Re: [PATCH net-next V2 1/3] net: Add GRO support for UDP encapsulating protocols

[Or Gerlitz <ogerlitz@mellanox.com>]

On 09/01/2014 05:12, Jerry Chu wrote:
> On Wed, Jan 8, 2014 at 12:02 AM, Or Gerlitz <ogerlitz@mellanox.com> wrote:
>> On 08/01/2014 00:11, Jerry Chu wrote:
>>> On Tue, Jan 7, 2014 at 12:37 PM, Or Gerlitz <or.gerlitz@gmail.com> wrote:
>>>> On Tue, Jan 7, 2014 at 10:32 PM, Eric Dumazet <eric.dumazet@gmail.com>
>>>> wrote:
>>>>> On Tue, 2014-01-07 at 22:19 +0200, Or Gerlitz wrote:
>>>>>> On Tue, Jan 7, 2014 at 6:33 PM, Eric Dumazet <eric.dumazet@gmail.com>
>>>>>> wrote:
>>>>>>> On Tue, 2014-01-07 at 17:29 +0200, Or Gerlitz wrote:
>>>>>>>> +
>>>>>>>> +#define MAX_UDP_PORT (1 << 16)
>>>>>>>> +extern const struct net_offload __rcu *udp_offloads[MAX_UDP_PORT];
>>>>>>> Thats 512 KB of memory.
>>>>>>> This will greatly impact forwarding performance of UDP packets with
>>>>>>> random ports, and will increase kernel memory size for embedded
>>>>>>> devices.
>>>>>> Re forwarding, are you referring to the case where the forwarded
>>>>>> packets are encapsulated? packets which are not encapusalted will be
>>>>>> flushed in the gro receive handler (this went out by mistake in V2 but
>>>>>> exists in V1)  if skb->encapsulation isn't set.
>>>>>>
>>>>> How do you know encapsulation must be tried for a given incoming
>>>>> packet ? NIC do not magically sets skb->encapsulation I think...
>>>> So here's the thing, per my understanding we want to GRO only received
>>>> **encapsulated** packets whose checksum status is != CHECKSUM_NONE
>>> What's wrong with GRO'ing pkts whose csum == CHECKSUM_NONE?
>>
>> I am not sure, intuitively it sounds a bit wrong to me, empirically, it
>> doesn't work for udp encapsulated  / vxlan
>> traffic, I got drops from the tcp stack in tcp_rcv_established() -- if
>> GRO-ed packets carry CHECKSUM_NONE
>> we arrive to the csum_error label, which means that
>> tcp_checksum_complete_user() failed for them
> This is odd because if pkts have been aggregated successfully,
> tcp4_gro_receive() should've skb_checksum() and turned CHECKSUM_NONE
> into CHECKSUM_UNNECESSARY. (I think i've already tested this
> case with my GRE-GRO patch on a NIC that sends up pkts w/ CHECKSUM_NONE.
>
> But granted there are a lot of csum related bugs in the current code. I just spent half a day scratching my head on a very low thruput number with my GRE patch over a GRE tunnel w/ csum flag on. I just tracked it down to be buggy TSO/GRE code that will produce bad csum on the tx side.

The "there are a lot of csum related bugs in the current code" comment 
sounds really bad, how do we get into a better place? can you point on 
the buggy TSO code that produced bad csum on the tx side?