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

From: Al Viro
Date: Thu Jan 12 2017 - 18:27:29 EST


On Thu, Jan 12, 2017 at 03:14:41PM -0800, Linus Torvalds wrote:
> On Thu, Jan 12, 2017 at 3:02 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > Looking at the callers of "do_splice_to()", we already have the
> > wait_for_space() in do_splice(), but we do *not* have it in the
> > do_splice_from() case when both the input and output file descriptors
> > are pipes.
>
> Bah. That case doesn't even trigger the new code. I was lazy with my
> grep. The two cases are "do_splice()" (which does have the
> wait-for-space) and splice_direct_to_actor(). And
> splice_direct_to_actor() shouldn't even need it, should it?
>
> So ignore that. But I think there is something about the EAGAIN.

It might, but I would really like to see where has that EAGAIN come
from. I see several possibilities:
* wait_for_space() with SPLICE_F_NONBLOCK in flags. Shouldn't
happen with 0 in the last argument of splice(2).
* default_file_splice_read() seeing pipe->nrbufs == pipe->buffers.
Shouldn't be possible after successful wait_for_space().
* vfs_readv() returning -EAGAIN. That might be possible,
actually - the damn thing has come from
13761 open("/dev/ptmx", O_RDWR|O_NOCTTY|O_NONBLOCK|O_CLOEXEC) = 5
and O_NONBLOCK had been present in open flags.

What I'd like to see is strace of the same thing on the working kernel,
ideally - just prior to the commit bisect has converged to.