[LU-15137] socklnd: decrement typed connection counters on close

Related — Fri, 19 Jan 2024 15:58:15 +0000

There appears to be a possibility that if some connections are being closed while some are still being created for the same control block, the race may result in LBUG in ksocknal_connect:

if ((wanted & BIT(SOCKLND_CONN_ANY)) != 0) {
    type = SOCKLND_CONN_ANY;
} else if ((wanted & BIT(SOCKLND_CONN_CONTROL)) != 0) {
    type = SOCKLND_CONN_CONTROL;
} else if ((wanted & BIT(SOCKLND_CONN_BULK_IN)) != 0 &&
           conn_cb->ksnr_blki_conn_count <= conn_cb->ksnr_blko_conn_count) {
    type = SOCKLND_CONN_BULK_IN;
} else {
    LASSERT ((wanted & BIT(SOCKLND_CONN_BULK_OUT)) != 0);
    type = SOCKLND_CONN_BULK_OUT;
}

This may happen if the previously created BULK_IN connection got closed, no more BULK_OUT connections are needed, but ksocknal_connect is trying again to create a BULK_IN connection. Because ksnr_blki_conn_count didn't get decremented on close, the assertion will fail.

Fix this by making sure counters are decremented on close.

Whamcloud Community JIRA

[LU-15137] socklnd: decrement typed connection counters on close