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_close1(struct vnode *vp, int flags, struct ucred *file_cred,
417 struct thread *td, bool keep_ref)
420 int error, lock_flags;
422 if (vp->v_type != VFIFO && (flags & FWRITE) == 0 &&
423 MNT_EXTENDED_SHARED(vp->v_mount))
424 lock_flags = LK_SHARED;
426 lock_flags = LK_EXCLUSIVE;
428 vn_start_write(vp, &mp, V_WAIT);
429 vn_lock(vp, lock_flags | LK_RETRY);
430 if (flags & FWRITE) {
431 VNASSERT(vp->v_writecount > 0, vp,
432 ("vn_close: negative writecount"));
433 VOP_ADD_WRITECOUNT(vp, -1);
434 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d",
435 __func__, vp, vp->v_writecount);
437 error = VOP_CLOSE(vp, flags, file_cred, td);
442 vn_finished_write(mp);
447 vn_close(struct vnode *vp, int flags, struct ucred *file_cred,
451 return (vn_close1(vp, flags, file_cred, td, false));
455 * Heuristic to detect sequential operation.
458 sequential_heuristic(struct uio *uio, struct file *fp)
461 ASSERT_VOP_LOCKED(fp->f_vnode, __func__);
462 if (fp->f_flag & FRDAHEAD)
463 return (fp->f_seqcount << IO_SEQSHIFT);
466 * Offset 0 is handled specially. open() sets f_seqcount to 1 so
467 * that the first I/O is normally considered to be slightly
468 * sequential. Seeking to offset 0 doesn't change sequentiality
469 * unless previous seeks have reduced f_seqcount to 0, in which
470 * case offset 0 is not special.
472 if ((uio->uio_offset == 0 && fp->f_seqcount > 0) ||
473 uio->uio_offset == fp->f_nextoff) {
475 * f_seqcount is in units of fixed-size blocks so that it
476 * depends mainly on the amount of sequential I/O and not
477 * much on the number of sequential I/O's. The fixed size
478 * of 16384 is hard-coded here since it is (not quite) just
479 * a magic size that works well here. This size is more
480 * closely related to the best I/O size for real disks than
481 * to any block size used by software.
483 fp->f_seqcount += howmany(uio->uio_resid, 16384);
484 if (fp->f_seqcount > IO_SEQMAX)
485 fp->f_seqcount = IO_SEQMAX;
486 return (fp->f_seqcount << IO_SEQSHIFT);
489 /* Not sequential. Quickly draw-down sequentiality. */
490 if (fp->f_seqcount > 1)
498 * Package up an I/O request on a vnode into a uio and do it.
501 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset,
502 enum uio_seg segflg, int ioflg, struct ucred *active_cred,
503 struct ucred *file_cred, ssize_t *aresid, struct thread *td)
510 struct vn_io_fault_args args;
511 int error, lock_flags;
513 auio.uio_iov = &aiov;
515 aiov.iov_base = base;
517 auio.uio_resid = len;
518 auio.uio_offset = offset;
519 auio.uio_segflg = segflg;
524 if ((ioflg & IO_NODELOCKED) == 0) {
525 if ((ioflg & IO_RANGELOCKED) == 0) {
526 if (rw == UIO_READ) {
527 rl_cookie = vn_rangelock_rlock(vp, offset,
530 rl_cookie = vn_rangelock_wlock(vp, offset,
536 if (rw == UIO_WRITE) {
537 if (vp->v_type != VCHR &&
538 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH))
541 if (MNT_SHARED_WRITES(mp) ||
542 ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount)))
543 lock_flags = LK_SHARED;
545 lock_flags = LK_EXCLUSIVE;
547 lock_flags = LK_SHARED;
548 vn_lock(vp, lock_flags | LK_RETRY);
552 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
554 if ((ioflg & IO_NOMACCHECK) == 0) {
556 error = mac_vnode_check_read(active_cred, file_cred,
559 error = mac_vnode_check_write(active_cred, file_cred,
564 if (file_cred != NULL)
568 if (do_vn_io_fault(vp, &auio)) {
569 args.kind = VN_IO_FAULT_VOP;
572 args.args.vop_args.vp = vp;
573 error = vn_io_fault1(vp, &auio, &args, td);
574 } else if (rw == UIO_READ) {
575 error = VOP_READ(vp, &auio, ioflg, cred);
576 } else /* if (rw == UIO_WRITE) */ {
577 error = VOP_WRITE(vp, &auio, ioflg, cred);
581 *aresid = auio.uio_resid;
583 if (auio.uio_resid && error == 0)
585 if ((ioflg & IO_NODELOCKED) == 0) {
588 vn_finished_write(mp);
591 if (rl_cookie != NULL)
592 vn_rangelock_unlock(vp, rl_cookie);
597 * Package up an I/O request on a vnode into a uio and do it. The I/O
598 * request is split up into smaller chunks and we try to avoid saturating
599 * the buffer cache while potentially holding a vnode locked, so we
600 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield()
601 * to give other processes a chance to lock the vnode (either other processes
602 * core'ing the same binary, or unrelated processes scanning the directory).
