Re: [PATCH net-next v1 3/3] net/mlx5e: TC, Add support for VxLAN GBP encap/decap flows offload

[Alexander Lobakin]

From: Gavin Li <gavinl@xxxxxxxxxx>
Date: Thu, 16 Feb 2023 16:40:33 +0800

> 
> On 2/16/2023 1:01 AM, Alexander Lobakin wrote:
>> External email: Use caution opening links or attachments
>>
>>
>> From: Gavin Li <gavinl@xxxxxxxxxx>
>> Date: Wed, 15 Feb 2023 16:30:04 +0800
>>
>>> On 2/15/2023 11:36 AM, Gavin Li wrote:
>>>> External email: Use caution opening links or attachments
>>>>
>>>>
>>>> On 2/14/2023 11:26 PM, Alexander Lobakin wrote:
>>>>> External email: Use caution opening links or attachments
>>>>>
>>>>>
>>>>> From: Gavin Li <gavinl@xxxxxxxxxx>
>>>>> Date: Tue, 14 Feb 2023 15:41:37 +0200
>> [...]
>>
>>>>>> @@ -96,6 +99,70 @@ static int mlx5e_gen_ip_tunnel_header_vxlan(char
>>>>>> buf[],
>>>>>>         udp->dest = tun_key->tp_dst;
>>>>>>         vxh->vx_flags = VXLAN_HF_VNI;
>>>>>>         vxh->vx_vni = vxlan_vni_field(tun_id);
>>>>>> +     if (tun_key->tun_flags & TUNNEL_VXLAN_OPT) {
>>>>>> +             md = ip_tunnel_info_opts((struct ip_tunnel_info
>>>>>> *)e->tun_info);
>>>>>> +             vxlan_build_gbp_hdr(vxh, tun_key->tun_flags,
>>>>>> +                                 (struct vxlan_metadata *)md);
>>>>> Maybe constify both ip_tunnel_info_opts() and vxlan_build_gbp_hdr()
>>>>> arguments instead of working around by casting away?
>>>> ACK. Sorry for the confusion---I misunderstood the comment.
>>> This ip_tunnel_info_opts is tricky to use const to annotate the arg
>>> because it will have to cast from const to non-const again upon
>>> returning.
>> It's okay to cast away for the `void *` returned.
>> Alternatively, use can convert it to a macro and use
>> __builtin_choose_expr() or _Generic to return const or non-const
>> depending on whether the argument is constant. That's what was recently
>> done for container_of() IIRC.
> 
> I've fixed vxlan_build_gbp_hdr in V2. For ip_tunnel_info_opts, it's
> confusing to me.
> 
> It would be as below after constifying the parameter.
> 
> static inline void *ip_tunnel_info_opts(const struct ip_tunnel_info *info)
> {
>     return (void *)(info + 1);
> }
> Is there any value gained by this change?

You wouldn't need to W/A it each time in each driver, just do it once in
the inline itself.
I did it once in __skb_header_pointer()[0] to be able to pass data
pointer as const to optimize code a bit and point out explicitly that
the function doesn't modify the packet anyhow, don't see any reason to
not do the same here.
Or, as I said, you can use macros + __builtin_choose_expr() or _Generic.
container_of_const() uses the latter[1]. A __builtin_choose_expr()
variant could rely on the __same_type() macro to check whether the
pointer passed from the driver const or not.

> 
>>
>>>>>> +     }
>>>>>> +
>>>>>> +     return 0;
>>>>>> +}
>> [...]
>>
>> Thanks,
>> Olek

[0]
https://elixir.bootlin.com/linux/v6.2-rc8/source/include/linux/skbuff.h#L3992
[1]
https://elixir.bootlin.com/linux/v6.2-rc8/source/include/linux/container_of.h#L33

Thanks,
Olek