ch4/init: Wait for ops on init comm to complete

If the special init comm used for roots only address exchange still has pending operations at destroy time, it could cause an assertion failure during MPI_INIT. Instead, we should wait for pending ops to complete to avoid a crash. Fixes pmodels#7200.
raffenet · Nov 7, 2024 · f50f7f1 · f50f7f1
1 parent 0afd4f5
commit f50f7f1
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 4 deletions.
diff --git a/src/mpid/ch4/include/mpidch4.h b/src/mpid/ch4/include/mpidch4.h
@@ -474,6 +474,6 @@ extern MPL_dbg_class MPIDI_CH4_DBG_MEMORY;
 
 /* routines only used during init */
 int MPIDI_create_init_comm(MPIR_Comm ** comm_ptr);
-void MPIDI_destroy_init_comm(MPIR_Comm ** comm_ptr);
+int MPIDI_destroy_init_comm(MPIR_Comm ** comm_ptr);
 
 #endif /* MPIDCH4_H_INCLUDED */
diff --git a/src/mpid/ch4/netmod/ofi/init_addrxchg.c b/src/mpid/ch4/netmod/ofi/init_addrxchg.c
@@ -180,7 +180,7 @@ int MPIDI_OFI_addr_exchange_root_ctx(void)
 
   fn_exit:
     if (init_comm && !mpi_errno) {
-        MPIDI_destroy_init_comm(&init_comm);
+        mpi_errno = MPIDI_destroy_init_comm(&init_comm);
     }
     return mpi_errno;
   fn_fail:

diff --git a/src/mpid/ch4/netmod/ucx/ucx_init.c b/src/mpid/ch4/netmod/ucx/ucx_init.c
@@ -140,7 +140,7 @@ static int initial_address_exchange(void)
 
   fn_exit:
     if (init_comm && !mpi_errno) {
-        MPIDI_destroy_init_comm(&init_comm);
+        mpi_errno = MPIDI_destroy_init_comm(&init_comm);
     }
     return mpi_errno;
   fn_fail:

diff --git a/src/mpid/ch4/src/init_comm.c b/src/mpid/ch4/src/init_comm.c
@@ -60,18 +60,24 @@ int MPIDI_create_init_comm(MPIR_Comm ** comm)
     goto fn_exit;
 }
 
-void MPIDI_destroy_init_comm(MPIR_Comm ** comm_ptr)
+int MPIDI_destroy_init_comm(MPIR_Comm ** comm_ptr)
 {
+    int mpi_errno = MPI_SUCCESS;
     int in_use;
     MPIR_Comm *comm = NULL;
     if (*comm_ptr != NULL) {
         comm = *comm_ptr;
         MPIDIU_release_lut(MPIDI_COMM(comm, map).irreg.lut.t);
         MPIDIG_destroy_comm(comm);
+        /* wait for any outstanding requests to complete */
+        MPIDIU_PROGRESS_WHILE(MPIR_Object_get_ref(comm) > 1, 0);
         MPIR_Object_release_ref(comm, &in_use);
         MPIR_Assertp(in_use == 0);
         MPII_COMML_FORGET(comm);
         MPIR_Handle_obj_free(&MPIR_Comm_mem, comm);
         *comm_ptr = NULL;
     }
+
+  fn_fail:
+    return mpi_errno;
 }