As Andrew Friedley pointed, my previous patch may cause deadlock if

mca_btl_openib_endpoint_connect_eager_rdma() is called recursively. He also
noticed that orte_pointer_array_add() can't fail because we allocate max number
of elements at init time. So just remove error handling and locking. No locking
 - no deadlocks.

This commit was SVN r12388.

ompi/mca/btl/openib/btl_openib.h

@@ -203,7 +203,6 @@ struct mca_btl_openib_module_t {
 
     size_t eager_rdma_frag_size; /**< length of eager frag */
     orte_pointer_array_t *eager_rdma_buffers; /**< RDMA buffers to poll */
-    opal_mutex_t eager_rdma_buffres_lock; /**< should be held while adding new rdma buffer */
     volatile int32_t eager_rdma_buffers_count; /**< number of RDMA buffers */
 
     mca_btl_base_module_error_cb_fn_t error_cb; /**< error handler */

ompi/mca/btl/openib/btl_openib_component.c

@@ -689,7 +689,6 @@ btl_openib_component_init(int *num_btl_modules,
                 mca_btl_openib_component.max_eager_rdma, 
                 0);
         openib_btl->eager_rdma_buffers_count = 0;
-        OBJ_CONSTRUCT(&openib_btl->eager_rdma_buffres_lock, opal_mutex_t);
 
         orte_pointer_array_init(&openib_btl->endpoints, 10, INT_MAX, 100);
         btls[i] = &openib_btl->super;

ompi/mca/btl/openib/btl_openib_endpoint.c

@@ -1228,24 +1228,16 @@ void mca_btl_openib_endpoint_connect_eager_rdma(
     opal_atomic_cmpset_ptr(&endpoint->eager_rdma_local.base.pval, (void*)1,
             buf);
 
-    /* lock is held during eager_rdma_buffers update in order to prevent hole
-     * in the array if mca_btl_openib_endpoint_send_eager_rdma() fails and
-     * another thread was creating eager RDMA buffer for another endpoint and
-     * allocated array index bigger then ours */
-    OPAL_THREAD_LOCK(&openib_btl->eager_rdma_buffres_lock);
-    if(orte_pointer_array_add(&index, openib_btl->eager_rdma_buffers, endpoint)
-            != ORTE_SUCCESS)
-	   goto cleanup;
-
     if(mca_btl_openib_endpoint_send_eager_rdma(endpoint) == 0) {
+        /* This can never fail because max number of entries allocated
+         * at init time */
+        orte_pointer_array_add(&index, openib_btl->eager_rdma_buffers,
+                endpoint);
         /* from this point progress function starts to poll new buffer */
         OPAL_THREAD_ADD32(&openib_btl->eager_rdma_buffers_count, 1);
-        OPAL_THREAD_UNLOCK(&openib_btl->eager_rdma_buffres_lock);
         return;
     }
 
-    orte_pointer_array_set_item(openib_btl->eager_rdma_buffers, index, NULL);
-    OPAL_THREAD_UNLOCK(&openib_btl->eager_rdma_buffres_lock);
 cleanup:
     openib_btl->super.btl_mpool->mpool_free(openib_btl->super.btl_mpool,
            buf, (mca_mpool_base_registration_t*)endpoint->eager_rdma_local.reg);