mpid/shm: Added support for moving copy buffer in MCDRAM

This commit introduces runtime control on shared memory object
placement, more specifically copy buffers used by CH3 for the
rendezvous protocol, in the heterogeneous memory system. Runtime
binding is provided through an additional CVAR:
MPIR_CVAR_SHMEM_COPYBUF_BIND_TYPE.
By default this is set to "DDR" but can be set to "MCDRAM" to move
the aforementioned object into MCDRAM.

src/mpid/ch3/channels/nemesis/src/mpid_nem_lmt_shm.c

@@ -14,6 +14,28 @@
 #define DBG_LMT(x)
 #endif
 
+/*
+=== BEGIN_MPI_T_CVAR_INFO_BLOCK ===
+
+categories:
+    - name        : SHMEM_RENDEZVOUS
+      description : >-
+        CVARs controlling shared memory object binding for Rendezvous protocol.
+
+cvars:
+    - name        : MPIR_CVAR_SHMEM_COPYBUF_BIND_TYPE
+      category    : SHMEM_RENDEZVOUS
+      type        : string
+      default     : "DDR"
+      class       : none
+      verbosity   : MPI_T_VERBOSITY_USER_BASIC
+      scope       : MPI_T_SCOPE_GROUP
+      description : >-
+        The memory type to be used for allocating memory in a KNL system.
+
+=== END_MPI_T_CVAR_INFO_BLOCK ===
+*/
+
 /* Progress queue */
 
 typedef struct lmt_shm_prog_element
@@ -179,6 +201,19 @@ int MPID_nem_lmt_shm_start_recv(MPIDI_VC_t *vc, MPIR_Request *req, MPL_IOV s_coo
         mpi_errno = MPID_nem_allocate_shm_region(&vc_ch->lmt_copy_buf, vc_ch->lmt_copy_buf_handle);
         if (mpi_errno) MPIR_ERR_POP(mpi_errno);
 
+#ifdef HAVE_HWLOC
+        /* bind shared memory to HBM */
+        if (MPIR_Process.mcdramset_is_valid &&
+            strcmp(MPIR_CVAR_SHMEM_COPYBUF_BIND_TYPE, "MCDRAM") == 0) {
+            hwloc_set_area_membind(MPIR_Process.topology, vc_ch->lmt_copy_buf,
+                                   sizeof(MPID_nem_copy_buf_t),
+                                   (MPIR_Process.bindset_is_valid) ?
+                                        MPIR_Process.mcdram_sibling : MPIR_Process.
+                                   mcdram_nodes, HWLOC_MEMBIND_BIND,
+                                   HWLOC_MEMBIND_STRICT | HWLOC_MEMBIND_BYNODESET);
+        }
+#endif
+
         vc_ch->lmt_copy_buf->sender_present.val   = 0;
         vc_ch->lmt_copy_buf->receiver_present.val = 0;
 
@@ -193,7 +228,7 @@ int MPID_nem_lmt_shm_start_recv(MPIDI_VC_t *vc, MPIR_Request *req, MPL_IOV s_coo
     /* send CTS with handle for copy buffer */
     mpi_errno = MPL_shm_hnd_get_serialized_by_ref((vc_ch->lmt_copy_buf_handle), &ser_lmt_copy_buf_handle);
     if(mpi_errno != MPI_SUCCESS) { MPIR_ERR_POP(mpi_errno); }
-    
+
     MPID_nem_lmt_send_CTS(vc, req, ser_lmt_copy_buf_handle, (int)strlen(ser_lmt_copy_buf_handle) + 1);
 
     queue_initially_empty = LMT_SHM_Q_EMPTY(vc_ch->lmt_queue) && vc_ch->lmt_active_lmt == NULL;
@@ -375,13 +410,13 @@ static int get_next_req(MPIDI_VC_t *vc)
         /* remote side chooses next transfer */
 
         OPA_read_barrier();
-        
+
         MPL_DBG_STMT(MPIDI_CH3_DBG_CHANNEL, VERBOSE, if (copy_buf->owner_info.val.remote_req_id == MPI_REQUEST_NULL)
                                                 MPL_DBG_MSG_D(MPIDI_CH3_DBG_CHANNEL, VERBOSE, "waiting for owner rank=%d", vc->pg_rank));
-            
+
         while (copy_buf->owner_info.val.remote_req_id == MPI_REQUEST_NULL) {
             int p = 0;
-            
+
             if (p == LMT_POLLS_BEFORE_GIVING_UP)
                 goto fn_exit;
             ++p;
@@ -470,7 +505,7 @@ static int lmt_shm_send_progress(MPIDI_VC_t *vc, MPIR_Request *req, int *done)
            waiting, otherwise wait for a bounded amount of time. */
         while (copy_buf->len[buf_num].val != 0) {
             int p = 0;
-            
+
             if (!copy_buf->receiver_present.val || p == LMT_POLLS_BEFORE_GIVING_UP) {
                 req->dev.segment_first = first;
                 vc_ch->lmt_buf_num = buf_num;
@@ -564,7 +599,7 @@ static int lmt_shm_recv_progress(MPIDI_VC_t *vc, MPIR_Request *req, int *done)
            waiting, otherwise wait for a bounded amount of time. */
         while ((len = copy_buf->len[buf_num].val) == 0) {
             int p = 0;
-            
+
             if (!copy_buf->sender_present.val || p == LMT_POLLS_BEFORE_GIVING_UP) {
                 req->dev.segment_first = first;
                 vc_ch->lmt_buf_num = buf_num;
@@ -727,7 +762,7 @@ static inline int lmt_shm_progress_vc(MPIDI_VC_t *vc, int *done)
         if (vc_ch->lmt_active_lmt == NULL)
         {
             /* couldn't find an appropriate request, try again later */
-            
+
             if (LMT_SHM_Q_EMPTY(vc_ch->lmt_queue))
                 *done = TRUE; /* There's nothing in the queue.  VC
                                  must have terminated */
@@ -956,7 +991,7 @@ static int MPID_nem_delete_shm_region(MPID_nem_copy_buf_t **buf_p, MPL_shm_hnd_t
     mpi_errno = MPL_shm_seg_remove(*handle_p);
     if (mpi_errno != MPI_SUCCESS) { MPIR_ERR_POP(mpi_errno); }
 
-    mpi_errno = MPID_nem_detach_shm_region(buf_p, *handle_p); 
+    mpi_errno = MPID_nem_detach_shm_region(buf_p, *handle_p);
     if (mpi_errno != MPI_SUCCESS) { MPIR_ERR_POP(mpi_errno); }
 
     mpi_errno = MPL_shm_hnd_finalize(handle_p);