Re: [PATCH V8] ethdev: fix one address occupies two entries in MAC addrs

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

在 2023/2/10 20:27, Ferruh Yigit 写道:
> On 2/10/2023 9:54 AM, lihuisong (C) wrote:
>> 在 2023/2/9 20:45, Ferruh Yigit 写道:
>>> On 2/4/2023 2:57 AM, lihuisong (C) wrote:
>>>> 在 2023/2/3 20:58, Ferruh Yigit 写道:
>>>>> On 2/3/2023 1:56 AM, lihuisong (C) wrote:
>>>>>> 在 2023/2/3 5:10, Thomas Monjalon 写道:
>>>>>>> 02/02/2023 19:09, Ferruh Yigit:
>>>>>>>> On 2/2/2023 12:36 PM, Huisong Li wrote:
>>>>>>>>> The dev->data->mac_addrs[0] will be changed to a new MAC address
>>>>>>>>> when
>>>>>>>>> applications modify the default MAC address by .mac_addr_set().
>>>>>>>>> However,
>>>>>>>>> if the new default one has been added as a non-default MAC
>>>>>>>>> address by
>>>>>>>>> .mac_addr_add(), the .mac_addr_set() doesn't remove it from the
>>>>>>>>> mac_addrs
>>>>>>>>> list. As a result, one MAC address occupies two entries in the
>>>>>>>>> list.
>>>>>>>>> Like:
>>>>>>>>> add(MAC1)
>>>>>>>>> add(MAC2)
>>>>>>>>> add(MAC3)
>>>>>>>>> add(MAC4)
>>>>>>>>> set_default(MAC3)
>>>>>>>>> default=MAC3, the rest of the list=MAC1, MAC2, MAC3, MAC4
>>>>>>>>> Note: MAC3 occupies two entries.
>>>>>>>>>
>>>>>>>>> In addition, some PMDs, such as i40e, ice, hns3 and so on, do
>>>>>>>>> remove
>>>>>>>>> the
>>>>>>>>> old default MAC when set default MAC. If user continues to do
>>>>>>>>> set_default(MAC5), and the mac_addrs list is default=MAC5,
>>>>>>>>> filters=(MAC1,
>>>>>>>>> MAC2, MAC3, MAC4). At this moment, user can still see MAC3 from the
>>>>>>>>> list,
>>>>>>>>> but packets with MAC3 aren't actually received by the PMD.
>>>>>>>>>
>>>>>>>>> So need to ensure that the new default address is removed from the
>>>>>>>>> rest of
>>>>>>>>> the list if the address was already in the list.
>>>>>>>>>
>>>>>>>> Same comment from past seems already valid, I am not looking to the
>>>>>>>> set
>>>>>>>> for a while, sorry if this is already discussed and decided,
>>>>>>>> if not, I am referring to the side effect that setting MAC addresses
>>>>>>>> cause to remove MAC addresses, think following case:
>>>>>>>>
>>>>>>>> add(MAC1) -> MAC1
>>>>>>>> add(MAC2) -> MAC1, MAC2
>>>>>>>> add(MAC3) -> MAC1, MAC2, MAC3
>>>>>>>> add(MAC4) -> MAC1, MAC2, MAC3, MAC4
>>>>>>>> set(MAC3) -> MAC3, MAC2, MAC4
>>>>>>>> set(MAC4) -> MAC4, MAC2
>>>>>>>> set(MAC2) -> MAC2
>>>>>>>>
>>>>>>>> I am not exactly clear what is the intention with set(),
>>>>>>> That's the problem, nobody is clear with the current behavior.
>>>>>>> The doc says "Set the default MAC address." and nothing else.
>>>>>> Indeed. But we can see the following information.
>>>>>>    From the ethdev layer, this set() API always replaces the old
>>>>>> default
>>>>>> address (index 0) without adding the old one.
>>>>>>    From the PMD layer, set() interface of some PMDs, such as i40e, ice,
>>>>>> hns3 and so on (as far as I know),
>>>>>> also do remove the hardware entry of the old default address.
>>>>> If we define behavior clearly, I think we can adapt PMD implementation
>>>>> according it, unless there is HW limitation.
>>>> Right. I think this is another point (issue 2/) to be discussed.
>>>> Namely, whether the old default address should be removed when set new
>>>> default one.
>>>> If we want to explicitly unify the behavior of all PMDs in ethdev layer
>>>> as described above,
>>>> there may be no problem if do the following:
>>>> 1) In the ethdev layer, remove the old default address if the old one is
>>>> exist.
