[openib-general] Dropping NETIF_F_SG since no checksum feature.

Roland Dreier rdreier at cisco.com
Tue Oct 10 20:42:20 PDT 2006


    David> Also, if you don't do checksumming on the card we MUST copy
    David> the data (be it from a user buffer, or from a filesystem
    David> page cache page) into a private buffer since if the data
    David> changes the checksum would become invalid, as I mentioned
    David> in another email earlier.

Yes, I get that now -- I replied to Michael's email before I read yours.

    David> Therefore, since we have to copy anyways, it always is
    David> better to checksum in parallel with the copy.

Yes.

    David> So the whole idea of SG without hw-checksum support is
    David> without much merit at all.

Well, on IB it is possible to implement a netdevice (IPoIB connected
mode, I assume that's what Michael is working on) with a large MTU
(64KB is a number thrown around, but really there's not any limit) but
no HW checksum capability.  Doing that in a practical way means we
need to allow non-linear skbs to be passed in.

On the other hand I'm not sure how useful such a netdevice would be --
will non-sendfile() paths generate big packets even if the MTU is 64KB?

Maybe GSO gives us all the real advantages of this anyway?

 - R.




More information about the general mailing list