[LU-20079] ldlm: kobject leak when client connect fails before obd_connect completes

• Type: Bug
• Resolution: Unresolved
• Priority: Minor
• Environment: RHEL9.7, RHEL10.1

Epic/Theme

ldlm leak,

• Severity: 3
• Assignee: Hiroshi Nishida
• Reporter: Hiroshi Nishida

Description

When a Lustre client mount fails before obd_connect() completes, cl_mgc_mgsexp stays NULL. lustre_stop_mgc(){{ skips {{obd_disconnect(), so client_disconnect_export() and ldlm_namespace_free_prior() are never called. client_obd_cleanup() then calls ldlm_namespace_free_post() directly, leaving stale sysfs kobjects under /sys/fs/lustre/ldlm/namespaces/ and /sys/fs/lustre/mgc/. Subsequent mount attempts fail with -EEXIST until reboot.

Root cause: client_connect_import() calls class_disconnect() on the error path instead of client_disconnect_export(), skipping ldlm_namespace_free_prior().

Fix: call ldlm_namespace_free_prior() explicitly in the error path of client_connect_import() after class_disconnect().

Affects: MGC, MDC, and OSC since all share client_connect_import() as their ->o_connect() method.

Verified on RHEL 9.7 and RHEL 10.1.