[Samba] winbindd requests failing with NT_STATUS_PIPE_BROKEN

[Micha Lenk]

Hi Samba experts,

I am currently investigating an issue with a bigger winbindd
installation which causes all following winbindd requests to fail until
winbindd is restarted. We use a slightly patched version of winbindd
based on Samba 3.5.8. In the used setup the winbindd is joined to an AD
domain and is quite busy with answering winbindd requests of type
PAM_AUTH_CRAP, LOOKUPNAME, GETUSERSIDS, and LOOKUPSID. Several AD users
are member of more than 128 groups. In this production environment the
issue happens a few times a day, however in our development environment
the issue is unreproducible.

All observed occurrences have in common that we see following messages
in debug output:

winbindd: Exceeding 200 client connections, removing idle connection.
Found 200 idle client connections, shutting down sock 122, pid 28456
accepted socket 122

Once the issue occurred, all further requests fail with the following
lines in debug output (loglevel 10):

Could not lookup sid S-1-5-21-1710691794-657421245-3729435916-358021:
NT_STATUS_PIPE_BROKEN
wb_request_done[28456:LOOKUPSID]: NT_STATUS_PIPE_BROKEN
winbind_client_response_written[28456:LOOKUPSID]: deliverd response to
client

Right now we don't understand really what is going on. So, the big
question is: What could cause the broken pipe message? And what
expression should I search the logfile for to get more information about
the possible reasons?

I would appreciate any assumptions about the reason that might trigger
the issue or suggestions on how to debug that issue further.

Regards,
Micha