Re: splice() from /dev/zero to a pipe does not work (5.9+)

[Kees Cook]

On Fri, May 07, 2021 at 07:05:51PM +0100, Colin Ian King wrote:
> Hi,
> 
> While doing some micro benchmarking with stress-ng I discovered that
> since linux 5.9 the splicing from /dev/zero to a pipe now fails with
> -EINVAL.
> 
> I bisected this down to the following commit:
> 
> 36e2c7421f02a22f71c9283e55fdb672a9eb58e7 is the first bad commit
> commit 36e2c7421f02a22f71c9283e55fdb672a9eb58e7
> Author: Christoph Hellwig <hch@xxxxxx>
> Date:   Thu Sep 3 16:22:34 2020 +0200
> 
>     fs: don't allow splice read/write without explicit ops
> 
> I'm not sure if this has been reported before, or if it's intentional
> behavior or not. As it stands, it's a regression in the stress-ng splice
> test case.

The general loss of generic splice read/write is known. Here's some
early conversations I was involved in:

https://lore.kernel.org/lkml/20200818200725.GA1081@xxxxxx/
https://lore.kernel.org/lkml/202009181443.C2179FB@keescook/
https://lore.kernel.org/lkml/20201005204517.2652730-1-keescook@xxxxxxxxxxxx/

And it's been getting re-implemented in individual places:

$ git log --oneline --no-merges --grep 36e2c742
42984af09afc jffs2: Hook up splice_write callback
a35d8f016e0b nilfs2: make splice write available again
f8ad8187c3b5 fs/pipe: allow sendfile() to pipe again
f2d6c2708bd8 kernfs: wire up ->splice_read and ->splice_write
9bb48c82aced tty: implement write_iter
dd78b0c483e3 tty: implement read_iter
14e3e989f6a5 proc mountinfo: make splice available again
c1048828c3db orangefs: add splice file operations
960f4f8a4e60 fs: 9p: add generic splice_write file operation
cf03f316ad20 fs: 9p: add generic splice_read file operations
06a17bbe1d47 afs: Fix copy_file_range()

So the question is likely, "do we want this for /dev/zero?"

-- 
Kees Cook