[PATCH v20 05/20] nvme-tcp: Add DDP offload control path

[Sagi Grimberg]

> Hi Sagi,
> 
> On 29/11/2023 15:52, Aurelien Aptel wrote:
>> Hi Sagi,
>>
>> Sagi Grimberg <sagi at grimberg.me> writes:
>>>> +     ok = ulp_ddp_query_limits(netdev, &ctrl->ddp_limits,
>>>> +                               ULP_DDP_NVME, ULP_DDP_CAP_NVME_TCP,
>>>> +                               ctrl->ctrl.opts->tls);
>>>> +     if (!ok) {
>>>
>>> please use a normal name (ret).
>>
>> Ok, we will rename to ret and make ulp_ddp_query_limits() return int 0
>> on success to be consistent with the name.
>>
>>> Plus, its strange that a query function receives a feature and returns
>>> true/false based on this. The query should return the limits, and the
>>> caller should look at the limits and see if it is appropriately
>>> supported.
>>
>> We are not sure how to proceed as this seems to conflict with what you
>> suggested in v12 [1] about hiding the details of checking supports in
>> the API. Limits just dictate some constants the nvme-layer should use
>> once we know it is supported.
>>
>> We can rename ulp_ddp_query_limits() to ulp_ddp_check_support(). This
>> function checks the support of the specified offload capability and also
>> returns the limitations of it.
>>
>> Alternatively, we can split it in 2 API functions (check_support
>> and query_limits).
>>
>> Let us know what you prefer.
>>
>> Thanks
>>
>> 1: 
>> https://lkml.kernel.org/netdev/bc5cd2a7-efc4-e4df-cae5-5c527dd704a6@grimberg.me/
> 
> We would like to submit another version by the end of this week and 
> hopefully progress with this series.
> To address your comment, I was thinking about something like:
> 
> +    if (!ulp_ddp_is_cap_active(netdev, ULP_DDP_CAP_NVME_TCP))
> +        goto err;
> +
> +    ret = ulp_ddp_get_limits(netdev, &ctrl->ddp_limits, ULP_DDP_NVME);
> +    if (ret)
> +        goto err;
> +
> +    if (ctrl->ctrl.opts->tls && !ctrl->ddp_limits.tls)
> +        goto err;
> 
> I would like to remind you that the community didn't want to add ulp_ddp 
> caps and limits context to the netdev structure (as we have for example 
> in the block layer q) and preferred the design of ops/cbs.

Seems fine to me.