605 vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred,
606 file_cred, aresid, td)
614 struct ucred *active_cred;
615 struct ucred *file_cred;
626 * Force `offset' to a multiple of MAXBSIZE except possibly
627 * for the first chunk, so that filesystems only need to
628 * write full blocks except possibly for the first and last
631 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE;
635 if (rw != UIO_READ && vp->v_type == VREG)
638 error = vn_rdwr(rw, vp, base, chunk, offset, segflg,
639 ioflg, active_cred, file_cred, &iaresid, td);
640 len -= chunk; /* aresid calc already includes length */
644 base = (char *)base + chunk;
645 kern_yield(PRI_USER);
648 *aresid = len + iaresid;
653 foffset_lock(struct file *fp, int flags)
658 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
660 #if OFF_MAX <= LONG_MAX
662 * Caller only wants the current f_offset value. Assume that
663 * the long and shorter integer types reads are atomic.
665 if ((flags & FOF_NOLOCK) != 0)
666 return (fp->f_offset);
670 * According to McKusick the vn lock was protecting f_offset here.
671 * It is now protected by the FOFFSET_LOCKED flag.
673 mtxp = mtx_pool_find(mtxpool_sleep, fp);
675 if ((flags & FOF_NOLOCK) == 0) {
676 while (fp->f_vnread_flags & FOFFSET_LOCKED) {
677 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING;
678 msleep(&fp->f_vnread_flags, mtxp, PUSER -1,
681 fp->f_vnread_flags |= FOFFSET_LOCKED;
689 foffset_unlock(struct file *fp, off_t val, int flags)
693 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed"));
695 #if OFF_MAX <= LONG_MAX
696 if ((flags & FOF_NOLOCK) != 0) {
697 if ((flags & FOF_NOUPDATE) == 0)
699 if ((flags & FOF_NEXTOFF) != 0)
705 mtxp = mtx_pool_find(mtxpool_sleep, fp);
707 if ((flags & FOF_NOUPDATE) == 0)
709 if ((flags & FOF_NEXTOFF) != 0)
711 if ((flags & FOF_NOLOCK) == 0) {
712 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0,
713 ("Lost FOFFSET_LOCKED"));
714 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING)
715 wakeup(&fp->f_vnread_flags);
716 fp->f_vnread_flags = 0;
722 foffset_lock_uio(struct file *fp, struct uio *uio, int flags)
725 if ((flags & FOF_OFFSET) == 0)
726 uio->uio_offset = foffset_lock(fp, flags);
730 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags)
733 if ((flags & FOF_OFFSET) == 0)
734 foffset_unlock(fp, uio->uio_offset, flags);
738 get_advice(struct file *fp, struct uio *uio)
743 ret = POSIX_FADV_NORMAL;
744 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG)
747 mtxp = mtx_pool_find(mtxpool_sleep, fp);
749 if (fp->f_advice != NULL &&
750 uio->uio_offset >= fp->f_advice->fa_start &&
751 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end)
752 ret = fp->f_advice->fa_advice;
758 * File table vnode read routine.
761 vn_read(fp, uio, active_cred, flags, td)
764 struct ucred *active_cred;
772 off_t offset, start, end;
774 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
776 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
779 if (fp->f_flag & FNONBLOCK)
781 if (fp->f_flag & O_DIRECT)
783 advice = get_advice(fp, uio);
784 vn_lock(vp, LK_SHARED | LK_RETRY);
787 case POSIX_FADV_NORMAL:
788 case POSIX_FADV_SEQUENTIAL:
789 case POSIX_FADV_NOREUSE:
790 ioflag |= sequential_heuristic(uio, fp);
792 case POSIX_FADV_RANDOM:
793 /* Disable read-ahead for random I/O. */
796 offset = uio->uio_offset;
799 error = mac_vnode_check_read(active_cred, fp->f_cred, vp);
802 error = VOP_READ(vp, uio, ioflag, fp->f_cred);
803 fp->f_nextoff = uio->uio_offset;
805 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
806 offset != uio->uio_offset) {
808 * Use POSIX_FADV_DONTNEED to flush clean pages and
809 * buffers for the backing file after a
810 * POSIX_FADV_NOREUSE read(2). To optimize the common
811 * case of using POSIX_FADV_NOREUSE with sequential
812 * access, track the previous implicit DONTNEED
813 * request and grow this request to include the
814 * current read(2) in addition to the previous
815 * DONTNEED. With purely sequential access this will
816 * cause the DONTNEED requests to continously grow to
817 * cover all of the previously read regions of the
818 * file. This allows filesystem blocks that are
819 * accessed by multiple calls to read(2) to be flushed
820 * once the last read(2) finishes.
823 end = uio->uio_offset - 1;
824 mtxp = mtx_pool_find(mtxpool_sleep, fp);
826 if (fp->f_advice != NULL &&
827 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
828 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
829 start = fp->f_advice->fa_prevstart;
830 else if (fp->f_advice->fa_prevstart != 0 &&
831 fp->f_advice->fa_prevstart == end + 1)
832 end = fp->f_advice->fa_prevend;
833 fp->f_advice->fa_prevstart = start;
834 fp->f_advice->fa_prevend = end;
837 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
843 * File table vnode write routine.
