindex,
1721 if (dst_m == NULL) {
1722 VM_OBJECT_WUNLOCK(dst_object);
1723 VM_OBJECT_RUNLOCK(object);
1724 vm_wait(dst_object);
1725 VM_OBJECT_WLOCK(dst_object);
1728 pmap_copy_page(src_m, dst_m);
1729 VM_OBJECT_RUNLOCK(object);
1730 dst_m->valid = VM_PAGE_BITS_ALL;
1731 dst_m->dirty = VM_PAGE_BITS_ALL;
1734 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1736 vm_page_xbusy(dst_m);
1737 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1738 ("invalid dst page %p", dst_m));
1740 VM_OBJECT_WUNLOCK(dst_object);
1743 * Enter it in the pmap. If a wired, copy-on-write
1744 * mapping is being replaced by a write-enabled
1745 * mapping, then wire that new mapping.
1747 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1748 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1751 * Mark it no longer busy, and put it on the active list.
1753 VM_OBJECT_WLOCK(dst_object);
1756 if (src_m != dst_m) {
1757 vm_page_lock(src_m);
1758 vm_page_unwire(src_m, PQ_INACTIVE);
1759 vm_page_unlock(src_m);
1760 vm_page_lock(dst_m);
1761 vm_page_wire(dst_m);
1762 vm_page_unlock(dst_m);
1764 KASSERT(dst_m->wire_count > 0,
1765 ("dst_m %p is not wired", dst_m));
1768 vm_page_lock(dst_m);
1769 vm_page_activate(dst_m);
1770 vm_page_unlock(dst_m);
1772 vm_page_xunbusy(dst_m);
1774 VM_OBJECT_WUNLOCK(dst_object);
1776 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1777 vm_object_deallocate(src_object);
1782 * Block entry into the machine-independent layer's page fault handler by
1783 * the calling thread. Subsequent calls to vm_fault() by that thread will
1784 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1785 * spurious page faults.
1788 vm_fault_disable_pagefaults(void)
1791 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1795 vm_fault_enable_pagefaults(int save)
1798 curthread_pflags_restore(save);