Context Navigation

source: trunk/pbs_drmaa/job.c @ 65

Revision 65, 16.0 KB checked in by mmamonski, 12 years ago (diff)
try to maintain binary compatibility with older version of Torque
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
211	#ifdef PBS_PROFESSIONAL
212	status = pbs_statjob( session->pbs_conn, self->job_id, NULL, NULL );
213	#else
214	status = pbs_statjob( session->pbs_conn, self->job_id, session->status_attrl, NULL );
215	#endif
216	fsd_log_info(( "pbs_statjob(fd=%d, job_id=%s, attribs={...}) =%p",
217	session->pbs_conn, self->job_id, (void*)status ));
218	if( status == NULL )
219	{
220
221	#ifndef PBS_PROFESSIONAL
222	if ( pbs_errno != PBSE_UNKJOBID )
223	fsd_log_error(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
224	else
225	fsd_log_debug(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
226	#else
227	# ifndef PBS_PROFESSIONAL_NO_LOG
228	if ( pbs_errno != PBSE_UNKJOBID )
229	fsd_log_error(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
230	else
231	fsd_log_debug(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
232	# else
233	if ( pbs_errno != PBSE_UNKJOBID )
234	fsd_log_error(("pbs_statjob error: %d", pbs_errno));
235	else
236	fsd_log_debug(("pbs_statjob error: %d", pbs_errno));
237	# endif
238	#endif
239
240	switch( pbs_errno )
241	{
242	case PBSE_UNKJOBID:
243	break;
244	case PBSE_PROTOCOL:
245	case PBSOLDE_PROTOCOL:
246	case PBSE_EXPIRED:
247	case PBSOLDE_EXPIRED:
248	if ( session->pbs_conn >= 0 )
249	pbs_disconnect( session->pbs_conn );
250	fsd_log_info(("Protocol error. Reconnecting..."));
251	retry_connect:
252	sleep(sleep_time++);
253	session->pbs_conn = pbs_connect( session->super.contact );
254	if( session->pbs_conn < 0 )
255	{
256	if (tries_left--) {
257	fsd_log_info(("%d tries left. Retrying...", tries_left));
258	goto retry_connect;
259	} else {
260	fsd_log_error(("No more tries left... Throwing exception"));
261	pbsdrmaa_exc_raise_pbs( "pbs_connect" );
262	}
263	}
264	else
265	{
266	goto retry;
267	}
268	default:
269	pbsdrmaa_exc_raise_pbs( "pbs_statjob" );
270	break;
271	case 0: /* ? */
272	fsd_exc_raise_code( FSD_ERRNO_INTERNAL_ERROR );
273	break;
274	}
275
276	}
277
278	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
279
280
281	if( status != NULL )
282	{
283	((pbsdrmaa_job_t*)self)->update( self, status );
284	}
285	else if( self->state < DRMAA_PS_DONE )
286	{
287	self->on_missing( self );
288	}
289	}
290	FINALLY
291	{
292	if( conn_lock )
293	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
294	if( status != NULL )
295	pbs_statfree( status );
296	}
297	END_TRY
298
299	fsd_log_return((""));
300	}
301
302
303	void
304	pbsdrmaa_job_update( fsd_job_t self, struct batch_status status )
305	{
306	struct attrl *attribs = status->attribs;
307	struct attrl *i = NULL;
308	char pbs_state = 0;
309	int exit_status = -2;
310	const char *cpu_usage = NULL;
311	const char *mem_usage = NULL;
312	const char *vmem_usage = NULL;
313	const char *walltime = NULL;
314	long unsigned int modify_time = 0;
315
316	fsd_log_enter(( "({job_id=%s})", self->job_id ));
317	#ifdef DEBUGGING
318	pbsdrmaa_dump_attrl( attribs, NULL );
319	#endif
320	fsd_assert( !strcmp( self->job_id, status->name ) );
321
322	for( i = attribs; i != NULL; i = i->next )
323	{
324	int attr;
325	attr = pbsdrmaa_pbs_attrib_by_name( i->name );
326	switch( attr )
327	{
328	case PBSDRMAA_ATTR_JOB_STATE:
329	pbs_state = i->value[0];
330	break;
331	case PBSDRMAA_ATTR_EXIT_STATUS:
332	exit_status = fsd_atoi( i->value );
333	break;
334	case PBSDRMAA_ATTR_RESOURCES_USED:
335	if( !strcmp( i->resource, "cput" ) )
336	cpu_usage = i->value;
337	else if( !strcmp( i->resource, "mem" ) )
338	mem_usage = i->value;
