coll/han/alltoallv: correct loop condition on empty sends

This change makes the loop condition more straightforward, and
corrects a logic error that was previously seen when a rank had 0
bytes to receive.

With the new logic the first pass through the receive-positing section
will walk through all datatypes, and never post a recv buffer.  Sender
logic remains unchanged.

Author: lrbison

ompi/mca/coll/han/coll_han_alltoallv.c

@@ -270,6 +270,7 @@ static int alltoallv_sendrecv_w(
     size_t nbytes_pack;
 
     int nsend_req_pending = 0;
+    int nrecv_req_pending = 0;
     opal_convertor_t send_convertor;
     opal_convertor_t recv_convertor;
     OBJ_CONSTRUCT( &send_convertor, opal_convertor_t );
@@ -323,7 +324,7 @@ static int alltoallv_sendrecv_w(
         int ii_more_sends_to_complete = nsend_req_pending > 0;
 
         int ii_more_recvs_to_post = recv_post_remaining_bytes > 0;
-        int ii_more_recvs_to_complete = recv_convertor_bytes_remaining > 0 || jtype_recv < ntypes_recv;
+        int ii_more_recvs_to_complete = nrecv_req_pending > 0;
 
 
         if ( !( ii_more_sends_to_post || ii_more_sends_to_complete ||
@@ -336,10 +337,22 @@ static int alltoallv_sendrecv_w(
             /* Common Case: */
             /* wait for any send or recv to complete */
             rc = ompi_request_wait_any(nreqs, requests, &jreq, MPI_STATUS_IGNORE);
-            if (rc != 0 || jreq == MPI_UNDEFINED) {
-                return 1;
+            if (rc != 0) {
+                opal_output_verbose(1, mca_coll_han_component.han_output,
+                    "ompi_request_wait_any returned error code %d in alltoallv_sendrecv_w (loop=%ld)\n",rc,jloop);
+                return rc;
             }
             have_completion = 1;
+            if (jreq == MPI_UNDEFINED) {
+                opal_output_verbose(1, mca_coll_han_component.han_output,
+                    "ERROR: no active requests to wait on! Loop=%ld: %d %d %d %d\n",
+                    jloop,
+                    ii_more_sends_to_post, ii_more_sends_to_complete,
+                    ii_more_recvs_to_post, ii_more_recvs_to_complete );
+                have_completion = 0;
+                jreq = jloop % nbufs;
+                return MPI_ERR_INTERN;
+            }
         } else {
             /* priming the loop: post sends or recvs while have_completion=0.
 
@@ -401,6 +414,7 @@ static int alltoallv_sendrecv_w(
 
         } else { /* recv request */
             if (have_completion) {
+                nrecv_req_pending--;
                 /* unpack data */
                 ssize_t buf_remain = buf_len;
                 size_t buf_converted = 0;
@@ -449,6 +463,7 @@ static int alltoallv_sendrecv_w(
 
                 /* update posted_recv_bytes */
                 recv_post_remaining_bytes -= bytes_to_post;
+                nrecv_req_pending++;
             } else {
                 requests[jreq] = MPI_REQUEST_NULL;
             }