Re: 4.9.0 regression in pipe-backed iov_iter with systemd-nspawn

From: Al Viro
Date: Thu Jan 12 2017 - 17:37:30 EST


On Thu, Jan 12, 2017 at 02:26:42PM -0800, Linus Torvalds wrote:
> On Thu, Jan 12, 2017 at 12:26 PM, Alan J. Wylie <alan@xxxxxxxxxxx> wrote:
> >
> > Strace shows that the processes are hanging in write() and read() calls.
>
> If this is splice-related, I'm assuming that they aren't actually the
> two ends of the same pipe, and there is somebody doing splice in the
> middle.
>
> I'm not seeing that process. I'm assuming it's systemd. Can you try
> to find it and strace that one too? Because that middle man is likely
> the one that has problems (and is not able to splice from one pipe to
> the other).
>
> Ugh. That one commit has had a lot of bugs in it already. We do not
> have good splice test coverage, because almost nobody uses it.

FWIW, I would really like to know what kind of files had been involved.
There are two paths that can lead to default_file_splice_read():
splice_direct_to_actor() -> do_splice_to() -> default_file_splice_read() and
do_splice() -> do_splice_to() -> default_file_splice_read().

The former only gets there for regular files and block devices. The latter
is guaranteed that file is not a pipe. So
* not a socket (have ->splice_read() of their own)
* not a pipe or FIFO (neither path allows those)
* not a block device (have ->splice_read() of their own)
* not a regular file on a normal local fs (ditto)

So what is it called for in that reproducer?