Dražen Kačar wrote:
> There is a serious flaw there. The pthread_setschedparam() function is
> being called by the main thread before other threads are created, so all
> threads will inherit the scheduling parameters, which we don't want. That
> call has to be moved from willRun() function to the rcvMainLoop() which is
> being executed by a new thread so it will afect the UDP thread only. You
> can check this with:
>
> ps -eLo "cmd lwp policy rtprio"
>
> I'm attaching a patch which does that, against rsyslog 5.6.2. (After
> applying you need to call autoreconf). It also has improved validation.
> Note that the allowed scheduling priority range is OS specific, so the
> code shouldn't expect that it will be non-negative.
I've put the patch to test on the server under load and there were several
threads with FIFO scheduling policy. One of them was receiving packets
from the UDP socket. Another was locking and unlocking mutex (or mutexes)
and strace didn't show anything else. Yet another thread was writing to
the log file. There was one or two more, but I didn't bother to check what
they were doing.
This is not good. It seems like the UDP receiver thread is spawning
other threads and then they inherit scheduling policy. That should be
prevented somehow.
--
.-. .-. Yes, I am an agent of Satan, but my duties are largely
(_ \ / _) ceremonial.
|
| [email protected]
_______________________________________________
rsyslog mailing list
http://lists.adiscon.net/mailman/listinfo/rsyslog
http://www.rsyslog.com
