ch4/ofi: use force in allocating pipeline cells

The pipeline algorithms may deadlock when run out of cells. Cells could
be locked in send or recv when the cells in the remote process is also locked
in different send/recv and preventing completion. Use
MPIDU_genq_private_pool_force_alloc_cell to allocate new pool blocks if necessary.

The rndv read/write protocol don't have the same issue since the
operations are one-sided and the chunk buffers used in read or write
should eventually complete.

Author: hzhou

src/mpid/ch4/netmod/ofi/ofi_pipeline.c

@@ -197,8 +197,8 @@ static int pipeline_send_poll(MPIX_Async_thing thing)
         }
 
         /* alloc a chunk */
-        MPIDU_genq_private_pool_alloc_cell(MPIDI_OFI_global.per_vci[p->vci_local].pipeline_pool,
-                                           &chunk_buf);
+        MPIDU_genq_private_pool_force_alloc_cell(MPIDI_OFI_global.
+                                                 per_vci[p->vci_local].pipeline_pool, &chunk_buf);
         if (!chunk_buf) {
             goto fn_exit;
         }
@@ -371,8 +371,8 @@ static int pipeline_recv_poll(MPIX_Async_thing thing)
         }
 
         /* alloc a chunk */
-        MPIDU_genq_private_pool_alloc_cell(MPIDI_OFI_global.per_vci[p->vci_local].pipeline_pool,
-                                           &chunk_buf);
+        MPIDU_genq_private_pool_force_alloc_cell(MPIDI_OFI_global.
+                                                 per_vci[p->vci_local].pipeline_pool, &chunk_buf);
         if (!chunk_buf) {
             goto fn_exit;
         }