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>
87 #include <sys/mutex.h>
89 #include <sys/racct.h>
90 #include <sys/refcount.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/signalvar.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysent.h>
96 #include <sys/vmmeter.h>
97 #include <sys/vnode.h>
99 #include <sys/ktrace.h>
103 #include <vm/vm_param.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_extern.h>
112 #include <vm/vm_reserv.h>
117 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
118 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
120 #define VM_FAULT_DONTNEED_MIN 1048576
127 vm_object_t first_object;
128 vm_pindex_t first_pindex;
130 vm_map_entry_t entry;
132 bool lookup_still_valid;
136 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
138 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
139 int backward, int forward, bool obj_locked);
141 static int vm_pfault_oom_attempts = 3;
142 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
143 &vm_pfault_oom_attempts, 0,
144 "Number of page allocation attempts in page fault handler before it "
145 "triggers OOM handling");
147 static int vm_pfault_oom_wait = 10;
148 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
149 &vm_pfault_oom_wait, 0,
150 "Number of seconds to wait for free pages before retrying "
151 "the page fault handler");
154 release_page(struct faultstate *fs)
157 vm_page_xunbusy(fs->m);
159 vm_page_deactivate(fs->m);
160 vm_page_unlock(fs->m);
165 unlock_map(struct faultstate *fs)
168 if (fs->lookup_still_valid) {
169 vm_map_lookup_done(fs->map, fs->entry);
170 fs->lookup_still_valid = false;
175 unlock_vp(struct faultstate *fs)
178 if (fs->vp != NULL) {
185 unlock_and_deallocate(struct faultstate *fs)
188 vm_object_pip_wakeup(fs->object);
189 VM_OBJECT_WUNLOCK(fs->object);
190 if (fs->object != fs->first_object) {
191 VM_OBJECT_WLOCK(fs->first_object);
192 vm_page_free(fs->first_m);
193 vm_object_pip_wakeup(fs->first_object);
194 VM_OBJECT_WUNLOCK(fs->first_object);
197 vm_object_deallocate(fs->first_object);
203 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
204 vm_prot_t fault_type, int fault_flags, bool set_wd)
208 if (((prot & VM_PROT_WRITE) == 0 &&
209 (fault_flags & VM_FAULT_DIRTY) == 0) ||
210 (m->oflags & VPO_UNMANAGED) != 0)
213 VM_OBJECT_ASSERT_LOCKED(m->object);
214 VM_PAGE_OBJECT_BUSY_ASSERT(m);
216 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
217 (fault_flags & VM_FAULT_WIRE) == 0) ||
218 (fault_flags & VM_FAULT_DIRTY) != 0;
221 vm_object_set_writeable_dirty(m->object);
224 * If two callers of vm_fault_dirty() with set_wd ==
225 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
226 * flag set, other with flag clear, race, it is
227 * possible for the no-NOSYNC thread to see m->dirty
228 * != 0 and not clear PGA_NOSYNC. Take vm_page lock
229 * around manipulation of PGA_NOSYNC and
230 * vm_page_dirty() call to avoid the race.
235 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
236 * if the page is already dirty to prevent data written with
237 * the expectation of being synced from not being synced.
238 * Likewise if this entry does not request NOSYNC then make
239 * sure the page isn't marked NOSYNC. Applications sharing
240 * data should use the same flags to avoid ping ponging.
242 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
244 vm_page_aflag_set(m, PGA_NOSYNC);
247 vm_page_aflag_clear(m, PGA_NOSYNC);
251 * If the fault is a write, we know that this page is being
252 * written NOW so dirty it explicitly to save on
253 * pmap_is_modified() calls later.
255 * Also, since the page is now dirty, we can possibly tell
256 * the pager to release any swap backing the page. Calling
257 * the pager requires a write lock on the object.
264 vm_pager_page_unswapped(m);
268 * Unlocks fs.first_object and fs.map on success.
271 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
272 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
275 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
276 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
283 MPASS(fs->vp == NULL);
284 vm_object_busy(fs->first_object);
285 m = vm_page_lookup(fs->first_object, fs->first_pindex);
286 /* A busy page can be mapped for read|execute access. */
287 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
288 vm_page_busied(m)) || !vm_page_all_valid(m)) {
294 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
295 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
297 if ((m->flags & PG_FICTITIOUS) == 0 &&
298 (m_super = vm_reserv_to_superpage(m)) != NULL &&
299 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
300 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
301 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
302 (pagesizes[m_super->psind] - 1)) && !wired &&
303 pmap_ps_enabled(fs->map->pmap)) {
304 flags = PS_ALL_VALID;
305 if ((prot & VM_PROT_WRITE) != 0) {
307 * Create a superpage mapping allowing write access
308 * only if none of the constituent pages are busy and
309 * all of them are already dirty (except possibly for
310 * the page that was faulted on).
312 flags |= PS_NONE_BUSY;
313 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
314 flags |= PS_ALL_DIRTY;
316 if (vm_page_ps_test(m_super, flags, m)) {
318 psind = m_super->psind;
319 vaddr = rounddown2(vaddr, pagesizes[psind]);
320 /* Preset the modified bit for dirty superpages. */
321 if ((flags & PS_ALL_DIRTY) != 0)
322 fault_type |= VM_PROT_WRITE;
326 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
327 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
328 if (rv != KERN_SUCCESS)
330 if (m_hold != NULL) {
334 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
335 if (psind == 0 && !wired)
336 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
337 VM_OBJECT_RUNLOCK(fs->first_object);
338 vm_map_lookup_done(fs->map, fs->entry);
339 curthread->td_ru.ru_minflt++;
342 vm_object_unbusy(fs->first_object);
347 vm_fault_restore_map_lock(struct faultstate *fs)
350 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
351 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
353 if (!vm_map_trylock_read(fs->map)) {
354 VM_OBJECT_WUNLOCK(fs->first_object);
355 vm_map_lock_read(fs->map);
356 VM_OBJECT_WLOCK(fs->first_object);
358 fs->lookup_still_valid = true;
362 vm_fault_populate_check_page(vm_page_t m)
366 * Check each page to ensure that the pager is obeying the
367 * interface: the page must be installed in the object, fully
368 * valid, and exclusively busied.
