Context Navigation

source: trunk/pbs_drmaa/job.c @ 71

Revision 71, 16.4 KB checked in by mmamonski, 12 years ago (diff)
PBS Pro PBSE_HISTJOBID handled properlly
Property svn:keywords set to `Id`

Line
1	/* $Id$ */
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>
28	#include <sys/stat.h>
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>
36	#include <pbs_drmaa/log_reader.h>
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
46	= "$Id$";
47	#endif
48
49
50	static void pbsdrmaa_job_control( fsd_job_t *self, int action );
51
52	static void pbsdrmaa_job_update_status( fsd_job_t *self );
53
54	static void pbsdrmaa_job_on_missing( fsd_job_t *self );
55
56	static void pbsdrmaa_job_on_missing_standard( fsd_job_t *self );
57
58	static void pbsdrmaa_job_update( fsd_job_t self, struct batch_status );
59
60	static int pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix);
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->missing_time = 0;
71	self->update = pbsdrmaa_job_update;
72	return (fsd_job_t*)self;
73	}
74
75
76	static void
77	pbsdrmaa_job_control( fsd_job_t *self, int action )
78	{
79	volatile bool conn_lock = false;
80	pbsdrmaa_session_t session = (pbsdrmaa_session_t)self->session;
81	const char *job_id = self->job_id;
82	const char *apicall = NULL;
83	int rc = PBSE_NONE;
84
85	fsd_log_enter(( "({job_id=%s}, action=%d)",
86	self->job_id, action ));
87
88	TRY
89	{
90	int tries_left = session->max_retries_count;
91	int sleep_time = 1;
92
93	conn_lock = fsd_mutex_lock( &self->session->drm_connection_mutex );
94
95	/TODO reconnect /
96	while ( true )
97	{
98	switch( action )
99	{
100	/*
101	* We cannot know whether we did suspend job
102	* in other way than remembering this inside DRMAA session.
103	*/
104	case DRMAA_CONTROL_SUSPEND:
105	apicall = "pbs_sigjob";
106	rc = pbs_sigjob( session->pbs_conn, (char*)job_id,
107	"SIGSTOP", NULL );
108	fsd_log_info(("pbs_sigjob(%s, SIGSTOP) =%d", job_id, rc));
109	if( rc == PBSE_NONE )
110	self->flags \|= FSD_JOB_SUSPENDED;
111	break;
112	case DRMAA_CONTROL_RESUME:
113	apicall = "pbs_sigjob";
114	rc = pbs_sigjob( session->pbs_conn, (char*)job_id,
115	"SIGCONT", NULL );
116	fsd_log_info(("pbs_sigjob(%s, SIGCONT) =%d", job_id, rc));
117	if( rc == PBSE_NONE )
118	self->flags &= ~FSD_JOB_SUSPENDED;
119	break;
120	case DRMAA_CONTROL_HOLD:
121	apicall = "pbs_holdjob";
122	rc = pbs_holdjob( session->pbs_conn, (char*)job_id,
123	USER_HOLD, NULL );
124	fsd_log_info(("pbs_sigjob(%s, SIGHOLD) =%d", job_id, rc));
125	if( rc == PBSE_NONE )
126	self->flags \|= FSD_JOB_HOLD;
127	break;
128	case DRMAA_CONTROL_RELEASE:
129	apicall = "pbs_rlsjob";
130	rc = pbs_rlsjob( session->pbs_conn, (char*)job_id,
131	USER_HOLD, NULL );
132	fsd_log_info(("pbs_rlsjob(%s) =%d", job_id, rc));
133	if( rc == PBSE_NONE )
134	self->flags &= FSD_JOB_HOLD;
135	break;
136	case DRMAA_CONTROL_TERMINATE:
137	apicall = "pbs_deljob";
138	rc = pbs_deljob( session->pbs_conn, (char*)job_id, NULL );
139	fsd_log_info(("pbs_deljob(%s) =%d", job_id, rc));
140	/* Torque:
141	* deldelay=N -- delay between SIGTERM and SIGKILL (default 0) */
142	if( rc == PBSE_NONE )
143	{
144	self->flags &= FSD_JOB_TERMINATED_MASK;
145	if( (self->flags & FSD_JOB_TERMINATED) == 0 )
