ompi-ipv6-link-local.patch

commit 8d1f0719776e92837ef3ab3f2b895f057e4a9c36
Author: Pieter Noordhuis <[email protected]>
Date:   Tue Aug 16 19:31:14 2016 -0700

    Support IPv6 link-local addresses
    
    Link-local addresses use a single prefix (fe:80) so a routing table
    doesn't help figuring out which interface to transmit it on. The Linux
    IPv6 implementation asks application developers to populate the
    `sin6_scope_id` field on the `sockaddr_in6` struct with the index of the
    interface to communicate on.

diff --git a/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp.h b/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp.h
index 5279c09..ae2c5c6 100644
--- a/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp.h
+++ b/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp.h
@@ -78,8 +78,10 @@ struct mca_btl_tcp_component_t {
     opal_event_t tcp6_recv_event;           /**< recv event for IPv6 listen socket */
     int tcp6_listen_sd;                     /**< IPv6 listen socket for incoming connection requests */
     unsigned short tcp6_listen_port;        /**< IPv6 listen port */
-    int tcp6_port_min;                      /**< IPv4 minimum port */
-    int tcp6_port_range;                    /**< IPv4 port range */
+    int tcp6_port_min;                      /**< IPv6 minimum port */
+    int tcp6_port_range;                    /**< IPv6 port range */
+    bool tcp6_use_link_local;               /**< Enable use of IPv6 link-local addresses */
+    int tcp6_link_local_scope_id;           /**< Kernel index of interface for link-local traffic */
 #endif
     /* Port range restriction */
 
diff --git a/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_component.c b/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_component.c
index 59a3a48..8bbbb96 100644
--- a/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_component.c
+++ b/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_component.c
@@ -160,6 +160,20 @@ static inline unsigned int mca_btl_tcp_param_register_uint(
     return *storage;
 }
 
+static inline unsigned int mca_btl_tcp_param_register_bool(
+    const char* param_name,
+    const char* help_string,
+    bool default_value,
+    int level,
+    bool *storage)
+{
+    *storage = default_value;
+    (void) mca_base_component_var_register(&mca_btl_tcp_component.super.btl_version,
+                                           param_name, help_string, MCA_BASE_VAR_TYPE_BOOL,
+                                           NULL, 0, 0, level,
+                                           MCA_BASE_VAR_SCOPE_READONLY, storage);
+    return *storage;
+}
 
 /*
  * Data structure for accepting connections.
@@ -254,6 +268,10 @@ static int mca_btl_tcp_component_register(void)
                                     (0x1 << 16) - mca_btl_tcp_component.tcp6_port_min - 1,
                                     OPAL_INFO_LVL_2, &mca_btl_tcp_component.tcp6_port_range );
     free(message);
+
+    mca_btl_tcp_param_register_bool("enable_ipv6_link_local",
+                                    "Whether to enable use of IPv6 link-local addresses (default: false)", false,
+                                    OPAL_INFO_LVL_2, &mca_btl_tcp_component.tcp6_use_link_local);
 #endif
 
     mca_btl_tcp_component.report_all_unfound_interfaces = false;
@@ -677,7 +695,7 @@ static int mca_btl_tcp_component_create_instances(void)
        and therefore we're done. */
     if (mca_btl_tcp_component.tcp_num_btls > 0) {
         ret = OMPI_SUCCESS;
-        goto cleanup;
+        goto check;
     }
 
     /* if the interface list was not specified by the user, create 
@@ -709,7 +727,91 @@ static int mca_btl_tcp_component_create_instances(void)
     }
     opal_argv_free(exclude);
 