371 MPASS(vm_page_all_valid(m));
372 MPASS(vm_page_xbusied(m));
376 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
382 VM_OBJECT_ASSERT_WLOCKED(object);
383 MPASS(first <= last);
384 for (pidx = first, m = vm_page_lookup(object, pidx);
385 pidx <= last; pidx++, m = vm_page_next(m)) {
386 vm_fault_populate_check_page(m);
388 vm_page_deactivate(m);
395 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
396 int fault_flags, boolean_t wired, vm_page_t *m_hold)
401 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
402 int i, npages, psind, rv;
404 MPASS(fs->object == fs->first_object);
405 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
406 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
407 MPASS(fs->first_object->backing_object == NULL);
408 MPASS(fs->lookup_still_valid);
410 pager_first = OFF_TO_IDX(fs->entry->offset);
411 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
416 * Call the pager (driver) populate() method.
418 * There is no guarantee that the method will be called again
419 * if the current fault is for read, and a future fault is
420 * for write. Report the entry's maximum allowed protection
423 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
424 fault_type, fs->entry->max_protection, &pager_first, &pager_last);
426 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
427 if (rv == VM_PAGER_BAD) {
429 * VM_PAGER_BAD is the backdoor for a pager to request
430 * normal fault handling.
432 vm_fault_restore_map_lock(fs);
433 if (fs->map->timestamp != fs->map_generation)
434 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
435 return (KERN_NOT_RECEIVER);
437 if (rv != VM_PAGER_OK)
438 return (KERN_FAILURE); /* AKA SIGSEGV */
440 /* Ensure that the driver is obeying the interface. */
441 MPASS(pager_first <= pager_last);
442 MPASS(fs->first_pindex <= pager_last);
443 MPASS(fs->first_pindex >= pager_first);
444 MPASS(pager_last < fs->first_object->size);
446 vm_fault_restore_map_lock(fs);
447 if (fs->map->timestamp != fs->map_generation) {
448 vm_fault_populate_cleanup(fs->first_object, pager_first,
450 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
454 * The map is unchanged after our last unlock. Process the fault.
456 * The range [pager_first, pager_last] that is given to the
457 * pager is only a hint. The pager may populate any range
458 * within the object that includes the requested page index.
459 * In case the pager expanded the range, clip it to fit into
462 map_first = OFF_TO_IDX(fs->entry->offset);
463 if (map_first > pager_first) {
464 vm_fault_populate_cleanup(fs->first_object, pager_first,
466 pager_first = map_first;
468 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
469 if (map_last < pager_last) {
470 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
472 pager_last = map_last;
474 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
476 pidx += npages, m = vm_page_next(&m[npages - 1])) {
477 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
478 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
479 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
481 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
482 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
483 !pmap_ps_enabled(fs->map->pmap) || wired))
488 npages = atop(pagesizes[psind]);
489 for (i = 0; i < npages; i++) {
490 vm_fault_populate_check_page(&m[i]);
491 vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
494 VM_OBJECT_WUNLOCK(fs->first_object);
495 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
496 (wired ? PMAP_ENTER_WIRED : 0), psind);
497 #if defined(__amd64__)
498 if (psind > 0 && rv == KERN_FAILURE) {
499 for (i = 0; i < npages; i++) {
500 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
501 &m[i], prot, fault_type |
502 (wired ? PMAP_ENTER_WIRED : 0), 0);
503 MPASS(rv == KERN_SUCCESS);
507 MPASS(rv == KERN_SUCCESS);
509 VM_OBJECT_WLOCK(fs->first_object);
511 for (i = 0; i < npages; i++) {
512 if ((fault_flags & VM_FAULT_WIRE) != 0) {
515 vm_page_change_lock(&m[i], &m_mtx);
516 vm_page_activate(&m[i]);
518 if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
522 vm_page_xunbusy(&m[i]);
527 curthread->td_ru.ru_majflt++;
528 return (KERN_SUCCESS);
531 static int prot_fault_translation;
532 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
533 &prot_fault_translation, 0,
534 "Control signal to deliver on protection fault");
536 /* compat definition to keep common code for signal translation */
537 #define UCODE_PAGEFLT 12
539 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
545 * Handle a page fault occurring at the given address,
546 * requiring the given permissions, in the map specified.
547 * If successful, the page is inserted into the
548 * associated physical map.
550 * NOTE: the given address should be truncated to the
551 * proper page address.
553 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
554 * a standard error specifying why the fault is fatal is returned.
556 * The map in question must be referenced, and remains so.
557 * Caller may hold no locks.
