Merge pull request #6687 from yosefe/topic/osc-ucx-fix-ud-self-deadlock

OSC/UCX: Fix deadlock with atomic lock

Author: yosefe

ompi/mca/osc/ucx/osc_ucx_active_target.c

@@ -272,6 +272,7 @@ int ompi_osc_ucx_post(struct ompi_group_t *group, int assert, struct ompi_win_t
                     ompi_osc_ucx_handle_incoming_post(module, &(module->state.post_state[j]), NULL, 0);
                 }
 
+                ucp_worker_progress(mca_osc_ucx_component.wpool->dflt_worker);
                 usleep(100);
             } while (1);
         }

ompi/mca/osc/ucx/osc_ucx_comm.c

@@ -240,7 +240,7 @@ static inline int start_atomicity(ompi_osc_ucx_module_t *module, int target) {
     uint64_t remote_addr = (module->state_addrs)[target] + OSC_UCX_STATE_ACC_LOCK_OFFSET;
     int ret = OMPI_SUCCESS;
 
-    while (result_value != TARGET_LOCK_UNLOCKED) {
+    for (;;) {
         ret = opal_common_ucx_wpmem_cmpswp(module->state_mem,
                                          TARGET_LOCK_UNLOCKED, TARGET_LOCK_EXCLUSIVE,
                                          target, &result_value, sizeof(result_value),
@@ -249,9 +249,12 @@ static inline int start_atomicity(ompi_osc_ucx_module_t *module, int target) {
             OSC_UCX_VERBOSE(1, "opal_common_ucx_mem_cmpswp failed: %d", ret);
             return OMPI_ERROR;
         }
-    }
+        if (result_value == TARGET_LOCK_UNLOCKED) {
+            return OMPI_SUCCESS;
+        }
 
-    return ret;
+        ucp_worker_progress(mca_osc_ucx_component.wpool->dflt_worker);
+    }
 }
 
 static inline int end_atomicity(ompi_osc_ucx_module_t *module, int target) {

ompi/mca/osc/ucx/osc_ucx_passive_target.c

@@ -42,6 +42,7 @@ static inline int start_shared(ompi_osc_ucx_module_t *module, int target) {
         } else {
             break;
         }
+        ucp_worker_progress(mca_osc_ucx_component.wpool->dflt_worker);
     }
 
     return ret;
@@ -58,17 +59,20 @@ static inline int start_exclusive(ompi_osc_ucx_module_t *module, int target) {
     uint64_t remote_addr = (module->state_addrs)[target] + OSC_UCX_STATE_LOCK_OFFSET;
     int ret = OMPI_SUCCESS;
 
-    while (result_value != TARGET_LOCK_UNLOCKED) {
+    for (;;) {
         ret = opal_common_ucx_wpmem_cmpswp(module->state_mem,
                                          TARGET_LOCK_UNLOCKED, TARGET_LOCK_EXCLUSIVE,
                                          target, &result_value, sizeof(result_value),
                                          remote_addr);
         if (OMPI_SUCCESS != ret) {
             return ret;
         }
-    }
+        if (result_value == TARGET_LOCK_UNLOCKED) {
+            return OMPI_SUCCESS;
+        }
 
-    return ret;
+        ucp_worker_progress(mca_osc_ucx_component.wpool->dflt_worker);
+    }
 }
 
 static inline int end_exclusive(ompi_osc_ucx_module_t *module, int target) {