| Message ID | cover.1681364951.git.wqu@suse.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | btrfs: reduce the duplicated reads during P/Q scrub | expand |
On Thu, Apr 13, 2023 at 01:57:16PM +0800, Qu Wenruo wrote: > [PROBLEM] > It's a known problem that btrfs scrub for RAID56 is pretty slow. > > [CAUSE] > One of the causes is that we read the same data stripes at least twice > during P/Q stripes scrub. > > This means, during a full fs scrub (one scrub process for each device), > there will be quite some extra seeks just because of this. > > [FIX] > The truth is, scrub stripes have a much better view of the data stripes. > As btrfs would firstly verify all data stripes, and only continue > scrubing the P/Q stripes if all the data stripes is fine after needed > repair. > > So this means, as long as there is no new RMW writes into the RAID56 > block group, we can re-use the scrub cache for P/Q verification. > > This patchset would fix it by: > > - Ensure the RAID56 block groups are marked read-only for scrub > This is to avoid RMW in to the block group, or scrub cache is no > longer reliable. > > - Introduce a new interface to pass cached pages to RAID56 cache > The only disadvantage is here we still need to do page copy, due to > the uncertain lifespan of an rbio. > > Qu Wenruo (2): > btrfs: scrub: try harder to mark RAID56 block groups read-only > btrfs: scrub: use recovered data stripes as cache to avoid unnecessary > read Added to misc-next, thanks. I'd like to batch it with the rest of the scrub but we also need to let it test for a while it so it'll be part of some rc.
[PROBLEM] It's a known problem that btrfs scrub for RAID56 is pretty slow. [CAUSE] One of the causes is that we read the same data stripes at least twice during P/Q stripes scrub. This means, during a full fs scrub (one scrub process for each device), there will be quite some extra seeks just because of this. [FIX] The truth is, scrub stripes have a much better view of the data stripes. As btrfs would firstly verify all data stripes, and only continue scrubing the P/Q stripes if all the data stripes is fine after needed repair. So this means, as long as there is no new RMW writes into the RAID56 block group, we can re-use the scrub cache for P/Q verification. This patchset would fix it by: - Ensure the RAID56 block groups are marked read-only for scrub This is to avoid RMW in to the block group, or scrub cache is no longer reliable. - Introduce a new interface to pass cached pages to RAID56 cache The only disadvantage is here we still need to do page copy, due to the uncertain lifespan of an rbio. Qu Wenruo (2): btrfs: scrub: try harder to mark RAID56 block groups read-only btrfs: scrub: use recovered data stripes as cache to avoid unnecessary read fs/btrfs/block-group.c | 16 +++++++++++++-- fs/btrfs/raid56.c | 46 ++++++++++++++++++++++++++++++++++++++++++ fs/btrfs/raid56.h | 3 +++ fs/btrfs/scrub.c | 16 ++++++++++++++- 4 files changed, 78 insertions(+), 3 deletions(-)