On 2019/11/28 下午7:24, David Sterba wrote:
> On Thu, Nov 28, 2019 at 07:36:41AM +0800, Qu Wenruo wrote:
>> On 2019/11/28 上午3:23, David Sterba wrote:
>>> On Tue, Nov 19, 2019 at 06:41:49PM +0800, Qu Wenruo wrote:
>>>> On 2019/11/19 下午6:05, Anand Jain wrote:
>>>>> On 11/7/19 2:27 PM, Qu Wenruo wrote:
>>>>>> [PROBLEM]
>>>>>> Btrfs degraded mount will fallback to SINGLE profile if there are not
>>>>>> enough devices:
>>>>>
>>>>>  Its better to keep it like this for now until there is a fix for the
>>>>>  write hole. Otherwise hitting the write hole bug in case of degraded
>>>>>  raid1 will be more prevalent.
>>>>
>>>> Write hole should be a problem for RAID5/6, not the degraded chunk
>>>> feature itself.
>>>>
>>>> Furthermore, this design will try to avoid allocating chunks using
>>>> missing devices.
>>>> So even for 3 devices RAID5, new chunks will be allocated by using
>>>> existing devices (2 devices RAID5), so no new write hole is introduced.
>>>
>>> That this would allow a 2 device raid5 (from expected 3) is similar to
>>> the reduced chunks, but now hidden because we don't have a detailed
>>> report for stripes on devices. And rebalance would be needed to make
>>> sure that's the filesystem is again 3 devices (and 1 device lost
>>> tolerant).
>>>
>>> This is different to the 1 device missing for raid1, where scrub can
>>> fix that (expected), but the balance is IMHO not.
>>>
>>> I'd suggest to allow allocation from missing devices only from the
>>> profiles with redundancy. For now.
>>
>> But RAID5 itself supports 2 devices, right?
>> And even 2 devices RAID5 can still tolerant 1 missing device.
> 
>> The tolerance hasn't changed in that case, just unbalanced disk usage then.
> 
> Ah right, the constraints are still fine. That the usage is unbalanced
> is something I'd still consider a problem because it's silently changing
> the layout from the one that was set by user.

Then it's the trade off between:
- Completely the same layout
- Just tolerance

I'd say, even without missing device, btrfs by its nature, it can't
ensure all chunks are allocated using the same device layout.

E.g. 4 disk raid5, 1T + 2T + 2T + 2T.
There will be a point where btrfs can only go 3 disks RAID5 other than 4.

> 
> As there are two conflicting ways to continue from the missing device state:
> 
> - try to use remaining devices to allow writes but change the layout
> - don't allow writes, let user/admin sort it out
> 
> I'd rather have more time to understand the implications and try to
> experiment with that.

Sure, no problem.

Thanks,
Qu