846 vn_write(fp, uio, active_cred, flags, td)
849 struct ucred *active_cred;
856 int error, ioflag, lock_flags;
858 off_t offset, start, end;
860 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p",
862 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET"));
864 if (vp->v_type == VREG)
867 if (vp->v_type == VREG && (fp->f_flag & O_APPEND))
869 if (fp->f_flag & FNONBLOCK)
871 if (fp->f_flag & O_DIRECT)
873 if ((fp->f_flag & O_FSYNC) ||
874 (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
877 if (vp->v_type != VCHR &&
878 (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
881 advice = get_advice(fp, uio);
883 if (MNT_SHARED_WRITES(mp) ||
884 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) {
885 lock_flags = LK_SHARED;
887 lock_flags = LK_EXCLUSIVE;
890 vn_lock(vp, lock_flags | LK_RETRY);
892 case POSIX_FADV_NORMAL:
893 case POSIX_FADV_SEQUENTIAL:
894 case POSIX_FADV_NOREUSE:
895 ioflag |= sequential_heuristic(uio, fp);
897 case POSIX_FADV_RANDOM:
898 /* XXX: Is this correct? */
901 offset = uio->uio_offset;
904 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
907 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred);
908 fp->f_nextoff = uio->uio_offset;
910 if (vp->v_type != VCHR)
911 vn_finished_write(mp);
912 if (error == 0 && advice == POSIX_FADV_NOREUSE &&
913 offset != uio->uio_offset) {
915 * Use POSIX_FADV_DONTNEED to flush clean pages and
916 * buffers for the backing file after a
917 * POSIX_FADV_NOREUSE write(2). To optimize the
918 * common case of using POSIX_FADV_NOREUSE with
919 * sequential access, track the previous implicit
920 * DONTNEED request and grow this request to include
921 * the current write(2) in addition to the previous
922 * DONTNEED. With purely sequential access this will
923 * cause the DONTNEED requests to continously grow to
924 * cover all of the previously written regions of the
927 * Note that the blocks just written are almost
928 * certainly still dirty, so this only works when
929 * VOP_ADVISE() calls from subsequent writes push out
930 * the data written by this write(2) once the backing
931 * buffers are clean. However, as compared to forcing
932 * IO_DIRECT, this gives much saner behavior. Write
933 * clustering is still allowed, and clean pages are
934 * merely moved to the cache page queue rather than
935 * outright thrown away. This means a subsequent
936 * read(2) can still avoid hitting the disk if the
937 * pages have not been reclaimed.
939 * This does make POSIX_FADV_NOREUSE largely useless
940 * with non-sequential access. However, sequential
941 * access is the more common use case and the flag is
945 end = uio->uio_offset - 1;
946 mtxp = mtx_pool_find(mtxpool_sleep, fp);
948 if (fp->f_advice != NULL &&
949 fp->f_advice->fa_advice == POSIX_FADV_NOREUSE) {
950 if (start != 0 && fp->f_advice->fa_prevend + 1 == start)
951 start = fp->f_advice->fa_prevstart;
952 else if (fp->f_advice->fa_prevstart != 0 &&
953 fp->f_advice->fa_prevstart == end + 1)
954 end = fp->f_advice->fa_prevend;
955 fp->f_advice->fa_prevstart = start;
956 fp->f_advice->fa_prevend = end;
959 error = VOP_ADVISE(vp, start, end, POSIX_FADV_DONTNEED);
967 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
968 * prevent the following deadlock:
970 * Assume that the thread A reads from the vnode vp1 into userspace
971 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is
972 * currently not resident, then system ends up with the call chain
973 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
974 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
975 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
976 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
977 * backed by the pages of vnode vp1, and some page in buf2 is not
978 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
980 * To prevent the lock order reversal and deadlock, vn_io_fault() does
981 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
982 * Instead, it first tries to do the whole range i/o with pagefaults
983 * disabled. If all pages in the i/o buffer are resident and mapped,
984 * VOP will succeed (ignoring the genuine filesystem errors).
985 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
986 * i/o in chunks, with all pages in the chunk prefaulted and held
987 * using vm_fault_quick_hold_pages().
989 * Filesystems using this deadlock avoidance scheme should use the
990 * array of the held pages from uio, saved in the curthread->td_ma,
991 * instead of doing uiomove(). A helper function
992 * vn_io_fault_uiomove() converts uiomove request into
993 * uiomove_fromphys() over td_ma array.
995 * Since vnode locks do not cover the whole i/o anymore, rangelocks
996 * make the current i/o request atomic with respect to other i/os and
1001 * Decode vn_io_fault_args and perform the corresponding i/o.
1004 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio,
1008 switch (args->kind) {
1009 case VN_IO_FAULT_FOP:
1010 return ((args->args.fop_args.doio)(args->args.fop_args.fp,
1011 uio, args->cred, args->flags, td));
1012 case VN_IO_FAULT_VOP:
1013 if (uio->uio_rw == UIO_READ) {
1014 return (VOP_READ(args->args.vop_args.vp, uio,
1015 args->flags, args->cred));
1016 } else if (uio->uio_rw == UIO_WRITE) {
1017 return (VOP_WRITE(args->args.vop_args.vp, uio,
1018 args->flags, args->cred));
1022 panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind,
1027 vn_io_fault_touch(char *base, const struct uio *uio)
1032 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1))
1038 vn_io_fault_prefault_user(const struct uio *uio)
1041 const struct iovec *iov;
1046 KASSERT(uio->uio_segflg == UIO_USERSPACE,
1047 ("vn_io_fault_prefault userspace"));
1051 resid = uio->uio_resid;
1052 base = iov->iov_base;
1055 error = vn_io_fault_touch(base, uio);
1058 if (len < PAGE_SIZE) {
1060 error = vn_io_fault_touch(base + len - 1, uio);
1065 if (++i >= uio->uio_iovcnt)
1067 iov = uio->uio_iov + i;
1068 base = iov->iov_base;
1080 * Common code for vn_io_fault(), agnostic to the kind of i/o request.