146	self->flags \|= FSD_JOB_TERMINATED \| FSD_JOB_ABORTED;
147	}
148	break;
149	}
150
151	retry_connect:
152	if ( rc == PBSE_NONE )
153	break;
154	else if (( rc == PBSE_INTERNAL \|\| rc == PBSE_PROTOCOL \|\| rc == PBSOLDE_PROTOCOL \|\| rc == PBSE_EXPIRED \|\| rc == PBSOLDE_EXPIRED) && (tries_left--))
155	{
156	if (rc == PBSE_PROTOCOL \|\| rc == PBSE_EXPIRED \|\| rc == PBSOLDE_PROTOCOL \|\| rc == PBSOLDE_EXPIRED)
157	{
158	if ( session->pbs_conn >= 0)
159	pbs_disconnect( session->pbs_conn );
160
161	sleep( sleep_time++ );
162
163	session->pbs_conn = pbs_connect( session->super.contact );
164
165	if (session->pbs_conn < 0)
166	goto retry_connect;
167
168	fsd_log_info(( "pbs_connect(%s) =%d", session->super.contact, session->pbs_conn ));
169	}
170	else /* PBSE_INTERNAL */
171	{
172	/*
173	* In PBS Pro pbs_sigjob raises internal server error (PBSE_INTERNAL)
174	* when job just changed its state to running.
175	*/
176	sleep( sleep_time++ );
177	}
178	fsd_log_debug(( "repeating request (%d of %d)", tries_left, session->max_retries_count));
179	}
180	else
181	pbsdrmaa_exc_raise_pbs( apicall );
182	} /* end while */
183	}
184	FINALLY
185	{
186	if( conn_lock )
187	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
188	}
189	END_TRY
190
191	fsd_log_return((""));
192	}
193
194
195	void
196	pbsdrmaa_job_update_status( fsd_job_t *self )
197	{
198	volatile bool conn_lock = false;
199	struct batch_status *volatile status = NULL;
200	pbsdrmaa_session_t session = (pbsdrmaa_session_t)self->session;
201	int tries_left = session->max_retries_count;
202	int sleep_time = 1;
203
204	fsd_log_enter(( "({job_id=%s})", self->job_id ));
205
206	TRY
207	{
208	conn_lock = fsd_mutex_lock( &self->session->drm_connection_mutex );
209	retry:
210	if (session->pbs_conn < 0) {
211	fsd_log_info(("No connection with pbs. Reconnecting"));
212	goto retry_connect;
213	}
214
215
216	#ifdef PBS_PROFESSIONAL
217	status = pbs_statjob( session->pbs_conn, self->job_id, NULL, NULL );
218	#else
219	status = pbs_statjob( session->pbs_conn, self->job_id, session->status_attrl, NULL );
220	#endif
221	fsd_log_info(( "pbs_statjob(fd=%d, job_id=%s, attribs={...}) =%p",
222	session->pbs_conn, self->job_id, (void*)status ));
223	if( status == NULL )
224	{
225
226	#ifndef PBS_PROFESSIONAL
227	if ( pbs_errno != PBSE_UNKJOBID )
228	fsd_log_error(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
229	else
230	fsd_log_debug(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
231	#else
232	# ifndef PBS_PROFESSIONAL_NO_LOG
233	if ( pbs_errno != PBSE_UNKJOBID && pbs_errno != PBSE_HISTJOBID )
234	fsd_log_error(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
235	else
236	fsd_log_debug(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
237	# else
238	if ( pbs_errno != PBSE_UNKJOBID && pbs_errno != PBSE_HISTJOBID )
239	fsd_log_error(("pbs_statjob error: %d", pbs_errno));
240	else
241	fsd_log_debug(("pbs_statjob error: %d", pbs_errno));
242	# endif
243	#endif
244
245	switch( pbs_errno )
246	{
247	case PBSE_UNKJOBID:
248	#ifdef PBS_PROFESSIONAL
249	case PBSE_HISTJOBID:
250	#endif
251	break;
252	case PBSE_PROTOCOL:
253	#if PBSOLDE_PROTOCOL != PBSE_PROTOCOL
254	case PBSOLDE_PROTOCOL:
255	#endif
256	case PBSE_EXPIRED:
