Restore mca_btl_self_send() to return 1

- Remove self from alternate btls.

- Force osc/rdma to take the non btl-atomic path for self.

Signed-off-by: Austen Lauria <[email protected]>

Author: awlauria

ompi/mca/osc/rdma/osc_rdma_component.c

@@ -266,7 +266,7 @@ static int ompi_osc_rdma_component_register (void)
                                             MCA_BASE_VAR_SCOPE_GROUP, &ompi_osc_rdma_btl_names);
     free(description_str);
 
-    ompi_osc_rdma_btl_alternate_names = "sm,self,tcp";
+    ompi_osc_rdma_btl_alternate_names = "sm,tcp";
     opal_asprintf(&description_str, "Comma-delimited list of alternate BTL component names to allow without verifying "
                   "connectivity (default: %s)", ompi_osc_rdma_btl_alternate_names);
     (void) mca_base_component_var_register (&mca_osc_rdma_component.super.osc_version, "alternate_btls", description_str,
@@ -521,7 +521,7 @@ static int allocate_state_single (ompi_osc_rdma_module_t *module, void **base, s
     my_peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
     my_peer->state = (uint64_t) (uintptr_t) module->state;
 
-    if (module->use_cpu_atomics) {
+    if (module->use_cpu_atomics || my_peer->rank == my_rank) {
         /* all peers are local or it is safe to mix cpu and nic atomics */
         my_peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_STATE;
     } else {
@@ -596,7 +596,7 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
 
     if (!module->single_node) {
         for (int i = 0 ; i < module->btls_in_use ; ++i) {
-            module->use_cpu_atomics = module->use_cpu_atomics && !!(module->selected_btls[i]->btl_flags & MCA_BTL_ATOMIC_SUPPORTS_GLOB);
+            module->use_cpu_atomics = module->use_cpu_atomics && !!(module->selected_btls[i]->btl_atomic_flags & MCA_BTL_ATOMIC_SUPPORTS_GLOB);
         }
     }
 
@@ -776,7 +776,7 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
             ex_peer = (ompi_osc_rdma_peer_extended_t *) peer;
 
             /* set up peer state */
-            if (module->use_cpu_atomics) {
+            if (module->use_cpu_atomics || peer->rank == my_rank) {
                 /* all peers are local or it is safe to mix cpu and nic atomics */
                 peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_STATE;
                 peer->state = (osc_rdma_counter_t) peer_state;
@@ -805,7 +805,7 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
             ompi_osc_module_add_peer (module, peer);
 
             if (MPI_WIN_FLAVOR_DYNAMIC == module->flavor) {
-                if (module->use_cpu_atomics && peer_rank == my_rank) {
+                if (peer_rank == my_rank) {
                     peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
                 }
                 /* nothing more to do */
@@ -821,15 +821,14 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
                 ex_peer->size = temp[i].size;
             }
 
-            if (module->use_cpu_atomics && (MPI_WIN_FLAVOR_ALLOCATE == module->flavor || peer_rank == my_rank)) {
+            if (module->use_cpu_atomics && (MPI_WIN_FLAVOR_ALLOCATE == module->flavor)) {
                 /* base is local and cpu atomics are available */
                 if (MPI_WIN_FLAVOR_ALLOCATE == module->flavor) {
                     ex_peer->super.base = (uintptr_t) module->segment_base + offset;
                 } else {
                     ex_peer->super.base = (uintptr_t) *base;
                 }
 
-                peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
                 offset += temp[i].size;
             } else {
                 ex_peer->super.base = peer_region->base;
@@ -838,6 +837,10 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
                     ex_peer->super.base_handle = (mca_btl_base_registration_handle_t *) peer_region->btl_handle_data;
                 }
             }
+
+            if(my_rank == peer_rank) {
+                peer->flags |= OMPI_OSC_RDMA_PEER_LOCAL_BASE;
+            }
         }
     } while (0);
 

ompi/mca/osc/rdma/osc_rdma_lock.h

@@ -86,10 +86,12 @@ static inline int ompi_osc_rdma_btl_fop (ompi_osc_rdma_module_t *module, uint8_t
 
