On Mon, 2014-12-15 at 19:14 +0100, Ralf Baechle wrote:

On Mon, Dec 01, 2014 at 04:09:36AM +0000, Ben Hutchings wrote:

quoted

quoted

 	/* Same as tx - compute csum of pseudo header  */
 	csum = hwsum +
-	       (ih->tot_len - (ih->ihl << 2)) +
-	       htons((uint16_t)ih->protocol) +
+	       (ih->tot_len - (ih->ihl << 2)) + ih->protocol +
 	       (ih->saddr >> 16) + (ih->saddr & 0xffff) +
 	       (ih->daddr >> 16) + (ih->daddr & 0xffff);

The pseudo-header is specified as:

                     +--------+--------+--------+--------+
                     |           Source Address          |
                     +--------+--------+--------+--------+
                     |         Destination Address       |
                     +--------+--------+--------+--------+
                     |  zero  |  PTCL  |    TCP Length   |
                     +--------+--------+--------+--------+

The current code zero-extends the protocol number to produce the 5th
16-bit word of the pseudo-header, then uses htons() to put it in
big-endian order, consistent with the other fields.  (Yes, it's doing
addition on big-endian words; this works even on little-endian machines
due to the way the checksum is specified.)

The driver should not be doing this at all, though.  It should set
skb->csum = hwsum; skb->ip_summed = CHECKSUM_COMPLETE; and let the
network stack adjust the hardware checksum.

Really?  The IOC3 isn't the exactly the smartest NIC around; it does add up
everything and the kitchen sink, that is ethernet headers, IP headers and
on RX the frame's trailing CRC.

That is almost exactly what CHECKSUM_COMPLETE means on receive; only the
CRC would need to be subtracted.  Then the driver can validate
{TCP,UDP}/IPv{4,6} checksums without any header parsing.

All that needs to be subtracted in software
which is what this does.  I think others NICs are all smarted and don't
need this particular piece of magic.

It may not be smart, but that allows it to cover more cases than most
smart network controllers!

On transmit, the driver should:
- Calculate the partial checksum of data up to offset csum_start
- Subtract this from the checksum at offset (csum_start + csum_offset)
It should set the NETIF_F_GEN_CSUM feature flag rather than
NETIF_F_IP_CSUM.  Then it will be able to generate {TCP,UDP}/IPv{4,6}
checksums.

Ben.

I agree with your other comment wrt. to htons().

-- 
Ben Hutchings
The two most common things in the universe are hydrogen and stupidity.