Re: [PATCHv2 RFC 4/4] Revert "virtio: make add_buf return capacity remaining:

[Rusty Russell <rusty@rustcorp.com.au>]

On Tue, 7 Jun 2011 18:54:57 +0300, "Michael S. Tsirkin" <mst@redhat.com> wrote:
> On Thu, Jun 02, 2011 at 06:43:25PM +0300, Michael S. Tsirkin wrote:
> > This reverts commit 3c1b27d5043086a485f8526353ae9fe37bfa1065.
> > The only user was virtio_net, and it switched to
> > min_capacity instead.
> > 
> > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> 
> It turns out another place in virtio_net: receive
> buf processing - relies on the old behaviour:
> 
> try_fill_recv:
> 	do {
> 		if (vi->mergeable_rx_bufs)
> 			err = add_recvbuf_mergeable(vi, gfp);
> 		else if (vi->big_packets)
> 			err = add_recvbuf_big(vi, gfp);
> 		else
> 			err = add_recvbuf_small(vi, gfp);
> 
> 		oom = err == -ENOMEM;
> 		if (err < 0)
> 			break;
> 		++vi->num;
> 	} while (err > 0);
> 
> The point is to avoid allocating a buf if
> the ring is out of space and we are sure
> add_buf will fail.
> 
> It works well for mergeable buffers and for big
> packets if we are not OOM. small packets and
> oom will do extra get_page/put_page calls
> (but maybe we don't care).
> 
> So this is RX, I intend to drop it from this patchset and focus on the
> TX side for starters.

We could do some hack where we get the capacity, and estimate how many
packets we need to fill it, then try to do that many.

I say hack, because knowing whether we're doing indirect buffers is a
layering violation.  But that's life when you're trying to do
microoptimizations.

Cheers,
Rusty.