Code review comment for lp:~jameinel/launchpad/lp-serve-child-hangup

Revision history for this message
Martin Pool (mbp) wrote :

On 17 February 2011 08:44, John A Meinel <email address hidden> wrote:
> John A Meinel has proposed merging lp:~jameinel/launchpad/lp-serve-child-hangup into lp:launchpad with lp:~jameinel/launchpad/lp-forking-serve-cleaner-childre as a prerequisite.
> Requested reviews:
>  Launchpad code reviewers (launchpad-reviewers)
> For more details, see:
> This branch builds on my earlier branch, which has the children spawned by LPForkingService clean themselves up more readily.
> This adds code so that the blocking calls will be unblocked after 2 minutes. At the moment, if the Conch server requests a fork, but is unable to actually connect to the child, that child ends up hanging forever on a file handle that it will never open.
> Originally, I got this working using a loop an O_NONBLOCK. But after doing some testing, fcntl.fcntl(...) was unable to actually change the file descriptor to be blocking again. It looks good in man, and it does unset the flag, but I still get EAGAIN failures in the smart server code.
> I don't really like spawning a thread to send SIGUSR1 back to myself, but it seemed the best tradeoff. If we want, we could even make it SIGTERM or something, since we know we are going to kill the process if the connection fails.

I don't really like that either, especially this is already getting to
be a fairly complex set of processes.

Why can't you make the file handles nonblocking and just leave them
so? Is it because they're later inherited by some other process?

If you do need a signal to interrupt things, how about just using alarm(2)?

« Back to merge proposal