Re: MAC address set requires decision

["lihuisong (C)" <lihuisong@huawei.com>]

在 2023/2/23 22:32, Bruce Richardson 写道:
> On Thu, Feb 23, 2023 at 02:18:59PM +0000, Ferruh Yigit wrote:
>> On 2/23/2023 4:32 AM, Hemant Agrawal wrote:
>>> On 22-Feb-23 11:25 PM, Honnappa Nagarahalli wrote:
>>>> -----Original Message-----
>>>>> From: Richardson, Bruce <bruce.richardson@intel.com>
>>>>> Sent: Thursday, February 16, 2023 9:05 AM
>>>>> To: Ferruh Yigit <ferruh.yigit@amd.com>; techboard@dpdk.org
>>>>> Cc: Huisong Li <lihuisong@huawei.com>; Chengwen Feng
>>>>> <fengchengwen@huawei.com>
>>>>> Subject: RE: MAC address set requires decision
>>>>>
>>>>> Alternative suggestions:
>>>>>
>>>>> 1. Don't allow "set" of mac address to value already in the list. The user must
>>>>> delete the entry manually first before adding it. Similarly, "add" fails if no
>>>>> default mac address is set. This ensures consistency by enforcing strict
>>>>> separation between the default mac address and the extra mac addresses.
>>>>> You can't have extra addresses without a default, and you can't have
>>>>> duplicates.
>>>>>
>>>>> 2. Always enforce overlap between the two lists - once default mac address is
>>>>> set (automatically adding it to the mac addresses list), you can only replace
>>>>> the default mac address by using an already-added one to the list. In this
>>>>> case, the default address is only really an index into the address list, and no
>>>>> deletion ever occurs.
>>>>>
>>>>> All the solutions below seem rather mixed to me, I'd rather see either strict
>>>>> overlap, or strict non-overlap. Both these cases make it that you need more
>>>>> calls to do certain tasks, e.g. with #2 to just replace mac address, you need to
>>>>> add, set, then delete, but we can always add new, clearly named APIs, to do
>>>>> these compound ops. On the plus side, with #2 we could make things doubly
>>>>> clear by changing the parameter type of "set" to be an index, rather than
>>>>> explicit mac, to make it clear what is happening, that you are choosing a
>>>>> default mac from a list of pre-configured options.
>>>>>
>>>>> Regards,
>>>>> /Bruce
>>> Both of the above option seems good.  The option #1 above is safe, where
>>> you are making the mac address set as independent of mac filtering. Also
>>> making sure that mac filter are not messed up. However, the application
>>> needs to add error handling now to delete and set.
>>>
>>> In the option #2,  I  assume, it will provide full backward
>>> compatibility i.e. the ethernet library can take care of the logic and
>>> application need not to implement anything extra ? If that is the case,
>>> it seems to be best.
>>>
>> I think #2 is not fully backward compatible,
>>
>> Previously set() replaces MAC in list[0], so multiple set() commands end
>> up with single MAC address. So device will filter only one MAC.
>>
>> With #2, after first set(), application will need to add() first, later
>> set() and del() old MAC, if I understand correctly.
>>
>> prev:
>> set(MAC1)
>> set(MAC2)
>> set(MAC3)
>>
>> becomes:
>> set(MAC1)
>> add(MAC2)
>> set(MAC2)
>> del(MAC1)
>> add(MAC3)
>> set(MAC3)
>> del(MAC2)
>>
>>
>> Hence I think this complicates application that wants to just update
>> default MAC.
>>
>>
> I agree. I tend to think that #1 is the simplest and most
> backward-compatible of the options. The only case where we end up with
> different behaviour is the problematic (and already ambiguous) case where
> one attempts to set a default mac that is already specified in the
> "alternative" mac list for the card. It keeps all the simple cases as-is.
>
Hi all,

"The user don't allow "set" of mac address to value already in the list. 
The user must
delete the entry manually first before adding it."
Is this the final decision? keep it what it was?

Do the ethdev framework prohibit the user from setting MAC address that 
are already in the list?
Or, why cannot the ethdev framework explicitly delete the address first 
before setting it, like patch [1]?

Whether there is a problem in this case is completely guaranteed by the 
user.
If the user forget or don't know, there is a another problem as the 
commit of patch [1] described.

[1] 
https://patches.dpdk.org/project/dpdk/patch/20230202123625.14975-1-lihuisong@huawei.com/

/Huisong
> .