osc/rdma: ensure fragment headers and the packed datatype are 8-byte aligned.

The datatype unpacking code assumes that the packed datatype buffer has the
same alignment as an OPAL_PTRDIFF_TYPE. This was not enforced by the rdma
one-sided component. I changed the ordering and sized of various osc/rdma
headers to ensure their sizes are a multiple of 8-bytes and modified the
fragment allocation call to ensure all headers are 8-byte aligned. While
not the cleanest way to handle this situation it should resolve the issue.

Fixes trac:4315

cmr=v1.7.5:reviewer=jsquyres

This commit was SVN r30974.

The following Trac tickets were found above:
  Ticket 4315 --> https://svn.open-mpi.org/trac/ompi/ticket/4315

ompi/mca/osc/rdma/osc_rdma_data_move.c

@@ -1360,7 +1360,9 @@ static inline int process_frag (ompi_osc_rdma_module_t *module,
             abort(); /* FIX ME */
         }
 
-        header = (ompi_osc_rdma_header_t *) ((uintptr_t) header + ret);
+        /* the next header will start on an 8-byte boundary. this is done to ensure
+         * that the packed datatype is properly aligned */
+        header = (ompi_osc_rdma_header_t *) (((uintptr_t) header + ret + 0x07) & ~0x07);
     }
 
     return OMPI_SUCCESS;

ompi/mca/osc/rdma/osc_rdma_frag.h

@@ -54,6 +54,9 @@ static inline int ompi_osc_rdma_frag_alloc(ompi_osc_rdma_module_t *module, int t
     ompi_osc_rdma_frag_t *curr = module->peers[target].active_frag;
     int ret;
 
+    /* ensure the next fragment will be allocated on an 8-byte boundary */
+    request_len = (request_len + 0x7) & ~0x7;
+
     if (request_len > mca_osc_rdma_component.buffer_size) {
         return OMPI_ERR_OUT_OF_RESOURCE;
     }
@@ -114,18 +117,19 @@ static inline int ompi_osc_rdma_frag_alloc(ompi_osc_rdma_module_t *module, int t
 /*
  * Note: module lock must be held for this operation
  */
-static inline int
-ompi_osc_rdma_frag_finish(ompi_osc_rdma_module_t *module,
-                            ompi_osc_rdma_frag_t* buffer)
+static inline int ompi_osc_rdma_frag_finish(ompi_osc_rdma_module_t *module,
+                                            ompi_osc_rdma_frag_t* buffer)
 {
-    int ret = OMPI_SUCCESS;
-
-    buffer->pending--;
-    if (0 == buffer->pending && 0 == buffer->remain_len) {
-        ret = ompi_osc_rdma_frag_start(module, buffer);
+    if (0 == --buffer->pending && 0 == buffer->remain_len) {
+        if (OPAL_LIKELY(buffer == module->peers[buffer->target].active_frag)) {
+            /* this is the active fragment. need to set the current fragment to null
+             * or it will be started multiple times */
+            module->peers[buffer->target].active_frag = NULL;
+        }
+        return ompi_osc_rdma_frag_start(module, buffer);
     }
 
-    return ret;
+    return OMPI_SUCCESS;
 }
 
 #endif

ompi/mca/osc/rdma/osc_rdma_header.h

@@ -78,9 +78,9 @@ struct ompi_osc_rdma_header_acc_t {
 
     uint16_t tag;
     uint32_t count;
+    uint32_t op;
     uint64_t len;
     uint64_t displacement;
-    uint32_t op;
 };
 typedef struct ompi_osc_rdma_header_acc_t ompi_osc_rdma_header_acc_t;
 
@@ -105,9 +105,9 @@ struct ompi_osc_rdma_header_get_acc_t {
 
     uint16_t tag;
     uint32_t count;
+    uint32_t op;
     uint64_t len;
     uint64_t displacement;
-    uint32_t op;
 };
 typedef struct ompi_osc_rdma_header_get_acc_t ompi_osc_rdma_header_get_acc_t;
 
@@ -172,6 +172,7 @@ struct ompi_osc_rdma_frag_header_t {
     uint16_t windx; /* cid of communicator backing window (our window id) */
     uint32_t source; /* rank in window of source process */
     uint16_t num_ops; /* number of operations in this buffer */
+    uint16_t pad[3]; /* ensure the fragment header is a multiple of 8 bytes */
 };
 typedef struct ompi_osc_rdma_frag_header_t ompi_osc_rdma_frag_header_t;