Context Navigation

source: trunk/pbs_drmaa/job.c @ 57

Revision 57, 15.2 KB checked in by mmamonski, 13 years ago (diff)
on site fixies
Property svn:keywords set to `Id`

Rev	Line
[12]	1	/* $Id$ */
[1]	2	/*
	3	* FedStage DRMAA for PBS Pro
	4	* Copyright (C) 2006-2009 FedStage Systems
	5	*
	6	* This program is free software: you can redistribute it and/or modify
	7	* it under the terms of the GNU General Public License as published by
	8	* the Free Software Foundation, either version 3 of the License, or
	9	* (at your option) any later version.
	10	*
	11	* This program is distributed in the hope that it will be useful,
	12	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	13	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
	14	* GNU General Public License for more details.
	15	*
	16	* You should have received a copy of the GNU General Public License
	17	* along with this program. If not, see <http://www.gnu.org/licenses/>.
	18	*/
	19
	20	#ifdef HAVE_CONFIG_H
	21	# include <config.h>
	22	#endif
	23
	24	#include <signal.h>
	25	#include <stdlib.h>
	26	#include <string.h>
	27	#include <unistd.h>
[57]	28	#include <sys/stat.h>
[1]	29
	30	#include <drmaa_utils/drmaa.h>
	31	#include <drmaa_utils/drmaa_util.h>
	32	#include <pbs_error.h>
	33	#include <pbs_ifl.h>
	34
	35	#include <pbs_drmaa/job.h>
[7]	36	#include <pbs_drmaa/log_reader.h>
[1]	37	#include <pbs_drmaa/pbs_attrib.h>
	38	#include <pbs_drmaa/session.h>
	39	#include <pbs_drmaa/util.h>
	40
	41	#ifndef lint
	42	static char rcsid[]
	43	# ifdef __GNUC__
	44	__attribute__ ((unused))
	45	# endif
[12]	46	= "$Id$";
[1]	47	#endif
	48
	49
[54]	50	static void pbsdrmaa_job_control( fsd_job_t *self, int action );
[1]	51
[54]	52	static void pbsdrmaa_job_update_status( fsd_job_t *self );
[1]	53
[54]	54	static void pbsdrmaa_job_on_missing( fsd_job_t *self );
[1]	55
[54]	56	static void pbsdrmaa_job_on_missing_standard( fsd_job_t *self );
[7]	57
[54]	58	static void pbsdrmaa_job_update( fsd_job_t self, struct batch_status );
[1]	59
[54]	60	static int pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix);
[1]	61
	62	fsd_job_t *
	63	pbsdrmaa_job_new( char *job_id )
	64	{
	65	pbsdrmaa_job_t self = (pbsdrmaa_job_t)fsd_job_new( job_id );
	66	fsd_realloc( self, 1, pbsdrmaa_job_t );
	67	self->super.control = pbsdrmaa_job_control;
	68	self->super.update_status = pbsdrmaa_job_update_status;
	69	self->super.on_missing = pbsdrmaa_job_on_missing;
	70	self->update = pbsdrmaa_job_update;
	71	return (fsd_job_t*)self;
	72	}
	73
	74
	75	static void
	76	pbsdrmaa_job_control( fsd_job_t *self, int action )
	77	{
	78	volatile bool conn_lock = false;
	79	pbsdrmaa_session_t session = (pbsdrmaa_session_t)self->session;
	80	const char *job_id = self->job_id;
	81	const char *apicall = NULL;
	82	int rc = PBSE_NONE;
	83
	84	fsd_log_enter(( "({job_id=%s}, action=%d)",
	85	self->job_id, action ));
	86
	87	TRY
	88	{
[45]	89	int tries_left = session->max_retries_count;
	90	int sleep_time = 1;
[1]	91
	92	conn_lock = fsd_mutex_lock( &self->session->drm_connection_mutex );
	93
	94	/TODO reconnect /
[45]	95	while ( true )
[1]	96	{
	97	switch( action )
	98	{
	99	/*
	100	* We cannot know whether we did suspend job
	101	* in other way than remembering this inside DRMAA session.
