On 10/15/2018 11:29 PM, Jason Gunthorpe wrote:
> On Wed, Oct 10, 2018 at 09:55:10AM +0300, Leon Romanovsky wrote:
>> From: Mark Bloch <markb@xxxxxxxxxxxx>
>>
>> A user can provide a hint which will be attached to the packet and
>> written to the CQE on receive. This can be used as a way to offload
>> operations into the HW, for example parsing a packet which is a tunneled
>> packet, and if so, pass 0x1 as the hint. The software can use that hint
>> to decapsulate the packet and parse only the inner headers thus saving
>> CPU cycles.
>>
>> Signed-off-by: Mark Bloch <markb@xxxxxxxxxxxx>
>> Reviewed-by: Yishai Hadas <yishaih@xxxxxxxxxxxx>
>> Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
>> ---
>> drivers/infiniband/hw/mlx5/flow.c | 15 ++++++++++++++-
>> include/uapi/rdma/mlx5_user_ioctl_cmds.h | 1 +
>> 2 files changed, 15 insertions(+), 1 deletion(-)
>
> Applied to for-next, thanks
>
>> diff --git a/drivers/infiniband/hw/mlx5/flow.c b/drivers/infiniband/hw/mlx5/flow.c
>> index 4ee4af450720..e57435cb6d96 100644
>> --- a/drivers/infiniband/hw/mlx5/flow.c
>> +++ b/drivers/infiniband/hw/mlx5/flow.c
>> @@ -153,6 +153,16 @@ static int UVERBS_HANDLER(MLX5_IB_METHOD_CREATE_FLOW)(
>> arr_flow_actions[i]->object);
>> }
>>
>> + ret = uverbs_copy_from(&flow_act.flow_tag, attrs,
>> + MLX5_IB_ATTR_CREATE_FLOW_TAG);
>> + if (!ret) {
>> + if (flow_act.flow_tag >= BIT(24)) {
>> + ret = -EINVAL;
>> + goto err_out;
>> + }
>> + flow_act.has_flow_tag = true;
>> + }
>
> I'm not super keen on this pattern though, uverbs_copy_from can fail
> with more than ENOENT, but since the attribute is inline it can't in
> this case. Little too tricky perhaps.
I know, the entire handler isn't very pretty (I'm mostly to blame :))
There was a lot of code/logic pushed recently, I hope/want to clean it up
a bit once things settle down.
>
> Jason
>
Mark
