Re: [PATCH v3 1/3] md/raid1,raid10: Do not set MD_BROKEN on failfast io failure

[Kenta Akagi <k@mgml.me>]

On 2025/09/09 10:02, Yu Kuai wrote:
> Hi,
> 
> 在 2025/09/09 1:29, Kenta Akagi 写道:
>> On 2025/09/08 10:20, Yu Kuai wrote:
>>>
>>>
>>> 在 2025/09/02 0:48, Kenta Akagi 写道:
>>>> In the current raid1_end_write_request implementation,
>>>> - md_error is called only in the Failfast case.
>>>> - Afterwards, if the rdev is not Faulty (that is, not Failfast,
>>>>     or Failfast but the last rdev — which originally was not expected
>>>>     MD_BROKEN in RAID1), R1BIO_WriteError is set.
>>>> In the suggested implementation, it seems that a non-Failfast write
>>>> failure will immediately mark the rdev as Faulty, without retries.
>>>
>>> I still prefer a common helper to unify the code, not sure if I still
>>> missing something ...
>>>
>>> In general, if bio failed, for read/write/metadata/resync should be the
>>> same:
>>>
>>> 1) failfast is set, and not last rdev, md_error();
>>> 2) otherwise, we should always retry;
>>>
>>> And I do believe it's the best to unify this by a common helper.
>>
>> Yes, I realized that my idea is bad. Your idea is best,
>> especially considering the error handling in super_written.
>> I'll implement a common helper.
>>
>> By the way, I think md_error should only be serialized on RAID1 and 10
>> for now. Serializing unnecessary personalities is inefficient and can
>> lead to unfavorable results. What do you think?
> 
> Just make code cleaner and I don't have preference here, md_error is
> super cold path I think.

Thank you for your advice.
Understood.

Thanks,
Akagi

> Thanks,
> Kuai
> 
>>
>> Thanks,
>> Akagi
>>
>>> Thanks,
>>> Kuai
>>>
>>>
>>
>> .
>>
> 
>