A first cut at the large data problem with TCP. As long

as the writev and readv support a sum larger than a uint32_t
this version will work. For the other OSes a different patch
is required. This patch is a slight modification of the one
proposed by @ggouaillardet.

Signed-off-by: George Bosilca <bosilca@icl.utk.edu>

opal/mca/btl/tcp/btl_tcp_frag.c

@@ -112,11 +112,11 @@ size_t mca_btl_tcp_frag_dump(mca_btl_tcp_frag_t* frag, char* msg, char* buf, siz
 
 bool mca_btl_tcp_frag_send(mca_btl_tcp_frag_t* frag, int sd)
 {
-    ssize_t cnt = -1;
+    ssize_t cnt;
     size_t i, num_vecs;
 
     /* non-blocking write, but continue if interrupted */
-    while(cnt < 0) {
+    do {
         cnt = writev(sd, frag->iov_ptr, frag->iov_cnt);
         if(cnt < 0) {
             switch(opal_socket_errno) {
@@ -140,7 +140,7 @@ bool mca_btl_tcp_frag_send(mca_btl_tcp_frag_t* frag, int sd)
                 return false;
             }
         }
-    }
+    } while(cnt < 0);
 
     /* if the write didn't complete - update the iovec state */
     num_vecs = frag->iov_cnt;
@@ -166,14 +166,14 @@ bool mca_btl_tcp_frag_send(mca_btl_tcp_frag_t* frag, int sd)
 bool mca_btl_tcp_frag_recv(mca_btl_tcp_frag_t* frag, int sd)
 {
     mca_btl_base_endpoint_t* btl_endpoint = frag->endpoint;
-    int i, num_vecs, dont_copy_data = 0;
     ssize_t cnt;
+    int32_t i, num_vecs, dont_copy_data = 0;
 
  repeat:
     num_vecs = frag->iov_cnt;
 #if MCA_BTL_TCP_ENDPOINT_CACHE
     if( 0 != btl_endpoint->endpoint_cache_length ) {
-        size_t length;
+        ssize_t length;
         /* It's strange at the first look but cnt have to be set to the full amount of data
          * available. After going to advance_iov_position we will use cnt to detect if there
          * is still some data pending.
@@ -208,8 +208,7 @@ bool mca_btl_tcp_frag_recv(mca_btl_tcp_frag_t* frag, int sd)
 #endif  /* MCA_BTL_TCP_ENDPOINT_CACHE */
 
     /* non-blocking read, but continue if interrupted */
-    cnt = -1;
+    do {
-    while( cnt < 0 ) {
         cnt = readv(sd, frag->iov_ptr, num_vecs);
         if( 0 < cnt ) goto advance_iov_position;
         if( cnt == 0 ) {
@@ -247,7 +246,7 @@ bool mca_btl_tcp_frag_recv(mca_btl_tcp_frag_t* frag, int sd)
             mca_btl_tcp_endpoint_close(btl_endpoint);
             return false;
         }
-    }
+    } while( cnt < 0 );
 
  advance_iov_position:
     /* if the read didn't complete - update the iovec state */

opal/mca/btl/tcp/btl_tcp_frag.h

@@ -53,8 +53,8 @@ struct mca_btl_tcp_frag_t {
     mca_btl_tcp_hdr_t hdr;
     struct iovec iov[MCA_BTL_TCP_FRAG_IOVEC_NUMBER + 1];
     struct iovec *iov_ptr;
-    size_t iov_cnt;
+    uint32_t iov_cnt;
-    size_t iov_idx;
+    uint32_t iov_idx;
     size_t size;
     uint16_t next_step;
     int rc;