560 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
561 int fault_flags, int *signo, int *ucode)
565 MPASS(signo == NULL || ucode != NULL);
567 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
568 ktrfault(vaddr, fault_type);
570 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
572 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
573 result == KERN_INVALID_ADDRESS ||
574 result == KERN_RESOURCE_SHORTAGE ||
575 result == KERN_PROTECTION_FAILURE ||
576 result == KERN_OUT_OF_BOUNDS,
577 ("Unexpected Mach error %d from vm_fault()", result));
579 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
582 if (result != KERN_SUCCESS && signo != NULL) {
585 case KERN_INVALID_ADDRESS:
587 *ucode = SEGV_MAPERR;
589 case KERN_RESOURCE_SHORTAGE:
593 case KERN_OUT_OF_BOUNDS:
597 case KERN_PROTECTION_FAILURE:
598 if (prot_fault_translation == 0) {
600 * Autodetect. This check also covers
601 * the images without the ABI-tag ELF
604 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
605 curproc->p_osrel >= P_OSREL_SIGSEGV) {
607 *ucode = SEGV_ACCERR;
610 *ucode = UCODE_PAGEFLT;
612 } else if (prot_fault_translation == 1) {
613 /* Always compat mode. */
615 *ucode = UCODE_PAGEFLT;
617 /* Always SIGSEGV mode. */
619 *ucode = SEGV_ACCERR;
623 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
632 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
633 int fault_flags, vm_page_t *m_hold)
635 struct faultstate fs;
637 struct domainset *dset;
638 vm_object_t next_object, retry_object;
639 vm_offset_t e_end, e_start;
640 vm_pindex_t retry_pindex;
641 vm_prot_t prot, retry_prot;
642 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount;
643 int locked, nera, oom, result, rv;
645 boolean_t wired; /* Passed by reference. */
646 bool dead, hardfault, is_first_object_locked;
648 VM_CNT_INC(v_vm_faults);
650 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
651 return (KERN_PROTECTION_FAILURE);
663 * Find the backing store object and offset into it to begin the
667 result = vm_map_lookup(&fs.map, vaddr, fault_type |
668 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
669 &fs.first_pindex, &prot, &wired);
670 if (result != KERN_SUCCESS) {
675 fs.map_generation = fs.map->timestamp;
677 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
678 panic("%s: fault on nofault entry, addr: %#lx",
679 __func__, (u_long)vaddr);
682 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
683 fs.entry->wiring_thread != curthread) {
684 vm_map_unlock_read(fs.map);
686 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
687 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
689 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
690 vm_map_unlock_and_wait(fs.map, 0);
692 vm_map_unlock(fs.map);
696 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
699 fault_type = prot | (fault_type & VM_PROT_COPY);
701 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
702 ("!wired && VM_FAULT_WIRE"));
705 * Try to avoid lock contention on the top-level object through
706 * special-case handling of some types of page faults, specifically,
707 * those that are both (1) mapping an existing page from the top-
708 * level object and (2) not having to mark that object as containing
709 * dirty pages. Under these conditions, a read lock on the top-level
710 * object suffices, allowing multiple page faults of a similar type to
711 * run in parallel on the same top-level object.
713 if (fs.vp == NULL /* avoid locked vnode leak */ &&
714 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
715 /* avoid calling vm_object_set_writeable_dirty() */
716 ((prot & VM_PROT_WRITE) == 0 ||
717 (fs.first_object->type != OBJT_VNODE &&
718 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
719 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
720 VM_OBJECT_RLOCK(fs.first_object);
721 if ((prot & VM_PROT_WRITE) == 0 ||
722 (fs.first_object->type != OBJT_VNODE &&
723 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
724 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
725 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
726 fault_flags, wired, m_hold);
727 if (rv == KERN_SUCCESS)
730 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
731 VM_OBJECT_RUNLOCK(fs.first_object);
732 VM_OBJECT_WLOCK(fs.first_object);
735 VM_OBJECT_WLOCK(fs.first_object);
739 * Make a reference to this object to prevent its disposal while we
740 * are messing with it. Once we have the reference, the map is free
741 * to be diddled. Since objects reference their shadows (and copies),
742 * they will stay around as well.
744 * Bump the paging-in-progress count to prevent size changes (e.g.
745 * truncation operations) during I/O.
747 vm_object_reference_locked(fs.first_object);
748 vm_object_pip_add(fs.first_object, 1);
750 fs.lookup_still_valid = true;
755 * Search for the page at object/offset.
757 fs.object = fs.first_object;
758 fs.pindex = fs.first_pindex;
761 * If the object is marked for imminent termination,
762 * we retry here, since the collapse pass has raced
763 * with us. Otherwise, if we see terminally dead
764 * object, return fail.
766 if ((fs.object->flags & OBJ_DEAD) != 0) {
767 dead = fs.object->type == OBJT_DEAD;
768 unlock_and_deallocate(&fs);
770 return (KERN_PROTECTION_FAILURE);
776 * See if page is resident
778 fs.m = vm_page_lookup(fs.object, fs.pindex);
781 * Wait/Retry if the page is busy. We have to do this
782 * if the page is either exclusive or shared busy
783 * because the vm_pager may be using read busy for
784 * pageouts (and even pageins if it is the vnode
785 * pager), and we could end up trying to pagein and
786 * pageout the same page simultaneously.
788 * We can theoretically allow the busy case on a read
789 * fault if the page is marked valid, but since such
790 * pages are typically already pmap'd, putting that
791 * special case in might be more effort then it is
792 * worth. We cannot under any circumstances mess
793 * around with a shared busied page except, perhaps,
796 if (vm_page_tryxbusy(fs.m) == 0) {
798 * Reference the page before unlocking and
799 * sleeping so that the page daemon is less
800 * likely to reclaim it.
