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(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
274 __ARM_ARCH >= 6) || defined(__i386__)) && VM_NRESERVLEVEL > 0
280 MPASS(fs->vp == NULL);
281 m = vm_page_lookup(fs->first_object, fs->first_pindex);
282 /* A busy page can be mapped for read|execute access. */
283 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
284 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
285 return (KERN_FAILURE);
288 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
289 __ARM_ARCH >= 6) || defined(__i386__)) && VM_NRESERVLEVEL > 0
290 if ((m->flags & PG_FICTITIOUS) == 0 &&
291 (m_super = vm_reserv_to_superpage(m)) != NULL &&
292 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
293 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
294 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
295 (pagesizes[m_super->psind] - 1)) &&
296 pmap_ps_enabled(fs->map->pmap)) {
297 flags = PS_ALL_VALID;
298 if ((prot & VM_PROT_WRITE) != 0) {
300 * Create a superpage mapping allowing write access
301 * only if none of the constituent pages are busy and
302 * all of them are already dirty (except possibly for
303 * the page that was faulted on).
305 flags |= PS_NONE_BUSY;
306 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
307 flags |= PS_ALL_DIRTY;
309 if (vm_page_ps_test(m_super, flags, m)) {
311 psind = m_super->psind;
312 vaddr = rounddown2(vaddr, pagesizes[psind]);
313 /* Preset the modified bit for dirty superpages. */
314 if ((flags & PS_ALL_DIRTY) != 0)
315 fault_type |= VM_PROT_WRITE;
319 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
320 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
321 if (rv != KERN_SUCCESS)
323 vm_fault_fill_hold(m_hold, m);
324 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
325 if (psind == 0 && !wired)
326 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
327 VM_OBJECT_RUNLOCK(fs->first_object);
328 vm_map_lookup_done(fs->map, fs->entry);
329 curthread->td_ru.ru_minflt++;
330 return (KERN_SUCCESS);
334 vm_fault_restore_map_lock(struct faultstate *fs)
337 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
338 MPASS(fs->first_object->paging_in_progress > 0);
340 if (!vm_map_trylock_read(fs->map)) {
341 VM_OBJECT_WUNLOCK(fs->first_object);
342 vm_map_lock_read(fs->map);
343 VM_OBJECT_WLOCK(fs->first_object);
345 fs->lookup_still_valid = true;
349 vm_fault_populate_check_page(vm_page_t m)
353 * Check each page to ensure that the pager is obeying the
354 * interface: the page must be installed in the object, fully
355 * valid, and exclusively busied.
358 MPASS(m->valid == VM_PAGE_BITS_ALL);
359 MPASS(vm_page_xbusied(m));
363 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
369 VM_OBJECT_ASSERT_WLOCKED(object);
370 MPASS(first <= last);
371 for (pidx = first, m = vm_page_lookup(object, pidx);
372 pidx <= last; pidx++, m = vm_page_next(m)) {
373 vm_fault_populate_check_page(m);
375 vm_page_deactivate(m);
382 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
383 int fault_flags, boolean_t wired, vm_page_t *m_hold)
388 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
389 int i, npages, psind, rv;
391 MPASS(fs->object == fs->first_object);
392 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
393 MPASS(fs->first_object->paging_in_progress > 0);
394 MPASS(fs->first_object->backing_object == NULL);
395 MPASS(fs->lookup_still_valid);
397 pager_first = OFF_TO_IDX(fs->entry->offset);
398 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
403 * Call the pager (driver) populate() method.
405 * There is no guarantee that the method will be called again
406 * if the current fault is for read, and a future fault is
407 * for write. Report the entry's maximum allowed protection
410 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
411 fault_type, fs->entry->max_protection, &pager_first, &pager_last);
413 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
414 if (rv == VM_PAGER_BAD) {
416 * VM_PAGER_BAD is the backdoor for a pager to request
417 * normal fault handling.
419 vm_fault_restore_map_lock(fs);
420 if (fs->map->timestamp != fs->map_generation)
421 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
422 return (KERN_NOT_RECEIVER);
424 if (rv != VM_PAGER_OK)
425 return (KERN_FAILURE); /* AKA SIGSEGV */
427 /* Ensure that the driver is obeying the interface. */
428 MPASS(pager_first <= pager_last);
429 MPASS(fs->first_pindex <= pager_last);
430 MPASS(fs->first_pindex >= pager_first);
431 MPASS(pager_last < fs->first_object->size);
433 vm_fault_restore_map_lock(fs);
434 if (fs->map->timestamp != fs->map_generation) {
435 vm_fault_populate_cleanup(fs->first_object, pager_first,
437 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
441 * The map is unchanged after our last unlock. Process the fault.
443 * The range [pager_first, pager_last] that is given to the
444 * pager is only a hint. The pager may populate any range
445 * within the object that includes the requested page index.