257	#if PBSOLDE_EXPIRED != PBSE_EXPIRED
258	case PBSOLDE_EXPIRED:
259	#endif
260	if ( session->pbs_conn >= 0 )
261	pbs_disconnect( session->pbs_conn );
262	fsd_log_info(("Protocol error. Reconnecting..."));
263	retry_connect:
264	sleep(sleep_time++);
265	session->pbs_conn = pbs_connect( session->super.contact );
266	if( session->pbs_conn < 0 )
267	{
268	if (tries_left--) {
269	fsd_log_info(("Retrying... (%d tries left)", tries_left));
270	goto retry_connect;
271	} else {
272	fsd_log_error(("No more tries left... Throwing exception"));
273	pbsdrmaa_exc_raise_pbs( "pbs_connect" );
274	}
275	}
276	else
277	{
278	goto retry;
279	}
280	default:
281	pbsdrmaa_exc_raise_pbs( "pbs_statjob" );
282	break;
283	case 0: /* ? */
284	fsd_exc_raise_code( FSD_ERRNO_INTERNAL_ERROR );
285	break;
286	}
287
288	}
289
290	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
291
292
293	if( status != NULL )
294	{
295	((pbsdrmaa_job_t*)self)->update( self, status );
296	}
297	else if( self->state < DRMAA_PS_DONE )
298	{
299	self->on_missing( self );
300	}
301	}
302	FINALLY
303	{
304	if( conn_lock )
305	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
306	if( status != NULL )
307	pbs_statfree( status );
308	}
309	END_TRY
310
311	fsd_log_return((""));
312	}
313
314
315	void
316	pbsdrmaa_job_update( fsd_job_t self, struct batch_status status )
317	{
318	struct attrl *attribs = status->attribs;
319	struct attrl *i = NULL;
320	char pbs_state = 0;
321	int exit_status = -2;
322	const char *cpu_usage = NULL;
323	const char *mem_usage = NULL;
324	const char *vmem_usage = NULL;
325	const char *walltime = NULL;
326	long unsigned int modify_time = 0;
327
328	fsd_log_enter(( "({job_id=%s})", self->job_id ));
329	#ifdef DEBUGGING
330	pbsdrmaa_dump_attrl( attribs, NULL );
331	#endif
332	fsd_assert( !strcmp( self->job_id, status->name ) );
333
334	for( i = attribs; i != NULL; i = i->next )
335	{
336	int attr;
337	attr = pbsdrmaa_pbs_attrib_by_name( i->name );
338	switch( attr )
339	{
340	case PBSDRMAA_ATTR_JOB_STATE:
341	pbs_state = i->value[0];
342	break;
343	case PBSDRMAA_ATTR_EXIT_STATUS:
344	exit_status = fsd_atoi( i->value );
345	break;
346	case PBSDRMAA_ATTR_RESOURCES_USED:
347	if( !strcmp( i->resource, "cput" ) )
348	cpu_usage = i->value;
349	else if( !strcmp( i->resource, "mem" ) )
350	mem_usage = i->value;
351	else if( !strcmp( i->resource, "vmem" ) )
352	vmem_usage = i->value;
353	else if( !strcmp( i->resource, "walltime" ) )
354	walltime = i->value;
355	break;
356	case PBSDRMAA_ATTR_QUEUE:
357	if (!self->queue)
358	self->queue = fsd_strdup(i->value);
359	break;
360	case PBSDRMAA_ATTR_ACCOUNT_NAME:
361	if (!self->project)
362	self->project = fsd_strdup(i->value);
363	break;
364	#ifndef PBS_PROFESSIONAL
365	case PBSDRMAA_ATTR_EXECUTION_HOST:
366	if (!self->execution_hosts) {
367	fsd_log_debug(("execution_hosts = %s", i->value));
368	self->execution_hosts = fsd_strdup(i->value);
369	}
370	break;
371	#else
372	case PBSDRMAA_ATTR_EXECUTION_VNODE:
373	if (!self->execution_hosts) {
374	fsd_log_debug(("execution_hosts = %s", i->value));
375	self->execution_hosts = fsd_strdup(i->value);
376	}
377	break;
378	#endif
379	case PBSDRMAA_ATTR_START_TIME:
380	{
381	long unsigned int start_time;
382	if (self->start_time == 0 && sscanf(i->value, "%lu", &start_time) == 1)
383	self->start_time = start_time;
384	break;
385	}
386	case PBSDRMAA_ATTR_MTIME:
387	if (sscanf(i->value, "%lu", &modify_time) != 1)
388	modify_time = 0;
389	break;
390	}
391	}
392
393	if( pbs_state )
394	fsd_log_debug(( "pbs_state: %c", pbs_state ));
395
396	if( exit_status != -2 )
397	{
398	fsd_log_debug(( "exit_status: %d", exit_status ));
399	self->exit_status = exit_status;
400	}
401	if(pbs_state){
402	switch( pbs_state )
403	{
404	case 'C': /* Job is completed after having run. */
405	self->flags &= FSD_JOB_TERMINATED_MASK;
406	self->flags \|= FSD_JOB_TERMINATED;
407	if (exit_status != -2) { /* has exit code */
408	if( self->exit_status == 0)
409	self->state = DRMAA_PS_DONE;
410	else
411	self->state = DRMAA_PS_FAILED;
412	} else {
413	self->state = DRMAA_PS_FAILED;
414	self->exit_status = -1;
415	}
416	if (modify_time != 0)
417	self->end_time = modify_time; /* take last modify time as end time */
418	else
419	self->end_time = time(NULL);
420
421	if (self->start_time == 0)
422	self->start_time = self->end_time;
423
424	break;
425	case 'E': /* Job is exiting after having run. - MM: ignore exiting state (transient state) - outputs might have not been transfered yet,
426	MM2: mark job as running if current job status is undetermined - fix "ps after job was ripped" */
427	if (self->state == DRMAA_PS_UNDETERMINED)
428	self->state = DRMAA_PS_RUNNING;
429	break;
430	case 'H': /* Job is held. */
431	self->state = DRMAA_PS_USER_ON_HOLD;
432	self->flags \|= FSD_JOB_HOLD;
433	break;
434	case 'Q': /* Job is queued, eligible to run or routed. */
435	case 'W': /* Job is waiting for its execution time to be reached. */
436	self->state = DRMAA_PS_QUEUED_ACTIVE;
437	self->flags &= ~FSD_JOB_HOLD;
438	break;
439	case 'R': /* Job is running. */
440	case 'T': /* Job is being moved to new location (?). */
441	{
442	if( self->flags & FSD_JOB_SUSPENDED )
443	self->state = DRMAA_PS_USER_SUSPENDED;
444	else
445	self->state = DRMAA_PS_RUNNING;
446	break;
447	}
448	case 'S': /* (Unicos only) job is suspend. */
449	self->state = DRMAA_PS_SYSTEM_SUSPENDED;
450	break;
451	case 0: default:
452	self->state = DRMAA_PS_UNDETERMINED;
453	break;
454	}
455	}
456	fsd_log_debug(( "job_ps: %s", drmaa_job_ps_to_str(self->state) ));
457
458	{
459	int hours, minutes, seconds;
460	long mem;
461	if( cpu_usage && sscanf( cpu_usage, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
462	{
463	self->cpu_usage = 60( 60hours + minutes ) + seconds;
464	fsd_log_debug(( "cpu_usage: %s=%lds", cpu_usage, self->cpu_usage ));
465	}
466	if( mem_usage && sscanf( mem_usage, "%ldkb", &mem ) == 1 )
467	{
468	self->mem_usage = 1024*mem;
469	fsd_log_debug(( "mem_usage: %s=%ldB", mem_usage, self->mem_usage ));
470	}
471	if( vmem_usage && sscanf( vmem_usage, "%ldkb", &mem ) == 1 )
472	{
473	self->vmem_usage = 1024*mem;
474	fsd_log_debug(( "vmem_usage: %s=%ldB", vmem_usage, self->vmem_usage ));
475	}
476	if( walltime && sscanf( walltime, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
477	{
478	self->walltime = 60( 60hours + minutes ) + seconds;
479	fsd_log_debug(( "walltime: %s=%lds", walltime, self->walltime ));
480	}
481	}
482	}
483
484	void
485	pbsdrmaa_job_on_missing( fsd_job_t *self )
