void gw_dm_failed ( void *_job_id )
{
gw_job_t * job;
int job_id;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer */
/* ----------------------------------------------------------- */
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("DM",'E',"Job %i does not exist (JOB_STATE_FAILED).\n",job_id);
free(_job_id);
return;
}
}
else
return;
/* ----------------------------------------------------------- */
/* 1.- Set state */
/* ----------------------------------------------------------- */
gw_log_print("DM",'I',"Job %i failed.\n",job->id);
gw_job_set_state(job, GW_JOB_STATE_FAILED, GW_FALSE);
gw_job_print(job,"DM",'I',"Job failed, history:\n");
gw_job_print_history(job);
job->exit_time = time(NULL);
if (job->history != NULL)
job->history->reason = GW_REASON_EXECUTION_ERROR;
if ( job->client_waiting > 0 )
gw_am_trigger(gw_dm.rm_am,"GW_RM_WAIT_SUCCESS", _job_id);
else
free(_job_id);
/* -------- Update Host & User running jobs -------- */
gw_user_pool_dec_running_jobs(job->user_id);
pthread_mutex_lock(&(job->history->host->mutex));
job->history->host->running_jobs--;
pthread_mutex_unlock(&(job->history->host->mutex));
pthread_mutex_unlock(&(job->mutex));
}
void gw_em_cancel(void *_job_id)
{
int job_id;
gw_job_t *job;
gw_em_mad_t *mad;
gw_em_state_t current_em_state;
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
free(_job_id);
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("EM",'E',"Job %s no longer exists (CANCEL).\n", job_id);
return;
}
}
else
return;
/* -------------------------------------------------------------------- */
current_em_state = job->em_state;
if ( issubmitted(current_em_state) )
{
gw_job_print (job,"EM",'I',"Cancelling job.\n");
mad = job->history->em_mad;
/* Warning! When in Migration Cancel, the previous MAD should be used */
if (job->job_state == GW_JOB_STATE_MIGR_CANCEL)
{
if (job->history->next == NULL)
{
gw_log_print("EM",'E',"Previous history record of job %i no longer exists\n", job_id);
pthread_mutex_unlock(&(job->mutex));
return;
}
else
mad = job->history->next->em_mad;
}
gw_em_mad_cancel(mad, job_id);
gw_log_print ("EM",'I',"Cancelling job %i.\n", job_id);
}
else
gw_log_print ("EM",'W',"Ignoring cancel request for job %i, will re-try.\n",
job_id);
/* -------------------------------------------------------------------- */
pthread_mutex_unlock(&(job->mutex));
}
int gw_tm_prolog_build_urls(gw_job_t * job,
char * src,
char * dst,
char ** src_url,
char ** dst_url)
{
int is_gsiftp;
char url_buffer[1024];
char *tmp;
if ( src[0] == '/' )
{
gw_job_print(job,"TM",'W',"\tSkipping file %s, absolute path.\n",src);
return 1;
}
is_gsiftp = strstr(src,"gsiftp://") != NULL;
if ( is_gsiftp )
{
*src_url = gw_job_substitute (src,job);
if (*src_url == NULL )
{
gw_job_print(job,"TM",'E',"\tSkipping file %s, parse error.\n",src);
return -1;
}
if ( dst == NULL )
{
tmp = strrchr(src,'/');
if ( tmp == NULL )
{
gw_job_print(job,"TM",'E',"\tSkipping file %s, no file in url.\n",src);
free(*src_url);
return -1;
}
else
tmp++;
}
else
tmp = dst;
snprintf(url_buffer,sizeof(char)*1024,"%s%s",job->history->rdir,tmp);
*dst_url = gw_job_substitute (url_buffer, job);
if ( dst_url == NULL )
{
gw_job_print(job,"TM",'E',"\tSkipping file %s, parse error.\n",url_buffer);
free(*src_url);
return -1;
}
}
else
{
if ( strstr(src,"file://") != NULL )
{
strncpy(url_buffer, src, sizeof(char) * 1024);
if ( dst == NULL )
{
tmp = strrchr(src,'/');
if ( tmp == NULL )
{
gw_job_print(job,"TM",'E',"\tSkipping file %s, no file in url.\n",src);
return -1;
}
else
tmp++;
}
else
{
tmp = dst;
}
}
else
{
snprintf(url_buffer,sizeof(char)*1024,"file://%s/%s",job->template.job_home, src);
if ( dst == NULL )
tmp = src;
else
tmp = dst;
}
void gw_dm_wrapper_done_cb ( void *_job_id )
{
gw_job_t * job;
int job_id;
time_t total;
time_t active;
time_t suspension;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer */
/* ----------------------------------------------------------- */
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("DM",'E',"Job %i does not exist (JOB_STATE_WRAPPER_DONE_CB).\n",job_id);
free(_job_id);
return;
}
}
else
return;
/* ----------------------------------------------------------- */
/* 1.- Set execution times & state transition */
/* ----------------------------------------------------------- */
job->em_state = GW_EM_STATE_INIT;
switch (job->job_state)
{
case GW_JOB_STATE_PRE_WRAPPER:
/* --------------- Update pre-wrapper stats -------------------- */
job->history->stats[PRE_WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_pre_wrapper_time(job->history);
active = job->history->stats[ACTIVE_TIME];
suspension = job->history->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Pre-Wrapper DONE:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
gw_job_print(job,"DM",'I',"\tTotal time : %i\n", total);
/* -------------- Transition to Wrapper state ------------------ */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_WRAPPER", _job_id);
break;
case GW_JOB_STATE_WRAPPER:
/* ----------------- Update wrapper stats ---------------------- */
job->history->stats[WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_wrapper_time(job->history);
active = job->history->stats[ACTIVE_TIME];
suspension = job->history->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Wrapper DONE:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
gw_job_print(job,"DM",'I',"\tTotal time : %i\n", total);
/* -------------- Free used slot from this host -------------- */
gw_host_dec_uslots(job->history->host, job->template.np);
void gw_em_listener(void *arg)
{
fd_set in_pipes;
int i,j;
int *job_id;
int greater, rc, rcm;
char c;
char info[GW_EM_MAX_INFO];
char s_job_id[GW_EM_MAX_JOB_ID];
char result[GW_EM_MAX_RESULT];
char action[GW_EM_MAX_ACTION];
char str[GW_EM_MAX_STRING];
int fd;
gw_job_t *job;
time_t now;
char contact_file[PATH_MAX];
FILE *file;
gw_em_mad_t *em_mad;
char *ptmp;
int *fds;
int num_fds;
gw_em_mad_t **em_mads;
pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL);
pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, NULL);
fds = (int *) malloc(sizeof(int)*gw_conf.number_of_users*GW_MAX_MADS);
em_mads = (gw_em_mad_t **) malloc(sizeof(gw_em_mad_t *) *
gw_conf.number_of_users * GW_MAX_MADS);
while (1)
{
greater = gw_user_pool_set_em_pipes(&in_pipes, fds, &num_fds, em_mads, gw_em.um_em_pipe_r);
rc = select( greater+1, &in_pipes, NULL, NULL, NULL);
if ( rc <= 0 )
continue;
for (i=0; i<num_fds; i++)
{
fd = fds[i];
if ( FD_ISSET(fd, &in_pipes) )
{
if ( fd == gw_em.um_em_pipe_r )
{
rc = read(fd, (void *) &c, sizeof(char));
#ifdef GWEMDEBUG
gw_log_print("EM",'D',"Updating MAD pipes (action is %c)\n",c);
#endif
continue;
}
#ifdef GWEMDEBUG
gw_log_print("EM",'D',"Reading from MAD pipe %i.\n",i);
#endif
j = 0;
do
{
rc = read(fd, (void *) &c, sizeof(char));
str[j++] = c;
}
while ((rc > 0) && (c != '\n') && (j < (GW_EM_MAX_STRING-1)));
str[j] = '\0';
if (rc <= 0)
{
gw_log_print("EM",'W',"Error reading MAD (%s) message\n",
em_mads[i]->name);
rcm = gw_em_mad_reload(em_mads[i]);
if ( rcm == 0 )
{
gw_log_print("EM",'I',"MAD (%s) successfully reloaded\n",
em_mads[i]->name);
gw_job_pool_em_recover(em_mads[i], &(gw_em.am));
}
else
{
gw_log_print("EM",'E',"Error reloading MAD (%s)\n",
em_mads[i]->name);
em_mads[i]->mad_em_pipe = -1;
}
continue;
}
sscanf(str,"%" GW2STR(GW_EM_MAX_ACTION) "s %"
GW2STR(GW_EM_MAX_JOB_ID) "s %"
GW2STR(GW_EM_MAX_RESULT) "s %"
GW2STR(GW_EM_MAX_INFO) "[^\n]",
action, s_job_id, result, info);
#ifdef GWEMDEBUG
