multipathd and RAID concurency.

[mtkaczyk]

Hello, I see the issue between multipathd and MD raid management. I have SLES 15 SP5 system configured with VROC RAID0 and native RAID0:

 # cat /proc/mdstat
Personalities : [raid1] [raid6] [raid5] [raid4] [raid0]
md127 : inactive nvme2n1[1](S) nvme0n1[0](S)
      2210 blocks super external:imsm

md126 : active raid0 nvme3n1[0]
      976630272 blocks super 1.2 512k chunks

md125 : active raid0 nvme2n1[1] nvme0n1[0]
      1953513472 blocks super external:/md127/0 128k chunks

unused devices: <none>

I determined that enabling multipathd causes that my arrays are not started after reboot:

# cat /proc/mdstat
Personalities : [raid6] [raid5] [raid4]
md125 : inactive nvme3n1[0]
      976630488 blocks super 1.2

md126 : inactive nvme0n1[1] nvme2n1[0]
      1953513472 blocks super external:/md127/0

md127 : inactive nvme2n1[1](S) nvme0n1[0](S)
      10402 blocks super external:imsm

It seems to be caused by concurrency between mdadm and this daemon on startup, the daemon doesn't respect metadata on the drives.

Obviously, I'm able to fix the issue by blacklisting devices in multipath config. Another workaround is to force MD modules loading to initrd image, so the raid will be started earlier. The main problem is if multipathd is enabled during the installation then after reboot to new OS raids are broken. The tool policy is aggressive because by default it claims every nvme device.

Can we do something to change this behavior? Here some ideas:

• check if there is the metadata before it starts manipulate the device like lvm2 does.
• change default behavior to mange only devices specified in config (probably it is not an option because the tool exists long time on market),
• create configuration file in installer to blacklist devices in raid arrays, (it needs to be addressed to each OSV separately so I would prefer to resolve it in the tool).

Thanks, Mariusz

[mtkaczyk]

@mwilck I can see you in recent contributors so I'm notifying you directly - SLES is affected, could you take a look?

[mwilck]

@mtkaczyk, Please open a SUSE bug.

[mwilck]

The tool policy is aggressive because by default it claims every nvme device.

It's true, we use this policy under SUSE. Other distributions are probably not affected. The solution is indeed to use blacklisting, and set find_multipaths option to yes or smart.

In your case I'd recommend to activate multipath after installation, which should "just work".

[mwilck]

This is related to SUSE only, not an upstream issue.