On Wed, 9 Jul 2003 12:06:57 -0400
Jeff Garzik <jgarzik@xxxxxxxxx> wrote:
> On Wed, Jul 09, 2003 at 05:53:55PM +0200, Andi Kleen wrote:
> > On Wed, 9 Jul 2003 11:25:53 -0400
> > Jeff Garzik <jgarzik@xxxxxxxxx> wrote:
> >
> > > I knew this at one time, but have forgotten it :)
> > >
> > > What is the reason for adding 16 to the dev_alloc_skb length?
> > > (and skb_reserve of the same length)
> >
> > For the skb_reserve alignment to align the IP header.
> >
> > But it's not clear it is still a good idea because it leads to cache line
> > misalignment of the beginning of the packet, forcing the card to do a
> > costly Read-Modify-Write cycle.
>
> Exactly. Ben H is running into this, and pondering direct use of
> alloc_skb for precisely this reason.
Problem with changing it is that the payload ends up misaligned.
And user space usually aligns the buffer passed to recvmsg. This
means csum_copy_to_user has to csum-copy unaligned->aligned, which will
be likely very slow.
Related problem is that the TCP/IP headers are unaligned, but if your CPU
has fast enough misalignment handling it shouldn't be too bad.
-Andi
|