339	else if( !strcmp( i->resource, "vmem" ) )
340	vmem_usage = i->value;
341	else if( !strcmp( i->resource, "walltime" ) )
342	walltime = i->value;
343	break;
344	case PBSDRMAA_ATTR_QUEUE:
345	if (!self->queue)
346	self->queue = fsd_strdup(i->value);
347	break;
348	case PBSDRMAA_ATTR_ACCOUNT_NAME:
349	if (!self->project)
350	self->project = fsd_strdup(i->value);
351	break;
352	#ifndef PBS_PROFESSIONAL
353	case PBSDRMAA_ATTR_EXECUTION_HOST:
354	if (!self->execution_hosts) {
355	fsd_log_debug(("execution_hosts = %s", i->value));
356	self->execution_hosts = fsd_strdup(i->value);
357	}
358	break;
359	#else
360	case PBSDRMAA_ATTR_EXECUTION_VNODE:
361	if (!self->execution_hosts) {
362	fsd_log_debug(("execution_hosts = %s", i->value));
363	self->execution_hosts = fsd_strdup(i->value);
364	}
365	break;
366	#endif
367	case PBSDRMAA_ATTR_START_TIME:
368	{
369	long unsigned int start_time;
370	if (self->start_time == 0 && sscanf(i->value, "%lu", &start_time) == 1)
371	self->start_time = start_time;
372	break;
373	}
374	case PBSDRMAA_ATTR_MTIME:
375	if (sscanf(i->value, "%lu", &modify_time) != 1)
376	modify_time = 0;
377	break;
378	}
379	}
380
381	if( pbs_state )
382	fsd_log_debug(( "pbs_state: %c", pbs_state ));
383
384	if( exit_status != -2 )
385	{
386	fsd_log_debug(( "exit_status: %d", exit_status ));
387	self->exit_status = exit_status;
388	}
389	if(pbs_state){
390	switch( pbs_state )
391	{
392	case 'C': /* Job is completed after having run. */
393	self->flags &= FSD_JOB_TERMINATED_MASK;
394	self->flags \|= FSD_JOB_TERMINATED;
395	if (exit_status != -2) { /* has exit code */
396	if( self->exit_status == 0)
397	self->state = DRMAA_PS_DONE;
398	else
399	self->state = DRMAA_PS_FAILED;
400	} else {
401	self->state = DRMAA_PS_FAILED;
402	self->exit_status = -1;
403	}
404	if (modify_time != 0)
405	self->end_time = modify_time; /* take last modify time as end time */
406	else
407	self->end_time = time(NULL);
408
409	if (self->start_time == 0)
410	self->start_time = self->end_time;
411
412	break;
413	case 'E': /* Job is exiting after having run. - MM: ignore exiting state (transient state) - outputs might have not been transfered yet,
414	MM2: mark job as running if current job status is undetermined - fix "ps after job was ripped" */
415	if (self->state == DRMAA_PS_UNDETERMINED)
416	self->state = DRMAA_PS_RUNNING;
417	break;
418	case 'H': /* Job is held. */
419	self->state = DRMAA_PS_USER_ON_HOLD;
420	self->flags \|= FSD_JOB_HOLD;
421	break;
422	case 'Q': /* Job is queued, eligible to run or routed. */
423	case 'W': /* Job is waiting for its execution time to be reached. */
424	self->state = DRMAA_PS_QUEUED_ACTIVE;
425	self->flags &= ~FSD_JOB_HOLD;
426	break;
427	case 'R': /* Job is running. */
428	case 'T': /* Job is being moved to new location (?). */
429	{
430	if( self->flags & FSD_JOB_SUSPENDED )
431	self->state = DRMAA_PS_USER_SUSPENDED;
432	else
433	self->state = DRMAA_PS_RUNNING;
434	break;
435	}
436	case 'S': /* (Unicos only) job is suspend. */
437	self->state = DRMAA_PS_SYSTEM_SUSPENDED;
438	break;
439	case 0: default:
440	self->state = DRMAA_PS_UNDETERMINED;
441	break;
442	}
443	}
444	fsd_log_debug(( "job_ps: %s", drmaa_job_ps_to_str(self->state) ));
445
446	{
447	int hours, minutes, seconds;
448	long mem;
449	if( cpu_usage && sscanf( cpu_usage, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
450	{
451	self->cpu_usage = 60( 60hours + minutes ) + seconds;
452	fsd_log_debug(( "cpu_usage: %s=%lds", cpu_usage, self->cpu_usage ));
453	}
454	if( mem_usage && sscanf( mem_usage, "%ldkb", &mem ) == 1 )
455	{
456	self->mem_usage = 1024*mem;
457	fsd_log_debug(( "mem_usage: %s=%ldB", mem_usage, self->mem_usage ));
458	}
