Re: [PATCH v2 RFC] 8139cp: Fix GSO MSS handling

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 1624 bytes --]

On Sun, 2015-09-27 at 22:37 -0700, Tom Herbert wrote:
> 
> Which drivers are doing this? It is up to the driver to determine
> whether a particular packet being sent can have checksum offloaded to
> the device. If it cannot offload the checksum it must call
> skb_checksum_help.

Not so.

A driver sets the NETIF_F_IP_CSUM feature to indicate that it can do
the checksum on Legacy IP TCP or UDP frames and *nothing* else.

It most certainly does not expect to be handed any other kind of packet
for checksumming, and bad things will often happen if it is. If drivers
*do* spot that they've been given something they don't handle, I see
BUG() calls and warnings, but I don't see any of them calling
skb_checksum_help() to silently cope. Many of them just feed it to the
hardware and don't even notice at all because it's the *hardware* which
decides whether to do a TCP or a UDP checksum. So who knows what'll
happen.

The check is supposed to be done in can_checksum_protocol(), called
from harmonize_features(). But as noted, that check has false positives
and lets some inappropriate packets through — for NETIF_F_IP_CSUM it
lets through *all* skbuffs with ->protocol == ETH_P_IP instead of only
TCP and UDP.

I originally couldn't see how to deal with this except by looking at
the contents of the packet, which sucked. But I think I've found a
somewhat more acceptable approach now:
http://lists.openwall.net/netdev/2015/09/25/85

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


[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5691 bytes --]