Ceph » CephFS

I run a metadata bench test, which about 30,000 files rename test in one directory, and always got mds slow request about wrlock failed message.
And when we got mds slow request, the tested directory hangs and never recover even reboot all mds.

My test ceph environment is 14.2.8, has 3 active mds, deployed on CentOS 7.5.

The test Calready attached, which test metadata likes below:
In a directory A includes 30,000 files about 100MB~4GB randomly.
The test will sort all files by name, and place a class of file in new created subdir of directory A.

I dumped directory A inode in all of them. The inode log attached.

In mds log, we see export operation, so we config mds argument(mds_bal_split_xxx/mds_bal_fragment_xxx) to disable dir fragment.
After that, we run the metadata bench scripts, and the problem do not reprouducted.

Through the mds log attached, we found the inest lock is always waiting stable. Maybe there is something wrong on rename lock operations?