On 6/11/25 9:10 AM, Ming Lei wrote:
On Wed, Jun 11, 2025 at 12:14:54PM +0000, Hazem Mohamed Abuelfotoh wrote:
This reverts commit e70c301faece15b618e54b613b1fd6ece3dd05b4.
Commit <e70c301faece> ("block: don't reorder requests in blk_add_rq_to_plug") reversed how requests are stored in the blk_plug list, this had significant impact on bio merging with requests exist on the plug list. This impact has been reported in [1] and could easily be reproducible using 4k randwrite fio benchmark on an NVME based SSD without having any filesystem on the disk.
My benchmark is:
fio --time_based --name=benchmark --size=50G --rw=randwrite \--runtime=60 --filename="/dev/nvme1n1" --ioengine=psync \ --randrepeat=0 --iodepth=1 --fsync=64 --invalidate=1 \ --verify=0 --verify_fatal=0 --blocksize=4k --numjobs=4 \ --group_reporting
On 1.9TiB SSD(180K Max IOPS) attached to i3.16xlarge AWS EC2 instance.
Kernel | fio (B.W MiB/sec) | I/O size (iostat) --------------+---------------------+-------------------- 6.15.1 | 362 | 2KiB 6.15.1+revert | 660 (+82%) | 4KiB --------------+---------------------+--------------------
I just run one quick test in my test VM, but can't reproduce it.
Also be curious, why does writeback produce so many 2KiB bios?
I was pondering that too, sounds like a misconfiguration of sorts. But even without that, in a quick synthetic test here locally, I do see a lot of missed merges that is solved with the alternative patch I sent out. I strongly suspect it'll fix this issue too.