]> git-server-git.apps.pok.os.sepia.ceph.com Git - ceph-ci.git/commit
tools/cephfs_mirror: Do remote fs sync once instead of fsync on each fd
authorKotresh HR <khiremat@redhat.com>
Sun, 15 Feb 2026 19:38:00 +0000 (01:08 +0530)
committerKotresh HR <khiremat@redhat.com>
Tue, 17 Feb 2026 20:10:51 +0000 (01:40 +0530)
commitdcd6a5985d96f329e46d7cf297463ab805c119e2
tree952f6965e39a315ce73bdcf7ce08be7fc9d6d2ff
parent4a16f34066962c45f21aaf856191412e509fcc2f
tools/cephfs_mirror: Do remote fs sync once instead of fsync on each fd

Doing remote fs sync once just before taking snapshot is faster than
doing fsync on each fd. Moreover, all the datasync threads use the
same sinlge libceph connection and doing ceph_fsync concurrently on
different fds on a single libcephfs connection could cause hang
as observed in testing as below.

-----
Thread 2 (Thread 0xffff644cc400 (LWP 74020) "d_replayer-0"):
0  0x0000ffff8e82656c in __futex_abstimed_wait_cancelable64 () from /lib64/libc.so.6
1  0x0000ffff8e828ff0 [PAC] in pthread_cond_wait@@GLIBC_2.17 () from /lib64/libc.so.6
2  0x0000ffff8fc90fd4 [PAC] in ceph::condition_variable_debug::wait ...
3  0x0000ffff9080fc9c in ceph::condition_variable_debug::wait<Client::wait_on_context_list ...
4  Client::wait_on_context_list ... at /lsandbox/upstream/ceph/src/client/Client.cc:4540
5  0x0000ffff9083fae8 in Client::_fsync ... at /lsandbox/upstream/ceph/src/client/Client.cc:13299
6  0x0000ffff90840278 in Client::_fsync ...
7  0x0000ffff90840514 in Client::fsync ... at /lsandbox/upstream/ceph/src/client/Client.cc:13042
8  0x0000ffff907f06e0 in ceph_fsync ... at /lsandbox/upstream/ceph/src/libcephfs.cc:316
9  0x0000aaaaad5b2f88 in cephfs::mirror::PeerReplayer::copy_to_remote ...
----

Fixes: https://tracker.ceph.com/issues/73452
Signed-off-by: Kotresh HR <khiremat@redhat.com>
src/tools/cephfs_mirror/PeerReplayer.cc