	102	*/
	103	case DRMAA_CONTROL_SUSPEND:
	104	apicall = "pbs_sigjob";
	105	rc = pbs_sigjob( session->pbs_conn, (char*)job_id,
	106	"SIGSTOP", NULL );
[10]	107	fsd_log_info(("pbs_sigjob(%s, SIGSTOP) =%d", job_id, rc));
[1]	108	if( rc == PBSE_NONE )
	109	self->flags \|= FSD_JOB_SUSPENDED;
	110	break;
	111	case DRMAA_CONTROL_RESUME:
	112	apicall = "pbs_sigjob";
	113	rc = pbs_sigjob( session->pbs_conn, (char*)job_id,
	114	"SIGCONT", NULL );
[10]	115	fsd_log_info(("pbs_sigjob(%s, SIGCONT) =%d", job_id, rc));
[1]	116	if( rc == PBSE_NONE )
	117	self->flags &= ~FSD_JOB_SUSPENDED;
	118	break;
	119	case DRMAA_CONTROL_HOLD:
	120	apicall = "pbs_holdjob";
	121	rc = pbs_holdjob( session->pbs_conn, (char*)job_id,
	122	USER_HOLD, NULL );
[10]	123	fsd_log_info(("pbs_sigjob(%s, SIGHOLD) =%d", job_id, rc));
[1]	124	if( rc == PBSE_NONE )
	125	self->flags \|= FSD_JOB_HOLD;
	126	break;
	127	case DRMAA_CONTROL_RELEASE:
	128	apicall = "pbs_rlsjob";
	129	rc = pbs_rlsjob( session->pbs_conn, (char*)job_id,
	130	USER_HOLD, NULL );
[10]	131	fsd_log_info(("pbs_rlsjob(%s) =%d", job_id, rc));
[1]	132	if( rc == PBSE_NONE )
	133	self->flags &= FSD_JOB_HOLD;
	134	break;
	135	case DRMAA_CONTROL_TERMINATE:
	136	apicall = "pbs_deljob";
	137	rc = pbs_deljob( session->pbs_conn, (char*)job_id, NULL );
[10]	138	fsd_log_info(("pbs_deljob(%s) =%d", job_id, rc));
[1]	139	/* Torque:
	140	* deldelay=N -- delay between SIGTERM and SIGKILL (default 0) */
	141	if( rc == PBSE_NONE )
	142	{
	143	self->flags &= FSD_JOB_TERMINATED_MASK;
	144	if( (self->flags & FSD_JOB_TERMINATED) == 0 )
	145	self->flags \|= FSD_JOB_TERMINATED \| FSD_JOB_ABORTED;
	146	}
	147	break;
	148	}
	149
[45]	150	retry_connect:
	151	if ( rc == PBSE_NONE )
[1]	152	break;
[45]	153	else if (( rc == PBSE_INTERNAL \|\| rc == PBSE_PROTOCOL \|\| rc == PBSE_EXPIRED) && (tries_left--))
[1]	154	{
[45]	155	if (rc == PBSE_PROTOCOL \|\| rc == PBSE_EXPIRED)
	156	{
	157	if ( session->pbs_conn >= 0)
	158	pbs_disconnect( session->pbs_conn );
	159
	160	sleep( sleep_time++ );
	161
	162	session->pbs_conn = pbs_connect( session->super.contact );
	163
	164	if (session->pbs_conn < 0)
	165	goto retry_connect;
	166
	167	fsd_log_info(( "pbs_connect(%s) =%d", session->super.contact, session->pbs_conn ));
	168	}
	169	else /* PBSE_INTERNAL */
	170	{
	171	/*
	172	* In PBS Pro pbs_sigjob raises internal server error (PBSE_INTERNAL)
	173	* when job just changed its state to running.
	174	*/
	175	sleep( sleep_time++ );
	176	}
	177	fsd_log_debug(( "repeating request (%d of %d)", tries_left, session->max_retries_count));
[1]	178	}
	179	else
	180	pbsdrmaa_exc_raise_pbs( apicall );
[45]	181	} /* end while */
[1]	182	}
	183	FINALLY
	184	{
	185	if( conn_lock )
	186	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
	187	}
	188	END_TRY
	189
	190	fsd_log_return((""));
	191	}
	192
	193
	194	void
	195	pbsdrmaa_job_update_status( fsd_job_t *self )
	196	{
	197	volatile bool conn_lock = false;
	198	struct batch_status *volatile status = NULL;
	199	pbsdrmaa_session_t session = (pbsdrmaa_session_t)self->session;
[45]	200	int tries_left = session->max_retries_count;
	201	int sleep_time = 1;
[1]	202
	203	fsd_log_enter(( "({job_id=%s})", self->job_id ));
[21]	204
[1]	205	TRY
	206	{
	207	conn_lock = fsd_mutex_lock( &self->session->drm_connection_mutex );
	208	retry:
[16]	209
[1]	210	#ifdef PBS_PROFESSIONAL
	211	status = pbs_statjob( session->pbs_conn, self->job_id, NULL, NULL );
	212	#else
	213	status = pbs_statjob( session->pbs_conn, self->job_id, session->status_attrl, NULL );
	214	#endif