802 vm_page_aflag_set(fs.m, PGA_REFERENCED);
803 if (fs.object != fs.first_object) {
804 if (!VM_OBJECT_TRYWLOCK(
806 VM_OBJECT_WUNLOCK(fs.object);
807 VM_OBJECT_WLOCK(fs.first_object);
808 VM_OBJECT_WLOCK(fs.object);
810 vm_page_free(fs.first_m);
811 vm_object_pip_wakeup(fs.first_object);
812 VM_OBJECT_WUNLOCK(fs.first_object);
816 if (fs.m == vm_page_lookup(fs.object,
818 vm_page_sleep_if_busy(fs.m, "vmpfw");
820 vm_object_pip_wakeup(fs.object);
821 VM_OBJECT_WUNLOCK(fs.object);
822 VM_CNT_INC(v_intrans);
823 vm_object_deallocate(fs.first_object);
828 * The page is marked busy for other processes and the
829 * pagedaemon. If it still isn't completely valid
830 * (readable), jump to readrest, else break-out ( we
833 if (!vm_page_all_valid(fs.m))
835 break; /* break to PAGE HAS BEEN FOUND */
837 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
840 * Page is not resident. If the pager might contain the page
841 * or this is the beginning of the search, allocate a new
842 * page. (Default objects are zero-fill, so there is no real
845 if (fs.object->type != OBJT_DEFAULT ||
846 fs.object == fs.first_object) {
847 if (fs.pindex >= fs.object->size) {
848 unlock_and_deallocate(&fs);
849 return (KERN_OUT_OF_BOUNDS);
852 if (fs.object == fs.first_object &&
853 (fs.first_object->flags & OBJ_POPULATE) != 0 &&
854 fs.first_object->shadow_count == 0) {
855 rv = vm_fault_populate(&fs, prot, fault_type,
856 fault_flags, wired, m_hold);
860 unlock_and_deallocate(&fs);
862 case KERN_RESOURCE_SHORTAGE:
863 unlock_and_deallocate(&fs);
865 case KERN_NOT_RECEIVER:
867 * Pager's populate() method
868 * returned VM_PAGER_BAD.
872 panic("inconsistent return codes");
877 * Allocate a new page for this object/offset pair.
879 * Unlocked read of the p_flag is harmless. At
880 * worst, the P_KILLED might be not observed
881 * there, and allocation can fail, causing
882 * restart and new reading of the p_flag.
884 dset = fs.object->domain.dr_policy;
886 dset = curthread->td_domain.dr_policy;
887 if (!vm_page_count_severe_set(&dset->ds_mask) ||
889 #if VM_NRESERVLEVEL > 0
890 vm_object_color(fs.object, atop(vaddr) -
893 alloc_req = P_KILLED(curproc) ?
894 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
895 if (fs.object->type != OBJT_VNODE &&
896 fs.object->backing_object == NULL)
897 alloc_req |= VM_ALLOC_ZERO;
898 fs.m = vm_page_alloc(fs.object, fs.pindex,
902 unlock_and_deallocate(&fs);
903 if (vm_pfault_oom_attempts < 0 ||
904 oom < vm_pfault_oom_attempts) {
907 vm_pfault_oom_wait * hz);
912 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
913 curproc->p_pid, curproc->p_comm);
914 vm_pageout_oom(VM_OOM_MEM_PF);
921 * At this point, we have either allocated a new page or found
922 * an existing page that is only partially valid.
924 * We hold a reference on the current object and the page is
929 * If the pager for the current object might have the page,
930 * then determine the number of additional pages to read and
931 * potentially reprioritize previously read pages for earlier
932 * reclamation. These operations should only be performed
933 * once per page fault. Even if the current pager doesn't
934 * have the page, the number of additional pages to read will
935 * apply to subsequent objects in the shadow chain.
937 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
938 !P_KILLED(curproc)) {
939 KASSERT(fs.lookup_still_valid, ("map unlocked"));
940 era = fs.entry->read_ahead;
941 behavior = vm_map_entry_behavior(fs.entry);
942 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
944 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
945 nera = VM_FAULT_READ_AHEAD_MAX;
946 if (vaddr == fs.entry->next_read)
947 vm_fault_dontneed(&fs, vaddr, nera);
948 } else if (vaddr == fs.entry->next_read) {
950 * This is a sequential fault. Arithmetically
951 * increase the requested number of pages in
952 * the read-ahead window. The requested
953 * number of pages is "# of sequential faults
954 * x (read ahead min + 1) + read ahead min"
956 nera = VM_FAULT_READ_AHEAD_MIN;
959 if (nera > VM_FAULT_READ_AHEAD_MAX)
960 nera = VM_FAULT_READ_AHEAD_MAX;
962 if (era == VM_FAULT_READ_AHEAD_MAX)
963 vm_fault_dontneed(&fs, vaddr, nera);
966 * This is a non-sequential fault.
972 * A read lock on the map suffices to update
973 * the read ahead count safely.
975 fs.entry->read_ahead = nera;
979 * Prepare for unlocking the map. Save the map
980 * entry's start and end addresses, which are used to
981 * optimize the size of the pager operation below.
982 * Even if the map entry's addresses change after
983 * unlocking the map, using the saved addresses is
986 e_start = fs.entry->start;
987 e_end = fs.entry->end;
991 * Call the pager to retrieve the page if there is a chance
992 * that the pager has it, and potentially retrieve additional
993 * pages at the same time.
995 if (fs.object->type != OBJT_DEFAULT) {
997 * Release the map lock before locking the vnode or
998 * sleeping in the pager. (If the current object has
999 * a shadow, then an earlier iteration of this loop
1000 * may have already unlocked the map.)
1004 if (fs.object->type == OBJT_VNODE &&
1005 (vp = fs.object->handle) != fs.vp) {
1007 * Perform an unlock in case the desired vnode
1008 * changed while the map was unlocked during a
1013 locked = VOP_ISLOCKED(vp);
1014 if (locked != LK_EXCLUSIVE)
1018 * We must not sleep acquiring the vnode lock
1019 * while we have the page exclusive busied or
1020 * the object's paging-in-progress count
1021 * incremented. Otherwise, we could deadlock.
1023 error = vget(vp, locked | LK_CANRECURSE |
1024 LK_NOWAIT, curthread);
1028 unlock_and_deallocate(&fs);
1029 error = vget(vp, locked | LK_RETRY |
1030 LK_CANRECURSE, curthread);
1034 ("vm_fault: vget failed"));
1039 KASSERT(fs.vp == NULL || !fs.map->system_map,
1040 ("vm_fault: vnode-backed object mapped by system map"));
1043 * Page in the requested page and hint the pager,
1044 * that it may bring up surrounding pages.