1081 * Uses vn_io_fault_doio() to make the call to an actual i/o function.
1082 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request
1083 * into args and call vn_io_fault1() to handle faults during the user
1084 * mode buffer accesses.
1087 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args,
1090 vm_page_t ma[io_hold_cnt + 2];
1091 struct uio *uio_clone, short_uio;
1092 struct iovec short_iovec[1];
1093 vm_page_t *prev_td_ma;
1095 vm_offset_t addr, end;
1098 int error, cnt, save, saveheld, prev_td_ma_cnt;
1100 if (vn_io_fault_prefault) {
1101 error = vn_io_fault_prefault_user(uio);
1103 return (error); /* Or ignore ? */
1106 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ;
1109 * The UFS follows IO_UNIT directive and replays back both
1110 * uio_offset and uio_resid if an error is encountered during the
1111 * operation. But, since the iovec may be already advanced,
1112 * uio is still in an inconsistent state.
1114 * Cache a copy of the original uio, which is advanced to the redo
1115 * point using UIO_NOCOPY below.
1117 uio_clone = cloneuio(uio);
1118 resid = uio->uio_resid;
1120 short_uio.uio_segflg = UIO_USERSPACE;
1121 short_uio.uio_rw = uio->uio_rw;
1122 short_uio.uio_td = uio->uio_td;
1124 save = vm_fault_disable_pagefaults();
1125 error = vn_io_fault_doio(args, uio, td);
1126 if (error != EFAULT)
1129 atomic_add_long(&vn_io_faults_cnt, 1);
1130 uio_clone->uio_segflg = UIO_NOCOPY;
1131 uiomove(NULL, resid - uio->uio_resid, uio_clone);
1132 uio_clone->uio_segflg = uio->uio_segflg;
1134 saveheld = curthread_pflags_set(TDP_UIOHELD);
1135 prev_td_ma = td->td_ma;
1136 prev_td_ma_cnt = td->td_ma_cnt;
1138 while (uio_clone->uio_resid != 0) {
1139 len = uio_clone->uio_iov->iov_len;
1141 KASSERT(uio_clone->uio_iovcnt >= 1,
1142 ("iovcnt underflow"));
1143 uio_clone->uio_iov++;
1144 uio_clone->uio_iovcnt--;
1147 if (len > io_hold_cnt * PAGE_SIZE)
1148 len = io_hold_cnt * PAGE_SIZE;
1149 addr = (uintptr_t)uio_clone->uio_iov->iov_base;
1150 end = round_page(addr + len);
1155 cnt = atop(end - trunc_page(addr));
1157 * A perfectly misaligned address and length could cause
1158 * both the start and the end of the chunk to use partial
1159 * page. +2 accounts for such a situation.
1161 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map,
1162 addr, len, prot, ma, io_hold_cnt + 2);
1167 short_uio.uio_iov = &short_iovec[0];
1168 short_iovec[0].iov_base = (void *)addr;
1169 short_uio.uio_iovcnt = 1;
1170 short_uio.uio_resid = short_iovec[0].iov_len = len;
1171 short_uio.uio_offset = uio_clone->uio_offset;
1173 td->td_ma_cnt = cnt;
1175 error = vn_io_fault_doio(args, &short_uio, td);
1176 vm_page_unhold_pages(ma, cnt);
1177 adv = len - short_uio.uio_resid;
1179 uio_clone->uio_iov->iov_base =
1180 (char *)uio_clone->uio_iov->iov_base + adv;
1181 uio_clone->uio_iov->iov_len -= adv;
1182 uio_clone->uio_resid -= adv;
1183 uio_clone->uio_offset += adv;
1185 uio->uio_resid -= adv;
1186 uio->uio_offset += adv;
1188 if (error != 0 || adv == 0)
1191 td->td_ma = prev_td_ma;
1192 td->td_ma_cnt = prev_td_ma_cnt;
1193 curthread_pflags_restore(saveheld);
1195 vm_fault_enable_pagefaults(save);
1196 free(uio_clone, M_IOV);
1201 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred,
1202 int flags, struct thread *td)
1207 struct vn_io_fault_args args;
1210 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write;
1212 foffset_lock_uio(fp, uio, flags);
1213 if (do_vn_io_fault(vp, uio)) {
1214 args.kind = VN_IO_FAULT_FOP;
1215 args.args.fop_args.fp = fp;
1216 args.args.fop_args.doio = doio;
1217 args.cred = active_cred;
1218 args.flags = flags | FOF_OFFSET;
1219 if (uio->uio_rw == UIO_READ) {
1220 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset,
1221 uio->uio_offset + uio->uio_resid);
1222 } else if ((fp->f_flag & O_APPEND) != 0 ||
1223 (flags & FOF_OFFSET) == 0) {
1224 /* For appenders, punt and lock the whole range. */
1225 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1227 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset,
1228 uio->uio_offset + uio->uio_resid);
1230 error = vn_io_fault1(vp, uio, &args, td);
1231 vn_rangelock_unlock(vp, rl_cookie);
1233 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td);
1235 foffset_unlock_uio(fp, uio, flags);
1240 * Helper function to perform the requested uiomove operation using
1241 * the held pages for io->uio_iov[0].iov_base buffer instead of
1242 * copyin/copyout. Access to the pages with uiomove_fromphys()
1243 * instead of iov_base prevents page faults that could occur due to
1244 * pmap_collect() invalidating the mapping created by
1245 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
1246 * object cleanup revoking the write access from page mappings.
