[PATCH] cifs: add .splice_write

[Steve French]

With
        gio copy -p source target
on smb3 mount with Andres's patch, I see a small improvement over localhost
interface, but likely would be larger over slow network

On Fri, Jan 26, 2018 at 3:25 PM, Steve French <smfrench at gmail.com> wrote:

> The patch did look like it improves the performance of file copy from
> Nautilus when I tried it, but would prefer experimenting tool where I
> could measure the improvement more precisely.  Is there a good command
> line tool which generates splice_write indirectly? As expected cp and
> dd don't seem to be affected by this patch.
>
> In looking back at the commit history, I noticed Jens's much older
> commit from 2007 (5ffc4ef45b3b0a57872f631b4e4ceb8ace0d7496)  which
> added the
> callouts to splice_read and more recently Al's commit
> 8d0207652cbe27d1f962050737848e5ad4671958 which
> introduced iter_file_splice_write a couple years ago (for cifs.ko he
> updated via commit
> 3dae8750c368f8ac11c3c8c2a28f56dcee865c01 to add write_iter support).
> Interesting that splice write wasn't
> added to cifs.ko but it does seem to help in the case you describe
> (Nautilus or other apps that use splice).
>
> To localhost (Samba server, cifs.ko on client with default mount
> options, current kernel rc9) on my laptop - my quick tests to showed >
> 500MB/sec performance with dd (1M block size), so will be curious what
> splice syscall results in.
>
> ---------- Forwarded message ----------
> From: Andrés Souto <kai670 at gmail.com>
> Date: Fri, Jan 26, 2018 at 12:13 PM
> Subject: Re: [PATCH] cifs: add .splice_write
> To: Steve French <sfrench at samba.org>
> Cc: linux-cifs at vger.kernel.org, linux-kernel at vger.kernel.org
>
>
> I want to explain a bit what is the problem this patch solves.
>
> Without it, I always obtain a transfer rate below 9MB/s when writing
> to a cifs share using splice syscall independently of the buffer size
> used. However, when reading from the cifs share, I obtain around 105
> MB/s.
>
> After applying the patch, the transfer rate I obtain writing to the
> cifs share depending on the buffer size is:
>
> blocksize (kB)| transfer rate (MB/s)
> --------------+----------------------
>            64 | 51.52
>           128 | 65.41
>           256 | 76.64
>           512 | 85.72
>          1024 | 90.21
>



-- 
Thanks,

Steve