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

[Al Viro]

On Fri, Jan 13, 2017 at 05:46:34PM -0800, Linus Torvalds wrote:
> On Fri, Jan 13, 2017 at 5:24 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > Why would advance by 0 change ->iov_offset here?
> 
> That's not my worry. Advancing by zero obviously doesn't change the position.
> 
> But the _truncation_ of the rest requires iov_offset to be zero in
> order to actually truncate everything.
> 
> So I was worrying about something updating it, and then wanting to
> truncate things on error.
> 
> But you bring up the kinds of cases I worried about:
> 
> > On error it does use iov_iter_advance(), pretty much as a way to
> > trigger pipe_truncate().  There we directly reset ->iov_offset to 0
> > and ->idx to its original value.
> 
> Ok, this was the part I worried about. And this
> 
> > However, theoretically it is possible that ->read_iter() instance does
> > successful copy_to_iter() and then decides to return an error.  This
> >         } else if (ret < 0) {
> >                 to.idx = idx;
> >                 to.iov_offset = 0;
> >                 iov_iter_advance(&to, 0); /* to free what was emitted */
> > in generic_file_splice_read() catches any such cases.
> 
> So I'm happy with that last patch then, and my worries are laid to rest.

OK.  Let's wait for Alan to confirm that the last variant works and
I'll send a pull request (with Cc: stable # v4.9).