diff --git a/src/common/switch.c b/src/common/switch.c
index 64dafde340853ee54f2437313963d6624d9c70e8..7b5c46234514792a883fe13cf632863bfebec4a1 100644
--- a/src/common/switch.c
+++ b/src/common/switch.c
@@ -97,6 +97,9 @@ typedef struct slurm_switch_ops {
int (*step_allocated) ( switch_jobinfo_t jobinfo,
char *nodelist );
int (*state_clear) ( void );
+ int (*slurmctld_init) ( void );
+ int (*slurmd_init) ( void );
+ int (*slurmd_step_init) ( void );
} slurm_switch_ops_t;
struct slurm_switch_context {
@@ -200,7 +203,10 @@ _slurm_switch_get_ops( slurm_switch_context_t c )
"switch_p_sprintf_node_info",
"switch_p_job_step_complete",
"switch_p_job_step_allocated",
- "switch_p_libstate_clear"
+ "switch_p_libstate_clear",
+ "switch_p_slurmctld_init",
+ "switch_p_slurmd_init",
+ "switch_p_slurmd_step_init"
};
int n_syms = sizeof( syms ) / sizeof( char * );
@@ -551,3 +557,27 @@ extern int switch_g_job_step_allocated(switch_jobinfo_t jobinfo,
return (*(g_context->ops.step_allocated))( jobinfo, nodelist );
}
+
+extern int switch_g_slurmctld_init(void)
+{
+ if ( switch_init() < 0 )
+ return SLURM_ERROR;
+
+ return (*(g_context->ops.slurmctld_init)) ();
+}
+
+extern int switch_g_slurmd_init(void)
+{
+ if ( switch_init() < 0 )
+ return SLURM_ERROR;
+
+ return (*(g_context->ops.slurmd_init)) ();
+}
+
+extern int switch_g_slurmd_step_init(void)
+{
+ if ( switch_init() < 0 )
+ return SLURM_ERROR;
+
+ return (*(g_context->ops.slurmd_step_init)) ();
+}
diff --git a/src/plugins/switch/elan/switch_elan.c b/src/plugins/switch/elan/switch_elan.c
index 68fc06e318b27c72cb2842028699d55e63ed1414..0d8c974fbf15be407ec379d2ff3bbcb3d170c9e0 100644
--- a/src/plugins/switch/elan/switch_elan.c
+++ b/src/plugins/switch/elan/switch_elan.c
@@ -806,3 +806,18 @@ extern int switch_p_job_step_allocated(switch_jobinfo_t jobinfo, char *nodelist)
{
return qsw_restore_jobinfo((qsw_jobinfo_t) jobinfo);
}
+
+extern int switch_p_slurmctld_init( void )
+{
+ return SLURM_SUCCESS;
+}
+
+extern int switch_p_slurmd_init( void )
+{
+ return SLURM_SUCCESS;
+}
+
+extern int switch_p_slurmd_step_init( void )
+{
+ return SLURM_SUCCESS;
+}
diff --git a/src/plugins/switch/federation/federation.c b/src/plugins/switch/federation/federation.c
index 3e4e831516f210b49b0a019741e51f0071d17175..c16d3e7dadaac85e138e536aa73d2d740269b014 100644
--- a/src/plugins/switch/federation/federation.c
+++ b/src/plugins/switch/federation/federation.c
@@ -68,6 +68,7 @@ char* fed_conf = NULL;
extern bool fed_need_state_save;
mode_t fed_umask;
+
/*
* Data structures specific to Federation
*
@@ -143,6 +144,9 @@ typedef struct {
*/
fed_libstate_t *fed_state = NULL;
pthread_mutex_t global_lock = PTHREAD_MUTEX_INITIALIZER;
+
+/* slurmd/slurmd_step global variables */
+hostlist_t adapter_list;
static fed_cache_entry_t lid_cache[FED_MAXADAPTERS];
@@ -171,6 +175,8 @@ static void _strip_cr_nl(char *line);
static void _strip_comments(char *line);
static int _set_up_adapter(fed_adapter_t *fed_adapter, char *adapter_name);
static int _parse_fed_file(hostlist_t *adapter_list);
+static void _init_adapter_cache(void);
+static int _fill_in_adapter_cache(void);
/* The _lock() and _unlock() functions are used to lock/unlock a
* global mutex. Used to serialize access to the global library
@@ -196,6 +202,51 @@ _unlock(void)
}
}
+int
+fed_slurmctld_init(void)
+{
+ return SLURM_SUCCESS;
+}
+
+int
+fed_slurmd_init(void)
+{
+ /*
+ * This is a work-around for the ntbl_* functions calling umask(0)
+ */
+ fed_umask = umask(0077);
+ umask(fed_umask);
+
+ /*_init_adapter_cache();*/
+
+ adapter_list = hostlist_create(NULL);
+ if (_parse_fed_file(&adapter_list) != SLURM_SUCCESS)
+ return SLURM_FAILURE;
+ assert(hostlist_count(adapter_list) <= FED_MAXADAPTERS);
+ return SLURM_SUCCESS;
+}
+
+int
+fed_slurmd_step_init(void)
+{
+ /*
+ * This is a work-around for the ntbl_* functions calling umask(0)
+ */
+ fed_umask = umask(0077);
+ umask(fed_umask);
+
+ _init_adapter_cache();
+
+ adapter_list = hostlist_create(NULL);
+ if (_parse_fed_file(&adapter_list) != SLURM_SUCCESS)
+ return SLURM_FAILURE;
+ assert(hostlist_count(adapter_list) <= FED_MAXADAPTERS);
+
+ _fill_in_adapter_cache();
+
+ return SLURM_SUCCESS;
+}
+
static char *
_lookup_fed_status_tab(int status)
{
@@ -258,13 +309,13 @@ char *fed_sprint_jobinfo(fed_jobinfo_t *j, char *buf,
/* The lid caching functions were created to avoid unnecessary
* function calls each time we need to load network tables on a node.
