RE: [Hang] t5562 subtest 8 on NonStop

From: Randall S. Becker <rsbecker@nexbridge.com>
Date: Feb 9, 2019, 20:22 UTC
Message-ID: <000b01d4c0b5$358b6b50$a0a241f0$@nexbridge.com>
In-Reply-To: <002301d4bfa9$158716e0$409544a0$@nexbridge.com>

On February 8, 2019 7:23, I wrote:

> We have suddenly encountered a hung git-http-backend in t5562 in the
> NonStop port. This is a new problem not seen before on the platform,
> surprisingly. I am wondering whether this is a result of not actually

having an

> apache2 server on-board. Is that a possibility and can that sub-test be
> bypassed if no apache2 is detected?
> 
> We also had subtest 15 and I am investigating, but it may depend on 8's

data

> (I had to kill the git-http-backend process, so maybe that's why).

I'm back to investigating subtest 8 hanging after dealing with the /dev/zero thing. What we end up with two git processes, a git-http-backend, a perl, and a few shell processes. The act.err file had:

fatal: request ended in the middle of the gzip stream

Which went through die() and should have unwound the whole thing, but didn't. The git-http-backend is waiting to clean up children and each git is polling for input. It seems like SIGCHLD did not get back through the process that started it, or its parent - but I don't quite follow the flow in this test. I have seen the perl port not like SIGCHLD sometimes, so that might be what is going on. Is it possible to skip the hanging subtest during a whole test run? We don't really have anyone using the http backend on platform (SSH is the generally preferred method to get to git on box), so I am tempted to ignore the ones that hang as conditional known breakages, if that is possible, so that we can get our CI builds back in operation. Thoughts? Advice?

Thanks, Randall

-- Brief whoami:
 NonStop developer since approximately 211288444200000000
 UNIX developer since approximately 421664400
-- In my real life, I talk too much.