Date: Mon, 10 Apr 2017 20:57:11 +0100
From: Al Viro <viro@ZenIV.linux.org.uk>
To: Dave Jones <davej@codemonkey.org.uk>,
        Linux Kernel <linux-kernel@vger.kernel.org>
Subject: Re: iov_iter_pipe warning.
Message-ID: <20170410195711.GD29622@ZenIV.linux.org.uk>
References: <20170321205901.mojmjd2fxmykq3df@codemonkey.org.uk>
 <20170410192800.GC29622@ZenIV.linux.org.uk>
 <20170410194206.loesu5licstif7or@codemonkey.org.uk>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20170410194206.loesu5licstif7or@codemonkey.org.uk>
User-Agent: Mutt/1.7.1 (2016-10-04)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1293
Lines: 29

On Mon, Apr 10, 2017 at 03:42:06PM -0400, Dave Jones wrote:
>  > Do you have a reliable reproducer?
> 
> Not reliable, but I'll see if I can find some time to narrow it down
> this week.  I've been working on better logging of "wtf just happened"
> the last few weeks, so it should get easier once I finish that work.

I would suggest
	a) slapping WARN_ON(pipe->nr_bufs); right before the loop
in splice_direct_to_actor().  Internal pipe should be empty when we
enter this function.
	b) the same WARN_ON() in the very end of the loop body.

We should have started with empty pipe.  We'd called ->splice_read()
and it had returned a positive number (in read_len).  That should be
the amount we'd pushed in there.  Then we call actor(), with
sd->total_len set to read_len.  Its return value is
	* positive (or we would've buggered off)
	* no less than read_len (ditto)
so it should have drained the pipe entirely, leaving it empty again.

Finding it not just non-empty, but full means that something's very
wrong.  The actor here is essentially ->splice_write(), and I'm really
curious which file is the target.  Actually, could you turn those
WARN_ON() into
	if (WARN_ON(pipe->nr_bufs))
		printk(KERN_ERR "->splice_write = %p",
			sd->u.file->f_op->splice_write);
and see which function it is?