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

[Al Viro]

On Fri, Jan 13, 2017 at 11:33:18AM -0800, Linus Torvalds wrote:

> What am I missing?

The fact that we want to free the _tail_, not the beginning of the
damn thing.

> Why is "pipe_advance()" written in that incomprehensible form? Why
> don't we do the pipe_buf_release() as we advance through it, instead
> of doing it at the end?
> 
> Also, the line
> 
>                 buf->len = size;
> 
> in that pipe_advance() function looks buggy. "size" is how much we
> *remove* from buf->len, shouldn't we update buf->len by subtracting
> size?

Because it's "truncate to size", not "throw everything up to that point
out".

We have some amount of data pushed into pipe (in this case - 0) and we
have some buffers allocated by ..._get_pages() past the end of it.
Some of that we want to keep (again, in this case - none) and have the next
copy_to_iter() go after those, the rest we discard.