LU-12249 lnet: fix list corruption

In shutdown the resend queues are cleared and freed. The monitor
thread state is set to shutdown. It is possible to get lnet_finalize()
called after the queues are freed. The code checks for ln_state to see
if we're shutting down. But in this case we should really be checking
ln_mt_state. The monitor thread is the one that matters in this case,
because it's the one which allocates and frees the resend queues.

Lustre-change: https://review.whamcloud.com/34778
Lustre-commit: d799ac910cd6c980b40c81b76eaefb65b88904d0

Signed-off-by: Amir Shehata <ashehata@whamcloud.com>
Change-Id: Ia077cec7a52ef5cd2e1b231437c6265ba9416b1b
Reviewed-by: Olaf Weber <olaf.weber@hpe.com>
Reviewed-by: Sebastien Buisson <sbuisson@ddn.com>
Reviewed-by: Chris Horn <hornc@cray.com>
Signed-off-by: Minh Diep <mdiep@whamcloud.com>
Reviewed-on: https://review.whamcloud.com/36035
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

diff --git a/lnet/lnet/lib-move.c b/lnet/lnet/lib-move.c
index 74e8ba7..b8b8434 100644 (file)
--- a/lnet/lnet/lib-move.c
+++ b/lnet/lnet/lib-move.c
@@ -3386,7 +3386,9 @@ lnet_monitor_thread(void *arg)
        lnet_prune_rc_data(1);
 
        /* Shutting down */
+       lnet_net_lock(LNET_LOCK_EX);
        the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
+       lnet_net_unlock(LNET_LOCK_EX);
 
        /* signal that the monitor thread is exiting */
        up(&the_lnet.ln_mt_signal);
@@ -3608,7 +3610,9 @@ int lnet_monitor_thr_start(void)
 
        sema_init(&the_lnet.ln_mt_signal, 0);
 
+       lnet_net_lock(LNET_LOCK_EX);
        the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
+       lnet_net_unlock(LNET_LOCK_EX);
        task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
        if (IS_ERR(task)) {
                rc = PTR_ERR(task);
@@ -3622,13 +3626,17 @@ int lnet_monitor_thr_start(void)
        return 0;
 
 clean_thread:
+       lnet_net_lock(LNET_LOCK_EX);
        the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
+       lnet_net_unlock(LNET_LOCK_EX);
        /* block until event callback signals exit */
        down(&the_lnet.ln_mt_signal);
        /* clean up */
        lnet_router_cleanup();
 free_mem:
+       lnet_net_lock(LNET_LOCK_EX);
        the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
+       lnet_net_unlock(LNET_LOCK_EX);
        lnet_rsp_tracker_clean();
        lnet_clean_local_ni_recoveryq();
        lnet_clean_peer_ni_recoveryq();
@@ -3652,7 +3660,9 @@ void lnet_monitor_thr_stop(void)
                return;
 
        LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
+       lnet_net_lock(LNET_LOCK_EX);
        the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
+       lnet_net_unlock(LNET_LOCK_EX);
 
        /* tell the monitor thread that we're shutting down */
        wake_up(&the_lnet.ln_mt_waitq);

diff --git a/lnet/lnet/lib-msg.c b/lnet/lnet/lib-msg.c
index 959e995..92bcd8c 100644 (file)
--- a/lnet/lnet/lib-msg.c
+++ b/lnet/lnet/lib-msg.c
@@ -600,7 +600,7 @@ lnet_health_check(struct lnet_msg *msg)
        bool lo = false;
 
        /* if we're shutting down no point in handling health. */
-       if (the_lnet.ln_state != LNET_STATE_RUNNING)
+       if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
                return -1;
 
        LASSERT(msg->msg_txni);
@@ -714,6 +714,12 @@ resend:
 
        lnet_net_lock(msg->msg_tx_cpt);
 
+       /* check again under lock */
+       if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
+               lnet_net_unlock(msg->msg_tx_cpt);
+               return -1;
+       }
+
        /*
         * remove message from the active list and reset it in preparation
         * for a resend. Two exception to this