Red Hat Bugzilla – Bug 591139
RDMA client shutdown is broken (client hangs)
Last modified: 2011-08-12 12:22:54 EDT
Description of problem:
Since the changes to the client shutdown in trunk r934503 rdma the client shutdown() callback will no longer be called when clients close an rdma connection.
This results in the client hanging whilst waiting for rdma connections to be completely closed.
Steps to Reproduce:
1. run "qpidd --auth no" (you will need to have a working rdma plugin)
To make sure that rdma is running, check that there is a line like:
2010-05-10 17:55:04 notice Rdma: Listening on RDMA port 5672
in the output.
2. run perftest -Prdma -b <IP address of IB interface>
3. Wait forever for exit after the test completes and print results.
This is fixed as of r954499 upstream
Verified on RHEL5 x86_64
But I am unable to reproduce described bug.
It is strange that I get the same (good, no hang) results on version
which should be buggy:
bash-3.2# rpm -qa | grep qpid
bash-3.2# qpid-perftest -Prdma -b 192.168.55.26
Processing 1 messages from qpid-perftest_sub_ready . done.
Sending start 1 times to qpid-perftest_pub_start
Processing 1 messages from qpid-perftest_pub_done . done.
Processing 1 messages from qpid-perftest_sub_done . done.
... [SNIP] ...
Total transfers/sec: 122993
Total Mbytes/sec: 120.111
I think that -2 had most of the fix in and you should try to reproduce with -1
So I have tried
none of above reproduce the bug how you describe it.
I am sure I use RDMA as I get troughput ~136 MB/s.
May the bug be connected with something else (libibverbs,
particular IB driver, etc.)? I run the broker without
store or acl, focusing on RDMA.
On which machine did you experience the reported bug?
What hardware did you use?
I use mrg26.lab.bos.redhat.com, which has (by lscpi)
Mellanox Technologies MT25204 [InfiniHost III Lx HCA]
i.e. the package containing its driver is libmthca.
More info: The machine (mrg26) I tried to reproduce this bug on
was running RHEL5 x86_64 with following MRG packages installed:
Though the bug is not reproducible, I am sure the described
bug is no longer valid in current versions. --> VERIFIED