[9]	215	fsd_log_info(( "pbs_statjob(fd=%d, job_id=%s, attribs={...}) =%p",
[1]	216	session->pbs_conn, self->job_id, (void*)status ));
	217	if( status == NULL )
	218	{
[21]	219
[25]	220	#ifndef PBS_PROFESSIONAL
[40]	221	if ( pbs_errno != PBSE_UNKJOBID )
	222	fsd_log_error(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
	223	else
	224	fsd_log_debug(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
[25]	225	#else
	226	# ifndef PBS_PROFESSIONAL_NO_LOG
[40]	227	if ( pbs_errno != PBSE_UNKJOBID )
	228	fsd_log_error(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
	229	else
	230	fsd_log_debug(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
[25]	231	# else
[40]	232	if ( pbs_errno != PBSE_UNKJOBID )
	233	fsd_log_error(("pbs_statjob error: %d", pbs_errno));
	234	else
	235	fsd_log_debug(("pbs_statjob error: %d", pbs_errno));
[25]	236	# endif
	237	#endif
[21]	238
[25]	239	switch( pbs_errno )
	240	{
	241	case PBSE_UNKJOBID:
	242	break;
	243	case PBSE_PROTOCOL:
	244	case PBSE_EXPIRED:
	245	if ( session->pbs_conn >= 0 )
	246	pbs_disconnect( session->pbs_conn );
[50]	247	fsd_log_info(("Protocol error. Reconnecting..."));
[45]	248	retry_connect:
	249	sleep(sleep_time++);
[25]	250	session->pbs_conn = pbs_connect( session->super.contact );
	251	if( session->pbs_conn < 0 )
[45]	252	{
[50]	253	if (tries_left--) {
	254	fsd_log_info(("%d tries left. Retrying...", tries_left));
[45]	255	goto retry_connect;
[50]	256	} else {
	257	fsd_log_error(("No more tries left... Throwing exception"));
[45]	258	pbsdrmaa_exc_raise_pbs( "pbs_connect" );
[50]	259	}
[45]	260	}
[25]	261	else
	262	{
	263	goto retry;
	264	}
	265	default:
	266	pbsdrmaa_exc_raise_pbs( "pbs_statjob" );
	267	break;
	268	case 0: /* ? */
	269	fsd_exc_raise_code( FSD_ERRNO_INTERNAL_ERROR );
	270	break;
[1]	271	}
[25]	272
[1]	273	}
	274
	275	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
	276
[25]	277
[1]	278	if( status != NULL )
	279	{
	280	((pbsdrmaa_job_t*)self)->update( self, status );
	281	}
[7]	282	else if( self->state < DRMAA_PS_DONE )
[25]	283	{
[1]	284	self->on_missing( self );
[25]	285	}
[1]	286	}
	287	FINALLY
	288	{
	289	if( conn_lock )
	290	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
	291	if( status != NULL )
	292	pbs_statfree( status );
	293	}
	294	END_TRY
	295
	296	fsd_log_return((""));
	297	}
	298
	299
	300	void
	301	pbsdrmaa_job_update( fsd_job_t self, struct batch_status status )
	302	{
	303	struct attrl *attribs = status->attribs;
	304	struct attrl *i = NULL;
	305	char pbs_state = 0;
	306	int exit_status = -2;
	307	const char *cpu_usage = NULL;
	308	const char *mem_usage = NULL;
	309	const char *vmem_usage = NULL;
	310	const char *walltime = NULL;
	311	long unsigned int modify_time = 0;
	312
	313	fsd_log_enter(( "({job_id=%s})", self->job_id ));
	314	#ifdef DEBUGGING
	315	pbsdrmaa_dump_attrl( attribs, NULL );
	316	#endif
	317	fsd_assert( !strcmp( self->job_id, status->name ) );
	318
	319	for( i = attribs; i != NULL; i = i->next )
	320	{
	321	int attr;
	322	attr = pbsdrmaa_pbs_attrib_by_name( i->name );
	323	switch( attr )
	324	{
	325	case PBSDRMAA_ATTR_JOB_STATE:
	326	pbs_state = i->value[0];
	327	break;
	328	case PBSDRMAA_ATTR_EXIT_STATUS:
[29]	329	exit_status = fsd_atoi( i->value );
[1]	330	break;
	331	case PBSDRMAA_ATTR_RESOURCES_USED:
	332	if( !strcmp( i->resource, "cput" ) )
	333	cpu_usage = i->value;
	334	else if( !strcmp( i->resource, "mem" ) )
	335	mem_usage = i->value;
	336	else if( !strcmp( i->resource, "vmem" ) )
	337	vmem_usage = i->value;