1248 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
1249 * instead of plain uiomove().
1252 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio)
1254 struct uio transp_uio;
1255 struct iovec transp_iov[1];
1261 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1262 uio->uio_segflg != UIO_USERSPACE)
1263 return (uiomove(data, xfersize, uio));
1265 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1266 transp_iov[0].iov_base = data;
1267 transp_uio.uio_iov = &transp_iov[0];
1268 transp_uio.uio_iovcnt = 1;
1269 if (xfersize > uio->uio_resid)
1270 xfersize = uio->uio_resid;
1271 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize;
1272 transp_uio.uio_offset = 0;
1273 transp_uio.uio_segflg = UIO_SYSSPACE;
1275 * Since transp_iov points to data, and td_ma page array
1276 * corresponds to original uio->uio_iov, we need to invert the
1277 * direction of the i/o operation as passed to
1278 * uiomove_fromphys().
1280 switch (uio->uio_rw) {
1282 transp_uio.uio_rw = UIO_READ;
1285 transp_uio.uio_rw = UIO_WRITE;
1288 transp_uio.uio_td = uio->uio_td;
1289 error = uiomove_fromphys(td->td_ma,
1290 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK,
1291 xfersize, &transp_uio);
1292 adv = xfersize - transp_uio.uio_resid;
1294 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) -
1295 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT);
1297 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1299 td->td_ma_cnt -= pgadv;
1300 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv;
1301 uio->uio_iov->iov_len -= adv;
1302 uio->uio_resid -= adv;
1303 uio->uio_offset += adv;
1308 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize,
1312 vm_offset_t iov_base;
1316 if ((td->td_pflags & TDP_UIOHELD) == 0 ||
1317 uio->uio_segflg != UIO_USERSPACE)
1318 return (uiomove_fromphys(ma, offset, xfersize, uio));
1320 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt));
1321 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize;
1322 iov_base = (vm_offset_t)uio->uio_iov->iov_base;
1323 switch (uio->uio_rw) {
1325 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma,
1329 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK,
1333 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT);
1335 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt,
1337 td->td_ma_cnt -= pgadv;
1338 uio->uio_iov->iov_base = (char *)(iov_base + cnt);
1339 uio->uio_iov->iov_len -= cnt;
1340 uio->uio_resid -= cnt;
1341 uio->uio_offset += cnt;
1347 * File table truncate routine.
1350 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred,
1362 * Lock the whole range for truncation. Otherwise split i/o
1363 * might happen partly before and partly after the truncation.
1365 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX);
1366 error = vn_start_write(vp, &mp, V_WAIT | PCATCH);
1369 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1370 if (vp->v_type == VDIR) {
1375 error = mac_vnode_check_write(active_cred, fp->f_cred, vp);
1379 error = vn_writechk(vp);
1382 vattr.va_size = length;
1383 if ((fp->f_flag & O_FSYNC) != 0)
1384 vattr.va_vaflags |= VA_SYNC;
1385 error = VOP_SETATTR(vp, &vattr, fp->f_cred);
1389 vn_finished_write(mp);
1391 vn_rangelock_unlock(vp, rl_cookie);
1396 * File table vnode stat routine.
1399 vn_statfile(fp, sb, active_cred, td)
1402 struct ucred *active_cred;
1405 struct vnode *vp = fp->f_vnode;
1408 vn_lock(vp, LK_SHARED | LK_RETRY);
1409 error = vn_stat(vp, sb, active_cred, fp->f_cred, td);
1416 * Stat a vnode; implementation for the stat syscall
1419 vn_stat(vp, sb, active_cred, file_cred, td)
1421 register struct stat *sb;
1422 struct ucred *active_cred;
1423 struct ucred *file_cred;
1427 register struct vattr *vap;
1432 error = mac_vnode_check_stat(active_cred, file_cred, vp);
1440 * Initialize defaults for new and unusual fields, so that file
1441 * systems which don't support these fields don't need to know
1444 vap->va_birthtime.tv_sec = -1;
1445 vap->va_birthtime.tv_nsec = 0;
1446 vap->va_fsid = VNOVAL;
1447 vap->va_rdev = NODEV;
1449 error = VOP_GETATTR(vp, vap, active_cred);
1454 * Zero the spare stat fields
1456 bzero(sb, sizeof *sb);
1459 * Copy from vattr table
1461 if (vap->va_fsid != VNOVAL)
1462 sb->st_dev = vap->va_fsid;
1464 sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0];
1465 sb->st_ino = vap->va_fileid;
1466 mode = vap->va_mode;
1467 switch (vap->va_type) {
1493 sb->st_nlink = vap->va_nlink;
1494 sb->st_uid = vap->va_uid;
1495 sb->st_gid = vap->va_gid;
1496 sb->st_rdev = vap->va_rdev;
1497 if (vap->va_size > OFF_MAX)
1499 sb->st_size = vap->va_size;
1500 sb->st_atim = vap->va_atime;
1501 sb->st_mtim = vap->va_mtime;
1502 sb->st_ctim = vap->va_ctime;
1503 sb->st_birthtim = vap->va_birthtime;
1506 * According to www.opengroup.org, the meaning of st_blksize is
1507 * "a filesystem-specific preferred I/O block size for this
1508 * object. In some filesystem types, this may vary from file
1510 * Use miminum/default of PAGE_SIZE (e.g. for VCHR).