446 * In case the pager expanded the range, clip it to fit into
449 map_first = OFF_TO_IDX(fs->entry->offset);
450 if (map_first > pager_first) {
451 vm_fault_populate_cleanup(fs->first_object, pager_first,
453 pager_first = map_first;
455 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
456 if (map_last < pager_last) {
457 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
459 pager_last = map_last;
461 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
463 pidx += npages, m = vm_page_next(&m[npages - 1])) {
464 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
465 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
466 __ARM_ARCH >= 6) || defined(__i386__)
468 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
469 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
470 !pmap_ps_enabled(fs->map->pmap)))
475 npages = atop(pagesizes[psind]);
476 for (i = 0; i < npages; i++) {
477 vm_fault_populate_check_page(&m[i]);
478 vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
481 VM_OBJECT_WUNLOCK(fs->first_object);
482 pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | (wired ?
483 PMAP_ENTER_WIRED : 0), psind);
484 VM_OBJECT_WLOCK(fs->first_object);
486 for (i = 0; i < npages; i++) {
487 vm_page_change_lock(&m[i], &m_mtx);
488 if ((fault_flags & VM_FAULT_WIRE) != 0)
491 vm_page_activate(&m[i]);
492 if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
496 vm_page_xunbusy_maybelocked(&m[i]);
501 curthread->td_ru.ru_majflt++;
502 return (KERN_SUCCESS);
508 * Handle a page fault occurring at the given address,
509 * requiring the given permissions, in the map specified.
510 * If successful, the page is inserted into the
511 * associated physical map.
513 * NOTE: the given address should be truncated to the
514 * proper page address.
516 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
517 * a standard error specifying why the fault is fatal is returned.
519 * The map in question must be referenced, and remains so.
520 * Caller may hold no locks.
523 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
530 if ((td->td_pflags & TDP_NOFAULTING) != 0)
531 return (KERN_PROTECTION_FAILURE);
533 if (map != kernel_map && KTRPOINT(td, KTR_FAULT))
534 ktrfault(vaddr, fault_type);
536 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags,
539 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND))
546 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
547 int fault_flags, vm_page_t *m_hold)
549 struct faultstate fs;
551 struct domainset *dset;
552 vm_object_t next_object, retry_object;
553 vm_offset_t e_end, e_start;
554 vm_pindex_t retry_pindex;
555 vm_prot_t prot, retry_prot;
556 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
557 int locked, nera, result, rv;
559 boolean_t wired; /* Passed by reference. */
560 bool dead, hardfault, is_first_object_locked;
562 VM_CNT_INC(v_vm_faults);
571 * Find the backing store object and offset into it to begin the
575 result = vm_map_lookup(&fs.map, vaddr, fault_type |
576 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
577 &fs.first_pindex, &prot, &wired);
578 if (result != KERN_SUCCESS) {
583 fs.map_generation = fs.map->timestamp;
585 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
586 panic("%s: fault on nofault entry, addr: %#lx",
587 __func__, (u_long)vaddr);
590 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
591 fs.entry->wiring_thread != curthread) {
592 vm_map_unlock_read(fs.map);
594 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
595 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
597 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
598 vm_map_unlock_and_wait(fs.map, 0);
600 vm_map_unlock(fs.map);
604 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
607 fault_type = prot | (fault_type & VM_PROT_COPY);
609 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
610 ("!wired && VM_FAULT_WIRE"));
613 * Try to avoid lock contention on the top-level object through
614 * special-case handling of some types of page faults, specifically,
615 * those that are both (1) mapping an existing page from the top-
616 * level object and (2) not having to mark that object as containing
617 * dirty pages. Under these conditions, a read lock on the top-level
618 * object suffices, allowing multiple page faults of a similar type to
619 * run in parallel on the same top-level object.
621 if (fs.vp == NULL /* avoid locked vnode leak */ &&
622 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
623 /* avoid calling vm_object_set_writeable_dirty() */
624 ((prot & VM_PROT_WRITE) == 0 ||
625 (fs.first_object->type != OBJT_VNODE &&
626 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
627 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
628 VM_OBJECT_RLOCK(fs.first_object);
629 if ((prot & VM_PROT_WRITE) == 0 ||
630 (fs.first_object->type != OBJT_VNODE &&
631 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
632 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
633 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
634 fault_flags, wired, m_hold);
635 if (rv == KERN_SUCCESS)
638 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
639 VM_OBJECT_RUNLOCK(fs.first_object);
640 VM_OBJECT_WLOCK(fs.first_object);
643 VM_OBJECT_WLOCK(fs.first_object);
647 * Make a reference to this object to prevent its disposal while we
648 * are messing with it. Once we have the reference, the map is free
649 * to be diddled. Since objects reference their shadows (and copies),
650 * they will stay around as well.
652 * Bump the paging-in-progress count to prevent size changes (e.g.
653 * truncation operations) during I/O.
655 vm_object_reference_locked(fs.first_object);
656 vm_object_pip_add(fs.first_object, 1);
658 fs.lookup_still_valid = true;
663 * Search for the page at object/offset.
665 fs.object = fs.first_object;
666 fs.pindex = fs.first_pindex;
669 * If the object is marked for imminent termination,
670 * we retry here, since the collapse pass has raced
671 * with us. Otherwise, if we see terminally dead
672 * object, return fail.
