On Tue, Oct 08, 2002 at 09:26:29PM -0700, Don Lewis wrote:
> On 8 Oct, Stefan Farfeleder wrote:
> > On Mon, Oct 07, 2002 at 03:48:45AM -0700, Terry Lambert wrote:
>
> > Following the advice from the spl* man page I turned the spl* calls to a
> > mutex and was surprised to see it working. My SMP -current survived a 'make
> > -j16 buildworld' with make using kqueue() (which it did not a single
> > time out of >30 times before). Further testings will follow tomorrow.
Building 6 worlds in a row with -j ranging from 4 to 128 didn't crash
it.
> > However, WITNESS complains (only once) about this:
> > lock order reversal
> > 1st 0xc662140c kqueue mutex (kqueue mutex) @
>/freebsd/current/src/sys/kern/kern_event.c:714
> > 2nd 0xc6727d00 pipe mutex (pipe mutex) @
>/freebsd/current/src/sys/kern/sys_pipe.c:1478
>
> That's pretty similar to the lock order reversal I've seen in the pipe
> code and it's interaction with sigio, which is not suprising since
> pipeselwakeup() calls both pgsigio() and KNOTE(), often while the pipe
> lock is held. Correctly fixing this doesn't look easy ...
Is it just a warning or does it pose a real problem?
I think the problem with the current code is that knote_{en,de}queue can
be executed in parallel (on another CPU, spl*() can't prevent that, can
it?) with kqueue_scan and that kq->kq_head thus can be corrupted.
Or am I totally wrong?
@Poul: Since you are the only person who reported a kernel crash too,
does the version with the mutex work for you?
Regards,
Stefan Farfelder
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-current" in the body of the message