>>>> 2) For PMD i40e, ice and hns3, remvoe the code of deleting the old
>>>> default address before adding the new one.
>>>>      For other PMDs, we probably don't need to do anything because they
>>>> have supported remove_addr() API.
>>>>      (Without explicitly removing the old default address, I don't
>>>> know if
>>>> their hardware or firmware
>>>>       removes the old one when set a new address. But, we explicitly
>>>> remove the old one in ethdev layer now,
>>>>       I'm not sure if this has an effect on these PMDs.)
>>>>>>>> if there is
>>>>>>>> single MAC I guess intention is to replace it with new one, but if
>>>>>>>> there
>>>>>>>> are multiple MACs and one of them are already in the list intention
>>>>>>>> may
>>>>>>>> be just to change the default MAC.
>>>>>>> The assumption in this patch is that "Set" means "Replace", not
>>>>>>> "Swap".
>>>>>>> So this patch takes the approach 1/ Replace and keep Unique.
>>>>>>>
>>>>>>>> If above assumption is correct, what about following:
>>>>>>>>
>>>>>>>> set(MAC) {
>>>>>>>>         if only_default_mac_exist
>>>>>>>>             replace_default_mac
>>>>>>>>
>>>>>>>>         if MAC exists in list
>>>>>>>>        swap MAC and list[0]
>>>>>>>>         else
>>>>>>>>        replace_default_mac
>>>>>>>> }
>>>>>>> This approach 2/ is a mix of Swap and Replace.
>>>>>>> The old default MAC destiny depends on whether
>>>>>>> we have added the new MAC as "secondary" before setting as new
>>>>>>> default.
>>>>>>>
>>>>>>>> This swap prevents removing MAC side affect, does it make sense?
>>>>>>> Another approach would be 3/ to do an "Always Swap"
>>>>>>> even if the new MAC didn't exist before,
>>>>>>> you keep the old default MAC as a secondary MAC.
>>>>>>>
>>>>>>> And the current approach 0/ is to Replace default MAC address
>>>>>>> without touching the secondary addresses at all.
>>>>>>>
>>>>>>> So we have 4 choices.
>>>>>>> We could vote, roll a dice, or find a strong argument?
>>>>>> According to the implement of set() in ethdev and PMD layer, it always
>>>>>> use "Replace", not "Swap".
>>>>>> If we use "Swap" now, the behavior of this API will be changed.
>>>>>> I'm not sure if the application can accept this change or has other
>>>>>> effects.
>>>>>>
>>>>> This patch is also changing behavior, because of implied remove
>>>>> address,
>>>>> same concern is valid with this patch.
>>>> Indeed, it changes the behavior.
>>>> But this patch only resolves the problem (issue 1/) that the entries of
>>>> the MAC address list possibly are not uniques.
>>>> Fixing it may be little impact on the application.
>>>>> As I checked again current implementation may have one more problem
>>>>> (this from reading code, I did not test this):
>>>>> add(MAC1) -> MAC1
>>>>> add(MAC2) -> MAC1, MAC2
>>>>> set(MAC2) -> MAC2, MAC2
>>>>> del(MAC2) -> FAILS
>>>>>
>>>>> This fails because `rte_eth_dev_mac_addr_remove()` can't remove default
>>>>> MAC, and it only tries to remove first address it finds, it can't find
>>>>> and remove second 'MAC2'.
>>>>> I wasn't too much bothered with wasting one MAC address slot, so wasn't
>>>>> sure if a change is required at all, but if above analysis is correct I
>>>>> think this is more serious problem to justify the change.
>>>> Your analysis is fully correct.
>>>>> I don't think always swap (option /3) is good idea, specially for
>>>>> single
>>>>> MAC address exists case, and current case has (option 0/) has mentioned
>>>>> problems.
>>>> +1
>>>>> Remaining ones are mix of swap and replace (option 2/) and this patch
>>>>> (option /1).
>>>>>
>>>>> I think mix of swap and replace (option 2/ above) has some benefits:
>>>>> - It always replaces default MAC
>>>>> - Prevents duplication MAC address in the list
>>>>> - Doesn't implicitly remove address from list
>>>> As far as I know, the first entry (index 0) always be the default
>>>> address in all PMDs,
>>>> but it's not documented. (So this patch did it, that's what was
>>>> discussed earlier).
