multipathd started leads to initrd blocked.

[lixiaokeng]

With 11f0440 commit, we meet a problem.

When there is a multipath device in initrd, multipathd will be actived becasue of dracut/modules.d/90multipath/multipathd-needshutdown.sh running. This lead to initrd-cleanup.service can't be run and switch root fails. The initrd is blocked.

https://github.com/dracutdevs/dracut/issues/2289

[mwilck]

I think the entire multipathd-needshutdown.sh logic is obsoleted by our queue_without_daemon feature. multipathd itself will take care that queuing is disabled when it quits.

@bmarzins, please double-check, I tend to overlook things lately ...

[mwilck]

If we still need multipathd-needshutdown.sh and multipathd-shutdown.sh (as noted above, I think we don't), we should call multipath with the -D option in these files.

[mwilck]

@lixiaokeng, please try if removing the line

Also=multipathd.socket

in multipathd.service fixes the issue.

[lixiaokeng]

We test this （https://github.com/dracutdevs/dracut/pull/2290/files），and it fix this issue.

[bmarzins]

I'm on the fence about the best way to solve this:

• Removing those dracut scripts cuts down on unnecessary work. But multipathd won't turn off queuing if it terminates abnormally, so I guess these are a backup to make sure that IO to these devices won't hang indefinitely, possibly hanging your system just when it hardest to figure out what's going wrong. I'm not sure if this corner case is enough to justify their existence.
• Disabling multipath.socket in the initramfs seems really sensible. If we need multipathd running in the initramfs, we should be starting it explicitly. Disabling the socket unit will catch any unintended startups. I doubt that this change will cause any problems. If some setup in some distribution is relying on a socket call to start up multipathd in the initramfs, that seems like their bug.
• Running the shutdown multipath commands run with -D seems like the right thing to do. But it also seems like we should be doing more.

In the end, I think I vote for doing the last two. Disabling the socket should be enough for the initramfs, but there's no point in multipathd trying to delegate work if we don't want it delegated. As for the first option, I may be letting my systemd shutdown ignorance show, but couldn't this also interfere with actual system shutdown. Accoring to the dracut-shutdown service man page: shutdown will try to umount every /oldroot mount and calls the various shutdown hooks from the dracut modules. If we are currently making certain that we disable queuing when we shut the system down, even if multipathd isn't currently running, then I feel we should probably continue to do so, just out of caution. You can override queue_without_daemon, and it's perfectly reasonable to do so in some cases, like if you are trying to update the multipath-tools binaries. If multipathd failed to restart after this, queuing would be left enabled.

[mwilck]

My previous comment about -D was premature. I double-checked, and the two dracut scripts in question just run multipath -l, which currently does't attempt to delegate the command, or access the socket. As noted in https://github.com/dracutdevs/dracut/pull/2290#issuecomment-1483939090, I don't understand how socket activation occured in the error case. It could have been multipath -u run from udev rules[^1], maybe, but does that run during shutdown?

You can override queue_without_daemon, and it's perfectly reasonable to do so in some cases, like if you are trying to update the multipath-tools binaries. If multipathd failed to restart after this, queuing would be left enabled.

OK. According to @lixiaokeng's comment above, disabling the socket in the initrd is sufficient to fix the issue. So we can leave the shutdown scripts untouched [^2].

[^1]: which doesn't care about -D [^2]: Although one might argue that turning off queue_without_daemon and then shutting down with no paths left is a corner case that qualifies as "shooting oneself in the foot", and deserves having to hit the big red button.

[lixiaokeng]

close