1046 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1047 P_KILLED(curproc)) {
1051 /* Is this a sequential fault? */
1057 * Request a cluster of pages that is
1058 * aligned to a VM_FAULT_READ_DEFAULT
1059 * page offset boundary within the
1060 * object. Alignment to a page offset
1061 * boundary is more likely to coincide
1062 * with the underlying file system
1063 * block than alignment to a virtual
1066 cluster_offset = fs.pindex %
1067 VM_FAULT_READ_DEFAULT;
1068 behind = ulmin(cluster_offset,
1069 atop(vaddr - e_start));
1070 ahead = VM_FAULT_READ_DEFAULT - 1 -
1073 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1075 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1077 if (rv == VM_PAGER_OK) {
1078 faultcount = behind + 1 + ahead;
1080 break; /* break to PAGE HAS BEEN FOUND */
1082 if (rv == VM_PAGER_ERROR)
1083 printf("vm_fault: pager read error, pid %d (%s)\n",
1084 curproc->p_pid, curproc->p_comm);
1087 * If an I/O error occurred or the requested page was
1088 * outside the range of the pager, clean up and return
1091 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1092 if (!vm_page_wired(fs.m))
1095 vm_page_xunbusy(fs.m);
1097 unlock_and_deallocate(&fs);
1098 return (KERN_OUT_OF_BOUNDS);
1102 * The requested page does not exist at this object/
1103 * offset. Remove the invalid page from the object,
1104 * waking up anyone waiting for it, and continue on to
1105 * the next object. However, if this is the top-level
1106 * object, we must leave the busy page in place to
1107 * prevent another process from rushing past us, and
1108 * inserting the page in that object at the same time
1111 if (fs.object != fs.first_object) {
1112 if (!vm_page_wired(fs.m))
1115 vm_page_xunbusy(fs.m);
1121 * We get here if the object has default pager (or unwiring)
1122 * or the pager doesn't have the page.
1124 if (fs.object == fs.first_object)
1128 * Move on to the next object. Lock the next object before
1129 * unlocking the current one.
1131 next_object = fs.object->backing_object;
1132 if (next_object == NULL) {
1134 * If there's no object left, fill the page in the top
1135 * object with zeros.
1137 if (fs.object != fs.first_object) {
1138 vm_object_pip_wakeup(fs.object);
1139 VM_OBJECT_WUNLOCK(fs.object);
1141 fs.object = fs.first_object;
1142 fs.pindex = fs.first_pindex;
1144 VM_OBJECT_WLOCK(fs.object);
1149 * Zero the page if necessary and mark it valid.
1151 if ((fs.m->flags & PG_ZERO) == 0) {
1152 pmap_zero_page(fs.m);
1154 VM_CNT_INC(v_ozfod);
1157 vm_page_valid(fs.m);
1158 /* Don't try to prefault neighboring pages. */
1160 break; /* break to PAGE HAS BEEN FOUND */
1162 KASSERT(fs.object != next_object,
1163 ("object loop %p", next_object));
1164 VM_OBJECT_WLOCK(next_object);
1165 vm_object_pip_add(next_object, 1);
1166 if (fs.object != fs.first_object)
1167 vm_object_pip_wakeup(fs.object);
1169 OFF_TO_IDX(fs.object->backing_object_offset);
1170 VM_OBJECT_WUNLOCK(fs.object);
1171 fs.object = next_object;
1175 vm_page_assert_xbusied(fs.m);
1178 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1183 * If the page is being written, but isn't already owned by the
1184 * top-level object, we have to copy it into a new page owned by the
1187 if (fs.object != fs.first_object) {
1189 * We only really need to copy if we want to write it.
1191 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1193 * This allows pages to be virtually copied from a
1194 * backing_object into the first_object, where the
1195 * backing object has no other refs to it, and cannot
1196 * gain any more refs. Instead of a bcopy, we just
1197 * move the page from the backing object to the
1198 * first object. Note that we must mark the page
1199 * dirty in the first object so that it will go out
1200 * to swap when needed.
1202 is_first_object_locked = false;
1205 * Only one shadow object
1207 (fs.object->shadow_count == 1) &&
1209 * No COW refs, except us
1211 (fs.object->ref_count == 1) &&
1213 * No one else can look this object up
1215 (fs.object->handle == NULL) &&
1217 * No other ways to look the object up
1219 ((fs.object->type == OBJT_DEFAULT) ||
1220 (fs.object->type == OBJT_SWAP)) &&
1221 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1223 * We don't chase down the shadow chain
1225 fs.object == fs.first_object->backing_object) {
1227 (void)vm_page_remove(fs.m);
1228 vm_page_replace_checked(fs.m, fs.first_object,
1229 fs.first_pindex, fs.first_m);
1230 vm_page_free(fs.first_m);
1231 vm_page_dirty(fs.m);
1232 #if VM_NRESERVLEVEL > 0
1234 * Rename the reservation.
1236 vm_reserv_rename(fs.m, fs.first_object,
1237 fs.object, OFF_TO_IDX(
1238 fs.first_object->backing_object_offset));
1242 VM_CNT_INC(v_cow_optim);
1245 * Oh, well, lets copy it.
1247 pmap_copy_page(fs.m, fs.first_m);
1248 vm_page_valid(fs.first_m);
1249 if (wired && (fault_flags &
1250 VM_FAULT_WIRE) == 0) {
1251 vm_page_wire(fs.first_m);
1252 vm_page_unwire(fs.m, PQ_INACTIVE);
1255 * We no longer need the old page or object.
1260 * fs.object != fs.first_object due to above
1263 vm_object_pip_wakeup(fs.object);
1264 VM_OBJECT_WUNLOCK(fs.object);
1267 * We only try to prefault read-only mappings to the
1268 * neighboring pages when this copy-on-write fault is
1269 * a hard fault. In other cases, trying to prefault
1270 * is typically wasted effort.
1272 if (faultcount == 0)
1276 * Only use the new page below...
1278 fs.object = fs.first_object;
1279 fs.pindex = fs.first_pindex;
1281 if (!is_first_object_locked)
1282 VM_OBJECT_WLOCK(fs.object);
1283 VM_CNT_INC(v_cow_faults);
1284 curthread->td_cow++;
1286 prot &= ~VM_PROT_WRITE;
1291 * We must verify that the maps have not changed since our last
1294 if (!fs.lookup_still_valid) {
1295 if (!vm_map_trylock_read(fs.map)) {
1297 unlock_and_deallocate(&fs);
1300 fs.lookup_still_valid = true;
1301 if (fs.map->timestamp != fs.map_generation) {
1302 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1303 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1306 * If we don't need the page any longer, put it on the inactive
1307 * list (the easiest thing to do here). If no one needs it,
1308 * pageout will grab it eventually.
1310 if (result != KERN_SUCCESS) {
1312 unlock_and_deallocate(&fs);
1315 * If retry of map lookup would have blocked then
1316 * retry fault from start.
1318 if (result == KERN_FAILURE)
1322 if ((retry_object != fs.first_object) ||
1323 (retry_pindex != fs.first_pindex)) {
1325 unlock_and_deallocate(&fs);
1330 * Check whether the protection has changed or the object has
1331 * been copied while we left the map unlocked. Changing from
1332 * read to write permission is OK - we leave the page
1333 * write-protected, and catch the write fault. Changing from
1334 * write to read permission means that we can't mark the page
1335 * write-enabled after all.
1338 fault_type &= retry_prot;
1341 unlock_and_deallocate(&fs);
1345 /* Reassert because wired may have changed. */
1346 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1347 ("!wired && VM_FAULT_WIRE"));
1352 * If the page was filled by a pager, save the virtual address that
1353 * should be faulted on next under a sequential access pattern to the
1354 * map entry. A read lock on the map suffices to update this address
1358 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1360 vm_page_assert_xbusied(fs.m);
1361 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1364 * Page must be completely valid or it is not fit to
1365 * map into user space. vm_pager_get_pages() ensures this.
1367 KASSERT(vm_page_all_valid(fs.m),
1368 ("vm_fault: page %p partially invalid", fs.m));
1369 VM_OBJECT_WUNLOCK(fs.object);
1372 * Put this page into the physical map. We had to do the unlock above
1373 * because pmap_enter() may sleep. We don't put the page
1374 * back on the active queue until later so that the pageout daemon
1375 * won't find it (yet).
1377 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1378 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1379 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1381 vm_fault_prefault(&fs, vaddr,
1382 faultcount > 0 ? behind : PFBAK,
1383 faultcount > 0 ? ahead : PFFOR, false);
1384 VM_OBJECT_WLOCK(fs.object);
1387 * If the page is not wired down, then put it where the pageout daemon
1390 if ((fault_flags & VM_FAULT_WIRE) != 0) {
1394 vm_page_activate(fs.m);
1395 vm_page_unlock(fs.m);
1397 if (m_hold != NULL) {
1401 vm_page_xunbusy(fs.m);
1404 * Unlock everything, and return
1406 unlock_and_deallocate(&fs);
1408 VM_CNT_INC(v_io_faults);
1409 curthread->td_ru.ru_majflt++;
1411 if (racct_enable && fs.object->type == OBJT_VNODE) {
1413 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1414 racct_add_force(curproc, RACCT_WRITEBPS,
1415 PAGE_SIZE + behind * PAGE_SIZE);
1416 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1418 racct_add_force(curproc, RACCT_READBPS,
1419 PAGE_SIZE + ahead * PAGE_SIZE);
1420 racct_add_force(curproc, RACCT_READIOPS, 1);
1422 PROC_UNLOCK(curproc);
1426 curthread->td_ru.ru_minflt++;
1428 return (KERN_SUCCESS);
1432 * Speed up the reclamation of pages that precede the faulting pindex within
1433 * the first object of the shadow chain. Essentially, perform the equivalent
1434 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1435 * the faulting pindex by the cluster size when the pages read by vm_fault()
1436 * cross a cluster-size boundary. The cluster size is the greater of the
1437 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1439 * When "fs->first_object" is a shadow object, the pages in the backing object
1440 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1441 * function must only be concerned with pages in the first object.
1444 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1446 vm_map_entry_t entry;
1447 vm_object_t first_object, object;
1448 vm_offset_t end, start;
1449 vm_page_t m, m_next;
1450 vm_pindex_t pend, pstart;
1453 object = fs->object;
1454 VM_OBJECT_ASSERT_WLOCKED(object);
1455 first_object = fs->first_object;
1456 if (first_object != object) {
1457 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1458 VM_OBJECT_WUNLOCK(object);
1459 VM_OBJECT_WLOCK(first_object);
1460 VM_OBJECT_WLOCK(object);
1463 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1464 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1465 size = VM_FAULT_DONTNEED_MIN;
1466 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1467 size = pagesizes[1];
1468 end = rounddown2(vaddr, size);
1469 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1470 (entry = fs->entry)->start < end) {
1471 if (end - entry->start < size)
1472 start = entry->start;
1475 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1476 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1478 m_next = vm_page_find_least(first_object, pstart);
1479 pend = OFF_TO_IDX(entry->offset) + atop(end -
1481 while ((m = m_next) != NULL && m->pindex < pend) {
1482 m_next = TAILQ_NEXT(m, listq);
1483 if (!vm_page_all_valid(m) ||
1488 * Don't clear PGA_REFERENCED, since it would
1489 * likely represent a reference by a different
1492 * Typically, at this point, prefetched pages
1493 * are still in the inactive queue. Only
1494 * pages that triggered page faults are in the
1498 if (!vm_page_inactive(m))
1499 vm_page_deactivate(m);
1504 if (first_object != object)
1505 VM_OBJECT_WUNLOCK(first_object);
1509 * vm_fault_prefault provides a quick way of clustering
1510 * pagefaults into a processes address space. It is a "cousin"
1511 * of vm_map_pmap_enter, except it runs at page fault time instead
1515 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1516 int backward, int forward, bool obj_locked)
1519 vm_map_entry_t entry;
1520 vm_object_t backing_object, lobject;
1521 vm_offset_t addr, starta;
1526 pmap = fs->map->pmap;
1527 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1532 if (addra < backward * PAGE_SIZE) {
1533 starta = entry->start;
1535 starta = addra - backward * PAGE_SIZE;
1536 if (starta < entry->start)
1537 starta = entry->start;
1541 * Generate the sequence of virtual addresses that are candidates for
1542 * prefaulting in an outward spiral from the faulting virtual address,
1543 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1544 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1545 * If the candidate address doesn't have a backing physical page, then
1546 * the loop immediately terminates.
1548 for (i = 0; i < 2 * imax(backward, forward); i++) {
1549 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1551 if (addr > addra + forward * PAGE_SIZE)
1554 if (addr < starta || addr >= entry->end)
1557 if (!pmap_is_prefaultable(pmap, addr))
1560 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1561 lobject = entry->object.vm_object;
1563 VM_OBJECT_RLOCK(lobject);
1564 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1565 lobject->type == OBJT_DEFAULT &&
1566 (backing_object = lobject->backing_object) != NULL) {
1567 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1568 0, ("vm_fault_prefault: unaligned object offset"));
1569 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1570 VM_OBJECT_RLOCK(backing_object);
1571 if (!obj_locked || lobject != entry->object.vm_object)
1572 VM_OBJECT_RUNLOCK(lobject);
1573 lobject = backing_object;
1576 if (!obj_locked || lobject != entry->object.vm_object)
1577 VM_OBJECT_RUNLOCK(lobject);
1580 if (vm_page_all_valid(m) &&
1581 (m->flags & PG_FICTITIOUS) == 0)
1582 pmap_enter_quick(pmap, addr, m, entry->protection);
1583 if (!obj_locked || lobject != entry->object.vm_object)
1584 VM_OBJECT_RUNLOCK(lobject);
1589 * Hold each of the physical pages that are mapped by the specified range of
1590 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1591 * and allow the specified types of access, "prot". If all of the implied
1592 * pages are successfully held, then the number of held pages is returned
1593 * together with pointers to those pages in the array "ma". However, if any
1594 * of the pages cannot be held, -1 is returned.
1597 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1598 vm_prot_t prot, vm_page_t *ma, int max_count)
1600 vm_offset_t end, va;
1603 boolean_t pmap_failed;
1607 end = round_page(addr + len);
1608 addr = trunc_page(addr);
1611 * Check for illegal addresses.
1613 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1616 if (atop(end - addr) > max_count)
1617 panic("vm_fault_quick_hold_pages: count > max_count");
1618 count = atop(end - addr);
1621 * Most likely, the physical pages are resident in the pmap, so it is
1622 * faster to try pmap_extract_and_hold() first.
1624 pmap_failed = FALSE;
1625 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1626 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1629 else if ((prot & VM_PROT_WRITE) != 0 &&
1630 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1632 * Explicitly dirty the physical page. Otherwise, the
1633 * caller's changes may go unnoticed because they are
1634 * performed through an unmanaged mapping or by a DMA
1637 * The object lock is not held here.
1638 * See vm_page_clear_dirty_mask().
1645 * One or more pages could not be held by the pmap. Either no
1646 * page was mapped at the specified virtual address or that
1647 * mapping had insufficient permissions. Attempt to fault in
1648 * and hold these pages.
1650 * If vm_fault_disable_pagefaults() was called,
1651 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1652 * acquire MD VM locks, which means we must not call
1653 * vm_fault(). Some (out of tree) callers mark
1654 * too wide a code area with vm_fault_disable_pagefaults()
1655 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1656 * the proper behaviour explicitly.
1658 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1659 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1661 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1662 if (*mp == NULL && vm_fault(map, va, prot,
1663 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1668 for (mp = ma; mp < ma + count; mp++)
1670 vm_page_unwire(*mp, PQ_INACTIVE);
1676 * vm_fault_copy_entry
1678 * Create new shadow object backing dst_entry with private copy of
1679 * all underlying pages. When src_entry is equal to dst_entry,
1680 * function implements COW for wired-down map entry. Otherwise,
1681 * it forks wired entry into dst_map.
1683 * In/out conditions:
1684 * The source and destination maps must be locked for write.
1685 * The source map entry must be wired down (or be a sharing map
1686 * entry corresponding to a main map entry that is wired down).
1689 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1690 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1691 vm_ooffset_t *fork_charge)
1693 vm_object_t backing_object, dst_object, object, src_object;
1694 vm_pindex_t dst_pindex, pindex, src_pindex;
1695 vm_prot_t access, prot;
1705 upgrade = src_entry == dst_entry;
1706 access = prot = dst_entry->protection;
1708 src_object = src_entry->object.vm_object;
1709 src_pindex = OFF_TO_IDX(src_entry->offset);
1711 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1712 dst_object = src_object;
1713 vm_object_reference(dst_object);
1716 * Create the top-level object for the destination entry. (Doesn't
1717 * actually shadow anything - we copy the pages directly.)
1719 dst_object = vm_object_allocate(OBJT_DEFAULT,
1720 atop(dst_entry->end - dst_entry->start));
1721 #if VM_NRESERVLEVEL > 0
1722 dst_object->flags |= OBJ_COLORED;
1723 dst_object->pg_color = atop(dst_entry->start);
1725 dst_object->domain = src_object->domain;
1726 dst_object->charge = dst_entry->end - dst_entry->start;
1729 VM_OBJECT_WLOCK(dst_object);
1730 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1731 ("vm_fault_copy_entry: vm_object not NULL"));
1732 if (src_object != dst_object) {
1733 dst_entry->object.vm_object = dst_object;
1734 dst_entry->offset = 0;
1735 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1737 if (fork_charge != NULL) {
1738 KASSERT(dst_entry->cred == NULL,
1739 ("vm_fault_copy_entry: leaked swp charge"));
1740 dst_object->cred = curthread->td_ucred;
1741 crhold(dst_object->cred);
1742 *fork_charge += dst_object->charge;
1743 } else if ((dst_object->type == OBJT_DEFAULT ||
1744 dst_object->type == OBJT_SWAP) &&
1745 dst_object->cred == NULL) {
1746 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1748 dst_object->cred = dst_entry->cred;
1749 dst_entry->cred = NULL;
1753 * If not an upgrade, then enter the mappings in the pmap as
1754 * read and/or execute accesses. Otherwise, enter them as
1757 * A writeable large page mapping is only created if all of
1758 * the constituent small page mappings are modified. Marking
1759 * PTEs as modified on inception allows promotion to happen
1760 * without taking potentially large number of soft faults.
1763 access &= ~VM_PROT_WRITE;
1766 * Loop through all of the virtual pages within the entry's
1767 * range, copying each page from the source object to the
1768 * destination object. Since the source is wired, those pages
1769 * must exist. In contrast, the destination is pageable.
1770 * Since the destination object doesn't share any backing storage
1771 * with the source object, all of its pages must be dirtied,
1772 * regardless of whether they can be written.
1774 for (vaddr = dst_entry->start, dst_pindex = 0;
1775 vaddr < dst_entry->end;
1776 vaddr += PAGE_SIZE, dst_pindex++) {
1779 * Find the page in the source object, and copy it in.
1780 * Because the source is wired down, the page will be
1783 if (src_object != dst_object)
1784 VM_OBJECT_RLOCK(src_object);
1785 object = src_object;
1786 pindex = src_pindex + dst_pindex;
1787 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1788 (backing_object = object->backing_object) != NULL) {
1790 * Unless the source mapping is read-only or
1791 * it is presently being upgraded from
1792 * read-only, the first object in the shadow
1793 * chain should provide all of the pages. In
1794 * other words, this loop body should never be
1795 * executed when the source mapping is already
1798 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1800 ("vm_fault_copy_entry: main object missing page"));
1802 VM_OBJECT_RLOCK(backing_object);
1803 pindex += OFF_TO_IDX(object->backing_object_offset);
1804 if (object != dst_object)
1805 VM_OBJECT_RUNLOCK(object);
1806 object = backing_object;
1808 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1810 if (object != dst_object) {
1812 * Allocate a page in the destination object.
1814 dst_m = vm_page_alloc(dst_object, (src_object ==
1815 dst_object ? src_pindex : 0) + dst_pindex,
1817 if (dst_m == NULL) {
1818 VM_OBJECT_WUNLOCK(dst_object);
1819 VM_OBJECT_RUNLOCK(object);
1820 vm_wait(dst_object);
1821 VM_OBJECT_WLOCK(dst_object);
1824 pmap_copy_page(src_m, dst_m);
1825 VM_OBJECT_RUNLOCK(object);
1826 dst_m->dirty = dst_m->valid = src_m->valid;
1829 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1831 if (dst_m->pindex >= dst_object->size) {
1833 * We are upgrading. Index can occur
1834 * out of bounds if the object type is
1835 * vnode and the file was truncated.
1837 vm_page_xunbusy(dst_m);
1841 VM_OBJECT_WUNLOCK(dst_object);
1844 * Enter it in the pmap. If a wired, copy-on-write
1845 * mapping is being replaced by a write-enabled
1846 * mapping, then wire that new mapping.
1848 * The page can be invalid if the user called
1849 * msync(MS_INVALIDATE) or truncated the backing vnode
1850 * or shared memory object. In this case, do not
1851 * insert it into pmap, but still do the copy so that
1852 * all copies of the wired map entry have similar
1855 if (vm_page_all_valid(dst_m)) {
1856 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1857 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1861 * Mark it no longer busy, and put it on the active list.
1863 VM_OBJECT_WLOCK(dst_object);
1866 if (src_m != dst_m) {
1867 vm_page_unwire(src_m, PQ_INACTIVE);
1868 vm_page_wire(dst_m);
1870 KASSERT(vm_page_wired(dst_m),
1871 ("dst_m %p is not wired", dst_m));
1874 vm_page_lock(dst_m);
1875 vm_page_activate(dst_m);
1876 vm_page_unlock(dst_m);
1878 vm_page_xunbusy(dst_m);
1880 VM_OBJECT_WUNLOCK(dst_object);
1882 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1883 vm_object_deallocate(src_object);
1888 * Block entry into the machine-independent layer's page fault handler by
1889 * the calling thread. Subsequent calls to vm_fault() by that thread will
1890 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1891 * spurious page faults.
1894 vm_fault_disable_pagefaults(void)
1897 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1901 vm_fault_enable_pagefaults(int save)
1904 curthread_pflags_restore(save);