	338	else if( !strcmp( i->resource, "walltime" ) )
	339	walltime = i->value;
	340	break;
	341	case PBSDRMAA_ATTR_QUEUE:
	342	if (!self->queue)
	343	self->queue = fsd_strdup(i->value);
	344	break;
	345	case PBSDRMAA_ATTR_ACCOUNT_NAME:
	346	if (!self->project)
	347	self->project = fsd_strdup(i->value);
	348	break;
[26]	349	#ifndef PBS_PROFESSIONAL
[1]	350	case PBSDRMAA_ATTR_EXECUTION_HOST:
	351	if (!self->execution_hosts) {
	352	fsd_log_debug(("execution_hosts = %s", i->value));
	353	self->execution_hosts = fsd_strdup(i->value);
	354	}
	355	break;
[26]	356	#else
	357	case PBSDRMAA_ATTR_EXECUTION_VNODE:
	358	if (!self->execution_hosts) {
	359	fsd_log_debug(("execution_hosts = %s", i->value));
	360	self->execution_hosts = fsd_strdup(i->value);
	361	}
	362	break;
	363	#endif
[1]	364	case PBSDRMAA_ATTR_START_TIME:
	365	{
	366	long unsigned int start_time;
	367	if (self->start_time == 0 && sscanf(i->value, "%lu", &start_time) == 1)
	368	self->start_time = start_time;
	369	break;
	370	}
	371	case PBSDRMAA_ATTR_MTIME:
	372	if (sscanf(i->value, "%lu", &modify_time) != 1)
	373	modify_time = 0;
	374	break;
	375	}
	376	}
	377
	378	if( pbs_state )
	379	fsd_log_debug(( "pbs_state: %c", pbs_state ));
	380
	381	if( exit_status != -2 )
	382	{
	383	fsd_log_debug(( "exit_status: %d", exit_status ));
	384	self->exit_status = exit_status;
	385	}
	386	if(pbs_state){
	387	switch( pbs_state )
	388	{
	389	case 'C': /* Job is completed after having run. */
	390	self->flags &= FSD_JOB_TERMINATED_MASK;
	391	self->flags \|= FSD_JOB_TERMINATED;
	392	if (exit_status != -2) { /* has exit code */
	393	if( self->exit_status == 0)
	394	self->state = DRMAA_PS_DONE;
	395	else
	396	self->state = DRMAA_PS_FAILED;
	397	} else {
	398	self->state = DRMAA_PS_FAILED;
	399	self->exit_status = -1;
	400	}
[22]	401	if (modify_time != 0)
	402	self->end_time = modify_time; /* take last modify time as end time */
	403	else
	404	self->end_time = time(NULL);
	405
	406	if (self->start_time == 0)
	407	self->start_time = self->end_time;
	408
[1]	409	break;
	410	case 'E': /* Job is exiting after having run. - MM: ignore exiting state (transient state) - outputs might have not been transfered yet,
	411	MM2: mark job as running if current job status is undetermined - fix "ps after job was ripped" */
	412	if (self->state == DRMAA_PS_UNDETERMINED)
	413	self->state = DRMAA_PS_RUNNING;
	414	break;
	415	case 'H': /* Job is held. */
	416	self->state = DRMAA_PS_USER_ON_HOLD;
	417	self->flags \|= FSD_JOB_HOLD;
	418	break;
	419	case 'Q': /* Job is queued, eligible to run or routed. */
	420	case 'W': /* Job is waiting for its execution time to be reached. */
	421	self->state = DRMAA_PS_QUEUED_ACTIVE;
	422	self->flags &= ~FSD_JOB_HOLD;
	423	break;
	424	case 'R': /* Job is running. */
	425	case 'T': /* Job is being moved to new location (?). */
	426	{
	427	if( self->flags & FSD_JOB_SUSPENDED )
	428	self->state = DRMAA_PS_USER_SUSPENDED;
	429	else
	430	self->state = DRMAA_PS_RUNNING;
	431	break;
	432	}
	433	case 'S': /* (Unicos only) job is suspend. */
	434	self->state = DRMAA_PS_SYSTEM_SUSPENDED;
	435	break;
	436	case 0: default:
	437	self->state = DRMAA_PS_UNDETERMINED;
	438	break;
	439	}
	440	}
	441	fsd_log_debug(( "job_ps: %s", drmaa_job_ps_to_str(self->state) ));
	442
	443	{
	444	int hours, minutes, seconds;
	445	long mem;
[12]	446	if( cpu_usage && sscanf( cpu_usage, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
[1]	447	{
	448	self->cpu_usage = 60( 60hours + minutes ) + seconds;
	449	fsd_log_debug(( "cpu_usage: %s=%lds", cpu_usage, self->cpu_usage ));
