diff options
| author | Sage Weil <sage@inktank.com> | 2013-06-18 21:24:16 -0700 |
|---|---|---|
| committer | Sage Weil <sage@inktank.com> | 2013-06-18 21:24:16 -0700 |
| commit | ffade3c85dfffa13a16edd9630a52d99eb8a413d (patch) | |
| tree | 6f4d6674d18bfc72faa8c28515c3c2c23458c157 | |
| parent | c14dd154584eef97075e4a260719b6cbe686d4c7 (diff) | |
| download | ceph-ffade3c85dfffa13a16edd9630a52d99eb8a413d.tar.gz | |
os/FileStore: use fdatasync(2) instead of sync_file_range(2)
The use of sync_file_range(2) on XFS screws up XFS' delicate ordering
of writeback and range zeroing; see #4976 and this thread:
http://oss.sgi.com/archives/xfs/2013-06/msg00066.html
Instead, replace all sync_file_range(2) calls with fdatasync(2), which
*does* do ordered writeback and should not leak unzeroed blocks.
Signed-off-by: Sage Weil <sage@inktank.com>
| -rw-r--r-- | src/os/FileStore.cc | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/src/os/FileStore.cc b/src/os/FileStore.cc index fec058edd17..966035f36bc 100644 --- a/src/os/FileStore.cc +++ b/src/os/FileStore.cc @@ -2957,7 +2957,7 @@ int FileStore::_write(coll_t cid, const hobject_t& oid, !m_filestore_flusher || !(async_done = queue_flusher(fd, offset, len, replica))) { if (should_flush && m_filestore_sync_flush) { - ::sync_file_range(fd, offset, len, SYNC_FILE_RANGE_WRITE); + ::fdatasync(fd); local_flush = true; } } @@ -3325,7 +3325,7 @@ void FileStore::flusher_entry() q.pop_front(); if (!stop && ep == sync_epoch) { dout(10) << "flusher_entry flushing+closing " << fd << " ep " << ep << dendl; - ::sync_file_range(fd, off, len, SYNC_FILE_RANGE_WRITE); + ::fdatasync(fd); if (replica && m_filestore_replica_fadvise) { int fa_r = posix_fadvise(fd, off, len, POSIX_FADV_DONTNEED); if (fa_r) { |