[omniORB] crash in openssl code

[Duncan Grisby]

On Sat, 2024-02-24 at 18:42 +0100, Michael Teske wrote:
> 
> That's good news. If I can be of any help let me know. From my
> experience on Linux, a mutex is not very expensive if it is not
> locked by another thread, but other systems may be different
> (I know that at least a few years ago it was rather expensive on
> OSX).
> I'll try to write a small test program next week which triggers the
> problem more often, so I can test more easily.

That would be helpful. I have tried to reproduce the crash, but I
haven't managed. I can see bidirectional GIOP causing it to call
SSL_write while it's blocked in SSL_read, but it has not once crashed.
This is with OpenSSL 3.1.1 on Fedora 39.

My performance concern is not just the lock, but also the overhead of
non-blocking sockets, which means it needs to call poll() or select()
to see when the socket is available, and potentially makes multiple
calls into OpenSSL instead of just one. A brief experiment that doesn't
properly implement the solution but does try to simulate the overheads
suggests that it is likely to add about 10% to the call times, so it is
not a trivial cost.

Duncan.

-- 
Duncan Grisby <duncan at grisby.org>