[LU-16990] Use NETWORK_TIMEOUT message status for TN_EVENT_TAG_RX_CANCEL

Tue, 22 Aug 2023 06:41:12 +0000

I noticed that after the MDS hit Inbound wait, it was failing transactions with REMOTE_TIMEOUT health status:

[Wed Mar  8 11:26:29 2023] LNet: 143125:0:(lib-msg.c:814:lnet_health_check()) Message from 17@kfi to 0@kfi exceeded message deadline by 2 seconds rc=-110 hstatus=REMOTE_TIMEOUT

Comes from this code:

static int kfilnd_tn_state_wait_timeout_tag_comp(struct kfilnd_transaction *tn,
                                                 enum tn_events event,
                                                 int status, bool *tn_released)
{
        KFILND_TN_DEBUG(tn, "%s event status %d", tn_event_to_str(event),
                        status);

        switch (event) {
        case TN_EVENT_TAG_RX_CANCEL:
                kfilnd_tn_status_update(tn, -ETIMEDOUT,
                                        LNET_MSG_STATUS_REMOTE_TIMEOUT);
                kfilnd_peer_tn_failed(tn->tn_kp, -ETIMEDOUT);
                break;

Current behavior means that the server never realizes its local NIC is not healthy:

[root@s-lmo-gaz38a ~]# lnetctl net show -v 2 | grep -e nid -e health
        - nid: 0@lo
          health stats:
              health value: 1000
        - nid: 172.18.2.3@tcp
          health stats:
              health value: 1000
        - nid: 17@kfi
          health stats:
              health value: 1000
[root@s-lmo-gaz38a ~]#

Whamcloud Community JIRA

[LU-16990] Use NETWORK_TIMEOUT message status for TN_EVENT_TAG_RX_CANCEL