Merge pull request #1837 from hjelmn/rdmacm_fix

btl/openib: fix rdmacm locking bug

opal/mca/btl/openib/btl_openib_component.c

@@ -496,6 +496,8 @@ static void btl_openib_control(mca_btl_base_module_t* btl,
            trigger credit management (because the rd_credits will
            still be negative), and Bad Things will happen. */
         if (ep->endpoint_posted_recvs) {
+            /* need to hold to lock for both send_cts and connected */
+            OPAL_THREAD_LOCK(&ep->endpoint_lock);
             if (!ep->endpoint_cts_sent) {
                 mca_btl_openib_endpoint_send_cts(ep);
             }

opal/mca/btl/openib/btl_openib_endpoint.c

@@ -537,13 +537,11 @@ void mca_btl_openib_endpoint_send_cts(mca_btl_openib_endpoint_t *endpoint)
     ctl_hdr->type = MCA_BTL_OPENIB_CONTROL_CTS;
 
     /* Send the fragment */
-    OPAL_THREAD_LOCK(&endpoint->endpoint_lock);
     if (OPAL_SUCCESS != mca_btl_openib_endpoint_post_send(endpoint, sc_frag)) {
         BTL_ERROR(("Failed to post CTS send"));
         mca_btl_openib_endpoint_invoke_error(endpoint);
     }
     endpoint->endpoint_cts_sent = true;
-    OPAL_THREAD_UNLOCK(&endpoint->endpoint_lock);
 }
 
 /*
@@ -588,6 +586,9 @@ void mca_btl_openib_endpoint_cpc_complete(mca_btl_openib_endpoint_t *endpoint)
                 OPAL_OUTPUT((-1, "cpc_complete to %s -- already got CTS, so marking endpoint as complete",
                              opal_get_proc_hostname(endpoint->endpoint_proc->proc_opal)));
                 mca_btl_openib_endpoint_connected(endpoint);
+            } else {
+                /* the caller hold the lock and expects us to drop it */
+                OPAL_THREAD_UNLOCK(&endpoint->endpoint_lock);
             }
         }
 

opal/mca/btl/openib/btl_openib_endpoint.h

@@ -342,8 +342,11 @@ int mca_btl_openib_endpoint_post_send(mca_btl_openib_endpoint_t*,
 void mca_btl_openib_endpoint_send_credits(mca_btl_base_endpoint_t*, const int);
 void mca_btl_openib_endpoint_connect_eager_rdma(mca_btl_openib_endpoint_t*);
 int mca_btl_openib_endpoint_post_recvs(mca_btl_openib_endpoint_t*);
+
+/* the endpoint lock must be held with OPAL_THREAD_LOCK for both CTS and cpc complete */
 void mca_btl_openib_endpoint_send_cts(mca_btl_openib_endpoint_t *endpoint);
 void mca_btl_openib_endpoint_cpc_complete(mca_btl_openib_endpoint_t*);
+
 void mca_btl_openib_endpoint_connected(mca_btl_openib_endpoint_t*);
 void mca_btl_openib_endpoint_init(mca_btl_openib_module_t*,
                                   mca_btl_base_endpoint_t*,

opal/mca/btl/openib/connect/btl_openib_connect_rdmacm.c

@@ -1246,6 +1246,7 @@ static void *local_endpoint_cpc_complete(void *context)
 
     OPAL_OUTPUT((-1, "MAIN local_endpoint_cpc_complete to %s",
                  opal_get_proc_hostname(endpoint->endpoint_proc->proc_opal)));
+    OPAL_THREAD_LOCK(&endpoint->endpoint_lock);
     mca_btl_openib_endpoint_cpc_complete(endpoint);
 
     return NULL;