459	if( vmem_usage && sscanf( vmem_usage, "%ldkb", &mem ) == 1 )
460	{
461	self->vmem_usage = 1024*mem;
462	fsd_log_debug(( "vmem_usage: %s=%ldB", vmem_usage, self->vmem_usage ));
463	}
464	if( walltime && sscanf( walltime, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
465	{
466	self->walltime = 60( 60hours + minutes ) + seconds;
467	fsd_log_debug(( "walltime: %s=%lds", walltime, self->walltime ));
468	}
469	}
470	}
471
472	void
473	pbsdrmaa_job_on_missing( fsd_job_t *self )
474	{
475	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t)self->session;
476	pbsdrmaa_job_t pbsself = (pbsdrmaa_job_t )self;
477
478	if (!pbsself->missing_time)
479	{
480	pbsself->missing_time = time(NULL);
481	}
482
483	fsd_log_info(("pbsdrmaa_job_on_missing: pbs_home=%s, wait_thread_started=%d, submit_time=%d, missing_time=%d",
484	pbssession->pbs_home,
485	pbssession->super.wait_thread_started,
486	self->submit_time,
487	pbsself->missing_time));
488
489	#define DRMAA_MAX_MISSING_TIME (30)
490	if( pbssession->pbs_home != NULL && pbssession->super.wait_thread_started && self->submit_time && (time(NULL) - pbsself->missing_time < DRMAA_MAX_MISSING_TIME))
491	fsd_log_info(("Job on missing but WT is running. Skipping...")); /* TODO: try to provide implementation that uses accounting/server log files */
492	else
493	pbsdrmaa_job_on_missing_standard( self );
494	}
495
496	void
497	pbsdrmaa_job_on_missing_standard( fsd_job_t *self )
498	{
499	fsd_drmaa_session_t *session = self->session;
500	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t )session;
501	int exit_status = -1;
502
503	fsd_log_enter(( "({job_id=%s})", self->job_id ));
504	fsd_log_warning(( "Job %s missing from DRM queue", self->job_id ));
505
506	fsd_log_info(( "job_on_missing: last job_ps: %s (0x%02x)", drmaa_job_ps_to_str(self->state), self->state));
507
508	if( (exit_status = pbsdrmaa_job_read_exit_status(self->job_id, pbssession->job_exit_status_file_prefix)) == 0 )
509	{
510	self->state = DRMAA_PS_DONE;
511	self->exit_status = exit_status;
512	}
513	else
514	{
515	self->state = DRMAA_PS_FAILED;
516	self->exit_status = exit_status;
517	}
518	fsd_log_info(("job_on_missing evaluation result: state=%d exit_status=%d", self->state, self->exit_status));
519
520	fsd_cond_broadcast( &self->status_cond);
521	fsd_cond_broadcast( &self->session->wait_condition );
522
523	fsd_log_return(( "; job_ps=%s, exit_status=%d", drmaa_job_ps_to_str(self->state), self->exit_status ));
524	}
525
526	int
527	pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix)
528	{
529	char status_file = NULL, start_file = NULL;
530	FILE *fhandle = NULL;
531	int exit_status = -1;
532
533	fsd_log_enter(("({job_id=%s, job_state_dir_prefix=%s})", job_id, job_state_dir_prefix));
534
535	status_file = fsd_asprintf("%s/%s.exitcode", job_state_dir_prefix, job_id);
536	start_file = fsd_asprintf("%s/%s.started", job_state_dir_prefix, job_id);
537
538	if ((fhandle = fopen(status_file, "r")) == NULL)
539	{
540	struct stat tmpstat;
541
542	fsd_log_error(("Failed to open job status file: %s", status_file));
543	if (stat(start_file, &tmpstat) == 0 && (tmpstat.st_mode & S_IFREG))
544	{
545	exit_status = 143; /* SIGTERM */
546	fsd_log_info(("But start file exist %s. Assuming that job was killed (exit_status=%d).", start_file, exit_status));
547	}
548	else
549	{
550	fsd_log_error(("Start file not found: %s", start_file));
551	}
552
553
554	}
555	else
556	{
557	(void)fscanf(fhandle, "%d", &exit_status); /on error exit_status == -1 /
558	fclose(fhandle);
559	}
560
561	fsd_free(status_file);
562	fsd_free(start_file);
563
564	return exit_status;
565	}
566

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

Download in other formats: