[ltt-dev] [PATCH] nfs: add support for splice writes

[Chuck Lever]

On Apr 21, 2009, at 1:35 PM, Trond Myklebust wrote:
> On Tue, 2009-04-21 at 10:48 -0400, Christoph Hellwig wrote:
>> On Mon, Apr 20, 2009 at 09:17:23PM +0530, Suresh Jayaraman wrote:
>>> +static ssize_t nfs_file_splice_write(struct pipe_inode_info *pipe,
>>> +				     struct file *filp, loff_t *ppos,
>>> +				     size_t count, unsigned int flags)
>>> +{
>>> +	struct dentry *dentry = filp->f_path.dentry;
>>> +
>>> +	dprintk("NFS splice_write(%s/%s, %lu@%Lu)\n",
>>> +		dentry->d_parent->d_name.name, dentry->d_name.name,
>>> +		(unsigned long) count, (unsigned long long) *ppos);
>>> +
>>> +	return generic_file_splice_write(pipe, filp, ppos, count, flags);
>>> +}
>>> +
>>
>> You need all calls from nfs_file_write, too:
>>
>> - most importantly the nfs_revalidate_file_size for O_APPEND
>
> Isn't O_APPEND illegal for splice_write()? It looks like it is from a
> quick perusal of do_splice_from().
>
>> - the nfs_do_fsync for sync writes
>
> generic_file_splice_write() calls generic_osync_inode(), which should
> ensure sync writes even with NFS.
> The one thing it won't do is propagate NFS write errors back to the
> caller. If we do care about this, then we should certainly test for
> nfs_need_sync_write() and then call nfs_do_fsync() (see
> nfs_file_write()).
>
>> - probably the stats increment
>
> We should talk to Chuck about this.

I don't know much about the characteristics of splice writes, so you  
can do one of two things:

1.  Create a separate BYTESWRITTEN counter for these
2.  Lump them into NORMALBYTESWRITTEN if these are effectively cached  
writes.

--
Chuck Lever
chuck[dot]lever[at]oracle[dot]com