674 if ((fs.object->flags & OBJ_DEAD) != 0) {
675 dead = fs.object->type == OBJT_DEAD;
676 unlock_and_deallocate(&fs);
678 return (KERN_PROTECTION_FAILURE);
684 * See if page is resident
686 fs.m = vm_page_lookup(fs.object, fs.pindex);
689 * Wait/Retry if the page is busy. We have to do this
690 * if the page is either exclusive or shared busy
691 * because the vm_pager may be using read busy for
692 * pageouts (and even pageins if it is the vnode
693 * pager), and we could end up trying to pagein and
694 * pageout the same page simultaneously.
696 * We can theoretically allow the busy case on a read
697 * fault if the page is marked valid, but since such
698 * pages are typically already pmap'd, putting that
699 * special case in might be more effort then it is
700 * worth. We cannot under any circumstances mess
701 * around with a shared busied page except, perhaps,
704 if (vm_page_busied(fs.m)) {
706 * Reference the page before unlocking and
707 * sleeping so that the page daemon is less
708 * likely to reclaim it.
710 vm_page_aflag_set(fs.m, PGA_REFERENCED);
711 if (fs.object != fs.first_object) {
712 if (!VM_OBJECT_TRYWLOCK(
714 VM_OBJECT_WUNLOCK(fs.object);
715 VM_OBJECT_WLOCK(fs.first_object);
716 VM_OBJECT_WLOCK(fs.object);
718 vm_page_lock(fs.first_m);
719 vm_page_free(fs.first_m);
720 vm_page_unlock(fs.first_m);
721 vm_object_pip_wakeup(fs.first_object);
722 VM_OBJECT_WUNLOCK(fs.first_object);
726 if (fs.m == vm_page_lookup(fs.object,
728 vm_page_sleep_if_busy(fs.m, "vmpfw");
730 vm_object_pip_wakeup(fs.object);
731 VM_OBJECT_WUNLOCK(fs.object);
732 VM_CNT_INC(v_intrans);
733 vm_object_deallocate(fs.first_object);
738 * Mark page busy for other processes, and the
739 * pagedaemon. If it still isn't completely valid
740 * (readable), jump to readrest, else break-out ( we
744 if (fs.m->valid != VM_PAGE_BITS_ALL)
746 break; /* break to PAGE HAS BEEN FOUND */
748 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
751 * Page is not resident. If the pager might contain the page
752 * or this is the beginning of the search, allocate a new
753 * page. (Default objects are zero-fill, so there is no real
756 if (fs.object->type != OBJT_DEFAULT ||
757 fs.object == fs.first_object) {
758 if (fs.pindex >= fs.object->size) {
759 unlock_and_deallocate(&fs);
760 return (KERN_PROTECTION_FAILURE);
763 if (fs.object == fs.first_object &&
764 (fs.first_object->flags & OBJ_POPULATE) != 0 &&
765 fs.first_object->shadow_count == 0) {
766 rv = vm_fault_populate(&fs, prot, fault_type,
767 fault_flags, wired, m_hold);
771 unlock_and_deallocate(&fs);
773 case KERN_RESOURCE_SHORTAGE:
774 unlock_and_deallocate(&fs);
776 case KERN_NOT_RECEIVER:
778 * Pager's populate() method
779 * returned VM_PAGER_BAD.
783 panic("inconsistent return codes");
788 * Allocate a new page for this object/offset pair.
790 * Unlocked read of the p_flag is harmless. At
791 * worst, the P_KILLED might be not observed
792 * there, and allocation can fail, causing
793 * restart and new reading of the p_flag.
795 dset = fs.object->domain.dr_policy;
797 dset = curthread->td_domain.dr_policy;
798 if (!vm_page_count_severe_set(&dset->ds_mask) ||
800 #if VM_NRESERVLEVEL > 0
801 vm_object_color(fs.object, atop(vaddr) -
804 alloc_req = P_KILLED(curproc) ?
805 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
806 if (fs.object->type != OBJT_VNODE &&
807 fs.object->backing_object == NULL)
808 alloc_req |= VM_ALLOC_ZERO;
809 fs.m = vm_page_alloc(fs.object, fs.pindex,
813 unlock_and_deallocate(&fs);
821 * At this point, we have either allocated a new page or found
822 * an existing page that is only partially valid.
824 * We hold a reference on the current object and the page is
829 * If the pager for the current object might have the page,
830 * then determine the number of additional pages to read and
831 * potentially reprioritize previously read pages for earlier
832 * reclamation. These operations should only be performed
833 * once per page fault. Even if the current pager doesn't
834 * have the page, the number of additional pages to read will
835 * apply to subsequent objects in the shadow chain.
837 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
838 !P_KILLED(curproc)) {
839 KASSERT(fs.lookup_still_valid, ("map unlocked"));
840 era = fs.entry->read_ahead;
841 behavior = vm_map_entry_behavior(fs.entry);
842 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
844 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
845 nera = VM_FAULT_READ_AHEAD_MAX;
846 if (vaddr == fs.entry->next_read)
847 vm_fault_dontneed(&fs, vaddr, nera);
848 } else if (vaddr == fs.entry->next_read) {
850 * This is a sequential fault. Arithmetically
851 * increase the requested number of pages in
852 * the read-ahead window. The requested
853 * number of pages is "# of sequential faults
854 * x (read ahead min + 1) + read ahead min"
856 nera = VM_FAULT_READ_AHEAD_MIN;
859 if (nera > VM_FAULT_READ_AHEAD_MAX)
860 nera = VM_FAULT_READ_AHEAD_MAX;
862 if (era == VM_FAULT_READ_AHEAD_MAX)
863 vm_fault_dontneed(&fs, vaddr, nera);
866 * This is a non-sequential fault.
