[LU-9063] hsm: race on the coordinator's state

• Type: Bug
• Resolution: Unresolved
• Priority: Minor
• Labels: HSM patch
• Severity: 3
• Assignee: Hongchao Zhang
• Reporter: CEA

Description

There is nothing to prevent a coordinator from restarting after it was shut down. This is a problem at cleanup time for example when unmounting an mdt.

The race is a bit difficult to trigger. The simplest way to do it is to add a delay in mdt_hsm_cdt_stop() right after the coordinator's state is set to CDT_STOPPED.

After applying the patch (cf. attachment), concurrently run:

while [ $(lctl get_param -n mdt.lustre-MDT0000.hsm_control) != "stopped" ]; do
    sleep 1
done
lctl set_param mdt.lustre-MDT0000.hsm_control enabled

and

lustre/tests/llmountcleanup.sh

This should trigger the following:
kernel:LustreError: 20570:0:(mdt_coordinator.c:391:hsm_cdt_procfs_fini()) ASSERTION( cdt->cdt_state == CDT_STOPPED ) failed:
kernel:LustreError: 20570:0:(mdt_coordinator.c:391:hsm_cdt_procfs_fini()) LBUG

Comments

[Cory Spitz]

And I'm guessing that https://review.whamcloud.com/c/fs/lustre-release/+/25170 can be abandoned.

[Cory Spitz]

Seems that this is an old issue that can be resolved. Quentin B. was probably right.

[Quentin Bouget (Inactive)]

I think this patch would be a fix: https://review.whamcloud.com/#/c/22667

[Gerrit Updater]

Vinayak ([e-mail]) uploaded a new patch: https://review.whamcloud.com/25269
Subject: LU-9063 tests: patch to create the race
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 3f97bdee4ae56a2e89c6a6eff13ba2aff52bc36b

[Gerrit Updater]

Vinayak ([e-mail]) uploaded a new patch: https://review.whamcloud.com/25170
Subject: LU-9063 hsm: protect cdt_state with mutex lock
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 52e08b6ec58240ce6f88f4ddbcdb3dd4216797bf

[Peter Jones]

Hongchao

Could you please advise on this one?

Thanks

Peter