git: a5e284038edc - main - open(2): Add O_DSYNC flag.
Thomas Munro
tmunro at FreeBSD.org
Fri Jan 8 00:16:58 UTC 2021
The branch main has been updated by tmunro:
URL: https://cgit.FreeBSD.org/src/commit/?id=a5e284038edc36b0447f1e6337419a3c0ea1788d
commit a5e284038edc36b0447f1e6337419a3c0ea1788d
Author: Thomas Munro <tmunro at FreeBSD.org>
AuthorDate: 2021-01-07 10:46:51 +0000
Commit: Thomas Munro <tmunro at FreeBSD.org>
CommitDate: 2021-01-08 00:15:56 +0000
open(2): Add O_DSYNC flag.
POSIX O_DSYNC means that writes include an implicit fdatasync(2), just
as O_SYNC implies fsync(2).
VOP_WRITE() functions that understand the new IO_DATASYNC flag can act
accordingly, but we'll still pass down IO_SYNC so that file systems that
don't understand it will continue to provide the stronger O_SYNC
behaviour.
Flag also applies to fcntl(2).
Reviewed by: kib, delphij
Differential Revision: https://reviews.freebsd.org/D25090
---
lib/libc/sys/fcntl.2 | 19 +++++++++++++++++--
lib/libc/sys/open.2 | 20 +++++++++++++-------
.../openzfs/include/os/freebsd/spl/sys/ccompile.h | 1 -
.../openzfs/include/os/freebsd/spl/sys/vnode.h | 2 ++
sys/kern/vfs_vnops.c | 7 +++++++
sys/sys/fcntl.h | 15 +++++++++------
sys/sys/vnode.h | 1 +
7 files changed, 49 insertions(+), 16 deletions(-)
diff --git a/lib/libc/sys/fcntl.2 b/lib/libc/sys/fcntl.2
index 9793024f48ed..33ad7a5673e1 100644
--- a/lib/libc/sys/fcntl.2
+++ b/lib/libc/sys/fcntl.2
@@ -28,7 +28,7 @@
.\" @(#)fcntl.2 8.2 (Berkeley) 1/12/94
.\" $FreeBSD$
.\"
-.Dd January 17, 2020
+.Dd January 6, 2021
.Dt FCNTL 2
.Os
.Sh NAME
@@ -196,7 +196,7 @@ The flags for the
.Dv F_GETFL
and
.Dv F_SETFL
-flags are as follows:
+commands are as follows:
.Bl -tag -width O_NONBLOCKX
.It Dv O_NONBLOCK
Non-blocking I/O; if no data is available to a
@@ -225,6 +225,21 @@ Enable the
signal to be sent to the process group
when I/O is possible, e.g.,
upon availability of data to be read.
+.It Dv O_SYNC
+Enable synchronous writes.
+Corresponds to the
+.Dv O_SYNC
+flag of
+.Xr open 2 .
+.Dv O_FSYNC
+is an historical synonym for
+.Dv O_SYNC .
+.It Dv O_DSYNC
+Enable synchronous data writes.
+Corresponds to the
+.Dv O_DSYNC
+flag of
+.Xr open 2 .
.El
.Pp
The seals that may be applied with
diff --git a/lib/libc/sys/open.2 b/lib/libc/sys/open.2
index d0918fb02eee..e43d012770df 100644
--- a/lib/libc/sys/open.2
+++ b/lib/libc/sys/open.2
@@ -181,8 +181,9 @@ O_EXCL error if create and file exists
O_SHLOCK atomically obtain a shared lock
O_EXLOCK atomically obtain an exclusive lock
O_DIRECT eliminate or reduce cache effects
-O_FSYNC synchronous writes
+O_FSYNC synchronous writes (historical synonym for O_SYNC)
O_SYNC synchronous writes
+O_DSYNC synchronous data writes
O_NOFOLLOW do not follow symlinks
O_NOCTTY ignored
O_TTY_INIT ignored
@@ -230,15 +231,18 @@ returns immediately.
The descriptor remains in non-blocking mode for subsequent operations.
.Pp
If
-.Dv O_FSYNC
+.Dv O_SYNC
is used in the mask, all writes will
immediately and synchronously be written to disk.
-.Pp
-.Dv O_SYNC
-is a synonym for
.Dv O_FSYNC
-required by
-.Tn POSIX .
+is an historical synonym for
+.Dv O_SYNC .
+.Pp
+If
+.Dv O_DSYNC
+is used in the mask, all data and metadata required to read the data will be
+synchronously written to disk, but changes to metadata such as file access and
+modification timestamps may be written later.
.Pp
If
.Dv O_NOFOLLOW
@@ -651,6 +655,8 @@ The
.Fn openat
function was introduced in
.Fx 8.0 .
+.Dv O_DSYNC
+appeared in 13.0.
