On 09/21/2015 05:41 PM, Or Gerlitz wrote:
> On Fri, Sep 18, 2015 at 5:46 AM, Doug Ledford <dledford@xxxxxxxxxx> wrote:
>> On 09/16/2015 11:56 AM, Bodong Wang wrote:
>>> Signed-off-by: Bodong Wang <bodong@xxxxxxxxxxxx>
>>> ---
>>> drivers/infiniband/hw/mlx4/main.c | 3 +++
>>> 1 file changed, 3 insertions(+)
>>>
>>> diff --git a/drivers/infiniband/hw/mlx4/main.c b/drivers/infiniband/hw/mlx4/main.c
>>> index 8be6db8..a70ca6a 100644
>>> --- a/drivers/infiniband/hw/mlx4/main.c
>>> +++ b/drivers/infiniband/hw/mlx4/main.c
>>> @@ -217,6 +217,9 @@ static int mlx4_ib_query_device(struct ib_device *ibdev,
>>> props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
>>> }
>>>
>>> + props->csum_cap.eth_csum_cap |= IB_CSUM_SUPPORT_RAW;
>>> + props->csum_cap.ib_csum_cap |= IB_CSUM_SUPPORT_UD;
>>> +
>
>> This patch highlights something I didn't think about on the previous
>> patch. Why separate eth/ib if you have per QP flags? The QP denotes
>> the ib/eth relationship without the need to separate it into two
>> different caps. In other words, you can never have an IB qp type on eth
>> because the only eth QP types we support other than RAW are all RDMA and
>> not IP. Really, there's enough spare bits in ib_device_cap_flags that
>> you could do away with the new caps entirely. Right now, we support UD
>> (which we already have a flag for), we can add two flags (for RAW and
>> RC) and that should cover all of the foreseeable options as that would
>> allow us to extend IP CSUM support to cover connected mode and cover all
>> of the current options. I don't see us doing IP traffic in any other
>> situation, so I thing that should suffice. Bits 25 and 26 could be used
>> for the two new bits. Then you just need to extend the bits to user space.
>
> Doug,
>
> The vendor may support the offload for a certain QP type only over
> certain link.
Exactly my point.
> E.g mlx4 support checksum for UD QPs only over IB but
> not over Eth,
As it should be. We are, after all, talking about IP embedded in UD
RDMA. Over IB that makes sense. Over Eth it makes no sense. If you
are going to do IP on Eth, then just do IP, don't do IP in UD. I see no
reason to support this construct.
> checksum for RC QPs isn't supported,
But could be for IB. The fact that it isn't is why there is an ongoing
effort to work around this issue.
> and RAW_PACKET QPs
> are available anyway only for Eth links.
Correct. And this will never be available on IB.
> But if this is what you think
> needs to be done, I guess we can do that.
Here's the only matrix of IP checksumming that makes sense:
1) UD over IB (because it is one of the supported IPoIB types)
2) RC over IB (same as above)
3) Raw ETH over Eth (because IP over Eth makes sense and is a common
type of packet to send on Raw Eth, but Raw Eth will never be sent on IB
as it isn't supported there at all)
Anything else would require adding more Raw ETH QPs elsewhere, or
expanding the IPoIB spec to include more connection types.
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: 0E572FDD
Attachment:
signature.asc
Description: OpenPGP digital signature
