getting rid of ->splice_write?

getting rid of ->splice_write?

[Christoph Hellwig]

Currently only /dev/null, fusedev and the socket code have a
splice_write implementation that isn't iter_file_splice_write, and
it seems like these three could easily be switched over if they
implemented a ->write_iter.

Similarly it seems to be like we could kill ->splice_read by
implementing an equivalent iteration over ->read_iter.

Re: getting rid of ->splice_write?

[Al Viro]

On Mon, Sep 22, 2014 at 10:30:53AM -0700, Christoph Hellwig wrote:
> Currently only /dev/null, fusedev and the socket code have a
> splice_write implementation that isn't iter_file_splice_write, and
> it seems like these three could easily be switched over if they
> implemented a ->write_iter.

Not really.  A minor nitpick is that you've missed port_fops_splice_write(),
but the real bitch isn't that and not even the sockets (see the fun with
iov_iter sendmsg/recvmsg work getting resurrected).  It's that NULL
->splice_write() means default_file_splice_write.  IOW, you'd need either
->write_iter() for _everything_ (with support of bvec-backed ones included,
since that's what iter_file_splice_write() will feed to ->write_iter()),
or you need to have do_splice_from() check if ->write_iter is NULL and
go for default_file_splice_write() instead of iter_file_splice_write().

The latter might be doable, but the former is really over the top - for
that we'd need to touch every driver instance of ->write() out there.
You want to do that, it's your funeral...

> Similarly it seems to be like we could kill ->splice_read by
> implementing an equivalent iteration over ->read_iter.

Hard to do.  I agree that we want to, but it'll take quite a bit of work
on iov_iter primitives, I'm afraid.  At the very least, we want a variant
of iov_iter that could steal pages.  Don't forget that a large part of
the rationale behind splice_read was the ability to play zero-copy games.

I'm not sure if it will happen this cycle; there's more than enough fun
on the net/* side.  _If_ that ends up done faster than I expect it to be,
->splice_read() is the obvious next target.

Re: getting rid of ->splice_write?

[Christoph Hellwig]

On Wed, Nov 05, 2014 at 06:49:45PM +0000, Al Viro wrote:
> Not really.  A minor nitpick is that you've missed port_fops_splice_write(),
> but the real bitch isn't that and not even the sockets (see the fun with
> iov_iter sendmsg/recvmsg work getting resurrected).  It's that NULL
> ->splice_write() means default_file_splice_write.  IOW, you'd need either
> ->write_iter() for _everything_ (with support of bvec-backed ones included,
> since that's what iter_file_splice_write() will feed to ->write_iter()),
> or you need to have do_splice_from() check if ->write_iter is NULL and
> go for default_file_splice_write() instead of iter_file_splice_write().
> 
> The latter might be doable, but the former is really over the top - for
> that we'd need to touch every driver instance of ->write() out there.
> You want to do that, it's your funeral...

The latter is what I thought off.  And yes, the socket work looks good,
especially if we can get rid of ->sendpage as well.  That'll require
passing new flags somewhere, the ones in the iocb added for
preadv2/pwritev2 might be usable.

> > Similarly it seems to be like we could kill ->splice_read by
> > implementing an equivalent iteration over ->read_iter.
> 
> Hard to do.  I agree that we want to, but it'll take quite a bit of work
> on iov_iter primitives, I'm afraid.  At the very least, we want a variant
> of iov_iter that could steal pages.  Don't forget that a large part of
> the rationale behind splice_read was the ability to play zero-copy games.
> 
> I'm not sure if it will happen this cycle; there's more than enough fun
> on the net/* side.  _If_ that ends up done faster than I expect it to be,
> ->splice_read() is the obvious next target.

And zero copy games would become a lot less nasty if they could go
straight through ->read_iter instead of the current abuses of splice
infrastructure.

Same for sendfile, btw.