Context Navigation

source: trunk/pbs_drmaa/job.c @ 55

Revision 55, 14.8 KB checked in by mmamonski, 14 years ago (diff)
on site fix
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
29	#include <drmaa_utils/drmaa.h>
30	#include <drmaa_utils/drmaa_util.h>
31	#include <pbs_error.h>
32	#include <pbs_ifl.h>
33
34	#include <pbs_drmaa/job.h>
35	#include <pbs_drmaa/log_reader.h>
36	#include <pbs_drmaa/pbs_attrib.h>
37	#include <pbs_drmaa/session.h>
38	#include <pbs_drmaa/util.h>
39
40	#ifndef lint
41	static char rcsid[]
42	# ifdef __GNUC__
43	__attribute__ ((unused))
44	# endif
45	= "$Id$";
46	#endif
47
48
49	static void pbsdrmaa_job_control( fsd_job_t *self, int action );
50
51	static void pbsdrmaa_job_update_status( fsd_job_t *self );
52
53	static void pbsdrmaa_job_on_missing( fsd_job_t *self );
54
55	static void pbsdrmaa_job_on_missing_standard( fsd_job_t *self );
56
57	static void pbsdrmaa_job_update( fsd_job_t self, struct batch_status );
58
59	static int pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix);
60
61	fsd_job_t *
62	pbsdrmaa_job_new( char *job_id )
63	{
64	pbsdrmaa_job_t self = (pbsdrmaa_job_t)fsd_job_new( job_id );
65	fsd_realloc( self, 1, pbsdrmaa_job_t );
66	self->super.control = pbsdrmaa_job_control;
67	self->super.update_status = pbsdrmaa_job_update_status;
68	self->super.on_missing = pbsdrmaa_job_on_missing;
69	self->update = pbsdrmaa_job_update;
70	return (fsd_job_t*)self;
71	}
72
73
74	static void
75	pbsdrmaa_job_control( fsd_job_t *self, int action )
76	{
77	volatile bool conn_lock = false;
78	pbsdrmaa_session_t session = (pbsdrmaa_session_t)self->session;
79	const char *job_id = self->job_id;
80	const char *apicall = NULL;
81	int rc = PBSE_NONE;
82
83	fsd_log_enter(( "({job_id=%s}, action=%d)",
84	self->job_id, action ));
85
86	TRY
87	{
88	int tries_left = session->max_retries_count;
89	int sleep_time = 1;
90
91	conn_lock = fsd_mutex_lock( &self->session->drm_connection_mutex );
92
93	/TODO reconnect /
94	while ( true )
95	{
96	switch( action )
97	{
98	/*
99	* We cannot know whether we did suspend job
100	* in other way than remembering this inside DRMAA session.
101	*/
102	case DRMAA_CONTROL_SUSPEND:
103	apicall = "pbs_sigjob";
104	rc = pbs_sigjob( session->pbs_conn, (char*)job_id,
105	"SIGSTOP", NULL );
106	fsd_log_info(("pbs_sigjob(%s, SIGSTOP) =%d", job_id, rc));
107	if( rc == PBSE_NONE )
108	self->flags \|= FSD_JOB_SUSPENDED;
109	break;
110	case DRMAA_CONTROL_RESUME:
111	apicall = "pbs_sigjob";
112	rc = pbs_sigjob( session->pbs_conn, (char*)job_id,
113	"SIGCONT", NULL );
114	fsd_log_info(("pbs_sigjob(%s, SIGCONT) =%d", job_id, rc));
115	if( rc == PBSE_NONE )
116	self->flags &= ~FSD_JOB_SUSPENDED;
117	break;
118	case DRMAA_CONTROL_HOLD:
119	apicall = "pbs_holdjob";
120	rc = pbs_holdjob( session->pbs_conn, (char*)job_id,
121	USER_HOLD, NULL );
122	fsd_log_info(("pbs_sigjob(%s, SIGHOLD) =%d", job_id, rc));
123	if( rc == PBSE_NONE )
124	self->flags \|= FSD_JOB_HOLD;
125	break;
126	case DRMAA_CONTROL_RELEASE:
127	apicall = "pbs_rlsjob";
128	rc = pbs_rlsjob( session->pbs_conn, (char*)job_id,
129	USER_HOLD, NULL );
130	fsd_log_info(("pbs_rlsjob(%s) =%d", job_id, rc));
131	if( rc == PBSE_NONE )
132	self->flags &= FSD_JOB_HOLD;
133	break;
134	case DRMAA_CONTROL_TERMINATE:
135	apicall = "pbs_deljob";
136	rc = pbs_deljob( session->pbs_conn, (char*)job_id, NULL );
137	fsd_log_info(("pbs_deljob(%s) =%d", job_id, rc));
138	/* Torque:
139	* deldelay=N -- delay between SIGTERM and SIGKILL (default 0) */
140	if( rc == PBSE_NONE )
141	{
142	self->flags &= FSD_JOB_TERMINATED_MASK;
143	if( (self->flags & FSD_JOB_TERMINATED) == 0 )
144	self->flags \|= FSD_JOB_TERMINATED \| FSD_JOB_ABORTED;
145	}
146	break;
147	}
148
149	retry_connect:
150	if ( rc == PBSE_NONE )
151	break;
152	else if (( rc == PBSE_INTERNAL \|\| rc == PBSE_PROTOCOL \|\| rc == PBSE_EXPIRED) && (tries_left--))
153	{
154	if (rc == PBSE_PROTOCOL \|\| rc == PBSE_EXPIRED)
155	{
156	if ( session->pbs_conn >= 0)
157	pbs_disconnect( session->pbs_conn );
158
159	sleep( sleep_time++ );
160
161	session->pbs_conn = pbs_connect( session->super.contact );
162
163	if (session->pbs_conn < 0)
164	goto retry_connect;
165
166	fsd_log_info(( "pbs_connect(%s) =%d", session->super.contact, session->pbs_conn ));
167	}
168	else /* PBSE_INTERNAL */
169	{
170	/*
171	* In PBS Pro pbs_sigjob raises internal server error (PBSE_INTERNAL)
172	* when job just changed its state to running.
173	*/
174	sleep( sleep_time++ );
175	}
176	fsd_log_debug(( "repeating request (%d of %d)", tries_left, session->max_retries_count));
177	}
178	else
179	pbsdrmaa_exc_raise_pbs( apicall );
180	} /* end while */
181	}
182	FINALLY
183	{
184	if( conn_lock )
185	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
186	}
187	END_TRY
188
189	fsd_log_return((""));
190	}
191
192
193	void
194	pbsdrmaa_job_update_status( fsd_job_t *self )
195	{
196	volatile bool conn_lock = false;
197	struct batch_status *volatile status = NULL;
198	pbsdrmaa_session_t session = (pbsdrmaa_session_t)self->session;
199	int tries_left = session->max_retries_count;
200	int sleep_time = 1;
201
202	fsd_log_enter(( "({job_id=%s})", self->job_id ));
203
204	TRY
205	{
206	conn_lock = fsd_mutex_lock( &self->session->drm_connection_mutex );
207	retry:
208
209	#ifdef PBS_PROFESSIONAL
210	status = pbs_statjob( session->pbs_conn, self->job_id, NULL, NULL );
211	#else
212	status = pbs_statjob( session->pbs_conn, self->job_id, session->status_attrl, NULL );
213	#endif
214	fsd_log_info(( "pbs_statjob(fd=%d, job_id=%s, attribs={...}) =%p",
215	session->pbs_conn, self->job_id, (void*)status ));
216	if( status == NULL )
217	{
218
219	#ifndef PBS_PROFESSIONAL
220	if ( pbs_errno != PBSE_UNKJOBID )
221	fsd_log_error(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
222	else
223	fsd_log_debug(("pbs_statjob error: %d, %s, %s", pbs_errno, pbse_to_txt(pbs_errno), pbs_strerror(pbs_errno)));
224	#else
225	# ifndef PBS_PROFESSIONAL_NO_LOG
226	if ( pbs_errno != PBSE_UNKJOBID )
227	fsd_log_error(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
228	else
229	fsd_log_debug(("pbs_statjob error: %d, %s", pbs_errno, pbse_to_txt(pbs_errno)));
230	# else
231	if ( pbs_errno != PBSE_UNKJOBID )
232	fsd_log_error(("pbs_statjob error: %d", pbs_errno));
233	else
234	fsd_log_debug(("pbs_statjob error: %d", pbs_errno));
235	# endif
236	#endif
237
238	switch( pbs_errno )
239	{
240	case PBSE_UNKJOBID:
241	break;
242	case PBSE_PROTOCOL:
243	case PBSE_EXPIRED:
244	if ( session->pbs_conn >= 0 )
245	pbs_disconnect( session->pbs_conn );
246	fsd_log_info(("Protocol error. Reconnecting..."));
247	retry_connect:
248	sleep(sleep_time++);
249	session->pbs_conn = pbs_connect( session->super.contact );
250	if( session->pbs_conn < 0 )
251	{
252	if (tries_left--) {
253	fsd_log_info(("%d tries left. Retrying...", tries_left));
254	goto retry_connect;
255	} else {
256	fsd_log_error(("No more tries left... Throwing exception"));
257	pbsdrmaa_exc_raise_pbs( "pbs_connect" );
258	}
259	}
260	else
261	{
262	goto retry;
263	}
264	default:
265	pbsdrmaa_exc_raise_pbs( "pbs_statjob" );
266	break;
267	case 0: /* ? */
268	fsd_exc_raise_code( FSD_ERRNO_INTERNAL_ERROR );
269	break;
270	}
271
272	}
273
274	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
275
276
277	if( status != NULL )
278	{
279	((pbsdrmaa_job_t*)self)->update( self, status );
280	}
281	else if( self->state < DRMAA_PS_DONE )
282	{
283	self->on_missing( self );
284	}
285	}
286	FINALLY
287	{
288	if( conn_lock )
289	conn_lock = fsd_mutex_unlock( &self->session->drm_connection_mutex );
290	if( status != NULL )
291	pbs_statfree( status );
292	}
293	END_TRY
294
295	fsd_log_return((""));
296	}
297
298
299	void
300	pbsdrmaa_job_update( fsd_job_t self, struct batch_status status )
301	{
302	struct attrl *attribs = status->attribs;
303	struct attrl *i = NULL;
304	char pbs_state = 0;
305	int exit_status = -2;
306	const char *cpu_usage = NULL;
307	const char *mem_usage = NULL;
308	const char *vmem_usage = NULL;
309	const char *walltime = NULL;
310	long unsigned int modify_time = 0;
311
312	fsd_log_enter(( "({job_id=%s})", self->job_id ));
313	#ifdef DEBUGGING
314	pbsdrmaa_dump_attrl( attribs, NULL );
315	#endif
316	fsd_assert( !strcmp( self->job_id, status->name ) );
317
318	for( i = attribs; i != NULL; i = i->next )
319	{
320	int attr;
321	attr = pbsdrmaa_pbs_attrib_by_name( i->name );
322	switch( attr )
323	{
324	case PBSDRMAA_ATTR_JOB_STATE:
325	pbs_state = i->value[0];
326	break;
327	case PBSDRMAA_ATTR_EXIT_STATUS:
328	exit_status = fsd_atoi( i->value );
329	break;
330	case PBSDRMAA_ATTR_RESOURCES_USED:
331	if( !strcmp( i->resource, "cput" ) )
332	cpu_usage = i->value;
333	else if( !strcmp( i->resource, "mem" ) )
334	mem_usage = i->value;
335	else if( !strcmp( i->resource, "vmem" ) )
336	vmem_usage = i->value;
337	else if( !strcmp( i->resource, "walltime" ) )
338	walltime = i->value;
339	break;
340	case PBSDRMAA_ATTR_QUEUE:
341	if (!self->queue)
342	self->queue = fsd_strdup(i->value);
343	break;
344	case PBSDRMAA_ATTR_ACCOUNT_NAME:
345	if (!self->project)
346	self->project = fsd_strdup(i->value);
347	break;
348	#ifndef PBS_PROFESSIONAL
349	case PBSDRMAA_ATTR_EXECUTION_HOST:
350	if (!self->execution_hosts) {
351	fsd_log_debug(("execution_hosts = %s", i->value));
352	self->execution_hosts = fsd_strdup(i->value);
353	}
354	break;
355	#else
356	case PBSDRMAA_ATTR_EXECUTION_VNODE:
357	if (!self->execution_hosts) {
358	fsd_log_debug(("execution_hosts = %s", i->value));
359	self->execution_hosts = fsd_strdup(i->value);
360	}
361	break;
362	#endif
363	case PBSDRMAA_ATTR_START_TIME:
364	{
365	long unsigned int start_time;
366	if (self->start_time == 0 && sscanf(i->value, "%lu", &start_time) == 1)
367	self->start_time = start_time;
368	break;
369	}
370	case PBSDRMAA_ATTR_MTIME:
371	if (sscanf(i->value, "%lu", &modify_time) != 1)
372	modify_time = 0;
373	break;
374	}
375	}
376
377	if( pbs_state )
378	fsd_log_debug(( "pbs_state: %c", pbs_state ));
379
380	if( exit_status != -2 )
381	{
382	fsd_log_debug(( "exit_status: %d", exit_status ));
383	self->exit_status = exit_status;
384	}
385	if(pbs_state){
386	switch( pbs_state )
387	{
388	case 'C': /* Job is completed after having run. */
389	self->flags &= FSD_JOB_TERMINATED_MASK;
390	self->flags \|= FSD_JOB_TERMINATED;
391	if (exit_status != -2) { /* has exit code */
392	if( self->exit_status == 0)
393	self->state = DRMAA_PS_DONE;
394	else
395	self->state = DRMAA_PS_FAILED;
396	} else {
397	self->state = DRMAA_PS_FAILED;
398	self->exit_status = -1;
399	}
400	if (modify_time != 0)
401	self->end_time = modify_time; /* take last modify time as end time */
402	else
403	self->end_time = time(NULL);
404
405	if (self->start_time == 0)
406	self->start_time = self->end_time;
407
408	break;
409	case 'E': /* Job is exiting after having run. - MM: ignore exiting state (transient state) - outputs might have not been transfered yet,
410	MM2: mark job as running if current job status is undetermined - fix "ps after job was ripped" */
411	if (self->state == DRMAA_PS_UNDETERMINED)
412	self->state = DRMAA_PS_RUNNING;
413	break;
414	case 'H': /* Job is held. */
415	self->state = DRMAA_PS_USER_ON_HOLD;
416	self->flags \|= FSD_JOB_HOLD;
417	break;
418	case 'Q': /* Job is queued, eligible to run or routed. */
419	case 'W': /* Job is waiting for its execution time to be reached. */
420	self->state = DRMAA_PS_QUEUED_ACTIVE;
421	self->flags &= ~FSD_JOB_HOLD;
422	break;
423	case 'R': /* Job is running. */
424	case 'T': /* Job is being moved to new location (?). */
425	{
426	if( self->flags & FSD_JOB_SUSPENDED )
427	self->state = DRMAA_PS_USER_SUSPENDED;
428	else
429	self->state = DRMAA_PS_RUNNING;
430	break;
431	}
432	case 'S': /* (Unicos only) job is suspend. */
433	self->state = DRMAA_PS_SYSTEM_SUSPENDED;
434	break;
435	case 0: default:
436	self->state = DRMAA_PS_UNDETERMINED;
437	break;
438	}
439	}
440	fsd_log_debug(( "job_ps: %s", drmaa_job_ps_to_str(self->state) ));
441
442	{
443	int hours, minutes, seconds;
444	long mem;
445	if( cpu_usage && sscanf( cpu_usage, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
446	{
447	self->cpu_usage = 60( 60hours + minutes ) + seconds;
448	fsd_log_debug(( "cpu_usage: %s=%lds", cpu_usage, self->cpu_usage ));
449	}
450	if( mem_usage && sscanf( mem_usage, "%ldkb", &mem ) == 1 )
451	{
452	self->mem_usage = 1024*mem;
453	fsd_log_debug(( "mem_usage: %s=%ldB", mem_usage, self->mem_usage ));
454	}
455	if( vmem_usage && sscanf( vmem_usage, "%ldkb", &mem ) == 1 )
456	{
457	self->vmem_usage = 1024*mem;
458	fsd_log_debug(( "vmem_usage: %s=%ldB", vmem_usage, self->vmem_usage ));
459	}
460	if( walltime && sscanf( walltime, "%d:%d:%d", &hours, &minutes, &seconds ) == 3 )
461	{
462	self->walltime = 60( 60hours + minutes ) + seconds;
463	fsd_log_debug(( "walltime: %s=%lds", walltime, self->walltime ));
464	}
465	}
466	}
467
468	void
469	pbsdrmaa_job_on_missing( fsd_job_t *self )
470	{
471	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t)self->session;
472
473	if( pbssession->pbs_home != NULL && pbssession->super.wait_thread_started && self->submit_time)
474	fsd_log_info(("Job on missing but WT is running. Skipping...")); /* TODO: try to provide implementation that uses accounting/server log files */
475	else
476	pbsdrmaa_job_on_missing_standard( self );
477	}
478
479	void
480	pbsdrmaa_job_on_missing_standard( fsd_job_t *self )
481	{
482	fsd_drmaa_session_t *session = self->session;
483	pbsdrmaa_session_t pbssession = (pbsdrmaa_session_t )session;
484	int exit_status = -1;
485
486	fsd_log_enter(( "({job_id=%s})", self->job_id ));
487	fsd_log_warning(( "Job %s missing from DRM queue", self->job_id ));
488
489	fsd_log_info(( "job_on_missing: last job_ps: %s (0x%02x)", drmaa_job_ps_to_str(self->state), self->state));
490
491	if( (exit_status = pbsdrmaa_job_read_exit_status(self->job_id, pbssession->job_exit_status_file_prefix)) == 0 )
492	{
493	self->state = DRMAA_PS_DONE;
494	self->exit_status = exit_status;
495	}
496	else
497	{
498	self->state = DRMAA_PS_FAILED;
499	self->exit_status = exit_status;
500	}
501
502	fsd_cond_broadcast( &self->status_cond);
503
504	fsd_log_return(( "; job_ps=%s, exit_status=%d", drmaa_job_ps_to_str(self->state), self->exit_status ));
505	}
506
507	int
508	pbsdrmaa_job_read_exit_status( const char job_id, const char job_state_dir_prefix)
509	{
510	char *status_file = NULL;
511	FILE *fhandle = NULL;
512	int exit_status = -1;
513
514	fsd_log_enter(("({job_id=%s, job_state_dir_prefix=%s})", job_id, job_state_dir_prefix));
515
516	status_file = fsd_asprintf("%s/%s.exitcode", job_state_dir_prefix, job_id);
517
518	if ((fhandle = fopen(status_file, "r")) == NULL)
519	{
520	fsd_log_error(("Failed to open job status file: %s", status_file));
521	}
522	else
523	{
524	(void)fscanf(fhandle, "%d", &exit_status); /on error exit_status == -1 /
525	fclose(fhandle);
526	}
527
528	fsd_free(status_file);
529
530	return exit_status;
531	}
532

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

Download in other formats: