Project

General

Profile

Bug #57985

mds: warning `clients failing to advance oldest client/flush tid` seen with some workloads

Added by Venky Shankar 3 months ago. Updated 6 days ago.

Status:
Triaged
Priority:
Normal
Assignee:
Category:
Correctness/Safety
Target version:
% Done:

0%

Source:
Q/A
Tags:
Backport:
pacific,quincy
Regression:
No
Severity:
3 - minor
Reviewed:
Affected Versions:
ceph-qa-suite:
Component(FS):
Client, MDS
Labels (FS):
Pull request ID:
Crash signature (v1):
Crash signature (v2):

Description

https://bugzilla.redhat.com/show_bug.cgi?id=2134709

Generally seen when the MDS is heavily loaded with I/Os. Interesting thing is that the client-id in the warning message is (more often than not) the ceph-mgr daemon (libcephfs instance in ceph-mgr), although the libcephfs instance is not the one that is doing heaving I/Os (but there seems to be a possible in-direct relation somehow) - which needs RCA/explanation. Furthermore, it is possible that there are bugs lurking around client tid management which could cause these warnings.

History

#1 Updated by Venky Shankar 3 months ago

  • Status changed from New to Triaged
  • Assignee set to Ramana Raja

#2 Updated by Venky Shankar about 2 months ago

  • Assignee changed from Ramana Raja to Venky Shankar

#3 Updated by Venky Shankar 18 days ago

Venky Shankar wrote:

https://bugzilla.redhat.com/show_bug.cgi?id=2134709

Generally seen when the MDS is heavily loaded with I/Os. Interesting thing is that the client-id in the warning message is (more often than not) the ceph-mgr daemon (libcephfs instance in ceph-mgr), although the libcephfs instance is not the one that is doing heaving I/Os (but there seems to be a possible in-direct relation somehow) - which needs RCA/explanation. Furthermore, it is possible that there are bugs lurking around client tid management which could cause these warnings.

I've started to look into this - it is not always the ceph-mgr daemon that gets reported in the warning. Maybe it was just coincidence. The tid management in the client and the mds looks sane too. What can be done for better debugability is to dump the oldest_tid that clients maintain. That would give some insights into which request was not acknowledged by the mds.

#4 Updated by Venky Shankar 13 days ago

#5 Updated by Venky Shankar 6 days ago

Enforce (stricter) client-id check in client limit test - https://github.com/ceph/ceph/pull/49844

Also available in: Atom PDF