Re: [PATCH] btrfs: repair all bad mirrors

[Christoph Hellwig <hch@lst.de>]

On Wed, Jun 22, 2022 at 01:14:30PM +0800, Qu Wenruo wrote:
>> We need to record at least one failed mirror to be able to repair it, and
>> with the design in this patch we can trivially walk back from the first
>> good mirror to the first bad one.
>
> Then in that case, I guess we can also just submit the good copy to all
> mirrors instead, no matter if it's corrupted or not?

Why would we submit it to a known good mirror?

> But considering repair_io_failure() is still synchronous submission,
> it's definitely going to be slower for RAID1C3/C4.

Yes, two or in the worst case three repair writes are going to be slower
than a single one.  But I think that is worth it for the improved
reliability.

> Just a small nitpick related to the failrec.
> Isn't the whole failrec facility going to be removed after the read
> repair code rework?

Yes.

> So I guess this patch itself is just to solve the test case failure, but
> will still be replaced by the new read repair rework?

I've shifted plans for repair a bit and plan to do more gradual work.
Eventually the failrec should go away in this form, though.