darshan-core.c 53.6 KB
Newer Older
1
/*
Shane Snyder's avatar
Shane Snyder committed
2
3
4
 * Copyright (C) 2015 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
5
6
 */

7
#define _XOPEN_SOURCE 500
8
#define _GNU_SOURCE
9

10
11
12
13
14
15
#include "darshan-runtime-config.h"

#include <stdio.h>
#ifdef HAVE_MNTENT_H
#include <mntent.h>
#endif
16
17
18
19
20
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <limits.h>
#include <pthread.h>
21
22
23
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/vfs.h>
24
#include <zlib.h>
25
#include <mpi.h>
26
#include <assert.h>
27

28
#include "uthash.h"
Shane Snyder's avatar
Shane Snyder committed
29
#include "darshan.h"
30
#include "darshan-core.h"
Shane Snyder's avatar
Shane Snyder committed
31
#include "darshan-dynamic.h"
32

33
extern char* __progname;
34
extern char* __progname_full;
35

36
/* internal variable delcarations */
37
static struct darshan_core_runtime *darshan_core = NULL;
38
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
39
static int my_rank = -1;
40
static int nprocs = -1;
41
static int darshan_mem_alignment = 1;
42

43
44
45
46
47
48
49
50
51
52
53
54
/* paths prefixed with the following directories are not traced by darshan */
char* darshan_path_exclusions[] = {
"/etc/",
"/dev/",
"/usr/",
"/bin/",
"/boot/",
"/lib/",
"/opt/",
"/sbin/",
"/sys/",
"/proc/",
55
"/var/",
56
57
58
NULL
};

59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
#ifdef DARSHAN_BGQ
extern void bgq_runtime_initialize();
#endif

/* array of init functions for modules which need to be statically
 * initialized by darshan at startup time
 */
void (*mod_static_init_fns[])(void) =
{
#ifdef DARSHAN_BGQ
    &bgq_runtime_initialize,
#endif
    NULL
};

Shane Snyder's avatar
Shane Snyder committed
74
75
76
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

77
78
79
80
81
82
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
83
    int block_size;
84
85
86
87
88
89
    char path[DARSHAN_MAX_MNT_PATH];
    char type[DARSHAN_MAX_MNT_TYPE];
};
static struct mnt_data mnt_data_array[DARSHAN_MAX_MNTS];
static int mnt_data_count = 0;

90
91
92
93
/* prototypes for internal helper functions */
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
94
95
96
97
98
99
    struct darshan_core_runtime* core);
static void darshan_get_exe_and_mounts_root(
    struct darshan_core_runtime *core, char* trailing_data,
    int space_left);
static char* darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core);
100
101
static void darshan_block_size_from_path(
    const char *path, int *block_size);
102
static void darshan_get_shared_records(
103
104
    struct darshan_core_runtime *core, darshan_record_id **shared_recs,
    int *shared_rec_cnt);
105
static int darshan_log_open_all(
106
    char *logfile_name, MPI_File *log_fh);
107
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
108
109
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
110
static int darshan_log_write_record_hash(
111
    MPI_File log_fh, struct darshan_core_runtime *core,
112
113
114
    uint64_t *inout_off);
static int darshan_log_append_all(
    MPI_File log_fh, struct darshan_core_runtime *core, void *buf,
Shane Snyder's avatar
Shane Snyder committed
115
    int count, uint64_t *inout_off);
Shane Snyder's avatar
Shane Snyder committed
116
117
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
118

119
120
/* *********************************** */

Shane Snyder's avatar
Shane Snyder committed
121
void darshan_core_initialize(int argc, char **argv)
122
{
123
    struct darshan_core_runtime *init_core = NULL;
124
125
126
    int i;
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
127
    char *envstr;
128
129
130
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int chars_left = 0;
131
132
    int ret;
    int tmpval;
133
134

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
135
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
136
137
138
139

    if(getenv("DARSHAN_INTERNAL_TIMING"))
        internal_timing_flag = 1;

140
    if(internal_timing_flag)
141
142
143
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
144
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
145
    {
146
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
147
148
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
149
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
150
151
152
153
154
155
156
157
158
159
160
        if(envstr)
        {
            ret = sscanf(envstr, "%d", &tmpval);
            /* silently ignore if the env variable is set poorly */
            if(ret == 1 && tmpval > 0)
            {
                darshan_mem_alignment = tmpval;
            }
        }
        else
        {
161
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
162
163
164
165
166
167
168
        }