1513 sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize);
1515 sb->st_flags = vap->va_flags;
1516 if (priv_check(td, PRIV_VFS_GENERATION))
1519 sb->st_gen = vap->va_gen;
1521 sb->st_blocks = vap->va_bytes / S_BLKSIZE;
1526 * File table vnode ioctl routine.
1529 vn_ioctl(fp, com, data, active_cred, td)
1533 struct ucred *active_cred;
1541 switch (vp->v_type) {
1546 vn_lock(vp, LK_SHARED | LK_RETRY);
1547 error = VOP_GETATTR(vp, &vattr, active_cred);
1550 *(int *)data = vattr.va_size - fp->f_offset;
1556 return (VOP_IOCTL(vp, com, data, fp->f_flag,
1565 * File table vnode poll routine.
1568 vn_poll(fp, events, active_cred, td)
1571 struct ucred *active_cred;
1579 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1580 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp);
1585 error = VOP_POLL(vp, events, fp->f_cred, td);
1590 * Acquire the requested lock and then check for validity. LK_RETRY
1591 * permits vn_lock to return doomed vnodes.
1594 _vn_lock(struct vnode *vp, int flags, char *file, int line)
1598 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
1599 ("vn_lock called with no locktype."));
1601 #ifdef DEBUG_VFS_LOCKS
1602 KASSERT(vp->v_holdcnt != 0,
1603 ("vn_lock %p: zero hold count", vp));
1605 error = VOP_LOCK1(vp, flags, file, line);
1606 flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */
1607 KASSERT((flags & LK_RETRY) == 0 || error == 0,
1608 ("LK_RETRY set with incompatible flags (0x%x) or an error occurred (%d)",
1611 * Callers specify LK_RETRY if they wish to get dead vnodes.
1612 * If RETRY is not set, we return ENOENT instead.
1614 if (error == 0 && vp->v_iflag & VI_DOOMED &&
1615 (flags & LK_RETRY) == 0) {
1620 } while (flags & LK_RETRY && error != 0);
1625 * File table vnode close routine.
1628 vn_closefile(fp, td)
1638 fp->f_ops = &badfileops;
1639 ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE;
1641 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref);
1643 if (__predict_false(ref)) {
1644 lf.l_whence = SEEK_SET;
1647 lf.l_type = F_UNLCK;
1648 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK);
1655 * Preparing to start a filesystem write operation. If the operation is
1656 * permitted, then we bump the count of operations in progress and
1657 * proceed. If a suspend request is in progress, we wait until the
1658 * suspension is over, and then proceed.
1661 vn_start_write_locked(struct mount *mp, int flags)
1665 mtx_assert(MNT_MTX(mp), MA_OWNED);
1669 * Check on status of suspension.
1671 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 ||
1672 mp->mnt_susp_owner != curthread) {
1673 mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ?
1674 (flags & PCATCH) : 0) | (PUSER - 1);
1675 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1676 if (flags & V_NOWAIT) {
1677 error = EWOULDBLOCK;
1680 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags,
1686 if (flags & V_XSLEEP)
1688 mp->mnt_writeopcount++;
1690 if (error != 0 || (flags & V_XSLEEP) != 0)
1697 vn_start_write(struct vnode *vp, struct mount **mpp, int flags)
1702 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1703 ("V_MNTREF requires mp"));
1707 * If a vnode is provided, get and return the mount point that
1708 * to which it will write.
1711 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1713 if (error != EOPNOTSUPP)
1718 if ((mp = *mpp) == NULL)
1722 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1724 * As long as a vnode is not provided we need to acquire a
1725 * refcount for the provided mountpoint too, in order to
1726 * emulate a vfs_ref().
1729 if (vp == NULL && (flags & V_MNTREF) == 0)
1732 return (vn_start_write_locked(mp, flags));
1736 * Secondary suspension. Used by operations such as vop_inactive
1737 * routines that are needed by the higher level functions. These
1738 * are allowed to proceed until all the higher level functions have
1739 * completed (indicated by mnt_writeopcount dropping to zero). At that
1740 * time, these operations are halted until the suspension is over.
1743 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags)
1748 KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL),
1749 ("V_MNTREF requires mp"));
1753 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) {
1755 if (error != EOPNOTSUPP)
1761 * If we are not suspended or have not yet reached suspended
1762 * mode, then let the operation proceed.
1764 if ((mp = *mpp) == NULL)
1768 * VOP_GETWRITEMOUNT() returns with the mp refcount held through
1770 * As long as a vnode is not provided we need to acquire a
1771 * refcount for the provided mountpoint too, in order to
1772 * emulate a vfs_ref().
1775 if (vp == NULL && (flags & V_MNTREF) == 0)
1777 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) {
1778 mp->mnt_secondary_writes++;
1779 mp->mnt_secondary_accwrites++;
1783 if (flags & V_NOWAIT) {
1786 return (EWOULDBLOCK);
1789 * Wait for the suspension to finish.
1791 error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP |
1792 ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0),
1801 * Filesystem write operation has completed. If we are suspending and this
1802 * operation is the last one, notify the suspender that the suspension is
1806 vn_finished_write(mp)
1813 mp->mnt_writeopcount--;
1814 if (mp->mnt_writeopcount < 0)
1815 panic("vn_finished_write: neg cnt");
1816 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1817 mp->mnt_writeopcount <= 0)
1818 wakeup(&mp->mnt_writeopcount);
1824 * Filesystem secondary write operation has completed. If we are
1825 * suspending and this operation is the last one, notify the suspender
1826 * that the suspension is now in effect.
1829 vn_finished_secondary_write(mp)
1836 mp->mnt_secondary_writes--;
1837 if (mp->mnt_secondary_writes < 0)
1838 panic("vn_finished_secondary_write: neg cnt");
1839 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 &&
1840 mp->mnt_secondary_writes <= 0)
1841 wakeup(&mp->mnt_secondary_writes);
1848 * Request a filesystem to suspend write operations.
1851 vfs_write_suspend(struct mount *mp, int flags)
1856 if (mp->mnt_susp_owner == curthread) {
1860 while (mp->mnt_kern_flag & MNTK_SUSPEND)
1861 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0);
1864 * Unmount holds a write reference on the mount point. If we
1865 * own busy reference and drain for writers, we deadlock with
1866 * the reference draining in the unmount path. Callers of
1867 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if
1868 * vfs_busy() reference is owned and caller is not in the
1871 if ((flags & VS_SKIP_UNMOUNT) != 0 &&
1872 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) {
1877 mp->mnt_kern_flag |= MNTK_SUSPEND;
1878 mp->mnt_susp_owner = curthread;
1879 if (mp->mnt_writeopcount > 0)
1880 (void) msleep(&mp->mnt_writeopcount,
1881 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0);
1884 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0)
1885 vfs_write_resume(mp, 0);
1890 * Request a filesystem to resume write operations.
1893 vfs_write_resume(struct mount *mp, int flags)
1897 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) {
1898 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner"));
1899 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 |
1901 mp->mnt_susp_owner = NULL;
1902 wakeup(&mp->mnt_writeopcount);
1903 wakeup(&mp->mnt_flag);
1904 curthread->td_pflags &= ~TDP_IGNSUSP;
1905 if ((flags & VR_START_WRITE) != 0) {
1907 mp->mnt_writeopcount++;
1910 if ((flags & VR_NO_SUSPCLR) == 0)
1912 } else if ((flags & VR_START_WRITE) != 0) {
1914 vn_start_write_locked(mp, 0);
1921 * Helper loop around vfs_write_suspend() for filesystem unmount VFS
1925 vfs_write_suspend_umnt(struct mount *mp)
1929 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0,
1930 ("vfs_write_suspend_umnt: recursed"));
1932 /* dounmount() already called vn_start_write(). */
1934 vn_finished_write(mp);
1935 error = vfs_write_suspend(mp, 0);
1937 vn_start_write(NULL, &mp, V_WAIT);
1941 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0)
1944 vn_start_write(NULL, &mp, V_WAIT);
1946 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2);
1947 wakeup(&mp->mnt_flag);
1949 curthread->td_pflags |= TDP_IGNSUSP;
1954 * Implement kqueues for files by translating it to vnode operation.
1957 vn_kqfilter(struct file *fp, struct knote *kn)
1960 return (VOP_KQFILTER(fp->f_vnode, kn));
1964 * Simplified in-kernel wrapper calls for extended attribute access.
1965 * Both calls pass in a NULL credential, authorizing as "kernel" access.
1966 * Set IO_NODELOCKED in ioflg if the vnode is already locked.
1969 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace,
1970 const char *attrname, int *buflen, char *buf, struct thread *td)
1976 iov.iov_len = *buflen;
1979 auio.uio_iov = &iov;
1980 auio.uio_iovcnt = 1;
1981 auio.uio_rw = UIO_READ;
1982 auio.uio_segflg = UIO_SYSSPACE;
1984 auio.uio_offset = 0;
1985 auio.uio_resid = *buflen;
1987 if ((ioflg & IO_NODELOCKED) == 0)
1988 vn_lock(vp, LK_SHARED | LK_RETRY);
1990 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
1992 /* authorize attribute retrieval as kernel */
1993 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL,
1996 if ((ioflg & IO_NODELOCKED) == 0)
2000 *buflen = *buflen - auio.uio_resid;
2007 * XXX failure mode if partially written?
2010 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace,
2011 const char *attrname, int buflen, char *buf, struct thread *td)
2018 iov.iov_len = buflen;
2021 auio.uio_iov = &iov;
2022 auio.uio_iovcnt = 1;
2023 auio.uio_rw = UIO_WRITE;
2024 auio.uio_segflg = UIO_SYSSPACE;
2026 auio.uio_offset = 0;
2027 auio.uio_resid = buflen;
2029 if ((ioflg & IO_NODELOCKED) == 0) {
2030 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2032 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2035 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2037 /* authorize attribute setting as kernel */
2038 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td);
2040 if ((ioflg & IO_NODELOCKED) == 0) {
2041 vn_finished_write(mp);
2049 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace,
2050 const char *attrname, struct thread *td)
2055 if ((ioflg & IO_NODELOCKED) == 0) {
2056 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0)
2058 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2061 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held");
2063 /* authorize attribute removal as kernel */
2064 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td);
2065 if (error == EOPNOTSUPP)
2066 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL,
2069 if ((ioflg & IO_NODELOCKED) == 0) {
2070 vn_finished_write(mp);
2078 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags,
2082 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp));
2086 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp)
2089 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino,
2094 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg,
2095 int lkflags, struct vnode **rvp)
2100 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get");
2102 ltype = VOP_ISLOCKED(vp);
2103 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED,
2104 ("vn_vget_ino: vp not locked"));
2105 error = vfs_busy(mp, MBF_NOWAIT);
2109 error = vfs_busy(mp, 0);
2110 vn_lock(vp, ltype | LK_RETRY);
2114 if (vp->v_iflag & VI_DOOMED) {
2120 error = alloc(mp, alloc_arg, lkflags, rvp);
2123 vn_lock(vp, ltype | LK_RETRY);
2124 if (vp->v_iflag & VI_DOOMED) {
2137 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio,
2138 const struct thread *td)
2141 if (vp->v_type != VREG || td == NULL)
2143 PROC_LOCK(td->td_proc);
2144 if ((uoff_t)uio->uio_offset + uio->uio_resid >
2145 lim_cur(td->td_proc, RLIMIT_FSIZE)) {
2146 kern_psignal(td->td_proc, SIGXFSZ);
2147 PROC_UNLOCK(td->td_proc);
2150 PROC_UNLOCK(td->td_proc);
2155 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred,
2162 vn_lock(vp, LK_SHARED | LK_RETRY);
2163 AUDIT_ARG_VNODE1(vp);
2166 return (setfmode(td, active_cred, vp, mode));
2170 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred,
2177 vn_lock(vp, LK_SHARED | LK_RETRY);
2178 AUDIT_ARG_VNODE1(vp);
2181 return (setfown(td, active_cred, vp, uid, gid));
2185 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end)
2189 if ((object = vp->v_object) == NULL)
2191 VM_OBJECT_WLOCK(object);
2192 vm_object_page_remove(object, start, end, 0);
2193 VM_OBJECT_WUNLOCK(object);
2197 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred)
2205 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA,
2206 ("Wrong command %lu", cmd));
2208 if (vn_lock(vp, LK_SHARED) != 0)
2210 if (vp->v_type != VREG) {
2214 error = VOP_GETATTR(vp, &va, cred);
2218 if (noff >= va.va_size) {
2222 bsize = vp->v_mount->mnt_stat.f_iosize;
2223 for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) {
2224 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL);
2225 if (error == EOPNOTSUPP) {
2229 if ((bnp == -1 && cmd == FIOSEEKHOLE) ||
2230 (bnp != -1 && cmd == FIOSEEKDATA)) {
2237 if (noff > va.va_size)
2239 /* noff == va.va_size. There is an implicit hole at the end of file. */
2240 if (cmd == FIOSEEKDATA)
2250 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td)
2255 off_t foffset, size;
2258 cred = td->td_ucred;
2260 foffset = foffset_lock(fp, 0);
2261 noneg = (vp->v_type != VCHR);
2267 (offset > 0 && foffset > OFF_MAX - offset))) {
2274 vn_lock(vp, LK_SHARED | LK_RETRY);
2275 error = VOP_GETATTR(vp, &vattr, cred);
2281 * If the file references a disk device, then fetch
2282 * the media size and use that to determine the ending
2285 if (vattr.va_size == 0 && vp->v_type == VCHR &&
2286 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0)
2287 vattr.va_size = size;
2289 (vattr.va_size > OFF_MAX ||
2290 (offset > 0 && vattr.va_size > OFF_MAX - offset))) {
2294 offset += vattr.va_size;
2299 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td);
2302 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td);
2307 if (error == 0 && noneg && offset < 0)
2311 VFS_KNOTE_UNLOCKED(vp, 0);
2312 *(off_t *)(td->td_retval) = offset;
2314 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0);
2319 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred,
2325 * Grant permission if the caller is the owner of the file, or
2326 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on
2327 * on the file. If the time pointer is null, then write
2328 * permission on the file is also sufficient.
2330 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes:
2331 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES
2332 * will be allowed to set the times [..] to the current
2335 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td);
2336 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0)
2337 error = VOP_ACCESS(vp, VWRITE, cred, td);