Re: [PATCH v7 1/2] btrfs: Fix metadata underflow caused by btrfs_reloc_clone_csum error

[Qu Wenruo <quwenruo@cn.fujitsu.com>]

At 03/13/2017 09:26 PM, Stefan Priebe - Profihost AG wrote:
>
> Am 13.03.2017 um 08:39 schrieb Qu Wenruo:
>>
>>
>> At 03/13/2017 03:26 PM, Stefan Priebe - Profihost AG wrote:
>>> Hi Qu,
>>>
>>> Am 13.03.2017 um 02:16 schrieb Qu Wenruo:
>>>
>>> But wasn't this part of the code identical in V5? Why does it only
>>> happen with V7?
>>
>> There are still difference, but just as you said, the related
>> part(checking if inode is free space cache inode) is identical across v5
>> and v7.
>
> But if i boot v7 it always happens. If i boot v5 it always works. Have
> done 5 repeatet tests.

I rechecked the code change between v7 and v5.

It turns out that, the code base may cause the problem.

In v7, the base is v4.11-rc1, which introduced quite a lot of 
btrfs_inode cleanup.

One of the difference is the parameter for btrfs_is_free_space_inode().

In v7, the parameter @inode changed from struct inode to struct btrfs_inode.

So in v7, we're passing BTRFS_I(inode) to btrfs_is_free_space_inode(), 
other than plain inode.

That's the most possible cause for me here.

So would you please paste the final patch applied to your tree?
Git diff or git format-patch can both handle it.

Thanks,
Qu

>
>> I'm afraid that's a rare race leading to NULL btrfs_inode->root, which
>> could happen in both v5 and v7.
>>
>> What's the difference between SUSE and mainline kernel?
>
> A lot ;-) But i don't think anything related.
>
>> Maybe some mainline kernel commits have already fixed it?
>
> May be no idea. But i haven't found any reason why v5 works.
>
> Stefan
>
>>
>> Thanks,
>> Qu
>>>