        /* avoid floating point errors on faulty input */
        if (darshan_mem_alignment < 1)
        {
            darshan_mem_alignment = 1;
        }
169
170

        /* allocate structure to track darshan_core_runtime information */
171
172
        init_core = malloc(sizeof(*init_core));
        if(init_core)
173
        {
174
            memset(init_core, 0, sizeof(*init_core));
175

176
177
178
179
            init_core->log_job.uid = getuid();
            init_core->log_job.start_time = time(NULL);
            init_core->log_job.nprocs = nprocs;
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();
180
181

            /* record exe and arguments */
182
            for(i=0; i<argc; i++)
183
            {
184
185
                chars_left = DARSHAN_EXE_LEN-strlen(init_core->exe);
                strncat(init_core->exe, argv[i], chars_left);
186
                if(i < (argc-1))
187
                {
188
189
                    chars_left = DARSHAN_EXE_LEN-strlen(init_core->exe);
                    strncat(init_core->exe, " ", chars_left);
190
191
192
193
194
195
196
197
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
198
199
200
201
                chars_left = DARSHAN_EXE_LEN-strlen(init_core->exe);
                strncat(init_core->exe, __progname_full, chars_left);
                chars_left = DARSHAN_EXE_LEN-strlen(init_core->exe);
                strncat(init_core->exe, " <unknown args>", chars_left);
202
203
204
205
206
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
207
                truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
208
                sprintf(&init_core->exe[truncate_offset], "%s",
209
210
                    truncate_string);
            }
211
212

            /* collect information about command line and mounted file systems */
213
            init_core->trailing_data = darshan_get_exe_and_mounts(init_core);
214

Shane Snyder's avatar
Shane Snyder committed
215
216
217
218
219
220
221
            /* bootstrap any modules with static initialization routines */
            i = 0;
            while(mod_static_init_fns[i])
            {
                (*mod_static_init_fns[i])();
                i++;
            }
222
        }
223
224
    }

225
226
227
228
229
    if(internal_timing_flag)
    {
        init_time = DARSHAN_MPI_CALL(PMPI_Wtime)() - init_start;
        DARSHAN_MPI_CALL(PMPI_Reduce)(&init_time, &init_max, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
230
        if(my_rank == 0)
231
        {
232
233
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
234
235
236
        }
    }

237
238
239
240
    /* if darshan was successfully initialized, set the global pointer */
    if(init_core)
        darshan_core = init_core;

241
242
243
    return;
}

Shane Snyder's avatar
Shane Snyder committed
244
void darshan_core_shutdown()
245
{
246
    int i;
247
    char *logfile_name;
248
    struct darshan_core_runtime *final_core;
249
    int internal_timing_flag = 0;
250
251
    char *envjobid;
    char *jobid_str;
252
    int jobid;
253
    struct tm *start_tm;
254
    time_t start_time_tmp;
255
256
    int ret = 0;
    int all_ret = 0;
257
258
    int64_t first_start_time;
    int64_t last_end_time;
259
260
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
261
262
263
    darshan_record_id *shared_recs;
    darshan_record_id *mod_shared_recs;
    int shared_rec_cnt = 0;
264
    double start_log_time;
265
266
267
268
269
270
271
    double open1, open2;
    double job1, job2;
    double rec1, rec2;
    double mod1[DARSHAN_MAX_MODS] = {0};
    double mod2[DARSHAN_MAX_MODS] = {0};
    double header1, header2;
    double tm_end;
272
    uint64_t gz_fp = 0;
273
274
    MPI_File log_fh;
    MPI_Status status;
275
276
277
278

    if(getenv("DARSHAN_INTERNAL_TIMING"))
        internal_timing_flag = 1;

279
280
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

Shane Snyder's avatar
Shane Snyder committed
281
    /* disable darhan-core while we shutdown */
282
    DARSHAN_CORE_LOCK();
283
    if(!darshan_core)
284
    {
285
        DARSHAN_CORE_UNLOCK();
286
287
        return;
    }
288
289
    final_core = darshan_core;
    darshan_core = NULL;
Shane Snyder's avatar
Shane Snyder committed
290

291
292
293
294
295
296
297
    final_core->comp_buf = malloc(DARSHAN_COMP_BUF_SIZE);
    if(!(final_core->comp_buf))
    {
        darshan_core_cleanup(final_core);
        return;
    }

