[LU-14027] Client recovery statemachine hangs in recovery disconnected during lock reply

• Type: Bug
• Resolution: Fixed
• Priority: Blocker
• Fix Version/s: Lustre 2.14.0, Lustre 2.12.7
• Affects Version/s: Lustre 2.14.0, Lustre 2.12.6
• Severity: 3
• Assignee: Oleg Drokin
• Reporter: Oleg Drokin

Description

LU-13600 introduced lock ratelimiting logic, but it did not take into account that if there's a disconnection in the REPLAY_LOCKS phase then yet unsent locks get stuck in the sending queue so the replay locks thread hangs with imp_replay_inflight elevated above zero.

The direct consequence from that is recovery state machine never advances from REPLAY to REPLAY_LOCKS status when imp_replay_inflight is non zero:

        if (imp->imp_state == LUSTRE_IMP_REPLAY) {
                CDEBUG(D_HA, "replay requested by %s\n",
                       obd2cli_tgt(imp->imp_obd));
                rc = ptlrpc_replay_next(imp, &inflight);
                if (inflight == 0 &&
                    atomic_read(&imp->imp_replay_inflight) == 0) {
                        import_set_state(imp, LUSTRE_IMP_REPLAY_LOCKS);
                        rc = ldlm_replay_locks(imp);
                        if (rc)
                                GOTO(out, rc);
                }
                rc = 0;
        }

To break this we either need to check import state in the replay locks thread before attempting any sending or make sure replay_one_lock() prepares resend requests in such a state that they are never stuck.

Issue Links

• is related to: LU-16943 replay-single test_135: error: set_param: param_path 'fail_val': No such file or directory
• is related to: LU-13600 limit number of RPCs in flight during recovery