gperciva
commented
9 months ago Added a third fix (not as embarrassing).
References (but nothing looks relevant):
- add LIBS to the build system was originally https://github.com/Tarsnap/libcperciva/pull/474
- using LIST for
dispatch.chttps://github.com/Tarsnap/spiped/pull/53
More importantly, we discussed shutdown() receiving an ENOTCONN back in https://github.com/Tarsnap/spiped/pull/350; the final opinion was:
Ok, I think this is ok; it looks to me like ENOTCONN should not (non-erroneously) happen here, even if we get RST.
If users start reporting ENOTCONN then we should probably treat it as a non-error.
We now have a report of ENOTCONN in #397 (although the cause was disguised by the problem in dispatch_shutdown()), so this PR makes it a non-error.
(I can reproduce the ENOTCONN using postgresql; or with our test suite if I add a few sleep(1) to our code. I haven't yet been able to reproduce ENOTCONN with test programs without modifying spiped code, but I think it's sufficiently clear what's happening.)
Two embarrassing mistakes here; the one in
dispatch_shutdown()is particularly bad.These were uncovered during the investigation into #397, but they don't resolve whatever is happening in that issue. (I can reproduce the problem, but I'm still trying to figure out exactly what's happening there.)