Re: net: BUG still has locks held in unix_stream_splice_read

From: Al Viro
Date: Sun Oct 09 2016 - 23:15:07 EST

On Mon, Oct 10, 2016 at 03:46:07AM +0100, Al Viro wrote:
> On Sun, Oct 09, 2016 at 12:06:14PM +0200, Dmitry Vyukov wrote:
> > I suspect this is:
> >
> > commit 25869262ef7af24ccde988867ac3eb1c3d4b88d4
> > Author: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
> > Date: Sat Sep 17 21:02:10 2016 -0400
> > skb_splice_bits(): get rid of callback
> > since pipe_lock is the outermost now, we don't need to drop/regain
> > socket locks around the call of splice_to_pipe() from skb_splice_bits(),
> > which kills the need to have a socket-specific callback; we can just
> > call splice_to_pipe() and be done with that.
> Unlikely, since that particular commit removes unlocking/relocking ->iolock
> around the call of splice_to_pipe(). Original would've retaken the same
> lock on the way out; it's not as if we could leave the syscall there.
> It might be splice-related, but I don't believe that you've got the right
> commit here.

It's not that commit, all right - it's "can't call unix_stream_read_generic()
with any locks held" stepped onto a couple of commits prior by
"splice: lift pipe_lock out of splice_to_pipe()". Could somebody explain
what is that about?

E.g what will happen if some code does a read on AF_UNIX socket with
some local mutex held? AFAICS, there are exactly two callers of
freezable_schedule_timeout() - this one and one in XFS; the latter is
in a kernel thread where we do have good warranties about the locking
environment, but here it's in the bleeding ->recvmsg/->splice_read and
for those assumption that caller doesn't hold any locks is pretty
strong, especially since it's not documented anywhere.

What's going on there?