872 * A read lock on the map suffices to update
873 * the read ahead count safely.
875 fs.entry->read_ahead = nera;
879 * Prepare for unlocking the map. Save the map
880 * entry's start and end addresses, which are used to
881 * optimize the size of the pager operation below.
882 * Even if the map entry's addresses change after
883 * unlocking the map, using the saved addresses is
886 e_start = fs.entry->start;
887 e_end = fs.entry->end;
891 * Call the pager to retrieve the page if there is a chance
892 * that the pager has it, and potentially retrieve additional
893 * pages at the same time.
895 if (fs.object->type != OBJT_DEFAULT) {
897 * Release the map lock before locking the vnode or
898 * sleeping in the pager. (If the current object has
899 * a shadow, then an earlier iteration of this loop
900 * may have already unlocked the map.)
904 if (fs.object->type == OBJT_VNODE &&
905 (vp = fs.object->handle) != fs.vp) {
907 * Perform an unlock in case the desired vnode
908 * changed while the map was unlocked during a
913 locked = VOP_ISLOCKED(vp);
914 if (locked != LK_EXCLUSIVE)
918 * We must not sleep acquiring the vnode lock
919 * while we have the page exclusive busied or
920 * the object's paging-in-progress count
921 * incremented. Otherwise, we could deadlock.
923 error = vget(vp, locked | LK_CANRECURSE |
924 LK_NOWAIT, curthread);
928 unlock_and_deallocate(&fs);
929 error = vget(vp, locked | LK_RETRY |
930 LK_CANRECURSE, curthread);
934 ("vm_fault: vget failed"));
939 KASSERT(fs.vp == NULL || !fs.map->system_map,
940 ("vm_fault: vnode-backed object mapped by system map"));
943 * Page in the requested page and hint the pager,
944 * that it may bring up surrounding pages.
946 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
951 /* Is this a sequential fault? */
957 * Request a cluster of pages that is
958 * aligned to a VM_FAULT_READ_DEFAULT
959 * page offset boundary within the
960 * object. Alignment to a page offset
961 * boundary is more likely to coincide
962 * with the underlying file system
963 * block than alignment to a virtual
966 cluster_offset = fs.pindex %
967 VM_FAULT_READ_DEFAULT;
968 behind = ulmin(cluster_offset,
969 atop(vaddr - e_start));
970 ahead = VM_FAULT_READ_DEFAULT - 1 -
973 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
975 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
977 if (rv == VM_PAGER_OK) {
978 faultcount = behind + 1 + ahead;
980 break; /* break to PAGE HAS BEEN FOUND */
982 if (rv == VM_PAGER_ERROR)
983 printf("vm_fault: pager read error, pid %d (%s)\n",
984 curproc->p_pid, curproc->p_comm);
987 * If an I/O error occurred or the requested page was
988 * outside the range of the pager, clean up and return
991 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
993 if (fs.m->wire_count == 0)
996 vm_page_xunbusy_maybelocked(fs.m);
997 vm_page_unlock(fs.m);
999 unlock_and_deallocate(&fs);
1000 return (rv == VM_PAGER_ERROR ? KERN_FAILURE :
1001 KERN_PROTECTION_FAILURE);
1005 * The requested page does not exist at this object/
1006 * offset. Remove the invalid page from the object,
1007 * waking up anyone waiting for it, and continue on to
1008 * the next object. However, if this is the top-level
1009 * object, we must leave the busy page in place to
1010 * prevent another process from rushing past us, and
1011 * inserting the page in that object at the same time
1014 if (fs.object != fs.first_object) {
1016 if (fs.m->wire_count == 0)
1019 vm_page_xunbusy_maybelocked(fs.m);
1020 vm_page_unlock(fs.m);
1026 * We get here if the object has default pager (or unwiring)
1027 * or the pager doesn't have the page.
1029 if (fs.object == fs.first_object)
1033 * Move on to the next object. Lock the next object before
1034 * unlocking the current one.
1036 next_object = fs.object->backing_object;
1037 if (next_object == NULL) {
1039 * If there's no object left, fill the page in the top
1040 * object with zeros.
1042 if (fs.object != fs.first_object) {
1043 vm_object_pip_wakeup(fs.object);
1044 VM_OBJECT_WUNLOCK(fs.object);
1046 fs.object = fs.first_object;
1047 fs.pindex = fs.first_pindex;
1049 VM_OBJECT_WLOCK(fs.object);
1054 * Zero the page if necessary and mark it valid.
