On Mon, Oct 5, 2015 at 4:16 AM, David Woodhouse [off-list ref] wrote:

On Tue, 2015-09-29 at 15:52 -0700, Tom Herbert wrote:

quoted

Please look at ixgbe_tx_csum in ixgbe driver. This one example of how
a driver can determine whether the checksum being offloaded is TCP or
UDP. The bug in this driver is...

I think it serves better as an example of why we don't *want* drivers
doing that kind of thing for themselves... :)

I propose we steal some high bits from csum_offset, as you suggested,
and use them to indicate a 'checksum type', which will include TCP and
UDP.

Then the filter in netif_skb_features() can trivially do the right
thing for NETIF_F_IP{V6,}_CSUM devices, so avoid feeding them packets
they can't handle.

You mentioned that you actually want to deprecate those feature flags —
which works for me, but it's kind of orthogonal. If we do that we'd
still want to provide generic functions that such drivers can use as
their .ndo_features_check() method. And we'd *still* want to do the
check based on a simple flag, rather than grubbing around in the packet
data. (And the drivers if they *are* asked to do the checksum will
sometimes care whether it's TCP vs. UDP too).

I don't think we want drivers calling skb_checksum_help() for
themselves; we want the pre-filter. Mainly because we *definitely*
don't want drivers calling gso_skb_segment() for themselves in the same
situation — see the comment I posted on Friday about the r8169 instance
of that. ('Re: [PATCH net-next 3/3] r8169: support IPv6').

David, here is what I am currently thinking the interface should be:

1) Drivers may advertise NETIF_F_HW_CSUM. The stack will indicate
checksum offload exclusively using the
CHECKSUM_PARTIAL/csum_start/csum_offset interface. No additional
interfaces (bits in skbuff should not be needed)
2) A driver may inspect packets via ndo_check to decide if it wants to
offload the checksum, if not cancels NETIF_F_HW_CSUM in the packet.
3) In driver xmit when CHECKSUM_PARTIAL is set the driver MUST
correctly resolve the checksum-- either by properly offloading to the
device or calling skb_checksum_help.
4) To help drivers for devices with limited offload capabilities we'll
define a helper function to check for typical restrictions (.e.g. IPv4
only, TCP/UDP only. no encapsulation, no IPv6 extension headers,
etc.). I am working on this helper function and will send RFC shortly.

Tom

--
David Woodhouse                            Open Source Technology Centre
David.Woodhouse@intel.com                              Intel Corporation