WIP Switch back to TCP when RDMA send fails (does not work)

src/client/messenger.cpp

@@ -192,9 +192,9 @@ void osd_messenger_t::init()
             auto cl = cl_it->second;
             cl_it++;
             if (!cl->osd_num && !cl->in_osd_num ||
-                cl->peer_state == PEER_CONNECTING ||
-                cl->peer_state == PEER_RDMA_CONNECTING_IN ||
-                cl->peer_state == PEER_STOPPED)
+                cl->peer_state != PEER_CONNECTED &&
+                cl->peer_state != PEER_RDMA_CONNECTING_OUT &&
+                cl->peer_state != PEER_RDMA)
             {
                 // Do not run keepalive on regular and unconnected clients
                 continue;
@@ -204,12 +204,23 @@ void osd_messenger_t::init()
                 cl->ping_time_remaining--;
                 if (!cl->ping_time_remaining)
                 {
-                    // Ping timed out, stop the client
+                    // Ping timed out
+                    if (cl->peer_state == PEER_RDMA_CONNECTING_OUT && cl->peer_fd >= 0)
+                    {
+                        fprintf(stderr, "Ping timed out for OSD %ju over RDMA (client %ju), switching back to TCP\n",
+                            cl->in_osd_num ? cl->in_osd_num : cl->osd_num, cl->client_id);
+                        delete cl->rdma_conn;
+                        cl->rdma_conn = NULL;
+                        cl->peer_state = PEER_CONNECTED;
+                    }
+                    else
+                    {
                         fprintf(stderr, "Ping timed out for OSD %ju (client %ju), disconnecting peer\n",
                             cl->in_osd_num ? cl->in_osd_num : cl->osd_num, cl->client_id);
                         clients_to_stop.push_back(cl->client_id);
                     }
                 }
+            }
             else if (cl->idle_time_remaining > 0)
             {
                 cl->idle_time_remaining--;
@@ -241,9 +252,19 @@ void osd_messenger_t::init()
                         if (failed)
                         {
                             auto fail_osd_num = cl->in_osd_num ? cl->in_osd_num : cl->osd_num;
+                            if (cl->peer_state == PEER_RDMA_CONNECTING_OUT && cl->peer_fd >= 0)
+                            {
+                                fprintf(stderr, "Ping failed for OSD %ju over RDMA (client %ju), switching back to TCP\n", fail_osd_num, cl->client_id);
+                                delete cl->rdma_conn;
+                                cl->rdma_conn = NULL;
+                                cl->peer_state = PEER_CONNECTED;
+                            }
+                            else
+                            {
                                 fprintf(stderr, "Ping failed for OSD %ju (client %ju), disconnecting peer\n", fail_osd_num, cl->client_id);
                                 stop_client(cl->client_id);
                             }
+                        }
                     };
                     cl->ping_time_remaining = osd_ping_timeout;
                     cl->idle_time_remaining = osd_idle_timeout;
@@ -830,6 +851,11 @@ void osd_messenger_t::check_peer_config(osd_client_t *cl)
                 cl->peer_state = (cl->rdma_close_tcp ? PEER_RDMA_CONNECTING_OUT : PEER_RDMA);
                 // Add the initial receive request
                 init_recv_rdma(cl);
+                // Check the connection by using a ping
+                cl->ping_time_remaining = osd_ping_timeout;
+                cl->idle_time_remaining = 0;
+                delete op;
+                return;
             }
         }
 #endif

src/client/msgr_rdma.cpp

@@ -643,6 +643,7 @@ static void try_recv_rdma_wr(osd_client_t *cl, void *buf)
 
 bool osd_messenger_t::init_recv_rdma(osd_client_t *cl)
 {
+    return true;
     auto rc = cl->rdma_conn;
     assert(!rc->recv_buf.buf);
     rc->recv_buf.buf = (uint8_t*)malloc_or_die(rc->max_msg * rc->max_recv);
@@ -698,6 +699,11 @@ void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
                 continue;
             }
             auto rc = cl->rdma_conn;
+            if (!rc)
+            {
+                // Connection destroyed (fallback to TCP)
+                continue;
+            }
             if (wc[i].status != IBV_WC_SUCCESS)
             {
                 fprintf(stderr, "RDMA work request (%s) failed for client %ju", is_send ? "send" : "recv", client_id);
@@ -727,6 +733,11 @@ void osd_messenger_t::handle_rdma_events(msgr_rdma_context_t *rdma_context)
                         close(cl->peer_fd);
                         cl->peer_fd = -1;
                     }
+                    if (cl->osd_num)
+                    {
+                        osd_peers[cl->osd_num] = cl;
+                        on_connect_peer(cl->osd_num, 0, cl->client_id);
+                    }
                 }
                 rc->cur_recv--;
                 if (!handle_read_buffer(cl, rc->recv_buffers[rc->next_recv_buf], wc[i].byte_len))

src/client/msgr_receive.cpp

@@ -489,6 +489,15 @@ void osd_messenger_t::handle_read(int result, osd_client_t *cl)
     {
         return;
     }
+    if (cl->peer_state == PEER_RDMA_CONNECTING_IN || cl->peer_state == PEER_RDMA_CONNECTING_OUT)
+    {
+        // Data received over TCP, abort RDMA connection
+        fprintf(stderr, "Data received for RDMA client %ju over TCP, falling back to TCP\n", cl->client_id);
+        delete cl->rdma_conn;
+        cl->rdma_conn = NULL;
+        cl->peer_state = PEER_CONNECTED;
+        assert(cl->peer_fd >= 0);
+    }
     if (cl->peer_state == PEER_STOPPED)
     {
         if (cl->refs <= 0)