Re: RAID1 two chunks of the same data on the same physical disk, one file keeps being corrupted

[ein <ein.net@gmail.com>]

On 13.01.2025 21:39, Qu Wenruo wrote:
> 在 2025/1/14 02:24, ein 写道:
>> On 29.07.2024 12:05, Qu Wenruo wrote:
>>> On 10.06.2024 16:56, ein wrote:
>>> In your case, I still do not believe it's hardware problem.
>>>
>>> > - it affects only one file, I have other much busier VMs, that one
>>> mostly stays idle,
>>>
>>> Due to btrfs' datacsum behavior, it's very sensitive to page content
>>> change during writeback.
>>>
>>> Normally this should not happen for buffered writes as btrfs has locked
>>> the page cache.
>>>
>>> But for Direct IO it's still very possible that one process submitted a
>>> direct IO, and when the IO was still under way, the user space changed
>>> the contents of that page.
>>>
>>> In that case, btrfs csum is calculated using that old contents, but the
>>> on-disk data is the new contents, causing the csum mismatch.
>>>
>>> So I'm wondering what's the workload inside the VM?
>>
>> As far as I know in such configuration there's no writeback:
>>
>> <disk type="file" device="disk">
>>    <driver name="qemu" type="qcow2" cache="none" discard="unmap"/>
>
> cache="none" means direct IO.
>
> Exactly the problem I mentioned, direct IO with data changed during
> writeback.
>
> You can change it to "cache=writeback" then it should resolve the false
> alert mismatch.
> (Or just simply change the disk image file to NODATASUM)
Hi Qu.
You were right, those errors still happened.
Switching to cache=writeback seemed to help for now.
Thank you.
>>    <source file="/var/lib/libvirt/images-red-btrfs/dell.qcow2" index="2"/>
>>    <backingStore/>
>>    <target dev="vda" bus="virtio"/>
>>    <alias name="virtio-disk0"/>
>>    <address type="pci" domain="0x0000" bus="0x00" slot="0x04"
>> function="0x0"/>
>> </disk>
>> [...]
>> <controller type="pci" index="0" model="pci-root">
>>    <alias name="pci.0"/>
>> </controller>
>>
>> This is mostly empty Win7 virtual machine with very small SQLite
>> database (100-500MiB) with some network monitoring tool.