- * fed_init_cache() simply initializes the cache to sane values and
+ * _init_cache() simply initializes the cache to sane values and
* needs to be called before any other cache functions are called.
*
- * Used by: slurmd
+ * Used by: slurmd/slurmd_step
*/
-void
-fed_init_cache(void)
+static void
+_init_adapter_cache(void)
{
int i;
@@ -273,14 +324,44 @@ fed_init_cache(void)
lid_cache[i].lid = -1;
lid_cache[i].network_id = -1;
}
+}
- /*
- * This is a work-around for the ntbl_* functions calling umask(0)
- */
- fed_umask = umask(0077);
+/* Use ntbl_adapter_resources to cache information about local adapters.
+ *
+ * Used by: slurmd_step
+ */
+static int
+_fill_in_adapter_cache(void)
+{
+ hostlist_iterator_t adapters;
+ char *adapter_name = NULL;
+ ADAPTER_RESOURCES res;
+ int num;
+ int rc;
+ int i;
+
+ adapters = hostlist_iterator_create(adapter_list);
+ for (i = 0; adapter_name = hostlist_next(adapters); i++) {
+ rc = ntbl_adapter_resources(NTBL_VERSION, adapter_name, &res);
+ if (rc != NTBL_SUCCESS)
+ return SLURM_ERROR;
+
+ num = adapter_name[3] - (int)'0';
+ assert(num < FED_MAXADAPTERS);
+ lid_cache[num].lid = res.lid;
+ lid_cache[num].network_id = res.network_id;
+ strncpy(lid_cache[num].name, adapter_name, FED_ADAPTERNAME_LEN);
+
+ free(res.window_list);
+ free(adapter_name);
+ }
+ hostlist_iterator_destroy(adapters);
umask(fed_umask);
+
+ return SLURM_SUCCESS;
}
+
/* Cache the lid and network_id of a given adapter. Ex: sni0 with lid 10
* gets cached in array index 0 with a lid = 10 and a name = sni0.
*
@@ -508,43 +589,29 @@ static int _parse_fed_file(hostlist_t *adapter_list)
* For all that exist, record vital adapter info plus status for all windows
* available on that adapter. Cache lid to adapter name mapping locally.
*
- * Is not thread-safe.
- *
* Used by: slurmd
*/
static int
_get_adapters(fed_adapter_t *list, int *count)
{
- static hostlist_t adapter_list = NULL;
- static hostlist_iterator_t adapter_iter;
- int i = 0;
+ hostlist_iterator_t adapter_iter;
char *adapter = NULL;
+ int i;
assert(list != NULL);
+ assert(adapter_list != NULL);
- if (adapter_list == NULL || hostlist_is_empty(adapter_list)) {
- int rc;
- adapter_list = hostlist_create(NULL);
- rc = _parse_fed_file(&adapter_list);
- if (rc != SLURM_SUCCESS)
- return rc;
- assert(hostlist_count(adapter_list) <= FED_MAXADAPTERS);
- adapter_iter = hostlist_iterator_create(adapter_list);
- }
- i=0;
- *count = hostlist_count(adapter_list);
- info("Number of adapters is = %d", *count);
- assert(*count > 0);
- //list = xmalloc(sizeof(fed_adapter_t) * (*count));
- while (adapter = hostlist_next(adapter_iter)) {
- if(_set_up_adapter(list + i, adapter)
- == SLURM_ERROR)
+ adapter_iter = hostlist_iterator_create(adapter_list);
+ for (i = 0; adapter = hostlist_next(adapter_iter); i++) {
+ if(_set_up_adapter(list + i, adapter) == SLURM_ERROR)
fatal("Failed to set up adapter %s.", adapter);
free(adapter);
- i++;
}
- hostlist_iterator_reset(adapter_iter);
+ hostlist_iterator_destroy(adapter_iter);
+ assert(i > 0);
+ *count = i;
+ info("Number of adapters is = %d", *count);
if(!*count)
slurm_seterrno_ret(ENOADAPTER);
diff --git a/src/plugins/switch/federation/federation.h b/src/plugins/switch/federation/federation.h
index 6c8328b2667eac52bfbdebca13af04778c1fb4f4..9370cabef30995b832bd4137db23447f5f70c077 100644
--- a/src/plugins/switch/federation/federation.h
+++ b/src/plugins/switch/federation/federation.h
@@ -81,7 +81,6 @@ fed_jobinfo_t *fed_copy_jobinfo(fed_jobinfo_t *jp);
void fed_free_jobinfo(fed_jobinfo_t *jp);
int fed_load_table(fed_jobinfo_t *jp, int uid, int pid);
int fed_init(void);
-void fed_init_cache(void);
int fed_unload_table(fed_jobinfo_t *jp);
int fed_unpack_libstate(fed_libstate_t *lp, Buf buffer);
int fed_get_jobinfo(fed_jobinfo_t *jp, int key, void *data);
diff --git a/src/plugins/switch/federation/switch_federation.c b/src/plugins/switch/federation/switch_federation.c
index 5c8652ab6050c7a213726e22e8ee7ba54c35c68d..3c5798f49262068176e412138c69016dfd757385 100644
--- a/src/plugins/switch/federation/switch_federation.c
+++ b/src/plugins/switch/federation/switch_federation.c
@@ -1,4 +1,4 @@
-/*****************************************************************************\
+/***************************************************************************** \
** switch_federation.c - Library routines for initiating jobs on IBM
** Federation
** $Id$
@@ -120,7 +120,6 @@ const uint32_t plugin_version = 100;
int init ( void )
{
verbose("%s loaded", plugin_name);
- fed_init_cache();
return SLURM_SUCCESS;
}
@@ -130,6 +129,21 @@ int fini ( void )
return SLURM_SUCCESS;
}
+int switch_p_slurmctld_init( void )
+{
+ return fed_slurmctld_init();
+}
+
+int switch_p_slurmd_init( void )
+{
+ return fed_slurmd_init();
+}
+
+int switch_p_slurmd_step_init( void )
+{
+ return fed_slurmd_step_init();
+}
+
/*
* switch functions for global state save/restore
*/
diff --git a/src/plugins/switch/none/switch_none.c b/src/plugins/switch/none/switch_none.c
index cdd43fa146ce20b03353ed99a4b10b7736091659..163d3ae583898b89e2c3c22c97eda950fdc2c1d9 100644
--- a/src/plugins/switch/none/switch_none.c
+++ b/src/plugins/switch/none/switch_none.c
@@ -284,3 +284,17 @@ extern int switch_p_job_step_allocated(switch_jobinfo_t jobinfo,
return SLURM_SUCCESS;
}
+extern int switch_p_slurmctld_init( void )
+{
+ return SLURM_SUCCESS;
+}
+
+extern int switch_p_slurmd_init( void )
+{
+ return SLURM_SUCCESS;
+}
+
+extern int switch_p_slurmd_step_init( void )
+{
+ return SLURM_SUCCESS;
+}
diff --git a/src/slurmd/slurmd.c b/src/slurmd/slurmd.c
index 0ca1f4921a5841a8d03cb40712b54c103c5cff6d..46ab3f4274e735209e4cbdf15b0ba171c9e9feab 100644
--- a/src/slurmd/slurmd.c
+++ b/src/slurmd/slurmd.c
@@ -195,6 +195,7 @@ main (int argc, char *argv[])
fatal("Unable to initialize interconnect.");
if (conf->cleanstart && switch_g_clear_node_state())
fatal("Unable to clear interconnect state.");
+ switch_g_slurmd_init();
_create_msg_socket();
diff --git a/src/slurmd/slurmd_step.c b/src/slurmd/slurmd_step.c
index f108b20267ffb057c65e58a87df422dbe2c2c851..b979e03ff866164dee1bfaad606f38bec16b6de1 100644
--- a/src/slurmd/slurmd_step.c
+++ b/src/slurmd/slurmd_step.c
@@ -93,6 +93,7 @@ main (int argc, char *argv[])
//log_alter(conf->log_opts, 0, NULL);
log_init(argv[0],conf->log_opts, LOG_DAEMON, conf->logfile);
g_slurmd_jobacct_init(conf->cf.job_acct_parameters);
+ switch_g_slurmd_step_init();
/* receive len of packed cli from main slurmd */
if((rc = read(STDIN_FILENO, &len, sizeof(int))) == -1) {