.Sh BUGS
The Open Group Extended API Set 2 specification requires that the test
for whether
diff --git a/sys/contrib/openzfs/include/os/freebsd/spl/sys/ccompile.h b/sys/contrib/openzfs/include/os/freebsd/spl/sys/ccompile.h
index a02e8f098540..524b81e68e61 100644
--- a/sys/contrib/openzfs/include/os/freebsd/spl/sys/ccompile.h
+++ b/sys/contrib/openzfs/include/os/freebsd/spl/sys/ccompile.h
@@ -168,7 +168,6 @@ extern "C" {
#define O_LARGEFILE 0
#define O_RSYNC 0
-#define O_DSYNC 0
#ifndef LOCORE
#ifndef HAVE_RPC_TYPES
diff --git a/sys/contrib/openzfs/include/os/freebsd/spl/sys/vnode.h b/sys/contrib/openzfs/include/os/freebsd/spl/sys/vnode.h
index 6a6146132765..fa7bbd88c6c8 100644
--- a/sys/contrib/openzfs/include/os/freebsd/spl/sys/vnode.h
+++ b/sys/contrib/openzfs/include/os/freebsd/spl/sys/vnode.h
@@ -127,7 +127,9 @@ vn_is_readonly(vnode_t *vp)
#define FCREAT O_CREAT
#define FTRUNC O_TRUNC
#define FEXCL O_EXCL
+#ifndef FDSYNC
#define FDSYNC FFSYNC
+#endif
#define FRSYNC FFSYNC
#define FSYNC FFSYNC
#define FOFFMAX 0x00
diff --git a/sys/kern/vfs_vnops.c b/sys/kern/vfs_vnops.c
index 2ca2cf124c23..de5cd68501a7 100644
--- a/sys/kern/vfs_vnops.c
+++ b/sys/kern/vfs_vnops.c
@@ -1066,6 +1066,13 @@ vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags,
if ((fp->f_flag & O_FSYNC) ||
(vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS)))
ioflag |= IO_SYNC;
+ /*
+ * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE()
+ * implementations that don't understand IO_DATASYNC fall back to full
+ * O_SYNC behavior.
+ */
+ if (fp->f_flag & O_DSYNC)
+ ioflag |= IO_SYNC | IO_DATASYNC;
mp = NULL;
if (vp->v_type != VCHR &&
(error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
diff --git a/sys/sys/fcntl.h b/sys/sys/fcntl.h
index 2f424d173949..e2597726c53b 100644
--- a/sys/sys/fcntl.h
+++ b/sys/sys/fcntl.h
@@ -141,8 +141,10 @@ typedef __pid_t pid_t;
return back */
#endif
+#define O_DSYNC 0x00800000 /* POSIX data sync */
+
/*
- * XXX missing O_DSYNC, O_RSYNC.
+ * XXX missing O_RSYNC.
*/
#ifdef _KERNEL
@@ -158,9 +160,9 @@ typedef __pid_t pid_t;
#define OFLAGS(fflags) ((fflags) & O_EXEC ? (fflags) : (fflags) - 1)
/* bits to save after open */
-#define FMASK (FREAD|FWRITE|FAPPEND|FASYNC|FFSYNC|FNONBLOCK|O_DIRECT|FEXEC)
+#define FMASK (FREAD|FWRITE|FAPPEND|FASYNC|FFSYNC|FDSYNC|FNONBLOCK|O_DIRECT|FEXEC)
/* bits settable by fcntl(F_SETFL, ...) */
-#define FCNTLFLAGS (FAPPEND|FASYNC|FFSYNC|FNONBLOCK|FRDAHEAD|O_DIRECT)
+#define FCNTLFLAGS (FAPPEND|FASYNC|FFSYNC|FDSYNC|FNONBLOCK|FRDAHEAD|O_DIRECT)
#if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \
defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4)
@@ -185,15 +187,16 @@ typedef __pid_t pid_t;
#define FAPPEND O_APPEND /* kernel/compat */
#define FASYNC O_ASYNC /* kernel/compat */
#define FFSYNC O_FSYNC /* kernel */
+#define FDSYNC O_DSYNC /* kernel */
#define FNONBLOCK O_NONBLOCK /* kernel */
#define FNDELAY O_NONBLOCK /* compat */
#define O_NDELAY O_NONBLOCK /* compat */
#endif
/*
- * We are out of bits in f_flag (which is a short). However,
- * the flag bits not set in FMASK are only meaningful in the
- * initial open syscall. Those bits can thus be given a
+ * Historically, we ran out of bits in f_flag (which was once a short).
+ * However, the flag bits not set in FMASK are only meaningful in the
+ * initial open syscall. Those bits were thus given a
* different meaning for fcntl(2).
*/
#if __BSD_VISIBLE
diff --git a/sys/sys/vnode.h b/sys/sys/vnode.h
index ff1323cb8719..0eadfec02313 100644
--- a/sys/sys/vnode.h
+++ b/sys/sys/vnode.h
@@ -326,6 +326,7 @@ struct vattr {
#define IO_NOMACCHECK 0x1000 /* MAC checks unnecessary */
#define IO_BUFLOCKED 0x2000 /* ffs flag; indir buf is locked */
#define IO_RANGELOCKED 0x4000 /* range locked */
+#define IO_DATASYNC 0x8000 /* do only data I/O synchronously */
#define IO_SEQMAX 0x7F /* seq heuristic max value */
#define IO_SEQSHIFT 16 /* seq heuristic in upper 16 bits */
More information about the dev-commits-src-all
mailing list