[PATCH v2] vfs: prevent copy_file_range to copy across devices

[Amir Goldstein]

On Mon, Feb 15, 2021 at 5:42 PM Luis Henriques <lhenriques at suse.de> wrote:
>
> Nicolas Boichat reported an issue when trying to use the copy_file_range
> syscall on a tracefs file.  It failed silently because the file content is
> generated on-the-fly (reporting a size of zero) and copy_file_range needs
> to know in advance how much data is present.
>
> This commit restores the cross-fs restrictions that existed prior to
> 5dae222a5ff0 ("vfs: allow copy_file_range to copy across devices") and
> removes generic_copy_file_range() calls from ceph, cifs, fuse, and nfs.
>
> Fixes: 5dae222a5ff0 ("vfs: allow copy_file_range to copy across devices")
> Link: https://lore.kernel.org/linux-fsdevel/20210212044405.4120619-1-drinkcat@chromium.org/
> Cc: Nicolas Boichat <drinkcat at chromium.org>
> Signed-off-by: Luis Henriques <lhenriques at suse.de>

Code looks ok.
You may add:

Reviewed-by: Amir Goldstein <amir73il at gmail.com>

I agree with Trond that the first paragraph of the commit message could
be improved.
The purpose of this change is to fix the change of behavior that
caused the regression.

Before v5.3, behavior was -EXDEV and userspace could fallback to read.
After v5.3, behavior is zero size copy.

It does not matter so much what makes sense for CFR to do in this
case (generic cross-fs copy).  What matters is that nobody asked for
this change and that it caused problems.

Thanks,
Amir.