Re: [PATCH 0/3] gfs2: stop using struct page in the log code

From: Tal Zussman

Date: Fri Nov 20 2026 - 13:22:54 EST


On 11/19/26 6:47 PM, Andreas Gruenbacher wrote:
> Tal,
>
> On Tue, Nov 17, 2026 at 8:46 AM Tal Zussman <tz2294@xxxxxxxxxxxx> wrote:
>> This series gets rid of the remaining struct page usage in the gfs2 log
>> code.
>>
>> The log headers, descriptors, revoke blocks, and escaped block copies
>> are allocated as pages from a mempool, and they end up in the same bios
>> as the page cache buffers being journaled. The completion handler then
>> calls page_folio() on every segment to figure out which kind it is
>> looking at. That won't work once memdescs land and those pool pages are
>> no longer folios.
>>
>> Patch 1 moves these blocks over to a kmalloc mempool, with their own
>> bios and completion handler. As a result each descriptor block is now
>> submitted as its own bio, so patch 2 adds a plug around the log flush to
>> let the block layer merge them back together. Patch 3 cleans up the
>> leftover page references in the log code.
>>
>> Tested with fio 4K write+fsync on gfs2 over virtio-blk and null_blk.
>> Without an I/O scheduler, patch 1 adds one write request per log flush
>> (196.7k to 229.5k for 32k fsyncs) and ~10 us to the p99.9 fsync latency
>> on null_blk (41 us to 51 us); the plug in patch 2 merges them and
>> restores both performance numbers to the baseline. With mq-deadline
>> nothing changes.
>
> thank you for these patches. Could you please rerun that test with a
> block size smaller than the page size?
>
> Thanks,
> Andreas
>

I ran it with a 1k block size on 4k pages, and the result is the same:

job (fsyncs) sched base p1 p2
---------------- ----------- -------- -------- --------
fsync4k (32768) none 199519 232296 199514
fsync64k (2048) none 15445 17503 15441
create (2000) none 12228 14234 12226
fsync4k (32768) mq-deadline 199509 200016 199505

p99.9 latency results are also the same.

Thanks,
Tal