Re: Kernel bug in async_xor_offs during RAID5 recovery

["Tkaczyk, Mariusz" <mariusz.tkaczyk@linux.intel.com>]

Hi Xiao,
I simplified the scenario.

On 06.05.2021 12:57, Xiao Ni wrote:
> Hi Oleksandr Shchirskyi
> 
> Can this only happen with PPL, imsm and nvme disks? My machine doesn't support 
> creating raid device with nvme devices.
Could you try to create array with IMSM_NO_PLATFORM=1?

> And rotational disks don't have /sys/block/nvme1n1/device/device/remove. What's 
> the meaning about setting 1
> to the remove file?
> 
> I tried to create the imsm raid with rotational disks and ppl. Then remove and 
> add disk to trigger recovery. It works
> well.
I verified that drive removal is not crucial here.
The main trick here is to impose the PPL recovery. I did that by following
scenario:
1. Create array (I was able to reproduce it with 1Gb size):
#mdadm -CR imsm -e imsm -n3 /dev/nvme[456]n1
#mdadm -CR vol2 -l5 -n3 /dev/nvme[456]n1 -z 1G -c64 --assume-clean 
--consistency-policy=ppl

2. Get mdmon pid:
# ps -ef | grep mdmon

3. Write data to the drive, and kill mdmon after that (this will force
array to stay dirty):
#dd of=/dev/md126 if=/dev/urandom bs=4M oflag=direct; kill -9 {mdmon_pid}

4. Stop Array:
# mdadm -Ss

5. Start Array:
# mdadm -As

Thanks,
Mariusz