hi, Christoph Hellwig,
On Mon, Jun 24, 2024 at 10:35:37AM +0200, Christoph Hellwig wrote:
> This is odd to say at least. Any chance you can check the value
> of /sys/block/$DEVICE/queue/rotational for the relevant device before
> and after this commit? And is this an ATA or NVMe SSD?
>
yeah, as Niklas mentioned, it's an ATA SSD.
I checked the /sys/block/$DEVICE/queue/rotational before and after this commit,
both show '0'. not sure if this is expected.
anyway, I noticed you send a patch [1]
so I applied this patch upon bd4a633b6f, and found the performance restored.
=========================================================================================
compiler/cpufreq_governor/disk/filesize/fs2/fs/iterations/kconfig/nr_directories/nr_files_per_directory/nr_threads/rootfs/sync_method/tbox_group/test_size/testcase:
gcc-13/performance/1SSD/9B/nfsv4/btrfs/1x/x86_64-rhel-8.3/16d/256fpd/32t/debian-12-x86_64-20240206.cgz/fsyncBeforeClose/lkp-ivb-2ep2/400M/fsmark
commit:
1122c0c1cc ("block: move cache control settings out of queue->flags")
bd4a633b6f ("block: move the nonrot flag to queue_limits")
e9a0f6a398 = bd4a633b6f + patch [1]
1122c0c1cc71f740 bd4a633b6f7c3c6b6ebc1a07317 e9a0f6a398f162d115d208ad95b
---------------- --------------------------- ---------------------------
%stddev %change %stddev %change %stddev
\ | \ | \
4177 ± 2% -64.7% 1475 -1.1% 4130 fsmark.files_per_sec
[1] https://lore.kernel.org/all/20240624173835.76753-1-hch@xxxxxx/