1056 if ((fs.m->flags & PG_ZERO) == 0) {
1057 pmap_zero_page(fs.m);
1059 VM_CNT_INC(v_ozfod);
1062 fs.m->valid = VM_PAGE_BITS_ALL;
1063 /* Don't try to prefault neighboring pages. */
1065 break; /* break to PAGE HAS BEEN FOUND */
1067 KASSERT(fs.object != next_object,
1068 ("object loop %p", next_object));
1069 VM_OBJECT_WLOCK(next_object);
1070 vm_object_pip_add(next_object, 1);
1071 if (fs.object != fs.first_object)
1072 vm_object_pip_wakeup(fs.object);
1074 OFF_TO_IDX(fs.object->backing_object_offset);
1075 VM_OBJECT_WUNLOCK(fs.object);
1076 fs.object = next_object;
1080 vm_page_assert_xbusied(fs.m);
1083 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1088 * If the page is being written, but isn't already owned by the
1089 * top-level object, we have to copy it into a new page owned by the
1092 if (fs.object != fs.first_object) {
1094 * We only really need to copy if we want to write it.
1096 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1098 * This allows pages to be virtually copied from a
1099 * backing_object into the first_object, where the
1100 * backing object has no other refs to it, and cannot
1101 * gain any more refs. Instead of a bcopy, we just
1102 * move the page from the backing object to the
1103 * first object. Note that we must mark the page
1104 * dirty in the first object so that it will go out
1105 * to swap when needed.
1107 is_first_object_locked = false;
1110 * Only one shadow object
1112 (fs.object->shadow_count == 1) &&
1114 * No COW refs, except us
1116 (fs.object->ref_count == 1) &&
1118 * No one else can look this object up
1120 (fs.object->handle == NULL) &&
1122 * No other ways to look the object up
1124 ((fs.object->type == OBJT_DEFAULT) ||
1125 (fs.object->type == OBJT_SWAP)) &&
1126 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1128 * We don't chase down the shadow chain
1130 fs.object == fs.first_object->backing_object) {
1132 vm_page_dequeue(fs.m);
1133 vm_page_remove(fs.m);
1134 vm_page_unlock(fs.m);
1135 vm_page_lock(fs.first_m);
1136 vm_page_replace_checked(fs.m, fs.first_object,
1137 fs.first_pindex, fs.first_m);
1138 vm_page_free(fs.first_m);
1139 vm_page_unlock(fs.first_m);
1140 vm_page_dirty(fs.m);
1141 #if VM_NRESERVLEVEL > 0
1143 * Rename the reservation.
1145 vm_reserv_rename(fs.m, fs.first_object,
1146 fs.object, OFF_TO_IDX(
1147 fs.first_object->backing_object_offset));
1150 * Removing the page from the backing object
1153 vm_page_xbusy(fs.m);
1156 VM_CNT_INC(v_cow_optim);
1159 * Oh, well, lets copy it.
1161 pmap_copy_page(fs.m, fs.first_m);
1162 fs.first_m->valid = VM_PAGE_BITS_ALL;
1163 if (wired && (fault_flags &
1164 VM_FAULT_WIRE) == 0) {
1165 vm_page_lock(fs.first_m);
1166 vm_page_wire(fs.first_m);
1167 vm_page_unlock(fs.first_m);
1170 vm_page_unwire(fs.m, PQ_INACTIVE);
1171 vm_page_unlock(fs.m);
1174 * We no longer need the old page or object.
1179 * fs.object != fs.first_object due to above
1182 vm_object_pip_wakeup(fs.object);
1183 VM_OBJECT_WUNLOCK(fs.object);
1186 * We only try to prefault read-only mappings to the
1187 * neighboring pages when this copy-on-write fault is
1188 * a hard fault. In other cases, trying to prefault
1189 * is typically wasted effort.
1191 if (faultcount == 0)
1195 * Only use the new page below...
1197 fs.object = fs.first_object;
1198 fs.pindex = fs.first_pindex;
1200 if (!is_first_object_locked)
1201 VM_OBJECT_WLOCK(fs.object);
1202 VM_CNT_INC(v_cow_faults);
1203 curthread->td_cow++;
1205 prot &= ~VM_PROT_WRITE;
1210 * We must verify that the maps have not changed since our last
1213 if (!fs.lookup_still_valid) {
1214 if (!vm_map_trylock_read(fs.map)) {
1216 unlock_and_deallocate(&fs);
1219 fs.lookup_still_valid = true;
1220 if (fs.map->timestamp != fs.map_generation) {
1221 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1222 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1225 * If we don't need the page any longer, put it on the inactive
1226 * list (the easiest thing to do here). If no one needs it,
1227 * pageout will grab it eventually.
1229 if (result != KERN_SUCCESS) {
1231 unlock_and_deallocate(&fs);
1234 * If retry of map lookup would have blocked then
1235 * retry fault from start.
1237 if (result == KERN_FAILURE)
1241 if ((retry_object != fs.first_object) ||
1242 (retry_pindex != fs.first_pindex)) {
1244 unlock_and_deallocate(&fs);
1249 * Check whether the protection has changed or the object has
1250 * been copied while we left the map unlocked. Changing from
1251 * read to write permission is OK - we leave the page
1252 * write-protected, and catch the write fault. Changing from
1253 * write to read permission means that we can't mark the page
1254 * write-enabled after all.
