Fix the TCL btl module endpoint matching during setup for the scenario

when running an MPI job spanning a node that has two TCP NICs and a
node that has one TCP NIC.  Previously, for the 2 NIC/module process,
we would return the first peer IP address if we couldn't find a subnet
match with any of the peer's published IP addresses -- this was to
support running OMPI across subnet boundaries.  Changed the behavior
to only do that behavior if the IP address we're trying to match is
public (i.e., not 10.x.y.z, 192.168.x.y, or 172.16.x.y) *and* any of
the remote peer's addresses are public (working on the assumption that
if we both have public addresses, they're routable to each other).

This definitely will not work in all scenarios, such as when we go to
WAN kinds of executions, and will need to be revisited at that time.

This commit was SVN r9119.

ompi/mca/btl/tcp/btl_tcp_proc.c

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
+ * Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana
  *                         University Research and Technology
  *                         Corporation.  All rights reserved.
  * Copyright (c) 2004-2005 The University of Tennessee and The University
@@ -18,6 +18,13 @@
 
 #include "ompi_config.h"
 
+#ifdef HAVE_ARPA_INET_H
+#include <arpa/inet.h>
+#endif
+#ifdef HAVE_NETINET_IN_H
+#include <netinet/in.h>
+#endif
+
 #include "orte/class/orte_proc_table.h"
 #include "ompi/mca/btl/base/btl_base_error.h"
 #include "ompi/mca/pml/base/pml_base_module_exchange.h"
@@ -27,6 +34,7 @@
 
 static void mca_btl_tcp_proc_construct(mca_btl_tcp_proc_t* proc);
 static void mca_btl_tcp_proc_destruct(mca_btl_tcp_proc_t* proc);
+static bool is_private_ipv4(struct in_addr *in);
 
 
 OBJ_CLASS_INSTANCE(
@@ -65,6 +73,28 @@ void mca_btl_tcp_proc_destruct(mca_btl_tcp_proc_t* proc)
 }
 
 
+/*
+ * Check to see if an IPv4 struct in_addr is public or private.  We
+ * can only do IPv4 here because some of the TCP BTL endpoint structs
+ * only hold the struct in_addr, not the upper-level sin_family that
+ * would indicate if the address is IPv6.
+ */
+static bool is_private_ipv4(struct in_addr *in)
+{
+    /* There are definitely ways to do this more efficiently, but
+       since this is not performance-critical code, it seems better to
+       use clear code (vs. clever code) */
+
+    uint32_t addr = ntohl((uint32_t) in->s_addr);
+    unsigned int a = (addr & 0xff000000) >> 24;
+    unsigned int b = (addr & 0x00ff0000) >> 16;
+
+    return ((10 == a) ||
+            (192 == a && 168 == b) ||
+            (172 == a && 16 == b)) ? true : false;
+}
+
+
 /*
  * Create a TCP process structure. There is a one-to-one correspondence
  * between a ompi_proc_t and a mca_btl_tcp_proc_t instance. We cache
@@ -141,11 +171,11 @@ int mca_btl_tcp_proc_insert(
     struct mca_btl_tcp_module_t *btl_tcp = btl_endpoint->endpoint_btl;
     size_t i;
     unsigned long net1;
-                                                                                                          
+
     /* insert into endpoint array */
     btl_endpoint->endpoint_proc = btl_proc;
     btl_proc->proc_endpoints[btl_proc->proc_endpoint_count++] = btl_endpoint;
-                                                                                                          
+
     net1 = btl_tcp->tcp_ifaddr.sin_addr.s_addr & btl_tcp->tcp_ifmask.sin_addr.s_addr;
 
     /*
@@ -171,12 +201,75 @@ int mca_btl_tcp_proc_insert(
         btl_endpoint->endpoint_addr->addr_inuse++;
         return OMPI_SUCCESS;
     }
-    /* No common interface.. grab the first one and hope for the best */ 
-    else if(btl_proc->proc_addr_count){
-      btl_endpoint->endpoint_addr = btl_proc->proc_addrs;
-      return OMPI_SUCCESS;
-    } else { 
-      return OMPI_ERR_UNREACH;
+
+    /* There was no common interface.  So what do we do?  For the
+       moment, we'll do enough to cover 2 common cases:
+
+       1. Running MPI processes on two computers that are not on the
+          same subnet, but still have routable addresses to each
+          other.  In this case, the above subnet matching will fail,
+          but since the addresses are routable, the
+          OS/networking/routers will make it all work ok.  So we need
+          to make this function *not* return OMPI_ERR_UNREACH.
+
+       2. Running MPI processes on a typical cluster configuration
+          where a head node has 2 TCP NICs (one public IP address one
+          private IP address) and all the back-end compute nodes have
+          only private IP addresses.  In this scenario, the MPI
+          process on the head node will have 2 TCP BTL modules (one
+          for the public, one for the private).  The module with the
+          private IP address will match the subnet and all will work
+          fine.  The module with the public IP address will not match
+          anything and fall through to here -- we want it to return
+          OMPI_ERR_UNREACH so that that module will effectively have
+          no peers that it can communicate with.
+
+       To support these two scenarios, do the following:
+
+       - if my address is private (10., 192.168., or 172.16.), return
+         UNREACH.
+       - if my address is public, return the first public address from
+         my peer (and hope for the best), or UNREACH if there are none
+         available.
+
+       This does not cover some other scenarios that we'll likely need
+       to support in the future, such as:
+
+       - Flat neighborhood networks -- where all the IP's in question
+         are private, the subnet masking won't necessarily match, but
+         they're routable to each other.
+       - Really large, private TCP-based clusters, such as a 1024 node
+         TCP-based cluster.  Depending on how the subnet masks are set
+         by the admins, there may be a subnet mask that effectively
+         spans the entire cluster, or (for example) subnet masks may
+         be set such that only nodes on the same switches are on the
+         same subnet.  This latter scenario will not be supported
+         by the above cases.
+
+       To support these kinds of scenarios, we really need "something
+       better", such as allowing the user to specify a config file
+       indicating which subnets are reachable by which interface, etc.
+    */
+
+    else {
+        /* If my address is private, return UNREACH */
+        if (is_private_ipv4(&(btl_tcp->tcp_ifaddr.sin_addr))) {
+            return OMPI_ERR_UNREACH;
+        }
+
+        /* Find the first public peer address */
+        for (i = 0; i < btl_proc->proc_addr_count; ++i) {
+            mca_btl_tcp_addr_t* endpoint_addr = btl_proc->proc_addrs + i;
+            if (!is_private_ipv4(&(endpoint_addr->addr_inet))) {
+                btl_endpoint->endpoint_addr = endpoint_addr;
+                btl_endpoint->endpoint_addr->addr_inuse++;
+                return OMPI_SUCCESS;
+            }
+        }
+
+        /* Didn't find any peer addresses that were public, so return
+           UNREACH */
+        return OMPI_ERR_UNREACH;
     }
 }