2 * Copyright (c) 1982, 1986, 1989, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
10 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org>
11 * Copyright (c) 2013, 2014 The FreeBSD Foundation
13 * Portions of this software were developed by Konstantin Belousov
14 * under sponsorship from the FreeBSD Foundation.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
46 #include <sys/param.h>
47 #include <sys/systm.h>
49 #include <sys/fcntl.h>
55 #include <sys/limits.h>
57 #include <sys/mount.h>
58 #include <sys/mutex.h>
59 #include <sys/namei.h>
60 #include <sys/vnode.h>
63 #include <sys/filio.h>
64 #include <sys/resourcevar.h>
65 #include <sys/rwlock.h>
67 #include <sys/sysctl.h>
68 #include <sys/ttycom.h>
70 #include <sys/syslog.h>
71 #include <sys/unistd.h>
73 #include <security/audit/audit.h>
74 #include <security/mac/mac_framework.h>
77 #include <vm/vm_extern.h>
79 #include <vm/vm_map.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_page.h>
83 static fo_rdwr_t vn_read;
84 static fo_rdwr_t vn_write;
85 static fo_rdwr_t vn_io_fault;
86 static fo_truncate_t vn_truncate;
87 static fo_ioctl_t vn_ioctl;
88 static fo_poll_t vn_poll;
89 static fo_kqfilter_t vn_kqfilter;
90 static fo_stat_t vn_statfile;
91 static fo_close_t vn_closefile;
93 struct fileops vnops = {
94 .fo_read = vn_io_fault,
95 .fo_write = vn_io_fault,
96 .fo_truncate = vn_truncate,
99 .fo_kqfilter = vn_kqfilter,
100 .fo_stat = vn_statfile,
101 .fo_close = vn_closefile,
102 .fo_chmod = vn_chmod,
103 .fo_chown = vn_chown,
104 .fo_sendfile = vn_sendfile,
106 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE
109 static const int io_hold_cnt = 16;
110 static int vn_io_fault_enable = 1;
111 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW,
112 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance");
113 static int vn_io_fault_prefault = 0;
114 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW,
115 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting");
116 static u_long vn_io_faults_cnt;
117 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD,
118 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers");
121 * Returns true if vn_io_fault mode of handling the i/o request should
125 do_vn_io_fault(struct vnode *vp, struct uio *uio)
129 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG &&
130 (mp = vp->v_mount) != NULL &&
131 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable);
135 * Structure used to pass arguments to vn_io_fault1(), to do either
136 * file- or vnode-based I/O calls.
138 struct vn_io_fault_args {
146 struct fop_args_tag {
150 struct vop_args_tag {
156 static int vn_io_fault1(struct vnode *vp, struct uio *uio,
157 struct vn_io_fault_args *args, struct thread *td);
160 vn_open(ndp, flagp, cmode, fp)
161 struct nameidata *ndp;
165 struct thread *td = ndp->ni_cnd.cn_thread;
167 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp));
171 * Common code for vnode open operations via a name lookup.
172 * Lookup the vnode and invoke VOP_CREATE if needed.
173 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine.
175 * Note that this does NOT free nameidata for the successful case,
176 * due to the NDINIT being done elsewhere.
179 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags,
180 struct ucred *cred, struct file *fp)
184 struct thread *td = ndp->ni_cnd.cn_thread;
186 struct vattr *vap = &vat;
191 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT |
192 O_EXCL | O_DIRECTORY))
194 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) {
195 ndp->ni_cnd.cn_nameiop = CREATE;
197 * Set NOCACHE to avoid flushing the cache when
198 * rolling in many files at once.
200 ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE;
201 if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0)
202 ndp->ni_cnd.cn_flags |= FOLLOW;
203 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
204 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
205 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
206 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
208 if ((error = namei(ndp)) != 0)
210 if (ndp->ni_vp == NULL) {
213 vap->va_mode = cmode;
215 vap->va_vaflags |= VA_EXCLUSIVE;
216 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) {
217 NDFREE(ndp, NDF_ONLY_PNBUF);
219 if ((error = vn_start_write(NULL, &mp,
220 V_XSLEEP | PCATCH)) != 0)
224 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0)
225 ndp->ni_cnd.cn_flags |= MAKEENTRY;
227 error = mac_vnode_check_create(cred, ndp->ni_dvp,
231 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp,
234 vn_finished_write(mp);
236 NDFREE(ndp, NDF_ONLY_PNBUF);
242 if (ndp->ni_dvp == ndp->ni_vp)
248 if (fmode & O_EXCL) {
255 ndp->ni_cnd.cn_nameiop = LOOKUP;
256 ndp->ni_cnd.cn_flags = ISOPEN |
257 ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF;
258 if (!(fmode & FWRITE))
259 ndp->ni_cnd.cn_flags |= LOCKSHARED;
260 if (!(vn_open_flags & VN_OPEN_NOAUDIT))
261 ndp->ni_cnd.cn_flags |= AUDITVNODE1;
262 if (vn_open_flags & VN_OPEN_NOCAPCHECK)
263 ndp->ni_cnd.cn_flags |= NOCAPCHECK;
264 if ((error = namei(ndp)) != 0)
268 error = vn_open_vnode(vp, fmode, cred, td, fp);
274 NDFREE(ndp, NDF_ONLY_PNBUF);
282 * Common code for vnode open operations once a vnode is located.
283 * Check permissions, and call the VOP_OPEN routine.