     if (OPAL_SUCCESS != ret) {
         if (OPAL_LIKELY(1 == ret)) {
-            *result = ((int64_t *) pending_op->op_buffer)[0];
             ret = OMPI_SUCCESS;
-            ompi_osc_rdma_atomic_complete (selected_btl, endpoint, pending_op->op_buffer,
-                                           pending_op->op_frag->handle, (void *) pending_op, NULL, OPAL_SUCCESS);
+            if(false == pending_op->op_complete) {
+                *result = ((int64_t *) pending_op->op_buffer)[0];
+                ompi_osc_rdma_atomic_complete (selected_btl, endpoint, pending_op->op_buffer,
+                                               pending_op->op_frag->handle, (void *) pending_op, NULL, OPAL_SUCCESS);
+            }
         } else {
             /* need to release here because ompi_osc_rdma_atomic_complete was not called */
             OBJ_RELEASE(pending_op);
@@ -161,8 +163,10 @@ static inline int ompi_osc_rdma_btl_op (ompi_osc_rdma_module_t *module, uint8_t
         /* need to release here because ompi_osc_rdma_atomic_complete was not called */
         OBJ_RELEASE(pending_op);
         if (OPAL_LIKELY(1 == ret)) {
-            if (cbfunc) {
-                cbfunc (cbdata, cbcontext, OMPI_SUCCESS);
+            if(false == pending_op->op_complete) {
+                if (cbfunc) {
+                    cbfunc (cbdata, cbcontext, OMPI_SUCCESS);
+                }
             }
             ret = OMPI_SUCCESS;
         }

ompi/mca/osc/rdma/osc_rdma_peer.c

@@ -71,6 +71,7 @@ static int ompi_osc_rdma_peer_btl_endpoint (struct ompi_osc_rdma_module_t *modul
         }
     }
 
+    OSC_RDMA_VERBOSE(MCA_BASE_VERBOSE_DEBUG, "WARNING - couldn't find a connection");
     /* unlikely but can happen when creating a peer for self */
     return OMPI_ERR_UNREACH;
 }

opal/mca/btl/base/btl_base_am_rdma.c

@@ -614,7 +614,7 @@ static int mca_btl_base_am_rdma_respond(mca_btl_base_module_t *btl,
     send_descriptor->des_cbfunc = NULL;
 
     int ret = btl->btl_send(btl, endpoint, send_descriptor, mca_btl_base_rdma_resp_tag());
-    if (OPAL_UNLIKELY(OPAL_SUCCESS != ret)) {
+    if (OPAL_UNLIKELY(ret < 0)) {
         *descriptor = send_descriptor;
     }
     return ret;
@@ -635,7 +635,7 @@ mca_btl_base_am_rmda_rdma_complete(mca_btl_base_module_t *btl,
     operation->is_completed = true;
     int ret = mca_btl_base_am_rdma_respond(operation->btl, operation->endpoint,
                                            &operation->descriptor, NULL, &operation->hdr);
-    if (OPAL_UNLIKELY(OPAL_SUCCESS != ret)) {
+    if (OPAL_UNLIKELY(ret < 0)) {
         BTL_VERBOSE(
             ("could not send a response. queueing the response for later. endpoint=%p, ret=%d",
              endpoint, ret));
@@ -781,7 +781,7 @@ static int mca_btl_base_am_rdma_progress(void)
             int ret = descriptor->btl->btl_send(descriptor->btl, descriptor->endpoint,
                                                 descriptor->descriptor,
                                                 mca_btl_base_rdma_tag(context->type));
-            if (OPAL_SUCCESS == ret) {
+            if (ret <= 1) {
                 opal_list_remove_item(&default_module.queued_initiator_descriptors,
                                       &descriptor->super);
             }
@@ -932,7 +932,7 @@ static void mca_btl_base_am_process_rdma(mca_btl_base_module_t *btl,
         abort();
     }
 
-    if (OPAL_SUCCESS != ret) {
+    if (ret < 0) {
         mca_btl_base_rdma_queue_operation(btl, desc->endpoint, descriptor, 0, hdr, operation);
     }
 }
@@ -995,7 +995,7 @@ static void mca_btl_base_am_process_atomic(mca_btl_base_module_t *btl,
 
     mca_btl_base_descriptor_t *descriptor = NULL;
     int ret = mca_btl_base_am_rdma_respond(btl, desc->endpoint, &descriptor, &atomic_response, hdr);
-    if (OPAL_SUCCESS != ret) {
+    if (ret < 0) {
         mca_btl_base_rdma_queue_operation(btl, desc->endpoint, descriptor, atomic_response, hdr,
                                           NULL);
     }

opal/mca/btl/self/btl_self.c

@@ -218,7 +218,7 @@ static int mca_btl_self_send(struct mca_btl_base_module_t *btl,
     if (btl_ownership) {
         mca_btl_self_free(btl, des);
     }
-    return OPAL_SUCCESS;
+    return 1;
 }
 
 static int mca_btl_self_sendi(struct mca_btl_base_module_t *btl,