298
    /* we also need to set which modules were registered on this process and
299
300
     * call into those modules and give them a chance to perform any necessary
     * pre-shutdown steps.
Shane Snyder's avatar
Shane Snyder committed
301
302
303
304
305
306
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
            local_mod_use[i] = 1;
307
            final_core->mod_array[i]->mod_funcs.begin_shutdown();
Shane Snyder's avatar
Shane Snyder committed
308
309
        }
    }
310
    DARSHAN_CORE_UNLOCK();
311
312
313
314

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
315
        darshan_core_cleanup(final_core);
316
317
318
        return;
    }

319
    /* set darshan job id/metadata and constuct log file name on rank 0 */
320
    if(my_rank == 0)
321
    {
322
        /* Use DARSHAN_JOBID_OVERRIDE for the env var for __DARSHAN_JOBID */
323
        envjobid = getenv(DARSHAN_JOBID_OVERRIDE);
324
        if(!envjobid)
325
        {
326
            envjobid = __DARSHAN_JOBID;
327
328
        }

329
        /* find a job id */
330
331
332
333
334
335
336
337
338
339
340
341
        jobid_str = getenv(envjobid);
        if(jobid_str)
        {
            /* in cobalt we can find it in env var */
            ret = sscanf(jobid_str, "%d", &jobid);
        }
        if(!jobid_str || ret != 1)
        {
            /* use pid as fall back */
            jobid = getpid();
        }

342
        final_core->log_job.jobid = (int64_t)jobid;
343

344
        /* if we are using any hints to write the log file, then record those
345
         * hints with the darshan job information
346
         */
347
        darshan_log_record_hints_and_ver(final_core);
348

349
        /* use human readable start time format in log filename */
350
        start_time_tmp = final_core->log_job.start_time;
351
        start_tm = localtime(&start_time_tmp);
352

353
354
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
355
356
357
358
359
360
361
362
363
    }

    /* broadcast log file name */
    DARSHAN_MPI_CALL(PMPI_Bcast)(logfile_name, PATH_MAX, MPI_CHAR, 0,
        MPI_COMM_WORLD);

    if(strlen(logfile_name) == 0)
    {
        /* failed to generate log file name */
364
        free(logfile_name);
365
        darshan_core_cleanup(final_core);
366
367
368
        return;
    }

369
    final_core->log_job.end_time = time(NULL);
370

371
372
373
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
374
375
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job.start_time, &first_start_time, 1, MPI_LONG_LONG, MPI_MIN, 0, MPI_COMM_WORLD);
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job.end_time, &last_end_time, 1, MPI_LONG_LONG, MPI_MAX, 0, MPI_COMM_WORLD);
376
377
    if(my_rank == 0)
    {
378
379
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
380
    }
381

382
383
384
    /* reduce the number of times a module was opened globally and bcast to everyone */   
    DARSHAN_MPI_CALL(PMPI_Allreduce)(local_mod_use, global_mod_use_count, DARSHAN_MAX_MODS, MPI_INT, MPI_SUM, MPI_COMM_WORLD);

385
    /* get a list of records which are shared across all processes */
386
    darshan_get_shared_records(final_core, &shared_recs, &shared_rec_cnt);
387

388
389
    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
390
    /* collectively open the darshan log file */
391
    ret = darshan_log_open_all(logfile_name, &log_fh);
392
393
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
394
395
396
397
398
399
400
401

    /* error out if unable to open log file */
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
402
403
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
404
405
        }
        free(logfile_name);
406
        darshan_core_cleanup(final_core);
407
408
409
        return;
    }

410
411
    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
412
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
413
    if(my_rank == 0)
414
    {
415
        void *pointers[2] = {&final_core->log_job, final_core->trailing_data};
416
        int lengths[2] = {sizeof(struct darshan_job), strlen(final_core->trailing_data)};
417
        int comp_buf_sz = 0;
418

419
        /* compress the job info and the trailing mount/exe data */
Shane Snyder's avatar
Shane Snyder committed
420
        all_ret = darshan_deflate_buffer(pointers, lengths, 2,
421
422
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
423
        {
424
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
425
            unlink(logfile_name);
426
        }
427
428
429
        else
        {
            /* write the job information, preallocing space for the log header */
Shane Snyder's avatar
Shane Snyder committed
430
            gz_fp += sizeof(struct darshan_header);
431
432
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, gz_fp,
                final_core->comp_buf, comp_buf_sz, MPI_BYTE, &status);
433
434
435
436
437
            if(all_ret != MPI_SUCCESS)
            {
                fprintf(stderr, "darshan library warning: unable to write job data to log file %s\n",
                        logfile_name);
                unlink(logfile_name);
Shane Snyder's avatar
Shane Snyder committed
438
                
