Set a hard limit on the TCP max fragment size.

Some OSes have hardcoded limits to prevent overflowing over an int32_t.
We can either detect this at configure (which might be a nicer but
incomplete solution), or always force the pipelined protocol over TCP.
As it only covers data larger than 1GB, no performance penalty is to be
expected.

Signed-off-by: George Bosilca <[email protected]>

Author: bosilca

opal/mca/btl/tcp/btl_tcp_component.c

@@ -303,7 +303,9 @@ static int mca_btl_tcp_component_register(void)
     mca_btl_tcp_module.super.btl_rndv_eager_limit = 64*1024;
     mca_btl_tcp_module.super.btl_max_send_size = 128*1024;
     mca_btl_tcp_module.super.btl_rdma_pipeline_send_length = 128*1024;
-    mca_btl_tcp_module.super.btl_rdma_pipeline_frag_size = INT_MAX;
+    /* Some OSes have hard coded limits on how many bytes can be manipulated by each writev operation.
+     * Force a reasonable limit, to prevent overflowing a 32-bit integer (limit comes from BSD and OS X) */
+    mca_btl_tcp_module.super.btl_rdma_pipeline_frag_size = ((1UL<<31) - 1024);
     mca_btl_tcp_module.super.btl_min_rdma_pipeline_size = 0;
     mca_btl_tcp_module.super.btl_flags = MCA_BTL_FLAGS_PUT |
                                        MCA_BTL_FLAGS_SEND_INPLACE |
@@ -320,7 +322,11 @@ static int mca_btl_tcp_component_register(void)
 
     mca_btl_base_param_register(&mca_btl_tcp_component.super.btl_version,
                                 &mca_btl_tcp_module.super);
-
+    if (mca_btl_tcp_module.super.btl_rdma_pipeline_frag_size > ((1UL<<31) - 1024) ) {
+        /* Assume a hard limit. A test in configure would be a better solution, but until then
+         * kicking-in the pipeline RDMA for extremely large data is good enough. */
+        mca_btl_tcp_module.super.btl_rdma_pipeline_frag_size = ((1UL<<31) - 1024);
+    }
     mca_btl_tcp_param_register_int ("disable_family", NULL, 0, OPAL_INFO_LVL_2,  &mca_btl_tcp_component.tcp_disable_family);
 
     return mca_btl_tcp_component_verify();

opal/mca/btl/tcp/btl_tcp_frag.c

@@ -173,7 +173,7 @@ bool mca_btl_tcp_frag_recv(mca_btl_tcp_frag_t* frag, int sd)
     num_vecs = frag->iov_cnt;
 #if MCA_BTL_TCP_ENDPOINT_CACHE
     if( 0 != btl_endpoint->endpoint_cache_length ) {
-        ssize_t length;
+        size_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.