>>>> The 'Swap' may be inappropriate. It may need to be discussed.
>>> Yes, index 0 always holds the default MAC, +1 to document this.
>>>
>>> In option /2, MAC swap is done only if the MAC address is already in the
>>> list.
>>>
>> Whether the 'swap' is required in this case depends on the assumption
>> that the
>> old default address needs to be 'secondary' address when set new default
>> one, right?
>
> We don't know if default address needs to be 'secondary' address or not,
> swap is just a way to both keep list unique and don't implicit remove
> any MAC address.
>
>
>> It is still the same issue as issue /2 to be discussed.>> Another option can be to store an enabled/disabled flag for each MAC
>>> address in the list.
>>> In set(MAC), if MAC is already in the list, the existing one in the list
>>> can be disabled. Next time default MAC changed, disabled MAC can be
>>> enabled again. Like:
>>>
>>> (d) => disabled
>>>
>>> add(MAC1) -> MAC1
>>> add(MAC2) -> MAC1, MAC2
>>> add(MAC3) -> MAC1, MAC2, MAC3
>>> add(MAC4) -> MAC1, MAC2, MAC3, MAC4
>>> set(MAC3) -> MAC3, MAC2, MAC3(d), MAC4
>>> set(MAC4) -> MAC4, MAC2, MAC3, MAC4(d)
>>> set(MAC2) -> MAC2, MAC2(d), MAC3, MAC4
>>> set(MAC5) -> MAC5, MAC2, MAC3, MAC4
>>>
>>> The ones disabled in the ethdev layer can be explicitly removed in
>>> drivers, and the ones enabled in the ethdev layer can be explicitly
>>> added back in drives. This simplifies the driver implementation a lot.
>>>
>> Good idea, but it's a little complicated.
>> If I understand correctly, all of the above are done under the condition
>> that 'swap' is accepted.
>
> Agree it is more complicated, and no this doesn't require swap.
>
> Instead of implicitly removing duplicated MAC address it keeps in the
> ethdev list as disabled, and adds it back when default MAC is changed.
>
>
>>>>> BUT, if the agreement is this patch (option 1/) I am OK with that
>>>>> too, I
>>>>> just want to make sure that it is discussed.
>>>>>
>>>>>
>>>>>> BTW, it seems that the ethernet port in kernel also replaces the old
>>>>>> address if we modify the one.
>>>>>> Use the test command: ifconfig eth0 hw ether new_mac
>>>>> For default MAC address it is more clear that intention is to replace
>>>>> it, but question is about what to do with the list of MAC addresses.
>>>> Hi Ferruh and Thomas,
>>>>
>>>> As mentioned above, they are actually two problems (issue /1 and
>>>> issue /2).
>>>> Can we deal with them separately?
>>>> #1 For issue /1, it's really a problem. This patch is responsible for
>>>> it.
>>>> #2 For issue /2, I will send a RFC to discuss as described above.
>>>>        It may require the participation of all PMD maintainers.
>>>>
>>>> What do you think?
>>>>
>>> Agree to separate fixing drivers (issue /2) and ethdev (issue /1), and
>>> drivers can be fixed after ethdev clarified.
>> Yeah, we are in the same boat.😁
>>> Sorry that this patch is taking time because expected behavior is not
>> I cannot understand the behavior determined by this patch is not expected.
>> Wouldn't it be simpler and more reasonable to make sure that the entries
>> in the MAC list are uniques?
>
> Yes, above part is OK. Only problem is it may cause shrinking the MAC
> list by setting MAC address, that may not be the expected behavior.
The root cause of shrinking the MAC list is that the old default address 
is removed when set MAC address,
rather than removing the address being already in the rest of the list.
>
>
>>> clear, and we are not getting enough feedback for it. Also issue is not
>>> a major or a blocking issue.
>> IMO, the last point to be discussed and have enough feedback is the
>> issue /2 instead of issue /1.
>> Whether the old default address needs to be removed or swapped when set
>> new one is the major or a blocking issue.
>>
>> We can't go back and forth between them. we should break the cycle.
>> For issue /1 (duplicate entries), please forget what to do with the old
>> default address. I'm sure you'll think it's ok.
>> Let's take a new discussion for the old default address in set() API to
>> get enough feedback.
>>
>> what do you think, Ferruh and Thomas?
> Does it work to present options to techboard, get a decision and proceed
> with it?
Agreed. At least, there's a decision soon.
> .