Re: [6.2][regression] after commit 947a629988f191807d2d22ba63ae18259bb645c5 btrfs volume periodical forced switch to readonly after a lot of disk writes

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

On 2022/12/29 08:08, Mikhail Gavrilov wrote:
> On Thu, Dec 29, 2022 at 4:31 AM Qu Wenruo <quwenruo.btrfs@gmx.com> wrote:
>>
>> Are you using qgroup? If so it may be worthy trying disabling qgroup.
> 
> I do not use quota.
> And looks like my distro does not use quita by default.
> ❯ btrfs qgroup show -f /
> ERROR: can't list qgroups: quotas not enabled
> 
>> But for newer kernel, qgroup hang should only happen when dropping large
>> snapshot, I don't know if podman pull would cause older snapshots to be
>> deleted...
> 
> It is not a regression, it also happened on older kernels.
> But it is really annoying when the browser process waits when "podman
> pull" writes changes to disk.
> In fact, I have been waiting for 5 years for caching of slow HDDs by
> using the cache on the SSD, but apparently I can’t wait.
> And I started slowly buying expensive large SSDs to replace the big
> HDD. I still can’t find time to connect D5 P5316 30.72 Tb to the
> primary workstation.
> I want to make a video review of it. I understand this is an expensive
> solution and not suitable for everyone, unlike an affordable SDD
> cache.
> 
This is really sad to hear that.

For now, I only have several guesses on how this could happen.

- Extra seeks between metadata and data chunks
   You can try with mixed block groups, but this needs mkfs time tuning.

- Extra inlined extents causing too much metadata overhead
   You can disable inline extents using max_inline=0 as mount options.
   But that only affects newly created files, not the existing ones.

Otherwise, using bcache may be a solution.

For now I'm not aware of any HDD specific tests, other than zoned 
devices, thus the performance problem can be really hard to debug.

Thanks,
Qu