gw_log_print("EM",'D',"MAD message received:\"%s %s %s %s\".\n",
action, s_job_id, result, info);
#endif
if (s_job_id[0] == '-')
continue;
job_id = (int *) malloc (sizeof(int));
*job_id = atoi(s_job_id);
job = gw_job_pool_get(*job_id, GW_TRUE);
if (job == NULL)
{
gw_log_print("EM",'W',"MAD message for job %s, but it does not exist: \"%s %s %s %s\".\n",
s_job_id,action, s_job_id, result, info);
free(job_id);
continue;
}
if (job->job_state != GW_JOB_STATE_PRE_WRAPPER
&& job->job_state != GW_JOB_STATE_WRAPPER
&& job->job_state != GW_JOB_STATE_MIGR_CANCEL
&& job->job_state != GW_JOB_STATE_STOP_CANCEL
&& job->job_state != GW_JOB_STATE_KILL_CANCEL)
{
gw_log_print("EM",'W',"MAD message for job %i but not in an execution state.\n",
*job_id);
free(job_id);
pthread_mutex_unlock(&(job->mutex));
continue;
}
else if ( job->em_state == GW_EM_STATE_HARD_KILL )
{
gw_log_print("EM",'W',"MAD message for job %i but it is being killed (hard).\n",
*job_id);
free(job_id);
pthread_mutex_unlock(&(job->mutex));
continue;
}
if (strcmp(action, "SUBMIT") == 0)
{
if (strcmp(result, "FAILURE") == 0)
{
gw_job_print(job, "EM",'E',"Job submission failed: %s\n",
info);
gw_log_print("EM",'E',"Submission of job %d failed: %s.\n",
job->id, info);
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_FAILED",
(void *) job_id);
}
else /* Save persistent job contact */
{
snprintf(contact_file, PATH_MAX-1,
GW_VAR_DIR "/%i/job.contact",
job->id);
file = fopen(contact_file, "w");
if ( file != NULL )
{
fprintf(file, "%s\n", info);
fclose(file);
}
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_PENDING",
(void *) job_id);
}
}
else if (strcmp(action, "CANCEL") == 0)
{
if (strcmp(result, "SUCCESS") == 0)
{
gw_job_print(job, "EM",'I',"Job cancel succeeded.\n");
gw_log_print("EM",'I',"Cancel of job %i succeeded.\n", *job_id);
}
else
{
gw_job_print(job, "EM",'E',"Job cancel failed (%s).\n",info);
gw_log_print("EM",'E',"Cancel of job %d failed: %s.\n",job->id, info);
}
}
else if (strcmp(action, "POLL") == 0)
{
if (strcmp(result, "SUCCESS") == 0)
{
now = time(NULL);
job->next_poll_time = now + gw_conf.poll_interval/2
+ gw_rand(gw_conf.poll_interval);
gw_job_print(job, "EM",'E',"Job poll OK (%s), will poll again in %d seconds.\n",
info, job->next_poll_time - now);
if (strcmp(info, "PENDING") == 0)
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_PENDING",
(void *) job_id);
else if (strcmp(info, "SUSPENDED") == 0)
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_SUSPENDED",
(void *) job_id);
else if (strcmp(info, "ACTIVE") == 0)
gw_am_trigger(&(gw_em.am),"GW_EM_STATE_ACTIVE",
(void *) job_id);
else if (strstr(info, "DONE") != NULL)
{
ptmp = strstr(info,"DONE:");
if ((ptmp != NULL) && (strlen(ptmp+5) > 0))/*No-wrapper mode*/
job->exit_code=atoi(ptmp+5);
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_DONE",
(void *) job_id);
}
else if (strcmp(info, "FAILED") == 0)
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_FAILED",
(void *) job_id);
}
else
{
job->history->failed_polls++;
em_mad = job->history->em_mad;
if ( job->history->failed_polls == 3 )
{
gw_job_print(job, "EM",'E',"Job poll failed (%s), assuming the job is done.\n",info);
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_DONE", (void *) job_id);
}
else
{
now = time(NULL);
job->next_poll_time = now + gw_conf.poll_interval*job->history->failed_polls
+ gw_rand(gw_conf.poll_interval*job->history->failed_polls);
gw_job_print(job, "EM",'E',"Job poll failed (%s), will poll again in %d seconds.\n",
info, job->next_poll_time - now);
free(job_id);
}
}
}
else if (strcmp(action, "RECOVER") == 0)
{
if (strcmp(result, "SUCCESS") == 0)
{
if (strcmp(info, "PENDING") == 0)
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_PENDING",
(void *) job_id);
else if (strcmp(info, "SUSPENDED") == 0)
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_SUSPENDED",
(void *) job_id);
else if (strcmp(info, "ACTIVE") == 0)
gw_am_trigger(&(gw_em.am),"GW_EM_STATE_ACTIVE",
(void *) job_id);
else if (strstr(info, "DONE") != NULL)
{
ptmp = strstr(info,"DONE:");
if ((ptmp != NULL) && (strlen(ptmp+5) > 0))/*No-wrapper mode*/
job->exit_code = atoi(ptmp+5);
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_DONE",
(void *) job_id);
}
else if (strcmp(info, "FAILED") == 0)
{
/* Do not retry */
job->history->tries= job->template.number_of_retries;
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_FAILED",
(void *) job_id);
}
}
else
{
gw_job_print(job,"EM",'E',"Job recover failed (%s), assuming the job is done.\n", info);
gw_log_print("EM",'E',"Recover of job %i failed.\n", *job_id);
gw_am_trigger(&(gw_em.am), "GW_EM_STATE_DONE", (void *) job_id);
}
}
void gw_dm_wrapper_failed_cb ( void *_job_id )
{
gw_job_t * job;
int job_id;
time_t total;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer */
/* ----------------------------------------------------------- */
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("DM",'E',"Job %i does not exist (JOB_STATE_WRAPPER_FAILED_CB).\n",job_id);
free(_job_id);
return;
}
}
else
return;
/* ----------------------------------------------------------- */
/* 1.- Set execution times & state transition */
/* ----------------------------------------------------------- */
job->em_state = GW_EM_STATE_INIT;
switch (job->job_state)
{
case GW_JOB_STATE_PRE_WRAPPER:
/* --------------- Update pre-wrapper stats -------------------- */
job->history->stats[PRE_WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_pre_wrapper_time(job->history);
gw_job_print(job,"DM",'E',"Pre-Wrapper failed:\n");
gw_job_print(job,"DM",'E',"\tTotal time : %i\n", total);
break;
case GW_JOB_STATE_WRAPPER:
/* ----------------- Update wrapper stats ---------------------- */
job->history->stats[WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_wrapper_time(job->history);
gw_job_print(job,"DM",'E',"Wrapper failed:\n");
gw_job_print(job,"DM",'E',"\tTotal time : %i\n", total);
/* ---------- We do not need to re-schedule this job --------- */
if ( job->reschedule == GW_TRUE )
{
job->reschedule = GW_FALSE;
gw_dm_mad_job_del(&gw_dm.dm_mad[0],job->id);
}
break;
default:
gw_log_print("DM",'E',"Wrapper failed callback in wrong job (%i) state.\n", job_id);
break;
}
/* ----------------------------------------------------------- */
/* 1.- State transtition */
/* ----------------------------------------------------------- */
/* -------------- Free used slot from this host -------------- */
if (job->history != NULL)
{
job->history->reason = GW_REASON_EXECUTION_ERROR;
gw_host_dec_uslots(job->history->host);
}
/* ----------------------------------------------------------- */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_EPILOG_FAIL", _job_id);
pthread_mutex_unlock(&(job->mutex));
}
void gw_dm_wrapper_done_cb ( void *_job_id )
{
gw_job_t * job;
int job_id;
time_t total;
time_t active;
time_t suspension;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer */
/* ----------------------------------------------------------- */
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("DM",'E',"Job %i does not exist (JOB_STATE_WRAPPER_DONE_CB).\n",job_id);
free(_job_id);
return;
}
}
else
return;
/* ----------------------------------------------------------- */
/* 1.- Set execution times & state transition */
/* ----------------------------------------------------------- */
job->em_state = GW_EM_STATE_INIT;
switch (job->job_state)
{
case GW_JOB_STATE_PRE_WRAPPER:
/* --------------- Update pre-wrapper stats -------------------- */
job->history->stats[PRE_WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_pre_wrapper_time(job->history);
active = job->history->stats[ACTIVE_TIME];
suspension = job->history->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Pre-Wrapper DONE:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
gw_job_print(job,"DM",'I',"\tTotal time : %i\n", total);
/* -------------- Transition to Wrapper state ------------------ */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_WRAPPER", _job_id);
break;
case GW_JOB_STATE_WRAPPER:
/* ----------------- Update wrapper stats ---------------------- */
job->history->stats[WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_wrapper_time(job->history);
active = job->history->stats[ACTIVE_TIME];
suspension = job->history->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Wrapper DONE:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
gw_job_print(job,"DM",'I',"\tTotal time : %i\n", total);
/* -------------- Free used slot from this host -------------- */
gw_host_dec_uslots(job->history->host);
/* ---------- We do not need to re-schedule this job --------- */
if ( job->reschedule == GW_TRUE )
{
job->reschedule = GW_FALSE;
gw_dm_mad_job_del(&gw_dm.dm_mad[0],job->id);
}
/* -------------- Transition to Epilog state ------------------ */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_EPILOG_STD", _job_id);
break;
case GW_JOB_STATE_STOP_CANCEL:
/* ----------------- Update wrapper stats ---------------------- */
job->history->stats[WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_wrapper_time(job->history);
active = job->history->stats[ACTIVE_TIME];
suspension = job->history->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Wrapper CANCELED:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
gw_job_print(job,"DM",'I',"\tTotal time : %i\n", total);
/* -------------- Free used slot from this host -------------- */
gw_host_dec_uslots(job->history->host);
/* ------------ Transition to Stop Epilog state --------------- */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_STOP_EPILOG", _job_id);
break;
case GW_JOB_STATE_KILL_CANCEL:
/* ----------------- Update wrapper stats ---------------------- */
job->history->stats[WRAPPER_EXIT_TIME] = time(NULL);
total = gw_job_history_get_wrapper_time(job->history);
active = job->history->stats[ACTIVE_TIME];
suspension = job->history->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Wrapper CANCELED:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
gw_job_print(job,"DM",'I',"\tTotal time : %i\n", total);
/* -------------- Free used slot from this host -------------- */
gw_host_dec_uslots(job->history->host);
/* ------------ Transition to Kill Epilog state ---------------- */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_KILL_EPILOG", _job_id);
break;
case GW_JOB_STATE_MIGR_CANCEL:
/* ----------- Update previous wrapper stats ------------------- */
job->history->next->stats[WRAPPER_EXIT_TIME] = time(NULL);
active = job->history->next->stats[ACTIVE_TIME];
suspension = job->history->next->stats[SUSPENSION_TIME];
gw_job_print(job,"DM",'I',"Wrapper CANCELED:\n");
gw_job_print(job,"DM",'I',"\tActive time : %i\n", active);
gw_job_print(job,"DM",'I',"\tSuspension time : %i\n", suspension);
/* -------------- Free used slot from previous host ------------ */
gw_host_dec_uslots(job->history->next->host);
/* ---------- Transition to Migration Prolog state ------------ */
gw_am_trigger(&(gw_dm.am), "GW_DM_STATE_MIGR_PROLOG", _job_id);
break;
default:
gw_log_print("DM",'E',"Wrapper done callback in wrong job (%i) state.\n", job_id);
free(_job_id);
break;
}
pthread_mutex_unlock(&(job->mutex));
}
void gw_dm_pending( void *_job_id )
{
gw_job_t *job;
int job_id;
gw_boolean_t failed;
gw_migration_reason_t reason;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer */
/* ----------------------------------------------------------- */
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("DM",'E',"Job %i does not exist (JOB_STATE_PENDING).\n",job_id);
free(_job_id);
return;
}
}
else
return;
/* ----------------------------------------------------------- */
/* 1.- Set state */
/* ----------------------------------------------------------- */
gw_job_print(job,"DM",'I',"Rescheduling job.\n");
gw_log_print("DM",'I',"Rescheduling job %d.\n", job->id);
gw_job_set_state (job, GW_JOB_STATE_PENDING, GW_FALSE);
/* -------- Update Host & User running jobs -------- */
gw_user_pool_dec_running_jobs(job->user_id);
gw_host_dec_rjobs(job->history->host);
/* ------------- Restart counter --------------- */
job->restarted++;
/* ------------- Notify the Scheduler --------------- */
if (job->history != NULL)
{
reason = GW_REASON_NONE;
failed = (job->history->reason == GW_REASON_EXECUTION_ERROR)
|| (job->history->reason == GW_REASON_PERFORMANCE);
if (failed)
{
gw_dm_mad_job_failed(&gw_dm.dm_mad[0], job->history->host->host_id,
job->user_id, job->history->reason);
}
gw_dm_mad_job_schedule(&gw_dm.dm_mad[0], job_id, job->array_id,
job->user_id, reason);
}
else
gw_log_print("DM",'E',"Rescheduling job %d, but no history records found.\n",
job->id);
/* ------------------------------------------------- */
free(_job_id);
pthread_mutex_unlock(&(job->mutex));
}
void gw_dm_zombie ( void *_job_id )
{
gw_job_t * job;
gw_array_t * array;
int job_id;
int task_id;
int array_id;
int rt;
char conf_filename[2048];
time_t prolog, epilog;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer */
/* ----------------------------------------------------------- */
if ( _job_id == NULL )
return;
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("DM",'E',"Job %i does not exist (JOB_STATE_ZOMBIE).\n",job_id);
free(_job_id);
return;
}
/* ----------------------------------------------------------- */
/* 0.- Update Job state */
/* ----------------------------------------------------------- */
switch (job->job_state)
{
case GW_JOB_STATE_EPILOG:
gw_job_set_state(job, GW_JOB_STATE_ZOMBIE, GW_FALSE);
gw_log_print("DM",'I',"Job %i done, with exit code %i.\n",job->id, job->exit_code);
job->history->reason = GW_REASON_NONE;
job->exit_time = time(NULL);
/* ------------- Print job history and send usage ------------ */
gw_job_print(job,"DM",'I',"Job done, history:\n");
gw_job_print_history(job);
gw_job_send_usage(job);
if ( job->client_waiting > 0 )
gw_am_trigger(gw_dm.rm_am,"GW_RM_WAIT_SUCCESS", _job_id);
else
{
if (gw_conf.dispose == GW_TRUE)
gw_am_trigger(&(gw_dm.am), "GW_DM_KILL", _job_id);
else
free(_job_id);
}
/* -------- Update User & Host running jobs -------- */
gw_user_pool_dec_running_jobs(job->user_id);
gw_host_dec_rjobs(job->history->host);
/* -------- Notify the scheduler -------- */
prolog = gw_job_history_get_prolog_time(job->history);
epilog = gw_job_history_get_epilog_time(job->history);
gw_dm_mad_job_success(&gw_dm.dm_mad[0],
job->history->host->host_id,
job->user_id,
(prolog + epilog),
job->history->stats[SUSPENSION_TIME],
job->history->stats[ACTIVE_TIME]);
pthread_mutex_unlock(&(job->mutex));
/* -------- Update other jobs dependencies -------- */
gw_job_pool_dep_check(job_id);
break;
case GW_JOB_STATE_KILL_EPILOG:
gw_job_set_state(job, GW_JOB_STATE_ZOMBIE, GW_FALSE);
job->exit_time = time(NULL);
/* ------------- Print job history and send usage ------------ */
gw_job_print(job,"DM",'I',"Job killed, history:\n");
gw_job_print_history(job);
gw_job_send_usage(job);
/* ---------------- Free job & Notify RM ---------------- */
array_id = job->array_id;
task_id = job->task_id;
/* -------- Update User & Host running jobs -------- */
gw_user_pool_dec_running_jobs(job->user_id);
gw_host_dec_rjobs(job->history->host);
sprintf(conf_filename, "%s/job.conf", job->directory);
unlink(conf_filename);
pthread_mutex_unlock(&(job->mutex));
/* ------------------------------------------------- */
gw_job_pool_free(job_id);
gw_log_print("DM",'I',"Job %i killed and freed.\n", job_id);
if (array_id != -1)
{
array = gw_array_pool_get_array(array_id,GW_TRUE);
if ( array != NULL )
{
rt = gw_array_del_task(array,task_id);
pthread_mutex_unlock(&(array->mutex));
if (rt == 0)
{
gw_array_pool_array_free(array_id);
gw_log_print("DM",'I',"Array %i freed\n",array_id);
}
}
else
gw_log_print("DM",'E',"Could not delete task %i from array %i.\n",
task_id, array_id);
}
gw_am_trigger(gw_dm.rm_am,"GW_RM_KILL_SUCCESS", _job_id);
break;
default:
gw_log_print("DM",'E',"Zombie callback in wrong job (%i) state.\n", job_id);
free(_job_id);
pthread_mutex_unlock(&(job->mutex));
break;
}
}
void gw_em_submit(void *_job_id)
{
int job_id;
gw_job_t *job;
char *rsl=NULL;
char *contact;
gw_job_state_t state;
char rsl_filename[2048];
FILE *fd;
time_t now;
/* ----------------------------------------------------------- */
/* 0.- Get job pointer, check if it exits and lock mutex */
/* ----------------------------------------------------------- */
if ( _job_id != NULL )
{
job_id = *( (int *) _job_id );
job = gw_job_pool_get(job_id, GW_TRUE);
if ( job == NULL )
{
gw_log_print("EM",'E',"Job %s no longer exists (PENDING).\n",
job_id);
return;
}
}
else
return;
if (job->history == NULL)
{
gw_log_print("EM",'E',"History of job %s doesn't exists\n",
job_id);
free(_job_id);
pthread_mutex_unlock(&(job->mutex));
return;
}
state = job->job_state;
/* ----------------------------------------------------------- */
/* 1.- Get execution MAD for this host */
/* ----------------------------------------------------------- */
job->em_state = GW_EM_STATE_INIT;
job->history->counter = -1;
if ( job->job_state == GW_JOB_STATE_PRE_WRAPPER )
{
contact = job->history->em_fork_rc;
rsl = (char *) job->history->em_mad->pre_wrapper_rsl((void *) job);
}
else
{
contact = job->history->em_rc;
rsl = (char *) job->history->em_mad->wrapper_rsl((void *) job);
}
if ( rsl == NULL )
{
job->em_state = GW_EM_STATE_FAILED;
gw_log_print("EM",'E',"Job %i failed, could not generate RSL.\n", job_id);
gw_am_trigger(gw_em.dm_am, "GW_DM_WRAPPER_FAILED", _job_id);
pthread_mutex_unlock(&(job->mutex));
return;
}
sprintf(rsl_filename, "%s/job.rsl.%i", job->directory,job->restarted);
fd = fopen(rsl_filename,"w");
if (fd != NULL )
{
gw_job_print(job,"EM",'I',"Submitting wrapper to %s, RSL used is in %s.\n",contact,rsl_filename);
fprintf(fd,"%s",rsl);
fclose(fd);
}
else
{
job->em_state = GW_EM_STATE_FAILED;
gw_log_print("EM",'E',"Job %i failed, could not open RSL file.\n", job_id);
gw_job_print(job,"EM",'E',"Job failed, could not open RSL file %s.\n",rsl_filename);
gw_am_trigger(gw_em.dm_am, "GW_DM_WRAPPER_FAILED", _job_id);
pthread_mutex_unlock(&(job->mutex));
return;
}
/* -------------------------------------------------------------------- */
now = time(NULL);
job->next_poll_time = now + gw_conf.poll_interval/2
+ gw_rand(gw_conf.poll_interval); /* randomize polls */
gw_job_print(job,"EM",'I',"Job will be polled in %d seconds.\n",
job->next_poll_time-now);
job->last_checkpoint_time = 0;
job->history->stats[LAST_SUSPENSION_TIME] = now;
job->history->stats[SUSPENSION_TIME] = 0;
job->history->stats[ACTIVE_TIME] = 0;
job->history->tries++;
/* -------------------------------------------------------------------- */
pthread_mutex_unlock(&(job->mutex));
gw_em_mad_submit(job->history->em_mad, job_id, contact, rsl_filename);
/* -------------------------------------------------------------------- */
free(_job_id);
free(rsl);
}
