Whamcloud - gitweb
LU-2934 lnet: Add LNet Router Priority parameter
[fs/lustre-release.git] / lnet / lnet / api-ni.c
index 41c4a64..ed5eba9 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, Whamcloud, Inc.
+ * Copyright (c) 2011, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -36,6 +36,9 @@
 
 #define DEBUG_SUBSYSTEM S_LNET
 #include <lnet/lib-lnet.h>
+#ifdef __KERNEL__
+#include <linux/log2.h>
+#endif
 
 #ifdef __KERNEL__
 #define D_LNI D_CONSOLE
@@ -44,6 +47,7 @@
 #endif
 
 lnet_t      the_lnet;                           /* THE state of the network */
+EXPORT_SYMBOL(the_lnet);
 
 #ifdef __KERNEL__
 
@@ -59,6 +63,10 @@ static char *routes = "";
 CFS_MODULE_PARM(routes, "s", charp, 0444,
                 "routes to non-local networks");
 
+static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
+CFS_MODULE_PARM(rnet_htable_size, "i", int, 0444,
+               "size of remote network hash table");
+
 char *
 lnet_get_routes(void)
 {
@@ -91,10 +99,10 @@ lnet_get_networks(void)
 void
 lnet_init_locks(void)
 {
-       cfs_spin_lock_init(&the_lnet.ln_eq_wait_lock);
+       spin_lock_init(&the_lnet.ln_eq_wait_lock);
        cfs_waitq_init(&the_lnet.ln_eq_waitq);
-       cfs_mutex_init(&the_lnet.ln_lnd_mutex);
-       cfs_mutex_init(&the_lnet.ln_api_mutex);
+       mutex_init(&the_lnet.ln_lnd_mutex);
+       mutex_init(&the_lnet.ln_api_mutex);
 }
 
 void
@@ -117,34 +125,16 @@ lnet_get_networks (void)
 {
         static char       default_networks[256];
         char             *networks = getenv ("LNET_NETWORKS");
-        char             *ip2nets  = getenv ("LNET_IP2NETS");
         char             *str;
         char             *sep;
         int               len;
         int               nob;
-        int               rc;
         cfs_list_t       *tmp;
 
-#ifdef NOT_YET
-        if (networks != NULL && ip2nets != NULL) {
-                LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
-                                   " 'LNET_IP2NETS' but not both at once\n");
-                return NULL;
-        }
-
-        if (ip2nets != NULL) {
-                rc = lnet_parse_ip2nets(&networks, ip2nets);
-                return (rc == 0) ? networks : NULL;
-        }
-#else
-        SET_BUT_UNUSED(ip2nets);
-        SET_BUT_UNUSED(rc);
-#endif
         if (networks != NULL)
                 return networks;
 
         /* In userland, the default 'networks=' is the list of known net types */
-
         len = sizeof(default_networks);
         str = default_networks;
         *str = 0;
@@ -155,13 +145,12 @@ lnet_get_networks (void)
 
                 nob = snprintf(str, len, "%s%s", sep,
                                libcfs_lnd2str(lnd->lnd_type));
+               if (nob >= len) {
+                       /* overflowed the string; leave it where it was */
+                       *str = 0;
+                       break;
+               }
                 len -= nob;
-                if (len < 0) {
-                        /* overflowed the string; leave it where it was */
-                        *str = 0;
-                        break;
-                }
-
                 str += nob;
                 sep = ",";
         }
@@ -207,23 +196,40 @@ void lnet_fini_locks(void)
 #endif
 
 static int
-lnet_create_locks(void)
+lnet_create_remote_nets_table(void)
 {
-       lnet_init_locks();
+       int             i;
+       cfs_list_t      *hash;
+
+       LASSERT(the_lnet.ln_remote_nets_hash == NULL);
+       LASSERT(the_lnet.ln_remote_nets_hbits > 0);
+       LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
+       if (hash == NULL) {
+               CERROR("Failed to create remote nets hash table\n");
+               return -ENOMEM;
+       }
 
-       the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
-       if (the_lnet.ln_res_lock == NULL)
-               goto failed;
+       for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
+               CFS_INIT_LIST_HEAD(&hash[i]);
+       the_lnet.ln_remote_nets_hash = hash;
+       return 0;
+}
 
-       the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
-       if (the_lnet.ln_net_lock == NULL)
-               goto failed;
+static void
+lnet_destroy_remote_nets_table(void)
+{
+       int             i;
+       cfs_list_t      *hash;
 
-       return 0;
+       if (the_lnet.ln_remote_nets_hash == NULL)
+               return;
 
- failed:
-       lnet_fini_locks();
-       return -ENOMEM;
+       for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
+               LASSERT(cfs_list_empty(&the_lnet.ln_remote_nets_hash[i]));
+
+       LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
+                   LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
+       the_lnet.ln_remote_nets_hash = NULL;
 }
 
 static void
@@ -242,6 +248,26 @@ lnet_destroy_locks(void)
        lnet_fini_locks();
 }
 
+static int
+lnet_create_locks(void)
+{
+       lnet_init_locks();
+
+       the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
+       if (the_lnet.ln_res_lock == NULL)
+               goto failed;
+
+       the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
+       if (the_lnet.ln_net_lock == NULL)
+               goto failed;
+
+       return 0;
+
+ failed:
+       lnet_destroy_locks();
+       return -ENOMEM;
+}
+
 void lnet_assert_wire_constants (void)
 {
         /* Wire protocol assertions generated by 'wirecheck'
@@ -368,6 +394,7 @@ lnet_register_lnd (lnd_t *lnd)
 
         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 }
+EXPORT_SYMBOL(lnet_register_lnd);
 
 void
 lnet_unregister_lnd (lnd_t *lnd)
@@ -383,6 +410,7 @@ lnet_unregister_lnd (lnd_t *lnd)
 