1257 fault_type &= retry_prot;
1260 unlock_and_deallocate(&fs);
1264 /* Reassert because wired may have changed. */
1265 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1266 ("!wired && VM_FAULT_WIRE"));
1271 * If the page was filled by a pager, save the virtual address that
1272 * should be faulted on next under a sequential access pattern to the
1273 * map entry. A read lock on the map suffices to update this address
1277 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1279 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1280 vm_page_assert_xbusied(fs.m);
1283 * Page must be completely valid or it is not fit to
1284 * map into user space. vm_pager_get_pages() ensures this.
1286 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1287 ("vm_fault: page %p partially invalid", fs.m));
1288 VM_OBJECT_WUNLOCK(fs.object);
1291 * Put this page into the physical map. We had to do the unlock above
1292 * because pmap_enter() may sleep. We don't put the page
1293 * back on the active queue until later so that the pageout daemon
1294 * won't find it (yet).
1296 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1297 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1298 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1300 vm_fault_prefault(&fs, vaddr,
1301 faultcount > 0 ? behind : PFBAK,
1302 faultcount > 0 ? ahead : PFFOR, false);
1303 VM_OBJECT_WLOCK(fs.object);
1307 * If the page is not wired down, then put it where the pageout daemon
1310 if ((fault_flags & VM_FAULT_WIRE) != 0)
1313 vm_page_activate(fs.m);
1314 if (m_hold != NULL) {
1318 vm_page_unlock(fs.m);
1319 vm_page_xunbusy(fs.m);
1322 * Unlock everything, and return
1324 unlock_and_deallocate(&fs);
1326 VM_CNT_INC(v_io_faults);
1327 curthread->td_ru.ru_majflt++;
1329 if (racct_enable && fs.object->type == OBJT_VNODE) {
1331 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1332 racct_add_force(curproc, RACCT_WRITEBPS,
1333 PAGE_SIZE + behind * PAGE_SIZE);
1334 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1336 racct_add_force(curproc, RACCT_READBPS,
1337 PAGE_SIZE + ahead * PAGE_SIZE);
1338 racct_add_force(curproc, RACCT_READIOPS, 1);
1340 PROC_UNLOCK(curproc);
1344 curthread->td_ru.ru_minflt++;
1346 return (KERN_SUCCESS);
1350 * Speed up the reclamation of pages that precede the faulting pindex within
1351 * the first object of the shadow chain. Essentially, perform the equivalent
1352 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1353 * the faulting pindex by the cluster size when the pages read by vm_fault()
1354 * cross a cluster-size boundary. The cluster size is the greater of the
1355 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1357 * When "fs->first_object" is a shadow object, the pages in the backing object
1358 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1359 * function must only be concerned with pages in the first object.
1362 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1364 vm_map_entry_t entry;
1365 vm_object_t first_object, object;
1366 vm_offset_t end, start;
1367 vm_page_t m, m_next;
1368 vm_pindex_t pend, pstart;
1371 object = fs->object;
1372 VM_OBJECT_ASSERT_WLOCKED(object);
1373 first_object = fs->first_object;
1374 if (first_object != object) {
1375 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1376 VM_OBJECT_WUNLOCK(object);
1377 VM_OBJECT_WLOCK(first_object);
1378 VM_OBJECT_WLOCK(object);
1381 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1382 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1383 size = VM_FAULT_DONTNEED_MIN;
1384 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1385 size = pagesizes[1];
1386 end = rounddown2(vaddr, size);
1387 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1388 (entry = fs->entry)->start < end) {
1389 if (end - entry->start < size)
1390 start = entry->start;
1393 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1394 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1396 m_next = vm_page_find_least(first_object, pstart);
1397 pend = OFF_TO_IDX(entry->offset) + atop(end -
1399 while ((m = m_next) != NULL && m->pindex < pend) {
1400 m_next = TAILQ_NEXT(m, listq);
1401 if (m->valid != VM_PAGE_BITS_ALL ||
1406 * Don't clear PGA_REFERENCED, since it would
1407 * likely represent a reference by a different
1410 * Typically, at this point, prefetched pages
1411 * are still in the inactive queue. Only
1412 * pages that triggered page faults are in the
1416 if (!vm_page_inactive(m))
1417 vm_page_deactivate(m);
1422 if (first_object != object)
1423 VM_OBJECT_WUNLOCK(first_object);
1427 * vm_fault_prefault provides a quick way of clustering
1428 * pagefaults into a processes address space. It is a "cousin"
1429 * of vm_map_pmap_enter, except it runs at page fault time instead
1433 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1434 int backward, int forward, bool obj_locked)
1437 vm_map_entry_t entry;
1438 vm_object_t backing_object, lobject;
1439 vm_offset_t addr, starta;
1444 pmap = fs->map->pmap;
1445 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1450 if (addra < backward * PAGE_SIZE) {
1451 starta = entry->start;
1453 starta = addra - backward * PAGE_SIZE;
1454 if (starta < entry->start)
1455 starta = entry->start;
1459 * Generate the sequence of virtual addresses that are candidates for
1460 * prefaulting in an outward spiral from the faulting virtual address,
1461 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1462 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1463 * If the candidate address doesn't have a backing physical page, then
1464 * the loop immediately terminates.
1466 for (i = 0; i < 2 * imax(backward, forward); i++) {
1467 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1469 if (addr > addra + forward * PAGE_SIZE)
1472 if (addr < starta || addr >= entry->end)
1475 if (!pmap_is_prefaultable(pmap, addr))
1478 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1479 lobject = entry->object.vm_object;
1481 VM_OBJECT_RLOCK(lobject);
1482 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1483 lobject->type == OBJT_DEFAULT &&
1484 (backing_object = lobject->backing_object) != NULL) {
1485 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1486 0, ("vm_fault_prefault: unaligned object offset"));
1487 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1488 VM_OBJECT_RLOCK(backing_object);
1489 if (!obj_locked || lobject != entry->object.vm_object)
1490 VM_OBJECT_RUNLOCK(lobject);
1491 lobject = backing_object;
1494 if (!obj_locked || lobject != entry->object.vm_object)
1495 VM_OBJECT_RUNLOCK(lobject);
1498 if (m->valid == VM_PAGE_BITS_ALL &&
1499 (m->flags & PG_FICTITIOUS) == 0)
1500 pmap_enter_quick(pmap, addr, m, entry->protection);
1501 if (!obj_locked || lobject != entry->object.vm_object)
1502 VM_OBJECT_RUNLOCK(lobject);
1507 * Hold each of the physical pages that are mapped by the specified range of
1508 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1509 * and allow the specified types of access, "prot". If all of the implied
1510 * pages are successfully held, then the number of held pages is returned
1511 * together with pointers to those pages in the array "ma". However, if any
1512 * of the pages cannot be held, -1 is returned.
1515 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1516 vm_prot_t prot, vm_page_t *ma, int max_count)
1518 vm_offset_t end, va;
1521 boolean_t pmap_failed;
1525 end = round_page(addr + len);
1526 addr = trunc_page(addr);
1529 * Check for illegal addresses.
1531 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1534 if (atop(end - addr) > max_count)
1535 panic("vm_fault_quick_hold_pages: count > max_count");
1536 count = atop(end - addr);
1539 * Most likely, the physical pages are resident in the pmap, so it is
1540 * faster to try pmap_extract_and_hold() first.
1542 pmap_failed = FALSE;
1543 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1544 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1547 else if ((prot & VM_PROT_WRITE) != 0 &&
1548 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1550 * Explicitly dirty the physical page. Otherwise, the
1551 * caller's changes may go unnoticed because they are
1552 * performed through an unmanaged mapping or by a DMA
1555 * The object lock is not held here.
1556 * See vm_page_clear_dirty_mask().
1563 * One or more pages could not be held by the pmap. Either no
1564 * page was mapped at the specified virtual address or that
1565 * mapping had insufficient permissions. Attempt to fault in
1566 * and hold these pages.
1568 * If vm_fault_disable_pagefaults() was called,
1569 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1570 * acquire MD VM locks, which means we must not call
1571 * vm_fault_hold(). Some (out of tree) callers mark
1572 * too wide a code area with vm_fault_disable_pagefaults()
1573 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1574 * the proper behaviour explicitly.
1576 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1577 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1579 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1580 if (*mp == NULL && vm_fault_hold(map, va, prot,
1581 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1586 for (mp = ma; mp < ma + count; mp++)
1589 vm_page_unhold(*mp);
1590 vm_page_unlock(*mp);
1597 * vm_fault_copy_entry
1599 * Create new shadow object backing dst_entry with private copy of
1600 * all underlying pages. When src_entry is equal to dst_entry,
1601 * function implements COW for wired-down map entry. Otherwise,
1602 * it forks wired entry into dst_map.
1604 * In/out conditions:
1605 * The source and destination maps must be locked for write.
1606 * The source map entry must be wired down (or be a sharing map
1607 * entry corresponding to a main map entry that is wired down).
1610 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1611 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1612 vm_ooffset_t *fork_charge)
1614 vm_object_t backing_object, dst_object, object, src_object;
1615 vm_pindex_t dst_pindex, pindex, src_pindex;
1616 vm_prot_t access, prot;
1626 upgrade = src_entry == dst_entry;
1627 access = prot = dst_entry->protection;
1629 src_object = src_entry->object.vm_object;
1630 src_pindex = OFF_TO_IDX(src_entry->offset);
1632 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1633 dst_object = src_object;
1634 vm_object_reference(dst_object);
1637 * Create the top-level object for the destination entry. (Doesn't
1638 * actually shadow anything - we copy the pages directly.)
