Mailing-List: contact dev-help@httpd.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@httpd.apache.org
Received-SPF: pass (nike.apache.org: domain of bojan@rexursive.com designates
 150.101.121.179 as permitted sender)
Subject: Re: worker MPM: close_worker_sockets race?
From: Bojan Smojver <bojan@rexursive.com>
To: dev@httpd.apache.org
In-Reply-To: <1256253140.9540.33.camel@shrek.rexursive.com>
References: <1256173732.9540.23.camel@shrek.rexursive.com>
	 <cc67648e0910220454j202eb399y951ca6217052397e@mail.gmail.com>
	 <1256245093.9540.31.camel@shrek.rexursive.com>
	 <1256253140.9540.33.camel@shrek.rexursive.com>
Content-Type: text/plain
Date: Fri, 30 Oct 2009 13:07:19 +1100
Message-Id: <1256868439.2488.21.camel@shrek.rexursive.com>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit

On Fri, 2009-10-23 at 10:12 +1100, Bojan Smojver wrote:
> Slower, but safer attached (not even compiled, so could be utterly
> bogus).

BTW, this approach still isn't safe. The socket can disappear (i.e. be
closed) between the test and the closing (or shutdown()). Worse, with
shutdown(), we could attempt to shut down the wrong FD, if we have fast
succession of close()/open() in various threads.

I think the correct approach would be to make sure that we actually
suspend worker threads before we do this. In other words, we send a
signal to each worker thread and their signal handler calls
sigsuspend(), thus waiting to be resumed. We also set thread specific
variables, notifying the listener thread about workers that have been
suspended.

Once the listener thread knows this (i.e. that all workers are
suspended), it shuts the sockets down and resumes the worker threads by
sending them a signal.

Workable?

-- 
Bojan