         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 }
+EXPORT_SYMBOL(lnet_unregister_lnd);
 
 void
 lnet_counters_get(lnet_counters_t *counters)
@@ -721,9 +749,12 @@ lnet_prepare(lnet_pid_t requested_pid)
        CFS_INIT_LIST_HEAD(&the_lnet.ln_nis);
        CFS_INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
        CFS_INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
-       CFS_INIT_LIST_HEAD(&the_lnet.ln_remote_nets);
        CFS_INIT_LIST_HEAD(&the_lnet.ln_routers);
 
+       rc = lnet_create_remote_nets_table();
+       if (rc != 0)
+               goto failed;
+
        the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 
        the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
@@ -813,6 +844,7 @@ lnet_unprepare (void)
                cfs_percpt_free(the_lnet.ln_counters);
                the_lnet.ln_counters = NULL;
        }
+       lnet_destroy_remote_nets_table();
 
        return 0;
 }
@@ -1019,7 +1051,6 @@ lnet_shutdown_lndnis (void)
        LASSERT(!the_lnet.ln_shutdown);
        LASSERT(the_lnet.ln_refcount == 0);
        LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
-       LASSERT(cfs_list_empty(&the_lnet.ln_remote_nets));
 
        lnet_net_lock(LNET_LOCK_EX);
        the_lnet.ln_shutdown = 1;       /* flag shutdown */
@@ -1345,15 +1376,24 @@ LNetInit(void)
        CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
 
 #ifdef __KERNEL__
+       /* The hash table size is the number of bits it takes to express the set
+        * ln_num_routes, minus 1 (better to under estimate than over so we
+        * don't waste memory). */
+       if (rnet_htable_size <= 0)
+               rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
+       else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
+               rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
+       the_lnet.ln_remote_nets_hbits = max_t(int, 1,
+                                          order_base_2(rnet_htable_size) - 1);
+
         /* All LNDs apart from the LOLND are in separate modules.  They
          * register themselves when their module loads, and unregister
          * themselves when their module is unloaded. */
 #else
+       the_lnet.ln_remote_nets_hbits = 8;
+
         /* Register LNDs
          * NB the order here determines default 'networks=' order */
-# ifdef CRAY_XT3
-        LNET_REGISTER_ULND(the_ptllnd);
-# endif
 # ifdef HAVE_LIBPTHREAD
         LNET_REGISTER_ULND(the_tcplnd);
 # endif
@@ -1361,6 +1401,7 @@ LNetInit(void)
         lnet_register_lnd(&the_lolnd);
         return 0;
 }
+EXPORT_SYMBOL(LNetInit);
 
 /**
  * Finalize LNet library.
@@ -1384,6 +1425,7 @@ LNetFini(void)
 
        the_lnet.ln_init = 0;
 }
+EXPORT_SYMBOL(LNetFini);
 
 /**
  * Set LNet PID and start LNet interfaces, routing, and forwarding.
@@ -1481,6 +1523,7 @@ LNetNIInit(lnet_pid_t requested_pid)
         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
         return rc;
 }
+EXPORT_SYMBOL(LNetNIInit);
 
 /**
  * Stop LNet interfaces, routing, and forwarding.
@@ -1520,6 +1563,7 @@ LNetNIFini()
         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
         return 0;
 }
+EXPORT_SYMBOL(LNetNIFini);
 
 /**
  * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
@@ -1555,8 +1599,8 @@ LNetCtl(unsigned int cmd, void *arg)
                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
 
         case IOC_LIBCFS_ADD_ROUTE:
-                rc = lnet_add_route(data->ioc_net, data->ioc_count,
-                                    data->ioc_nid);
+               rc = lnet_add_route(data->ioc_net, data->ioc_count,
+                                   data->ioc_nid, data->ioc_priority);
                 return (rc != 0) ? rc : lnet_check_routes();
 
         case IOC_LIBCFS_DEL_ROUTE:
@@ -1565,7 +1609,8 @@ LNetCtl(unsigned int cmd, void *arg)
         case IOC_LIBCFS_GET_ROUTE:
                 return lnet_get_route(data->ioc_count,
                                       &data->ioc_net, &data->ioc_count,
-                                      &data->ioc_nid, &data->ioc_flags);
+                                     &data->ioc_nid, &data->ioc_flags,
+                                     &data->ioc_priority);
         case IOC_LIBCFS_NOTIFY_ROUTER:
                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
                                    cfs_time_current() -
@@ -1639,6 +1684,7 @@ LNetCtl(unsigned int cmd, void *arg)
         }
         /* not reached */
 }
+EXPORT_SYMBOL(LNetCtl);
 
 /**
  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
@@ -1679,6 +1725,7 @@ LNetGetId(unsigned int index, lnet_process_id_t *id)
        lnet_net_unlock(cpt);
        return rc;
 }
+EXPORT_SYMBOL(LNetGetId);
 
 /**
  * Print a string representation of handle \a h into buffer \a str of
@@ -1689,6 +1736,7 @@ LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
 {
         snprintf(str, len, LPX64, h.cookie);
 }
+EXPORT_SYMBOL(LNetSnprintHandle);
 
 static int
 lnet_create_ping_info(void)
@@ -2041,12 +2089,8 @@ lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_i
         for (i = 0; i < n_ids; i++) {
                 tmpid.pid = info->pi_pid;
                 tmpid.nid = info->pi_ni[i].ns_nid;
-#ifdef __KERNEL__
-                if (cfs_copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
+               if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
                         goto out_1;
-#else
-                ids[i] = tmpid;
-#endif
         }
         rc = info->pi_nnis;