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

[Al Viro]

On Fri, Jan 13, 2017 at 12:08:44PM -0800, Linus Torvalds wrote:
> On Fri, Jan 13, 2017 at 11:33 AM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > This function looks so broken that I must be missing something. Why
> > doesn't pipe_advance() just look like the following:
> >
> >   static void pipe_advance(struct iov_iter *i, size_t size)
> >   {
> ...
> >                 pipe_buf_release(pipe, buf);
> >                 pipe->nrbufs--;
> ...
> 
> I think this part needs to update "curbufs" too, so something like
> 
>                 pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
> 
> although I think that "idx" has to track curbuf here anyway, so I
> guess it could just be combined with the idx update and look something
> like
> 
>                 pipe->curbuf = idx = next_idx(idx, pipe);
> 
> in there. Otherwise we get out of sync with the pipe state.

You are looking at the wrong end of that cyclic buffer.  ->curbuf is where
the data begins (and it might be well prior to anything we'd pushed there -
pipe might've been non-empty).  Then we have ->nrbufs allocated buffers and
i->idx points to the place where copy_to_iter() will put the data.

We want pipe_advance() to
	* move the point where copy_to_iter() would go by this much
	* free all preallocated buffers past that point.

->curbuf is for pipe readers; we are dealing with writing to pipe here.