- cleanup:
+ check:
+#if OPAL_ENABLE_IPV6
+    /* If using IPv6 link-local addresses is OK, we need to verify
+     * we have only a single BTL instance with a link-local address,
+     * or they will be ambiguous. */
+    if (!mca_btl_tcp_component.tcp6_use_link_local) {
+        goto cleanup;
+    }
+
+    {
+        int link_local_ifkindex = -1;
+        unsigned int btl_index;
+
+        for (btl_index = 0;
+             btl_index < mca_btl_tcp_component.tcp_num_btls;
+             btl_index++) {
+            for (if_index = opal_ifbegin();
+                 if_index >= 0;
+                 if_index = opal_ifnext(if_index)) {
+                /* IF_NAMESIZE is defined in opal/util/if.h */
+                char if_name[IF_NAMESIZE];
+                struct sockaddr_in6 ss;
+
+                if (opal_ifindextokindex(if_index) !=
+                    mca_btl_tcp_component.tcp_btls[btl_index]->tcp_ifkindex) {
+                    continue;
+                }
+
+                ret = opal_ifindextoaddr(if_index,
+                                         (struct sockaddr*) &ss,
+                                         sizeof(ss));
+                if (ret != OPAL_SUCCESS) {
+                    opal_output (0,
+                                 "btl_tcp_component: "
+                                 "unable to get address for "
+                                 "index %i (kernel index %i)",
+                                 if_index,
+                                 opal_ifindextokindex(if_index));
+                    goto cleanup;
+                }
+
+                /* Ignore non-IPv6 addresses */
+                if (ss.sin6_family != AF_INET6) {
+                    continue;
+                }
+
+                /* Ignore addresses other than link-local */
+                if (ss.sin6_scope_id != 0x20) {
+                    continue;
+                }
+
+                /*
+                 * Error if there are multiple interfaces with
+                 * a link-local address (they will be ambiguous).
+                 */
+                if (link_local_ifkindex >= 0) {
+                    opal_output (0,
+                                 "btl_tcp_component: "
+                                 "multiple link-local addresses found");
+                    ret = OMPI_ERROR;
+                    goto cleanup;
+                }
+
+                link_local_ifkindex = opal_ifindextokindex(if_index);
+                opal_ifindextoname(if_index, if_name, sizeof(if_name));
+                opal_output (0,
+                             "btl_tcp_component: "
+                             "using %s for link-local traffic",
+                             if_name);
+            }
+        }
+
+        if (link_local_ifkindex < 0) {
+            opal_output (0,
+                         "btl_tcp_component: "
+                         "no link-local addresses found");
+            ret = OMPI_ERROR;
+            goto cleanup;
+        }
+
+        mca_btl_tcp_component.tcp6_link_local_scope_id = link_local_ifkindex;
+    }
+#endif
+
+cleanup:
     if (NULL != kindexes) {
         free(kindexes);
     }
diff --git a/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_proc.c b/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_proc.c
index 89aee88..21e4e50 100644
--- a/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_proc.c
+++ b/1.10.2/src/openmpi-1.10.2/ompi/mca/btl/tcp/btl_tcp_proc.c
@@ -798,6 +798,18 @@ bool mca_btl_tcp_proc_tosocks(mca_btl_tcp_addr_t* proc_addr,
             inaddr->sin6_port = proc_addr->addr_port;
             inaddr->sin6_scope_id = 0;
             inaddr->sin6_flowinfo = 0;
+
+            /*
+             * If this is a link-local address AND the component is configured
+             * to allow link-local addresses for BTL traffic, set the
+             * scope_id so the kernel passes it to the right network interface.
+             */
+            if ((inaddr->sin6_addr.s6_addr[0] & 0xff) == 0xfe &&
+                (inaddr->sin6_addr.s6_addr[1] & 0xc0) == 0x80 &&
+                mca_btl_tcp_component.tcp6_use_link_local) {
+                inaddr->sin6_scope_id =
+                    mca_btl_tcp_component.tcp6_link_local_scope_id;
+            }
         }
         break;
 #endif
diff --git a/1.10.2/src/openmpi-1.10.2/opal/mca/if/linux_ipv6/if_linux_ipv6.c b/1.10.2/src/openmpi-1.10.2/opal/mca/if/linux_ipv6/if_linux_ipv6.c
index 2832371..ffa0b8e 100644
--- a/1.10.2/src/openmpi-1.10.2/opal/mca/if/linux_ipv6/if_linux_ipv6.c
+++ b/1.10.2/src/openmpi-1.10.2/opal/mca/if/linux_ipv6/if_linux_ipv6.c
@@ -118,8 +118,8 @@ static int if_linux_ipv6_open(void)
                                 addrbyte[8], addrbyte[9], addrbyte[10], addrbyte[11],
                                 addrbyte[12], addrbyte[13], addrbyte[14], addrbyte[15], scope);
 
-            /* Only interested in global (0x00) scope */
-            if (scope != 0x00)  {
+            /* Only interested in global (0x00) and link-local (0x20) scope */
+            if (scope != 0x00 && scope != 0x20)  {
                 opal_output_verbose(1, opal_if_base_framework.framework_output,
                                     "skipping interface %2x%2x:%2x%2x:%2x%2x:%2x%2x:%2x%2x:%2x%2x:%2x%2x:%2x%2x scope %x\n",
                                     addrbyte[0], addrbyte[1], addrbyte[2], addrbyte[3],
diff --git a/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.c b/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.c
index 498a42d..5b453ca 100644
--- a/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.c
+++ b/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.c
@@ -438,6 +438,17 @@ static int tcp_component_register(void)
                                           OPAL_INFO_LVL_6,
                                           MCA_BASE_VAR_SCOPE_READONLY,
                                           &mca_oob_tcp_component.skip_version_check);
+
+#if OPAL_ENABLE_IPV6
+    mca_oob_tcp_component.tcp6_use_link_local = false;
+    mca_oob_tcp_component.tcp6_link_local_scope_id = -1;
+    (void)mca_base_component_var_register(component, "enable_ipv6_link_local",
+                                          "Whether to enable use of IPv6 link-local addresses (default: false)",
+                                          MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
+                                          OPAL_INFO_LVL_2,
+                                          MCA_BASE_VAR_SCOPE_READONLY,
+                                          &mca_oob_tcp_component.tcp6_use_link_local);
+#endif
     return ORTE_SUCCESS;
 }
 
