7200742: (se) Selector.select does not block when starting Coherence (sol11u1)

[Alan Bateman]

I need a reviewer for a change that resolves a regression in the Solaris 
/dev/poll based Selector. The regression stems from the changes for 
7169050 [1] whereby a POLLREMOVE is no longer written to the driver when 
changing events. As the events are OR'ed in the kernel it creates the 
potential for spinning when changing the interest ops from say OP_WRITE 
to OP_READ and vice versa.

It's somewhat surprising that the issue wasn't caught by any of the 
tests but they tend not to notice performance issues.  It probably 
hasn't been noticed by many servers either as typically events are 
handed over to worker threads and that typically involves setting the 
interest ops to 0 before the hand off (changing the interest ops to 0 
causes a POLLREMOVE to be written). The issue only arises when changing 
the interest ops without changing to 0.

The webrev with the changes is here and I've got a test that exercises 
all possible transitions.

http://cr.openjdk.java.net/~alanb/7200742/webrev/

As 7169050 was back-ported to jdk7u it means that the changes here will 
need to be back-ported too.

Thanks,

Alan.

[1] http://hg.openjdk.java.net/jdk8/tl/jdk/rev/21703d431217