diff options
| author | Krutika Dhananjay <kdhananj@redhat.com> | 2016-03-29 18:36:08 +0530 |
|---|---|---|
| committer | Jeff Darcy <jdarcy@redhat.com> | 2016-04-11 12:48:53 -0700 |
| commit | c272c71391cea9db817f4e7e38cfc25a7cff8bd5 (patch) | |
| tree | e7f6b47d9cc44fa6fb9812a7738e2a2d2f1cfdce | |
| parent | d129d4eea33aae5db24dba17adcb04e9d4829817 (diff) | |
features/shard: Make o-direct writes work with sharding
With files opened with o-direct, the expectation is that
the IO performed on the fds is byte aligned wrt the sector size
of the underlying device. With files getting sharded, a single
write from the application could be broken into more than one write
falling on different shards which _might_ cause the original byte alignment
property to be lost. To get around this, shard translator will send fsync
on odirect writes to emulate o-direct-like behavior in the backend.
Change-Id: Ie8a6c004df215df78deff5cf4bcc698b4e17a7ae
BUG: 1322214
Signed-off-by: Krutika Dhananjay <kdhananj@redhat.com>
Reviewed-on: http://review.gluster.org/13846
Smoke: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Pranith Kumar Karampuri <pkarampu@redhat.com>
NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org>
CentOS-regression: Gluster Build System <jenkins@build.gluster.com>
| -rw-r--r-- | xlators/features/shard/src/shard.c | 6 |
1 files changed, 6 insertions, 0 deletions
diff --git a/xlators/features/shard/src/shard.c b/xlators/features/shard/src/shard.c index 154411682f3..c5f80b198b2 100644 --- a/xlators/features/shard/src/shard.c +++ b/xlators/features/shard/src/shard.c @@ -3591,6 +3591,7 @@ shard_common_inode_write_do (call_frame_t *frame, xlator_t *this) shard_local_t *local = NULL; struct iovec *vec = NULL; gf_boolean_t wind_failed = _gf_false; + gf_boolean_t odirect = _gf_false; off_t orig_offset = 0; off_t shard_offset = 0; off_t vec_offset = 0; @@ -3621,6 +3622,9 @@ shard_common_inode_write_do (call_frame_t *frame, xlator_t *this) return 0; } + if ((fd->flags & O_DIRECT) && (local->fop == GF_FOP_WRITE)) + odirect = _gf_true; + while (cur_block <= last_block) { if (wind_failed) { shard_common_inode_write_do_cbk (frame, @@ -3678,6 +3682,8 @@ shard_common_inode_write_do (call_frame_t *frame, xlator_t *this) NULL, NULL); goto next; } + if (odirect) + local->flags |= O_SYNC; } shard_common_inode_write_wind (frame, this, anon_fd, |