Re: [PATCH stable/6.18-6.17] md: add check_new_feature module parameter

在 2025/12/18 14:30, Greg KH 写道:
> On Wed, Dec 17, 2025 at 10:31:30PM +0500, Roman Mamedov wrote:
>> On Thu, 18 Dec 2025 01:11:43 +0800
>> "Yu Kuai" <[email protected]> wrote:
>>
>>> Hi,
>>>
>>> 在 2025/12/17 22:04, Greg KH 写道:
>>>> On Wed, Dec 17, 2025 at 09:05:13PM +0800, [email protected] wrote:
>>>>> From: Li Nan <[email protected]>
>>>>>
>>>>> commit 9c47127a807da3e36ce80f7c83a1134a291fc021 upstream.
>>>>>
>>>>> Raid checks if pad3 is zero when loading superblock from disk. Arrays
>>>>> created with new features may fail to assemble on old kernels as pad3
>>>>> is used.
>>>>>
>>>>> Add module parameter check_new_feature to bypass this check.
>>>> This is a new feature, why does it need to go to stable kernels?
>>>>
>>>> And a module parameter?  Ugh, this isn't the 1990's anymore, this is not
>>>> good and will be a mess over time (think multiple devices...)
>>>
>>> Nan didn't mention the background. We won't backport the new feature to stable
>>> kernels(Although this fix a data lost problem in the case array is created
>>> with disks in different lbs, anyone is interested can do this). However, this
>>> backport is just used to provide a possible solution for user to still assemble
>>> arrays after switching to old LTS kernels when they are using the default lbs.
>>
>> This is still a bad scenario. Original problem:
>>
>> - Boot into a new kernel once, reboot into the old one, the existing array no
>>    longer works.
>>
>> After this patch:
>>
>> - Same. Unless you know how, where and which module parameter to add, to
>>    be passed to md module on load. Might be not convenient if the root FS
>>    didn't assemble and mount and is inaccessible.
>>
>> Not ideal whatsoever.
>>
>> Wouldn't it be possible to implement minimal *automatic* recognition (and
>> ignoring) of those newly utilized bits instead?
> 
> Yes, that should be done instead.
> 
> And again, a module parameter does not work for multiple devices in a
> system, the upstream change should also be reverted.
> 
> thanks,
> 
> greg k-h
> 
> .

We propose the following fix for this issue. After fix, md arrays created
on old kernels won't be affected by this feature.

https://lore.kernel.org/linux-raid/[email protected]/T/#mb205fb97ab4af629cae9db8dfd236ceaa93f14ad

The method is:
 > only set lbs by default for new array, for assembling the array still
 > left the lbs field unset, in this case the data loss problem is not fixed,
 > we should also print a warning and guide users to set lbs to fix the 
problem,
 > with the notification the array will not be assembled in old kernels.

-- 
Thanks,
Nan