SASADA Koichi <ko1 / atdot.net> wrote:
> (2011/09/20 2:22), Tomoyuki Chikanaga wrote:
> > I guess this is because SIGVTALRM from send from ubf_select() is received by blocking thread after GVL release but before enter poll(2)/ppoll(2).
> 
> Thanks.  I understand.  How about it?

I'm not sure issues like this are completely avoidable, but your patch
seems to help avoid it.

I needed ~15K tries to reproduce blocking on the original test case
(no SMP on this Xen VM).  So far I'm at 65K loop interations and still
going...  I will update if I notice it blocked again.