On Aug 31, 2021, at 8:49 AM, Alex Zhuravlev [off-list ref] wrote:

Hi,

Not so long ago we noticed that journal replay can take quite a lot (hours)
in cases where many journaled blocks were freed during a short period.

It may be worthwhile to mention this was a case with a 4GB journal size.

I benchmarked hash table used by revoke code, basically it’s lookup+insert
like jbd2 does at replay:

1048576 records - 95 seconds
2097152 records - 580 seconds

Then I benchmarked rhashtable:
1048576 records - 2 seconds
2097152 records - 3 seconds
4194304 records - 7 seconds

So, here is a patch replacing existing fixed-size hash table with rhashtable, please have a look.

Thanks, Alex

Alex,
the patch looks good from both a performance standpoint, as well as
a good reduction in lines (and possibly memory, for the cases where
there are fewer entries in the hash than the static table size).

I did notice some lines are using 8-space indents instead of tabs.
That can be fixed if there are any other comments, and you resubmit
without [RFC].

Cheers, Andreas