1640 dst_object = vm_object_allocate(OBJT_DEFAULT,
1641 atop(dst_entry->end - dst_entry->start));
1642 #if VM_NRESERVLEVEL > 0
1643 dst_object->flags |= OBJ_COLORED;
1644 dst_object->pg_color = atop(dst_entry->start);
1646 dst_object->domain = src_object->domain;
1647 dst_object->charge = dst_entry->end - dst_entry->start;
1650 VM_OBJECT_WLOCK(dst_object);
1651 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1652 ("vm_fault_copy_entry: vm_object not NULL"));
1653 if (src_object != dst_object) {
1654 dst_entry->object.vm_object = dst_object;
1655 dst_entry->offset = 0;
1657 if (fork_charge != NULL) {
1658 KASSERT(dst_entry->cred == NULL,
1659 ("vm_fault_copy_entry: leaked swp charge"));
1660 dst_object->cred = curthread->td_ucred;
1661 crhold(dst_object->cred);
1662 *fork_charge += dst_object->charge;
1663 } else if ((dst_object->type == OBJT_DEFAULT ||
1664 dst_object->type == OBJT_SWAP) &&
1665 dst_object->cred == NULL) {
1666 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1668 dst_object->cred = dst_entry->cred;
1669 dst_entry->cred = NULL;
1673 * If not an upgrade, then enter the mappings in the pmap as
1674 * read and/or execute accesses. Otherwise, enter them as
1677 * A writeable large page mapping is only created if all of
1678 * the constituent small page mappings are modified. Marking
1679 * PTEs as modified on inception allows promotion to happen
1680 * without taking potentially large number of soft faults.
1683 access &= ~VM_PROT_WRITE;
1686 * Loop through all of the virtual pages within the entry's
1687 * range, copying each page from the source object to the
1688 * destination object. Since the source is wired, those pages
1689 * must exist. In contrast, the destination is pageable.
1690 * Since the destination object doesn't share any backing storage
1691 * with the source object, all of its pages must be dirtied,
1692 * regardless of whether they can be written.
1694 for (vaddr = dst_entry->start, dst_pindex = 0;
1695 vaddr < dst_entry->end;
1696 vaddr += PAGE_SIZE, dst_pindex++) {
1699 * Find the page in the source object, and copy it in.
1700 * Because the source is wired down, the page will be
1703 if (src_object != dst_object)
1704 VM_OBJECT_RLOCK(src_object);
1705 object = src_object;
1706 pindex = src_pindex + dst_pindex;
1707 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1708 (backing_object = object->backing_object) != NULL) {
1710 * Unless the source mapping is read-only or
1711 * it is presently being upgraded from
1712 * read-only, the first object in the shadow
1713 * chain should provide all of the pages. In
1714 * other words, this loop body should never be
1715 * executed when the source mapping is already
1718 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1720 ("vm_fault_copy_entry: main object missing page"));
1722 VM_OBJECT_RLOCK(backing_object);
1723 pindex += OFF_TO_IDX(object->backing_object_offset);
1724 if (object != dst_object)
1725 VM_OBJECT_RUNLOCK(object);
1726 object = backing_object;
1728 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1730 if (object != dst_object) {
1732 * Allocate a page in the destination object.
1734 dst_m = vm_page_alloc(dst_object, (src_object ==
1735 dst_object ? src_pindex : 0) + dst_pindex,
1737 if (dst_m == NULL) {
1738 VM_OBJECT_WUNLOCK(dst_object);
1739 VM_OBJECT_RUNLOCK(object);
1740 vm_wait(dst_object);
1741 VM_OBJECT_WLOCK(dst_object);
1744 pmap_copy_page(src_m, dst_m);
1745 VM_OBJECT_RUNLOCK(object);
1746 dst_m->valid = VM_PAGE_BITS_ALL;
1747 dst_m->dirty = VM_PAGE_BITS_ALL;
1750 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1752 if (dst_m->pindex >= dst_object->size)
1754 * We are upgrading. Index can occur
1755 * out of bounds if the object type is
1756 * vnode and the file was truncated.
1759 vm_page_xbusy(dst_m);
1760 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL,
1761 ("invalid dst page %p", dst_m));
1763 VM_OBJECT_WUNLOCK(dst_object);
1766 * Enter it in the pmap. If a wired, copy-on-write
1767 * mapping is being replaced by a write-enabled
1768 * mapping, then wire that new mapping.
1770 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1771 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1774 * Mark it no longer busy, and put it on the active list.
1776 VM_OBJECT_WLOCK(dst_object);
1779 if (src_m != dst_m) {
1780 vm_page_lock(src_m);
1781 vm_page_unwire(src_m, PQ_INACTIVE);
1782 vm_page_unlock(src_m);
1783 vm_page_lock(dst_m);
1784 vm_page_wire(dst_m);
1785 vm_page_unlock(dst_m);
1787 KASSERT(dst_m->wire_count > 0,
1788 ("dst_m %p is not wired", dst_m));
1791 vm_page_lock(dst_m);
1792 vm_page_activate(dst_m);
1793 vm_page_unlock(dst_m);
1795 vm_page_xunbusy(dst_m);
1797 VM_OBJECT_WUNLOCK(dst_object);
1799 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1800 vm_object_deallocate(src_object);
1805 * Block entry into the machine-independent layer's page fault handler by
1806 * the calling thread. Subsequent calls to vm_fault() by that thread will
1807 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1808 * spurious page faults.
1811 vm_fault_disable_pagefaults(void)
1814 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1818 vm_fault_enable_pagefaults(int save)
1821 curthread_pflags_restore(save);