@@ -566,6 +577,12 @@ static bool component_available(void)
             opal_argv_append_nosize(&mca_oob_tcp_component.ipv4conns, opal_net_get_hostname((struct sockaddr*) &my_ss));
         } else if (AF_INET6 == my_ss.ss_family) {
 #if OPAL_ENABLE_IPV6
+            /* If this address has link local scope, capture its interface index */
+            if (mca_oob_tcp_component.tcp6_use_link_local &&
+                ((struct sockaddr_in6*) &my_ss)->sin6_scope_id == 0x20) {
+                mca_oob_tcp_component.tcp6_link_local_scope_id = kindex;
+            }
+
             opal_output_verbose(10, orte_oob_base_framework.framework_output,
                                 "%s oob:tcp:init adding %s to our list of %s connections",
                                 ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
diff --git a/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.h b/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.h
index d8d47a2..03182b4 100644
--- a/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.h
+++ b/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_component.h
@@ -62,9 +62,11 @@ typedef struct {
 
 #if OPAL_ENABLE_IPV6
     /* IPv6 support */
-    bool               disable_ipv6_family;  /**< disable this AF */
-    char**             tcp6_static_ports;    /**< Static ports - IPV6 */
-    char**             tcp6_dyn_ports;       /**< Dynamic ports - IPV6 */
+    bool               disable_ipv6_family;      /**< disable this AF */
+    char**             tcp6_static_ports;        /**< Static ports - IPV6 */
+    char**             tcp6_dyn_ports;           /**< Dynamic ports - IPV6 */
+    bool               tcp6_use_link_local;      /**< Enable use of IPv6 link-local addresses */
+    int                tcp6_link_local_scope_id; /**< Kernel index of interface for link-local traffic */
     char**             ipv6conns;
     char**             ipv6ports;
 #endif
diff --git a/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_connection.c b/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_connection.c
index 124dc9d..71565ee 100644
--- a/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_connection.c
+++ b/1.10.2/src/openmpi-1.10.2/orte/mca/oob/tcp/oob_tcp_connection.c
@@ -144,6 +144,19 @@ void mca_oob_tcp_peer_try_connect(int fd, short args, void *cbdata)
                 continue;
             }
 
+#if OPAL_ENABLE_IPV6
+            if (addr->addr.ss_family == AF_INET6) {
+                struct sockaddr_in6* inaddr = (struct sockaddr_in6*) &addr->addr;
+
+                if ((inaddr->sin6_addr.s6_addr[0] & 0xff) == 0xfe &&
+                    (inaddr->sin6_addr.s6_addr[1] & 0xc0) == 0x80 &&
+                    mca_oob_tcp_component.tcp6_use_link_local) {
+                    inaddr->sin6_scope_id =
+                      mca_oob_tcp_component.tcp6_link_local_scope_id;
+              }
+            }
+#endif
+
             addrlen = addr->addr.ss_family == AF_INET6
                           ? sizeof(struct sockaddr_in6)
                           : sizeof(struct sockaddr_in);
@@ -172,6 +185,25 @@ void mca_oob_tcp_peer_try_connect(int fd, short args, void *cbdata)
                 CLOSE_THE_SOCKET(peer->sd);
                 continue;
             }
+            /* When testing use of IPv6 link-local addresses, the 4.0
+             * kernel would immediately return EADDRNOTAVAIL when
+             * connecting to a link-local address on the same host.
+             * This appears to be a transient problem that only
+             * manifests for a short period of time after calling
+             * listen(2) on the server side of the socket. Therefore,
+             * inserting a small delay on the client side fixes the
+             * problem. Since establishing these connections only
+             * happens at initialization time, a delay is acceptable.
+             */
+            if (EADDRNOTAVAIL == opal_socket_errno) {
+                opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
+                                    "%s connection to %s returned EADDRNOTAVAIL - retrying after delay",
+                                    ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
+                                    ORTE_NAME_PRINT(&peer->name));
+                CLOSE_THE_SOCKET(peer->sd);
+                sleep(1);
+                continue;
+            }
             if (rc < 0) {
                 opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
                                     "%s connection to %s returned %d (%d, %s)",