Re: [PATCH rdma-next v1 11/11] RDMA/mlx5: Add support to get MR resource in RAW format

[Maor Gottlieb <maorg@mellanox.com>]

On 6/2/2020 4:23 PM, Leon Romanovsky wrote:
> On Tue, Jun 02, 2020 at 09:27:02AM -0300, Jason Gunthorpe wrote:
>> On Tue, Jun 02, 2020 at 09:21:18AM +0300, Leon Romanovsky wrote:
>>> On Mon, Jun 01, 2020 at 09:26:46AM -0300, Jason Gunthorpe wrote:
>>>> On Sun, May 31, 2020 at 12:54:14PM +0300, Leon Romanovsky wrote:
>>>>> On Fri, May 29, 2020 at 08:31:21PM -0300, Jason Gunthorpe wrote:
>>>>>> On Wed, May 27, 2020 at 04:54:08PM +0300, Leon Romanovsky wrote:
>>>>>>> From: Maor Gottlieb <maorg@mellanox.com>
>>>>>>>
>>>>>>> Add support to get MR (mkey) resource dump in RAW format.
>>>>>>>
>>>>>>> Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
>>>>>>> Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
>>>>>>>   drivers/infiniband/hw/mlx5/restrack.c | 3 ++-
>>>>>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/drivers/infiniband/hw/mlx5/restrack.c b/drivers/infiniband/hw/mlx5/restrack.c
>>>>>>> index 9e1389b8dd9f..834886536127 100644
>>>>>>> +++ b/drivers/infiniband/hw/mlx5/restrack.c
>>>>>>> @@ -116,7 +116,8 @@ int mlx5_ib_fill_res_mr_entry(struct sk_buff *msg,
>>>>>>>   	struct nlattr *table_attr;
>>>>>>>
>>>>>>>   	if (raw)
>>>>>>> -		return -EOPNOTSUPP;
>>>>>>> +		return fill_res_raw(msg, mr->dev, MLX5_SGMT_TYPE_PRM_QUERY_MKEY,
>>>>>>> +				    mlx5_mkey_to_idx(mr->mmkey.key));
>>>>>> None of the raw functions actually share any code with the non raw
>>>>>> part, why are the in the same function? In fact all the implemenations
>>>>>> just call some other function for raw.
>>>>>>
>>>>>> To me this looks like they should should all be a new op
>>>>>> 'fill_raw_res_mr_entry' and drop the 'bool'
>>>>> I don't think that this is right approach, we already created ops per-objects
>>>>> o remove API multiplexing. Extra de-duplication will create too much ops
>>>>> without any real benefit.
>>>> If there is no code sharing then they should not be in the same
>>>> function at all. More ops is not really a problem.
>>> Logically they are the same, user asks to get object property, driver returns.
>> I'm starting to think it is also a mistake to have the same netlink op
>> and trigger it by an inbound attribute. Are there other examples of
>> that in netlink? Feels wrong
> I have no idea, don't see it in devlink.c

Jason, do you mean trigger the raw by inbound attribute? I don't see a 
reason why not do that. Netlink attributes used for input and output. 
What regarding the driver ops, let's converge with the decision, I am 
okay with the both approaches.

>
>> Jason