	450	}
	451	if( mem_usage && sscanf( mem_usage, "%ldkb", &mem ) == 1 )
	452	{
	453	self->mem_usage = 1024*mem;
	454	fsd_log_debug(( "mem_usage: %s=%ldB", mem_usage, self->mem_usage ));
	455	}
	456	if( vmem_usage && sscanf( vmem_usage, "%ldkb", &mem ) == 1 )
	457	{
	458	self->vmem_usage = 1024*mem;
	459	fsd_log_debug(( "vmem_usage: %s=%ldB", vmem_usage, self->vmem_usage ));
	460	}
[12]	461	if( walltime && sscanf( walltime, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
[1]	462	{
	463	self->walltime = 60( 60hours + minutes ) + seconds;
	464	fsd_log_debug(( "walltime: %s=%lds", walltime, self->walltime ));
	465	}
	466	}
	467	}
	468
	469	void
	470	pbsdrmaa_job_on_missing( fsd_job_t *self )
	471	{
[7]	472	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t)self->session;
[12]	473
[54]	474	if( pbssession->pbs_home != NULL && pbssession->super.wait_thread_started && self->submit_time)
[40]	475	fsd_log_info(("Job on missing but WT is running. Skipping...")); /* TODO: try to provide implementation that uses accounting/server log files */
[33]	476	else
[7]	477	pbsdrmaa_job_on_missing_standard( self );
	478	}
	479
	480	void
	481	pbsdrmaa_job_on_missing_standard( fsd_job_t *self )
	482	{
[1]	483	fsd_drmaa_session_t *session = self->session;
[54]	484	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t )session;
	485	int exit_status = -1;
[7]	486
[1]	487	fsd_log_enter(( "({job_id=%s})", self->job_id ));
[12]	488	fsd_log_warning(( "Job %s missing from DRM queue", self->job_id ));
[1]	489
[54]	490	fsd_log_info(( "job_on_missing: last job_ps: %s (0x%02x)", drmaa_job_ps_to_str(self->state), self->state));
[1]	491
[54]	492	if( (exit_status = pbsdrmaa_job_read_exit_status(self->job_id, pbssession->job_exit_status_file_prefix)) == 0 )
[7]	493	{
[1]	494	self->state = DRMAA_PS_DONE;
[54]	495	self->exit_status = exit_status;
[7]	496	}
[1]	497	else
[54]	498	{
[1]	499	self->state = DRMAA_PS_FAILED;
[54]	500	self->exit_status = exit_status;
[7]	501	}
[1]	502
[12]	503	fsd_cond_broadcast( &self->status_cond);
	504
[54]	505	fsd_log_return(( "; job_ps=%s, exit_status=%d", drmaa_job_ps_to_str(self->state), self->exit_status ));
[7]	506	}
	507
[54]	508	int
	509	pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix)
	510	{
[57]	511	char status_file = NULL, start_file = NULL;
[54]	512	FILE *fhandle = NULL;
	513	int exit_status = -1;
	514
	515	fsd_log_enter(("({job_id=%s, job_state_dir_prefix=%s})", job_id, job_state_dir_prefix));
	516
[55]	517	status_file = fsd_asprintf("%s/%s.exitcode", job_state_dir_prefix, job_id);
[57]	518	start_file = fsd_asprintf("%s/%s.started", job_state_dir_prefix, job_id);
[54]	519
	520	if ((fhandle = fopen(status_file, "r")) == NULL)
	521	{
[56]	522	struct stat tmpstat;
	523
[54]	524	fsd_log_error(("Failed to open job status file: %s", status_file));
[56]	525	if (stat(start_file, &tmpstat) == 0 && (tmpstat.st_mode & S_IFREG))
	526	{
	527	exit_status = 143; /* SIGTERM */
[57]	528	fsd_log_info(("But start file exist %s. Assuming that job was killed (exit_status=%d).", start_file, exit_status));
[56]	529	}
[57]	530	else
	531	{
	532	fsd_log_error(("Start file not found: %s", start_file));
	533	}
	534
	535
[54]	536	}
	537	else
	538	{
	539	(void)fscanf(fhandle, "%d", &exit_status); /on error exit_status == -1 /
	540	fclose(fhandle);
	541	}
	542
	543	fsd_free(status_file);
[56]	544	fsd_free(start_file);
[54]	545
	546	return exit_status;
	547	}
	548

Note: See TracBrowser for help on using the repository browser.

Download in other formats: