Alexander Bergolth wrote:
> Btw. I've found out that clients on AIX (5.2., oslevel 07) also show the
> same problem.
> However Windows Clients (Version 1.4.0008) don't suffer from the delays.

I find this quite interesting.   This provides a bit more weight to my
theory that the actual bug is not in the RX patches that were applied
between .82 and .84.  I believe there is a race condition in the RX
library on non-Windows platforms.  The patches to RX to use thread local
queues and statistics gathering increases the parallelism and therefore
the likelihood that the race condition will be triggered.

I suspect that this race condition is the cause of both these delays
and perhaps the Solaris 10 deadlock.   I believe the delays are caused
by inbound response packets not being properly associated with the
outbound request packets thereby forcing a retry.   Race conditions are
very hard to track down.

The other problem with a 50+ second delay when the client changes the
source ip-addr/port pair has been traced to a lock on the host entry
that is being held across an RPC sent to the client to determine if the
client is still available on the previous ip-addr/port.  A patch for
this part is being worked on.

Jeffrey Altman


Attachment: smime.p7s
Description: S/MIME Cryptographic Signature

Reply via email to