286 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred,
287 struct thread *td, struct file *fp)
292 int error, have_flock, lock_flags, type;
294 if (vp->v_type == VLNK)
296 if (vp->v_type == VSOCK)
298 if (vp->v_type != VDIR && fmode & O_DIRECTORY)
301 if (fmode & (FWRITE | O_TRUNC)) {
302 if (vp->v_type == VDIR)
310 if ((fmode & O_APPEND) && (fmode & FWRITE))
313 error = mac_vnode_check_open(cred, vp, accmode);
317 if ((fmode & O_CREAT) == 0) {
318 if (accmode & VWRITE) {
319 error = vn_writechk(vp);
324 error = VOP_ACCESS(vp, accmode, cred, td);
329 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
330 vn_lock(vp, LK_UPGRADE | LK_RETRY);
331 if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0)
334 if (fmode & (O_EXLOCK | O_SHLOCK)) {
335 KASSERT(fp != NULL, ("open with flock requires fp"));
336 lock_flags = VOP_ISLOCKED(vp);
338 lf.l_whence = SEEK_SET;
341 if (fmode & O_EXLOCK)
346 if ((fmode & FNONBLOCK) == 0)
348 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type);
349 have_flock = (error == 0);
350 vn_lock(vp, lock_flags | LK_RETRY);
351 if (error == 0 && vp->v_iflag & VI_DOOMED)
354 * Another thread might have used this vnode as an
355 * executable while the vnode lock was dropped.
356 * Ensure the vnode is still able to be opened for
357 * writing after the lock has been obtained.
359 if (error == 0 && accmode & VWRITE)
360 error = vn_writechk(vp);
364 lf.l_whence = SEEK_SET;
368 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf,
371 vn_start_write(vp, &mp, V_WAIT);
372 vn_lock(vp, lock_flags | LK_RETRY);
373 (void)VOP_CLOSE(vp, fmode, cred, td);
374 vn_finished_write(mp);
375 /* Prevent second close from fdrop()->vn_close(). */
377 fp->f_ops= &badfileops;
380 fp->f_flag |= FHASLOCK;
382 if (fmode & FWRITE) {
383 VOP_ADD_WRITECOUNT(vp, 1);
384 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d",
385 __func__, vp, vp->v_writecount);
387 ASSERT_VOP_LOCKED(vp, "vn_open_vnode");
392 * Check for write permissions on the specified vnode.
393 * Prototype text segments cannot be written.
397 register struct vnode *vp;
400 ASSERT_VOP_LOCKED(vp, "vn_writechk");
402 * If there's shared text associated with
403 * the vnode, try to free it up once. If
404 * we fail, we can't allow writing.
416 vn_close(vp, flags, file_cred, td)
417 register struct vnode *vp;
419 struct ucred *file_cred;
423 int error, lock_flags;
425 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
426 MNT_EXTENDED_SHARED(vp->v_mount))
427 lock_flags = LK_SHARED;
429 lock_flags = LK_EXCLUSIVE;
431 vn_start_write(vp, &mp, V_WAIT);
432 vn_lock(vp, lock_flags | LK_RETRY);
433 if (flags & FWRITE) {
434 VNASSERT(vp->v_writecount > 0, vp,
435 ("vn_close: negative writecount"));
436 VOP_ADD_WRITECOUNT(vp, -1);
437 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
438 __func__, vp, vp->v_writecount);
440 error = VOP_CLOSE(vp, flags, file_cred, td);
442 vn_finished_write(mp);
447 * Heuristic to detect sequential operation.
450 sequential_heuristic(struct uio *uio, struct file *fp)
453 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
454 if (fp->f_flag & FRDAHEAD)
455 return (fp->f_seqcount << IO_SEQSHIFT);
458 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
459 * that the first I/O is normally considered to be slightly
460 * sequential. Seeking to offset 0 doesn't change sequentiality
461 * unless previous seeks have reduced f_seqcount to 0, in which
462 * case offset 0 is not special.
464 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
465 uio->uio_offset == fp->f_nextoff) {
467 * f_seqcount is in units of fixed-size blocks so that it
468 * depends mainly on the amount of sequential I/O and not
469 * much on the number of sequential I/O's. The fixed size
470 * of 16384 is hard-coded here since it is (not quite) just
471 * a magic size that works well here. This size is more
472 * closely related to the best I/O size for real disks than
473 * to any block size used by software.
475 fp->f_seqcount += howmany(uio->uio_resid, 16384);
476 if (fp->f_seqcount > IO_SEQMAX)
477 fp->f_seqcount = IO_SEQMAX;
478 return (fp->f_seqcount << IO_SEQSHIFT);
481 /* Not sequential. Quickly draw-down sequentiality. */
482 if (fp->f_seqcount > 1)
490 * Package up an I/O request on a vnode into a uio and do it.
493 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
494 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
495 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
502 struct vn_io_fault_args args;
503 int error, lock_flags;
505 auio.uio_iov = &aiov;
507 aiov.iov_base = base;
509 auio.uio_resid = len;
510 auio.uio_offset = offset;
511 auio.uio_segflg = segflg;
516 if ((ioflg & IO_NODELOCKED) == 0) {
517 if ((ioflg & IO_RANGELOCKED) == 0) {
518 if (rw == UIO_READ) {
519 rl_cookie = vn_rangelock_rlock(vp, offset,
522 rl_cookie = vn_rangelock_wlock(vp, offset,
528 if (rw == UIO_WRITE) {
529 if (vp->v_type != VCHR &&
530 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
533 if (MNT_SHARED_WRITES(mp) ||
534 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
535 lock_flags = LK_SHARED;
537 lock_flags = LK_EXCLUSIVE;
539 lock_flags = LK_SHARED;
540 vn_lock(vp, lock_flags | LK_RETRY);
544 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
546 if ((ioflg & IO_NOMACCHECK) == 0) {
548 error = mac_vnode_check_read(active_cred, file_cred,
551 error = mac_vnode_check_write(active_cred, file_cred,
556 if (file_cred != NULL)
560 if (do_vn_io_fault(vp, &auio)) {
561 args.kind = VN_IO_FAULT_VOP;
564 args.args.vop_args.vp = vp;
565 error = vn_io_fault1(vp, &auio, &args, td);
566 } else if (rw == UIO_READ) {
567 error = VOP_READ(vp, &auio, ioflg, cred);
568 } else /* if (rw == UIO_WRITE) */ {
569 error = VOP_WRITE(vp, &auio, ioflg, cred);
573 *aresid = auio.uio_resid;
575 if (auio.uio_resid && error == 0)
577 if ((ioflg & IO_NODELOCKED) == 0) {
580 vn_finished_write(mp);
583 if (rl_cookie != NULL)
584 vn_rangelock_unlock(vp, rl_cookie);
589 * Package up an I/O request on a vnode into a uio and do it. The I/O
590 * request is split up into smaller chunks and we try to avoid saturating
591 * the buffer cache while potentially holding a vnode locked, so we
592 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
593 * to give other processes a chance to lock the vnode (either other processes
594 * core'ing the same binary, or unrelated processes scanning the directory).
597 vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred,
598 file_cred, aresid, td)
606 struct ucred *active_cred;
607 struct ucred *file_cred;
618 * Force `offset' to a multiple of MAXBSIZE except possibly
619 * for the first chunk, so that filesystems only need to
620 * write full blocks except possibly for the first and last
623 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
627 if (rw != UIO_READ && vp->v_type == VREG)
630 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
631 ioflg, active_cred, file_cred, &iaresid, td);
632 len -= chunk; /* aresid calc already includes length */
636 base = (char *)base + chunk;
637 kern_yield(PRI_USER);
640 *aresid = len + iaresid;
645 foffset_lock(struct file *fp, int flags)
650 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
652 #if OFF_MAX <= LONG_MAX
654 * Caller only wants the current f_offset value. Assume that
655 * the long and shorter integer types reads are atomic.
657 if ((flags & FOF_NOLOCK) != 0)
658 return (fp->f_offset);
662 * According to McKusick the vn lock was protecting f_offset here.
663 * It is now protected by the FOFFSET_LOCKED flag.
665 mtxp = mtx_pool_find(mtxpool_sleep, fp);
667 if ((flags & FOF_NOLOCK) == 0) {
668 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
669 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
670 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
673 fp->f_vnread_flags |= FOFFSET_LOCKED;
681 foffset_unlock(struct file *fp, off_t val, int flags)
685 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
687 #if OFF_MAX <= LONG_MAX
688 if ((flags & FOF_NOLOCK) != 0) {
689 if ((flags & FOF_NOUPDATE) == 0)
691 if ((flags & FOF_NEXTOFF) != 0)
697 mtxp = mtx_pool_find(mtxpool_sleep, fp);
699 if ((flags & FOF_NOUPDATE) == 0)
701 if ((flags & FOF_NEXTOFF) != 0)
703 if ((flags & FOF_NOLOCK) == 0) {
704 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
705 ("Lost FOFFSET_LOCKED"));
706 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
707 wakeup(&fp->f_vnread_flags);
708 fp->f_vnread_flags = 0;
714 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
717 if ((flags & FOF_OFFSET) == 0)
718 uio->uio_offset = foffset_lock(fp, flags);
722 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
725 if ((flags & FOF_OFFSET) == 0)
726 foffset_unlock(fp, uio->uio_offset, flags);
730 get_advice(struct file *fp, struct uio *uio)
735 ret = POSIX_FADV_NORMAL;
736 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
739 mtxp = mtx_pool_find(mtxpool_sleep, fp);
741 if (fp->f_advice != NULL &&
742 uio->uio_offset >= fp->f_advice->fa_start &&
743 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
744 ret = fp->f_advice->fa_advice;
750 * File table vnode read routine.
753 vn_read(fp, uio, active_cred, flags, td)
756 struct ucred *active_cred;
764 off_t offset, start, end;
766 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
768 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
771 if (fp->f_flag & FNONBLOCK)
773 if (fp->f_flag & O_DIRECT)
775 advice = get_advice(fp, uio);
776 vn_lock(vp, LK_SHARED | LK_RETRY);
779 case POSIX_FADV_NORMAL:
780 case POSIX_FADV_SEQUENTIAL:
781 case POSIX_FADV_NOREUSE:
782 ioflag |= sequential_heuristic(uio, fp);
784 case POSIX_FADV_RANDOM:
785 /* Disable read-ahead for random I/O. */
788 offset = uio->uio_offset;
791 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
794 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
795 fp->f_nextoff = uio->uio_offset;
797 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
798 offset != uio->uio_offset) {
800 * Use POSIX_FADV_DONTNEED to flush clean pages and
801 * buffers for the backing file after a
802 * POSIX_FADV_NOREUSE read(2). To optimize the common
803 * case of using POSIX_FADV_NOREUSE with sequential
804 * access, track the previous implicit DONTNEED
805 * request and grow this request to include the
806 * current read(2) in addition to the previous
807 * DONTNEED. With purely sequential access this will
808 * cause the DONTNEED requests to continously grow to
809 * cover all of the previously read regions of the
810 * file. This allows filesystem blocks that are
811 * accessed by multiple calls to read(2) to be flushed
812 * once the last read(2) finishes.
815 end = uio->uio_offset - 1;
816 mtxp = mtx_pool_find(mtxpool_sleep, fp);
818 if (fp->f_advice != NULL &&
819 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
820 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
821 start = fp->f_advice->fa_prevstart;
822 else if (fp->f_advice->fa_prevstart != 0 &&
823 fp->f_advice->fa_prevstart == end + 1)
824 end = fp->f_advice->fa_prevend;
825 fp->f_advice->fa_prevstart = start;
826 fp->f_advice->fa_prevend = end;
829 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
835 * File table vnode write routine.
838 vn_write(fp, uio, active_cred, flags, td)
841 struct ucred *active_cred;
848 int error, ioflag, lock_flags;
850 off_t offset, start, end;
852 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
854 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
856 if (vp->v_type == VREG)
859 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
861 if (fp->f_flag & FNONBLOCK)
863 if (fp->f_flag & O_DIRECT)
865 if ((fp->f_flag & O_FSYNC) ||
866 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
869 if (vp->v_type != VCHR &&
870 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
873 advice = get_advice(fp, uio);
875 if (MNT_SHARED_WRITES(mp) ||
876 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
877 lock_flags = LK_SHARED;
879 lock_flags = LK_EXCLUSIVE;
882 vn_lock(vp, lock_flags | LK_RETRY);
884 case POSIX_FADV_NORMAL:
885 case POSIX_FADV_SEQUENTIAL:
886 case POSIX_FADV_NOREUSE:
887 ioflag |= sequential_heuristic(uio, fp);
889 case POSIX_FADV_RANDOM:
890 /* XXX: Is this correct? */
893 offset = uio->uio_offset;
896 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
899 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
900 fp->f_nextoff = uio->uio_offset;
902 if (vp->v_type != VCHR)
903 vn_finished_write(mp);
904 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
905 offset != uio->uio_offset) {
907 * Use POSIX_FADV_DONTNEED to flush clean pages and
908 * buffers for the backing file after a
909 * POSIX_FADV_NOREUSE write(2). To optimize the
910 * common case of using POSIX_FADV_NOREUSE with
911 * sequential access, track the previous implicit
912 * DONTNEED request and grow this request to include
913 * the current write(2) in addition to the previous
914 * DONTNEED. With purely sequential access this will
915 * cause the DONTNEED requests to continously grow to
916 * cover all of the previously written regions of the
919 * Note that the blocks just written are almost
920 * certainly still dirty, so this only works when
921 * VOP_ADVISE() calls from subsequent writes push out
922 * the data written by this write(2) once the backing
923 * buffers are clean. However, as compared to forcing
924 * IO_DIRECT, this gives much saner behavior. Write
925 * clustering is still allowed, and clean pages are
926 * merely moved to the cache page queue rather than
927 * outright thrown away. This means a subsequent
928 * read(2) can still avoid hitting the disk if the
929 * pages have not been reclaimed.
931 * This does make POSIX_FADV_NOREUSE largely useless
932 * with non-sequential access. However, sequential
933 * access is the more common use case and the flag is
937 end = uio->uio_offset - 1;
938 mtxp = mtx_pool_find(mtxpool_sleep, fp);
940 if (fp->f_advice != NULL &&
941 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
942 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
943 start = fp->f_advice->fa_prevstart;
944 else if (fp->f_advice->fa_prevstart != 0 &&
945 fp->f_advice->fa_prevstart == end + 1)
946 end = fp->f_advice->fa_prevend;
947 fp->f_advice->fa_prevstart = start;
948 fp->f_advice->fa_prevend = end;
951 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
959 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
960 * prevent the following deadlock:
962 * Assume that the thread A reads from the vnode vp1 into userspace
963 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
964 * currently not resident, then system ends up with the call chain
965 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
966 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
967 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
968 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
969 * backed by the pages of vnode vp1, and some page in buf2 is not
970 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
972 * To prevent the lock order reversal and deadlock, vn_io_fault() does
973 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
974 * Instead, it first tries to do the whole range i/o with pagefaults
975 * disabled. If all pages in the i/o buffer are resident and mapped,
976 * VOP will succeed (ignoring the genuine filesystem errors).
977 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
978 * i/o in chunks, with all pages in the chunk prefaulted and held
979 * using vm_fault_quick_hold_pages().
981 * Filesystems using this deadlock avoidance scheme should use the
982 * array of the held pages from uio, saved in the curthread->td_ma,
983 * instead of doing uiomove(). A helper function
984 * vn_io_fault_uiomove() converts uiomove request into
985 * uiomove_fromphys() over td_ma array.
987 * Since vnode locks do not cover the whole i/o anymore, rangelocks
988 * make the current i/o request atomic with respect to other i/os and
993 * Decode vn_io_fault_args and perform the corresponding i/o.
996 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1000 switch (args->kind) {
1001 case VN_IO_FAULT_FOP:
1002 return ((args->args.fop_args.doio)(args->args.fop_args.fp,
1003 uio, args->cred, args->flags, td));
1004 case VN_IO_FAULT_VOP:
1005 if (uio->uio_rw == UIO_READ) {
1006 return (VOP_READ(args->args.vop_args.vp, uio,
1007 args->flags, args->cred));
1008 } else if (uio->uio_rw == UIO_WRITE) {
1009 return (VOP_WRITE(args->args.vop_args.vp, uio,
1010 args->flags, args->cred));
1014 panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1019 vn_io_fault_touch(char *base, const struct uio *uio)
1024 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1030 vn_io_fault_prefault_user(const struct uio *uio)
1033 const struct iovec *iov;
1038 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1039 ("vn_io_fault_prefault userspace"));
1043 resid = uio->uio_resid;
1044 base = iov->iov_base;
1047 error = vn_io_fault_touch(base, uio);
1050 if (len < PAGE_SIZE) {
1052 error = vn_io_fault_touch(base + len - 1, uio);
1057 if (++i >= uio->uio_iovcnt)
1059 iov = uio->uio_iov + i;
1060 base = iov->iov_base;
1072 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1073 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1074 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1075 * into args and call vn_io_fault1() to handle faults during the user
1076 * mode buffer accesses.
1079 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1082 vm_page_t ma[io_hold_cnt + 2];
1083 struct uio *uio_clone, short_uio;
1084 struct iovec short_iovec[1];
1085 vm_page_t *prev_td_ma;
1087 vm_offset_t addr, end;
1090 int error, cnt, save, saveheld, prev_td_ma_cnt;
1092 if (vn_io_fault_prefault) {
1093 error = vn_io_fault_prefault_user(uio);
1095 return (error); /* Or ignore ? */
1098 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1101 * The UFS follows IO_UNIT directive and replays back both
1102 * uio_offset and uio_resid if an error is encountered during the
1103 * operation. But, since the iovec may be already advanced,
1104 * uio is still in an inconsistent state.
1106 * Cache a copy of the original uio, which is advanced to the redo
1107 * point using UIO_NOCOPY below.
1109 uio_clone = cloneuio(uio);
1110 resid = uio->uio_resid;
1112 short_uio.uio_segflg = UIO_USERSPACE;
1113 short_uio.uio_rw = uio->uio_rw;
1114 short_uio.uio_td = uio->uio_td;
1116 save = vm_fault_disable_pagefaults();
1117 error = vn_io_fault_doio(args, uio, td);
1118 if (error != EFAULT)
1121 atomic_add_long(&vn_io_faults_cnt, 1);
1122 uio_clone->uio_segflg = UIO_NOCOPY;
1123 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1124 uio_clone->uio_segflg = uio->uio_segflg;
1126 saveheld = curthread_pflags_set(TDP_UIOHELD);
1127 prev_td_ma = td->td_ma;
1128 prev_td_ma_cnt = td->td_ma_cnt;
1130 while (uio_clone->uio_resid != 0) {
1131 len = uio_clone->uio_iov->iov_len;
1133 KASSERT(uio_clone->uio_iovcnt >= 1,
1134 ("iovcnt underflow"));
1135 uio_clone->uio_iov++;
1136 uio_clone->uio_iovcnt--;
1139 if (len > io_hold_cnt * PAGE_SIZE)
1140 len = io_hold_cnt * PAGE_SIZE;
1141 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1142 end = round_page(addr + len);
1147 cnt = atop(end - trunc_page(addr));
1149 * A perfectly misaligned address and length could cause
1150 * both the start and the end of the chunk to use partial
1151 * page. +2 accounts for such a situation.
1153 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1154 addr, len, prot, ma, io_hold_cnt + 2);
1159 short_uio.uio_iov = &short_iovec[0];
1160 short_iovec[0].iov_base = (void *)addr;
1161 short_uio.uio_iovcnt = 1;
1162 short_uio.uio_resid = short_iovec[0].iov_len = len;
1163 short_uio.uio_offset = uio_clone->uio_offset;
1165 td->td_ma_cnt = cnt;
1167 error = vn_io_fault_doio(args, &short_uio, td);
1168 vm_page_unhold_pages(ma, cnt);
1169 adv = len - short_uio.uio_resid;
1171 uio_clone->uio_iov->iov_base =
1172 (char *)uio_clone->uio_iov->iov_base + adv;
1173 uio_clone->uio_iov->iov_len -= adv;
1174 uio_clone->uio_resid -= adv;
1175 uio_clone->uio_offset += adv;
1177 uio->uio_resid -= adv;
1178 uio->uio_offset += adv;
1180 if (error != 0 || adv == 0)
1183 td->td_ma = prev_td_ma;
1184 td->td_ma_cnt = prev_td_ma_cnt;
1185 curthread_pflags_restore(saveheld);
1187 vm_fault_enable_pagefaults(save);
1188 free(uio_clone, M_IOV);
1193 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1194 int flags, struct thread *td)
1199 struct vn_io_fault_args args;
1202 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1204 foffset_lock_uio(fp, uio, flags);
1205 if (do_vn_io_fault(vp, uio)) {
1206 args.kind = VN_IO_FAULT_FOP;
1207 args.args.fop_args.fp = fp;
1208 args.args.fop_args.doio = doio;
1209 args.cred = active_cred;
1210 args.flags = flags | FOF_OFFSET;
1211 if (uio->uio_rw == UIO_READ) {
1212 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1213 uio->uio_offset + uio->uio_resid);
1214 } else if ((fp->f_flag & O_APPEND) != 0 ||
1215 (flags & FOF_OFFSET) == 0) {
1216 /* For appenders, punt and lock the whole range. */
1217 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1219 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1220 uio->uio_offset + uio->uio_resid);
1222 error = vn_io_fault1(vp, uio, &args, td);
1223 vn_rangelock_unlock(vp, rl_cookie);
1225 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1227 foffset_unlock_uio(fp, uio, flags);
1232 * Helper function to perform the requested uiomove operation using
1233 * the held pages for io->uio_iov[0].iov_base buffer instead of
1234 * copyin/copyout. Access to the pages with uiomove_fromphys()
1235 * instead of iov_base prevents page faults that could occur due to
1236 * pmap_collect() invalidating the mapping created by
1237 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1238 * object cleanup revoking the write access from page mappings.
1240 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1241 * instead of plain uiomove().
1244 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1246 struct uio transp_uio;
1247 struct iovec transp_iov[1];
1253 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1254 uio->uio_segflg != UIO_USERSPACE)
1255 return (uiomove(data, xfersize, uio));
1257 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1258 transp_iov[0].iov_base = data;
1259 transp_uio.uio_iov = &transp_iov[0];
1260 transp_uio.uio_iovcnt = 1;
1261 if (xfersize > uio->uio_resid)
1262 xfersize = uio->uio_resid;
1263 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1264 transp_uio.uio_offset = 0;
1265 transp_uio.uio_segflg = UIO_SYSSPACE;
1267 * Since transp_iov points to data, and td_ma page array
1268 * corresponds to original uio->uio_iov, we need to invert the
1269 * direction of the i/o operation as passed to
1270 * uiomove_fromphys().
1272 switch (uio->uio_rw) {
1274 transp_uio.uio_rw = UIO_READ;
1277 transp_uio.uio_rw = UIO_WRITE;
1280 transp_uio.uio_td = uio->uio_td;
1281 error = uiomove_fromphys(td->td_ma,
1282 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1283 xfersize, &transp_uio);
1284 adv = xfersize - transp_uio.uio_resid;
1286 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1287 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1289 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1291 td->td_ma_cnt -= pgadv;
1292 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1293 uio->uio_iov->iov_len -= adv;
1294 uio->uio_resid -= adv;
1295 uio->uio_offset += adv;
1300 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1304 vm_offset_t iov_base;
1308 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1309 uio->uio_segflg != UIO_USERSPACE)
1310 return (uiomove_fromphys(ma, offset, xfersize, uio));
1312 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1313 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1314 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1315 switch (uio->uio_rw) {
1317 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1321 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1325 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1327 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1329 td->td_ma_cnt -= pgadv;
1330 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1331 uio->uio_iov->iov_len -= cnt;
1332 uio->uio_resid -= cnt;
1333 uio->uio_offset += cnt;
1339 * File table truncate routine.
1342 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1354 * Lock the whole range for truncation. Otherwise split i/o
1355 * might happen partly before and partly after the truncation.
1357 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1358 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1361 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1362 if (vp->v_type == VDIR) {
1367 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1371 error = vn_writechk(vp);
1374 vattr.va_size = length;
1375 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1379 vn_finished_write(mp);
1381 vn_rangelock_unlock(vp, rl_cookie);
1386 * File table vnode stat routine.
1389 vn_statfile(fp, sb, active_cred, td)
1392 struct ucred *active_cred;
1395 struct vnode *vp = fp->f_vnode;
1398 vn_lock(vp, LK_SHARED | LK_RETRY);
1399 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1406 * Stat a vnode; implementation for the stat syscall
1409 vn_stat(vp, sb, active_cred, file_cred, td)
1411 register struct stat *sb;
1412 struct ucred *active_cred;
1413 struct ucred *file_cred;
1417 register struct vattr *vap;
1422 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1430 * Initialize defaults for new and unusual fields, so that file
1431 * systems which don't support these fields don't need to know
1434 vap->va_birthtime.tv_sec = -1;
1435 vap->va_birthtime.tv_nsec = 0;
1436 vap->va_fsid = VNOVAL;
1437 vap->va_rdev = NODEV;
1439 error = VOP_GETATTR(vp, vap, active_cred);
1444 * Zero the spare stat fields
1446 bzero(sb, sizeof *sb);
1449 * Copy from vattr table
1451 if (vap->va_fsid != VNOVAL)
1452 sb->st_dev = vap->va_fsid;
1454 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1455 sb->st_ino = vap->va_fileid;
1456 mode = vap->va_mode;
1457 switch (vap->va_type) {
1483 sb->st_nlink = vap->va_nlink;
1484 sb->st_uid = vap->va_uid;
1485 sb->st_gid = vap->va_gid;
1486 sb->st_rdev = vap->va_rdev;
1487 if (vap->va_size > OFF_MAX)
1489 sb->st_size = vap->va_size;
1490 sb->st_atim = vap->va_atime;
1491 sb->st_mtim = vap->va_mtime;
1492 sb->st_ctim = vap->va_ctime;
1493 sb->st_birthtim = vap->va_birthtime;
1496 * According to www.opengroup.org, the meaning of st_blksize is
1497 * "a filesystem-specific preferred I/O block size for this
1498 * object. In some filesystem types, this may vary from file
1500 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1503 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1505 sb->st_flags = vap->va_flags;
1506 if (priv_check(td, PRIV_VFS_GENERATION))
1509 sb->st_gen = vap->va_gen;
1511 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1516 * File table vnode ioctl routine.
1519 vn_ioctl(fp, com, data, active_cred, td)
1523 struct ucred *active_cred;
1531 switch (vp->v_type) {
1536 vn_lock(vp, LK_SHARED | LK_RETRY);
1537 error = VOP_GETATTR(vp, &vattr, active_cred);
1540 *(int *)data = vattr.va_size - fp->f_offset;
1546 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1555 * File table vnode poll routine.
1558 vn_poll(fp, events, active_cred, td)
1561 struct ucred *active_cred;
1569 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1570 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1575 error = VOP_POLL(vp, events, fp->f_cred, td);
1580 * Acquire the requested lock and then check for validity. LK_RETRY
1581 * permits vn_lock to return doomed vnodes.
1584 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1588 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1589 ("vn_lock called with no locktype."));
1591 #ifdef DEBUG_VFS_LOCKS
1592 KASSERT(vp->v_holdcnt != 0,
1593 ("vn_lock %p: zero hold count", vp));
1595 error = VOP_LOCK1(vp, flags, file, line);
1596 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1597 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1598 ("LK_RETRY set with incompatible flags (0x%x) or an error occured (%d)",
1601 * Callers specify LK_RETRY if they wish to get dead vnodes.
1602 * If RETRY is not set, we return ENOENT instead.
1604 if (error == 0 && vp->v_iflag & VI_DOOMED &&
1605 (flags & LK_RETRY) == 0) {
1610 } while (flags & LK_RETRY && error != 0);
1615 * File table vnode close routine.
1618 vn_closefile(fp, td)
1627 fp->f_ops = &badfileops;
1629 if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK)
1632 error = vn_close(vp, fp->f_flag, fp->f_cred, td);
1634 if (fp->f_type == DTYPE_VNODE && fp->f_flag & FHASLOCK) {
1635 lf.l_whence = SEEK_SET;
1638 lf.l_type = F_UNLCK;
1639 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1646 * Preparing to start a filesystem write operation. If the operation is
1647 * permitted, then we bump the count of operations in progress and
1648 * proceed. If a suspend request is in progress, we wait until the
1649 * suspension is over, and then proceed.
1652 vn_start_write_locked(struct mount *mp, int flags)
1656 mtx_assert(MNT_MTX(mp), MA_OWNED);
1660 * Check on status of suspension.
1662 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1663 mp->mnt_susp_owner != curthread) {
1664 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1665 (flags & PCATCH) : 0) | (PUSER - 1);
1666 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1667 if (flags & V_NOWAIT) {
1668 error = EWOULDBLOCK;
1671 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1677 if (flags & V_XSLEEP)
1679 mp->mnt_writeopcount++;
1681 if (error != 0 || (flags & V_XSLEEP) != 0)
1688 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1693 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1694 ("V_MNTREF requires mp"));
1698 * If a vnode is provided, get and return the mount point that
1699 * to which it will write.
1702 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1704 if (error != EOPNOTSUPP)
1709 if ((mp = *mpp) == NULL)
1713 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1715 * As long as a vnode is not provided we need to acquire a
1716 * refcount for the provided mountpoint too, in order to
1717 * emulate a vfs_ref().
1720 if (vp == NULL && (flags & V_MNTREF) == 0)
1723 return (vn_start_write_locked(mp, flags));
1727 * Secondary suspension. Used by operations such as vop_inactive
1728 * routines that are needed by the higher level functions. These
1729 * are allowed to proceed until all the higher level functions have
1730 * completed (indicated by mnt_writeopcount dropping to zero). At that
1731 * time, these operations are halted until the suspension is over.
1734 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1739 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1740 ("V_MNTREF requires mp"));
1744 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1746 if (error != EOPNOTSUPP)
1752 * If we are not suspended or have not yet reached suspended
1753 * mode, then let the operation proceed.
1755 if ((mp = *mpp) == NULL)
1759 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1761 * As long as a vnode is not provided we need to acquire a
1762 * refcount for the provided mountpoint too, in order to
1763 * emulate a vfs_ref().
1766 if (vp == NULL && (flags & V_MNTREF) == 0)
1768 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1769 mp->mnt_secondary_writes++;
1770 mp->mnt_secondary_accwrites++;
1774 if (flags & V_NOWAIT) {
1777 return (EWOULDBLOCK);
1780 * Wait for the suspension to finish.
1782 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1783 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1792 * Filesystem write operation has completed. If we are suspending and this
1793 * operation is the last one, notify the suspender that the suspension is
1797 vn_finished_write(mp)
1804 mp->mnt_writeopcount--;
1805 if (mp->mnt_writeopcount < 0)
1806 panic("vn_finished_write: neg cnt");
1807 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1808 mp->mnt_writeopcount <= 0)
1809 wakeup(&mp->mnt_writeopcount);
1815 * Filesystem secondary write operation has completed. If we are
1816 * suspending and this operation is the last one, notify the suspender
1817 * that the suspension is now in effect.
1820 vn_finished_secondary_write(mp)
1827 mp->mnt_secondary_writes--;
1828 if (mp->mnt_secondary_writes < 0)
1829 panic("vn_finished_secondary_write: neg cnt");
1830 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1831 mp->mnt_secondary_writes <= 0)
1832 wakeup(&mp->mnt_secondary_writes);
1839 * Request a filesystem to suspend write operations.
1842 vfs_write_suspend(struct mount *mp, int flags)
1847 if (mp->mnt_susp_owner == curthread) {
1851 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1852 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1855 * Unmount holds a write reference on the mount point. If we
1856 * own busy reference and drain for writers, we deadlock with
1857 * the reference draining in the unmount path. Callers of
1858 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1859 * vfs_busy() reference is owned and caller is not in the
1862 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1863 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1868 mp->mnt_kern_flag |= MNTK_SUSPEND;
1869 mp->mnt_susp_owner = curthread;
1870 if (mp->mnt_writeopcount > 0)
1871 (void) msleep(&mp->mnt_writeopcount,
1872 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1875 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1876 vfs_write_resume(mp, 0);
1881 * Request a filesystem to resume write operations.
1884 vfs_write_resume(struct mount *mp, int flags)
1888 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1889 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1890 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1892 mp->mnt_susp_owner = NULL;
1893 wakeup(&mp->mnt_writeopcount);
1894 wakeup(&mp->mnt_flag);
1895 curthread->td_pflags &= ~TDP_IGNSUSP;
1896 if ((flags & VR_START_WRITE) != 0) {
1898 mp->mnt_writeopcount++;
1901 if ((flags & VR_NO_SUSPCLR) == 0)
1903 } else if ((flags & VR_START_WRITE) != 0) {
1905 vn_start_write_locked(mp, 0);
1912 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1916 vfs_write_suspend_umnt(struct mount *mp)
1920 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1921 ("vfs_write_suspend_umnt: recursed"));
1923 /* dounmount() already called vn_start_write(). */
1925 vn_finished_write(mp);
1926 error = vfs_write_suspend(mp, 0);
1928 vn_start_write(NULL, &mp, V_WAIT);
1932 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1935 vn_start_write(NULL, &mp, V_WAIT);
1937 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1938 wakeup(&mp->mnt_flag);
1940 curthread->td_pflags |= TDP_IGNSUSP;
1945 * Implement kqueues for files by translating it to vnode operation.
1948 vn_kqfilter(struct file *fp, struct knote *kn)
1951 return (VOP_KQFILTER(fp->f_vnode, kn));
1955 * Simplified in-kernel wrapper calls for extended attribute access.
1956 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1957 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1960 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1961 const char *attrname, int *buflen, char *buf, struct thread *td)
1967 iov.iov_len = *buflen;
1970 auio.uio_iov = &iov;
1971 auio.uio_iovcnt = 1;
1972 auio.uio_rw = UIO_READ;
1973 auio.uio_segflg = UIO_SYSSPACE;
1975 auio.uio_offset = 0;
1976 auio.uio_resid = *buflen;
1978 if ((ioflg & IO_NODELOCKED) == 0)
1979 vn_lock(vp, LK_SHARED | LK_RETRY);
1981 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1983 /* authorize attribute retrieval as kernel */
1984 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1987 if ((ioflg & IO_NODELOCKED) == 0)
1991 *buflen = *buflen - auio.uio_resid;
1998 * XXX failure mode if partially written?
2001 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2002 const char *attrname, int buflen, char *buf, struct thread *td)
2009 iov.iov_len = buflen;
2012 auio.uio_iov = &iov;
2013 auio.uio_iovcnt = 1;
2014 auio.uio_rw = UIO_WRITE;
2015 auio.uio_segflg = UIO_SYSSPACE;
2017 auio.uio_offset = 0;
2018 auio.uio_resid = buflen;
2020 if ((ioflg & IO_NODELOCKED) == 0) {
2021 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2023 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2026 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2028 /* authorize attribute setting as kernel */
2029 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2031 if ((ioflg & IO_NODELOCKED) == 0) {
2032 vn_finished_write(mp);
2040 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2041 const char *attrname, struct thread *td)
2046 if ((ioflg & IO_NODELOCKED) == 0) {
2047 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2049 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2052 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2054 /* authorize attribute removal as kernel */
2055 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2056 if (error == EOPNOTSUPP)
2057 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2060 if ((ioflg & IO_NODELOCKED) == 0) {
2061 vn_finished_write(mp);
2069 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2073 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2077 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2080 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2085 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2086 int lkflags, struct vnode **rvp)
2091 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2093 ltype = VOP_ISLOCKED(vp);
2094 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2095 ("vn_vget_ino: vp not locked"));
2096 error = vfs_busy(mp, MBF_NOWAIT);
2100 error = vfs_busy(mp, 0);
2101 vn_lock(vp, ltype | LK_RETRY);
2105 if (vp->v_iflag & VI_DOOMED) {
2111 error = alloc(mp, alloc_arg, lkflags, rvp);
2114 vn_lock(vp, ltype | LK_RETRY);
2115 if (vp->v_iflag & VI_DOOMED) {
2128 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2129 const struct thread *td)
2132 if (vp->v_type != VREG || td == NULL)
2134 PROC_LOCK(td->td_proc);
2135 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2136 lim_cur(td->td_proc, RLIMIT_FSIZE)) {
2137 kern_psignal(td->td_proc, SIGXFSZ);
2138 PROC_UNLOCK(td->td_proc);
2141 PROC_UNLOCK(td->td_proc);
2146 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2153 vn_lock(vp, LK_SHARED | LK_RETRY);
2154 AUDIT_ARG_VNODE1(vp);
2157 return (setfmode(td, active_cred, vp, mode));
2161 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2168 vn_lock(vp, LK_SHARED | LK_RETRY);
2169 AUDIT_ARG_VNODE1(vp);
2172 return (setfown(td, active_cred, vp, uid, gid));
2176 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2180 if ((object = vp->v_object) == NULL)
2182 VM_OBJECT_WLOCK(object);
2183 vm_object_page_remove(object, start, end, 0);
2184 VM_OBJECT_WUNLOCK(object);
2188 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2196 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2197 ("Wrong command %lu", cmd));
2199 if (vn_lock(vp, LK_SHARED) != 0)
2201 if (vp->v_type != VREG) {
2205 error = VOP_GETATTR(vp, &va, cred);
2209 if (noff >= va.va_size) {
2213 bsize = vp->v_mount->mnt_stat.f_iosize;
2214 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2215 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2216 if (error == EOPNOTSUPP) {
2220 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2221 (bnp != -1 && cmd == FIOSEEKDATA)) {
2228 if (noff > va.va_size)
2230 /* noff == va.va_size. There is an implicit hole at the end of file. */
2231 if (cmd == FIOSEEKDATA)
2241 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2246 off_t foffset, size;
2249 cred = td->td_ucred;
2251 foffset = foffset_lock(fp, 0);
2252 noneg = (vp->v_type != VCHR);
2258 (offset > 0 && foffset > OFF_MAX - offset))) {
2265 vn_lock(vp, LK_SHARED | LK_RETRY);
2266 error = VOP_GETATTR(vp, &vattr, cred);
2272 * If the file references a disk device, then fetch
2273 * the media size and use that to determine the ending
2276 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2277 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2278 vattr.va_size = size;
2280 (vattr.va_size > OFF_MAX ||
2281 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2285 offset += vattr.va_size;
2290 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2293 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2298 if (error == 0 && noneg && offset < 0)
2302 VFS_KNOTE_UNLOCKED(vp, 0);
2303 *(off_t *)(td->td_retval) = offset;
2305 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2310 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2316 * Grant permission if the caller is the owner of the file, or
2317 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2318 * on the file. If the time pointer is null, then write
2319 * permission on the file is also sufficient.
2321 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2322 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2323 * will be allowed to set the times [..] to the current
2326 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2327 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2328 error = VOP_ACCESS(vp, VWRITE, cred, td);