Re: Massive filesystem corruption since kernel 5.2 (ARCH)

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

On 2019/8/27 下午2:34, Swâmi Petaramesh wrote:
> Hi Qu,
>
> Le 27/08/2019 à 08:21, Qu Wenruo a écrit :
>> If free space cache is invalid but passes its csum check, it's
>> completely *possible* to break metadata CoW, thus leads to transid mismatch.
>>
>> You can go v2 space cache which uses metadata CoW to protect its space
>> cache, thus in theory it should be a little safer than V1 space cache.
>>
>> Or you can just disable space cache using nospace_cache mount option, as
>> it's just an optimization. It's also recommended to clean existing cache
>> by using "btrfs check --clear-space-cache v1".
>>
>> I'd prefer to do a "btrfs check --readonly" anyway (which also checks
>> free space cache), then go nospace_cache if you're concerned.
>
> I will leave for travel shortly, so I will be unable to perform further
> tests on this machine for a week, but I'll do when I'm back.
>
> Should I understand your statement as an advice to clear the space cache
> even though the kernel said it has rebuilt it,

Rebuild only happens when kernel detects such mismatch by comparing the
block group free space (recorded in block group item) and free space cache.

If those numbers (along with other things like csum and generation)
match, we don't have a way to detect wrong free space cache at all.

So if kernel is already complaining about free space cache, then no
matter whatever the reason is, you'd better take extra care about the
free space cache.

Although another possible cause is, your extent tree is already
corrupted thus the free space number in block group item is already
incorrect.
You can only determine that by running btrfs check --readonly.

> or to use the V2 space
> cache generally speaking, on any machine that I use (I had understood it
> was useful only on multi-TB filesystems...)

10GiB is enough to create large enough block groups to utilize free
space cache.
So you can't really escape from free space cache.

Thanks,
Qu

>
> Thanks.
>
> ॐ
>