Re: [PATCH 11/11] net: xilinx: Show csum in bootlog

[Michal Simek <monstr@monstr.eu>]

On 10/05/2012 03:51 PM, Ben Hutchings wrote:
> On Fri, 2012-10-05 at 11:35 +0200, Michal Simek wrote:
>> On 10/04/2012 09:15 PM, Ben Hutchings wrote:
>>> On Thu, 2012-10-04 at 20:14 +0200, Michal Simek wrote:
>>>> Just show current setting in bootlog.
>>> [...]
>>>> --- a/drivers/net/ethernet/xilinx/ll_temac_main.c
>>>> +++ b/drivers/net/ethernet/xilinx/ll_temac_main.c
>>>> @@ -1052,12 +1052,14 @@ static int __devinit temac_of_probe(struct platform_device *op)
>>>>    	/* Setup checksum offload, but default to off if not specified */
>>>>    	lp->temac_features = 0;
>>>>    	p = (__be32 *)of_get_property(op->dev.of_node, "xlnx,txcsum", NULL);
>>>> +	dev_info(&op->dev, "TX_CSUM %d\n", be32_to_cpup(p));
>>>>    	if (p && be32_to_cpu(*p)) {
>>>>    		lp->temac_features |= TEMAC_FEATURE_TX_CSUM;
>>>>    		/* Can checksum TCP/UDP over IPv4. */
>>>>    		ndev->features |= NETIF_F_IP_CSUM;
>>>>    	}
>>>>    	p = (__be32 *)of_get_property(op->dev.of_node, "xlnx,rxcsum", NULL);
>>>> +	dev_info(&op->dev, "RX_CSUM %d\n", be32_to_cpup(p));
>>> [...]
>>>
>>> Is there any particular reason you think this needs to be logged by
>>> default, rather than letting users run ethtool -k?  I suggest using
>>> dev_dbg() instead.
>>
>> Ok. I have looked at it and there are missing some bits in ndev->features.
>>
>> Can you please check that my setting is correct?
>>
>> It is SG DMA ip/driver.
>> ndev->features = NETIF_F_FRAGLIST | NETIF_F_SG
>
> NETIF_F_SG only; NETIF_F_FRAGLIST means you can handle skbs chained
> through the frag_list pointer.

The driver is able to handle skb fragments too. temac_start_xmit


>> With two options for csum on RX/TX. They can be selected independently.
>> tx Partial csum over IPv4. -> NETIF_F_IP_CSUM
>> tx Full csum. -> NETIF_F_HW_CSUM
>
> NETIF_F_IP_CSUM means you have hardware checksum generation for TCP/IPv4
> and UDP/IPv4 only (XAE_FEATURE_FULL_TX_CSUM).
>
> NETIF_F_HW_CSUM means you have generic TCP-style checksum generation
> using the csum_start and csum_offset fields of the skb
> (XAE_FEATURE_PARTIAL_TX_CSUM).  By the way, you're actually testing
> XAE_FEATURE_PARTIAL_RX_CSUM in axienet_start_xmit()...

We have used non mainline ll_temac driver for a long time but we will
move to this mainline version soon. It is on my todo list to clean this
driver and test all these options.
Also performance tests will be necessary to do.


>> rx Full csum -> NETIF_F_RXCSUM
>>
>> Is there any option to support partial csum?
>
> There is no need to differentiate these in the device features.  For TX
> the stack needs to know whether to use a software fallback before
> passing the skb to you, but on RX it looks at the ip_summed field of
> each skb you pass up.

Hardware can be setup asymmetrically. It means enable CSUM only on RX or TX.
All combination are valid.
The point here is if Linux is not able to handle this then we have to
create logic in the driver to support these options too.

Thanks,
Michal




-- 
Michal Simek, Ing. (M.Eng)
w: www.monstr.eu p: +42-0-721842854
Maintainer of Linux kernel 2.6 Microblaze Linux - http://www.monstr.eu/fdt/
Microblaze U-BOOT custodian