perf: weird si_code results on overflow signal

[Vince Weaver]

Hello

I just compiled up a fresh git kernel and all of the perf_event_test 
overflow tests are failing.

The reason is that instead of getting POLL_IN or POLL_HUP sources as 
expected, they are getting weird results in si_code of "-5".

I haven't had time to bisect this, but I do notice that some major changes 
to si_code handling went in with
	64a76d0d64bea159da997c002a916ffc03f98bfc

Vince