486	{
487	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t)self->session;
488	pbsdrmaa_job_t pbsself = (pbsdrmaa_job_t )self;
489
490	if (!pbsself->missing_time)
491	{
492	pbsself->missing_time = time(NULL);
493	}
494
495	fsd_log_info(("pbsdrmaa_job_on_missing: pbs_home=%s, wait_thread_started=%d, submit_time=%d, missing_time=%d",
496	pbssession->pbs_home,
497	pbssession->super.wait_thread_started,
498	(int)self->submit_time,
499	(int)pbsself->missing_time));
500
501	#define DRMAA_MAX_MISSING_TIME (30)
502	if( pbssession->pbs_home != NULL && pbssession->super.wait_thread_started && self->submit_time && (time(NULL) - pbsself->missing_time < DRMAA_MAX_MISSING_TIME))
503	fsd_log_info(("Job on missing but WT is running. Skipping...")); /* TODO: try to provide implementation that uses accounting/server log files */
504	else
505	pbsdrmaa_job_on_missing_standard( self );
506	}
507
508	void
509	pbsdrmaa_job_on_missing_standard( fsd_job_t *self )
510	{
511	fsd_drmaa_session_t *session = self->session;
512	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t )session;
513	int exit_status = -1;
514
515	fsd_log_enter(( "({job_id=%s})", self->job_id ));
516	fsd_log_warning(( "Job %s missing from DRM queue", self->job_id ));
517
518	fsd_log_info(( "job_on_missing: last job_ps: %s (0x%02x)", drmaa_job_ps_to_str(self->state), self->state));
519
520	if( (exit_status = pbsdrmaa_job_read_exit_status(self->job_id, pbssession->job_exit_status_file_prefix)) == 0 )
521	{
522	self->state = DRMAA_PS_DONE;
523	self->exit_status = exit_status;
524	}
525	else
526	{
527	self->state = DRMAA_PS_FAILED;
528	self->exit_status = exit_status;
529	}
530	fsd_log_info(("job_on_missing evaluation result: state=%d exit_status=%d", self->state, self->exit_status));
531
532	fsd_cond_broadcast( &self->status_cond);
533	fsd_cond_broadcast( &self->session->wait_condition );
534
535	fsd_log_return(( "; job_ps=%s, exit_status=%d", drmaa_job_ps_to_str(self->state), self->exit_status ));
536	}
537
538	int
539	pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix)
540	{
541	char status_file = NULL, start_file = NULL;
542	FILE *fhandle = NULL;
543	int exit_status = -1;
544
545	fsd_log_enter(("({job_id=%s, job_state_dir_prefix=%s})", job_id, job_state_dir_prefix));
546
547	status_file = fsd_asprintf("%s/%s.exitcode", job_state_dir_prefix, job_id);
548	start_file = fsd_asprintf("%s/%s.started", job_state_dir_prefix, job_id);
549
550	if ((fhandle = fopen(status_file, "r")) == NULL)
551	{
552	struct stat tmpstat;
553
554	fsd_log_error(("Failed to open job status file: %s", status_file));
555	if (stat(start_file, &tmpstat) == 0 && (tmpstat.st_mode & S_IFREG))
556	{
557	exit_status = 143; /* SIGTERM */
558	fsd_log_info(("But start file exist %s. Assuming that job was killed (exit_status=%d).", start_file, exit_status));
559	}
560	else
561	{
562	fsd_log_error(("Start file not found: %s", start_file));
563	}
564
565
566	}
567	else
568	{
569	(void)fscanf(fhandle, "%d", &exit_status); /on error exit_status == -1 /
570	fclose(fhandle);
571	}
572
573	fsd_free(status_file);
574	fsd_free(start_file);
575
576	return exit_status;
577	}
578

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

Download in other formats: