W. Trevor King c82166afab cmd/podman/utils.go: Cancel-able resize writes ...

On Thu, Jun 28, 2018 at 03:48:26AM -0700, Marco Vedovati wrote [1]:
> The root cause is a deadlock between two channel writes made by two
> different goroutines:
>
> 1. `resizeTty() : go func(){} : sendUpdate()` is sending a resize
>    message thru `resize` right at the beginning, but the channel is
>    never read if some startup error occurs.
>
> 2. Upon program termination, `startAttachCtr() : defer func(){} ` is
>    telling the goroutine in "1." to stop via the `resizeTerminate`
>    channel. But that guy is still waiting for the write to `resize`
>    to complete so the the termination message is never read.
>
> I think the go deadlock detection does not kick in because not all
> goroutines are seen as asleep. E.g. `os/signal Notify()` is enough
> to have the deadlock not detected.

333ab8c2 (Fix podman hangs when detecting startup error in container
attached mode, 2018-06-27, #1010) addressed this with a deferred
drain.  This commit adjusts that approach to use a single select to
cover "have we been canceled?", "has there been a resize signal?", and
(when we have one) "can we write the most recent resize event to the
resize channel?".

A side benefit to this approach is that if we have a slow resize
consumer and several resize signals, the resizeTty function will keep
updating its local resizeEvent.  Once the resize channel is able to
accept, only the most-recent event will be written.  Previously we'd
have written one resize event for every received signal, even if the
resize consumer was falling behind.

[1]: https://github.com/projectatomic/libpod/pull/1010#issuecomment-400994436

Signed-off-by: W. Trevor King <wking@tremily.us>

Closes: #1018
Approved by: rhatdan

2018-06-28 15:53:02 +00:00

..

podman

cmd/podman/utils.go: Cancel-able resize writes

2018-06-28 15:53:02 +00:00