Re: [PATCHv2 net] i40e: Implement ndo_gso_check()

From: Jesse Gross
Date: Tue Dec 02 2014 - 13:27:21 EST


On Mon, Dec 1, 2014 at 4:09 PM, Tom Herbert <therbert@xxxxxxxxxx> wrote:
> On Mon, Dec 1, 2014 at 3:53 PM, Jesse Gross <jesse@xxxxxxxxxx> wrote:
>> On Mon, Dec 1, 2014 at 3:47 PM, Tom Herbert <therbert@xxxxxxxxxx> wrote:
>>> On Mon, Dec 1, 2014 at 3:35 PM, Joe Stringer <joestringer@xxxxxxxxxx> wrote:
>>>> On 21 November 2014 at 09:59, Joe Stringer <joestringer@xxxxxxxxxx> wrote:
>>>>> On 20 November 2014 16:19, Jesse Gross <jesse@xxxxxxxxxx> wrote:
>>>>>> I don't know if we need to have the check at all for IPIP though -
>>>>>> after all the driver doesn't expose support for it all (actually it
>>>>>> doesn't expose GRE either). This raises kind of an interesting
>>>>>> question about the checks though - it's pretty easy to add support to
>>>>>> the driver for a new GSO type (and I imagine that people will be
>>>>>> adding GRE soon) and forget to update the check.
>>>>>
>>>>> If the check is more conservative, then testing would show that it's
>>>>> not working and lead people to figure out why (and update the check).
>>>>
>>>> More concretely, one suggestion would be something like following at
>>>> the start of each gso_check():
>>>>
>>>> + const int supported = SKB_GSO_TCPV4 | SKB_GSO_TCPV6 | SKB_GSO_FCOE |
>>>> + SKB_GSO_UDP | SKB_GSO_UDP_TUNNEL;
>>>> +
>>>> + if (skb_shinfo(skb)->gso_type & ~supported)
>>>> + return false;
>>>
>>> This should already be handled by net_gso_ok.
>>
>> My original point wasn't so much that this isn't handled at the moment
>> but that it's easy to add a supported GSO type but then forget to
>> update this check - i.e. if a driver already supports UDP_TUNNEL and
>> adds support for GRE with the same constraints. It seems not entirely
>> ideal that this function is acting as a blacklist rather than a
>> whitelist.
>
> Agreed, it would be nice to have all the checking logic in one place.
> If all the drivers end up implementing ndo_gso_check then we could
> potentially get rid of the GSO types as features. This probably
> wouldn't be a bad thing since we already know that the features
> mechanism doesn't scale (for instance there's no way to indicate that
> certain combinations of GSO types are supported by a device).

This crossed my mind and I agree that it's pretty clear that the
features mechanism isn't scaling very well. Presumably, the logical
extension of this is that each driver would have a function that looks
at a packet and returns a set of offload operations that it can
support rather than exposing a set of protocols. However, it seems
like it would probably result in a bunch of duplicate code in each
driver.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/