439
            }
440
            gz_fp += comp_buf_sz;
441
        }
442
443
    }

444
445
446
447
448
    /* error out if unable to write job information */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&all_ret, 1, MPI_INT, 0, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        free(logfile_name);
449
        darshan_core_cleanup(final_core);
450
451
        return;
    }
452
453
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
454

455
456
    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
457
    /* write the record name->id hash to the log file */
Shane Snyder's avatar
Shane Snyder committed
458
    final_core->log_header.rec_map.off = gz_fp;
459
    ret = darshan_log_write_record_hash(log_fh, final_core, &gz_fp);
Shane Snyder's avatar
Shane Snyder committed
460
    final_core->log_header.rec_map.len = gz_fp - final_core->log_header.rec_map.off;
461

462
    /* error out if unable to write record hash */
463
464
465
466
467
468
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
469
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
470
                logfile_name);
471
            unlink(logfile_name);
472
473
        }
        free(logfile_name);
474
        darshan_core_cleanup(final_core);
475
476
        return;
    }
Shane Snyder's avatar
Shane Snyder committed
477
478
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
479

480
481
482
    mod_shared_recs = malloc(shared_rec_cnt * sizeof(darshan_record_id));
    assert(mod_shared_recs);

483
    /* loop over globally used darshan modules and:
484
     *      - perform shared file reductions, if possible
485
     *      - get final output buffer
486
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
487
     *      - append compressed buffer to log file
488
489
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
490
     */
491
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
492
    {
493
        struct darshan_core_module* this_mod = final_core->mod_array[i];
494
        struct darshan_core_record_ref *ref = NULL;
495
        int mod_shared_rec_cnt = 0;
496
        void* mod_buf = NULL;
497
        int mod_buf_sz = 0;
498
        int j;
499

500
        if(global_mod_use_count[i] == 0)
501
502
        {
            if(my_rank == 0)
503
504
505
506
            {
                final_core->log_header.mod_map[i].off = 0;
                final_core->log_header.mod_map[i].len = 0;
            }
507
            continue;
508
        }
509
510
 
        if(internal_timing_flag)
511
            mod1[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
512

513
        /* set the shared file list for this module */
514
515
        memset(mod_shared_recs, 0, shared_rec_cnt * sizeof(darshan_record_id));
        for(j = 0; j < shared_rec_cnt; j++)
516
517
518
519
        {
            HASH_FIND(hlink, final_core->rec_hash, &shared_recs[j],
                sizeof(darshan_record_id), ref);
            assert(ref);
520
            if(DARSHAN_MOD_FLAG_ISSET(ref->global_mod_flags, i))
521
            {
522
                mod_shared_recs[mod_shared_rec_cnt++] = shared_recs[j];
523
            }
524
        }
525

526
527
528
529
530
        /* if module is registered locally, get the corresponding output buffer
         * 
         * NOTE: this function can be used to run collective operations across
         * modules, if there are file records shared globally.
         */
531
        if(this_mod)
532
        {
533
534
            this_mod->mod_funcs.get_output_data(MPI_COMM_WORLD, mod_shared_recs,
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
535
536
        }

537
        /* append this module's data to the darshan log */
Shane Snyder's avatar
Shane Snyder committed
538
539
540
541
        final_core->log_header.mod_map[i].off = gz_fp;
        ret = darshan_log_append_all(log_fh, final_core, mod_buf, mod_buf_sz, &gz_fp);
        final_core->log_header.mod_map[i].len =
            gz_fp - final_core->log_header.mod_map[i].off;
542

543
        /* error out if the log append failed */
544
545
546
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
547
        {
548
549
550
551
552
553
554
555
            if(my_rank == 0)
            {
                fprintf(stderr,
                    "darshan library warning: unable to write %s module data to log file %s\n",
                    darshan_module_names[i], logfile_name);
                unlink(logfile_name);
            }
            free(logfile_name);
556
            darshan_core_cleanup(final_core);
557
            return;
558
559
560
        }

        /* shutdown module if registered locally */
561
        if(this_mod)
562
563
564
        {
            this_mod->mod_funcs.shutdown();
        }
565
566
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
567
568
    }

569
570
    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
571
572
573
574
575
576
    /* write out log header, after running 2 reduction on header variables:
     *  1) reduce 'partial_flag' variable to determine which modules ran out
     *     of memory for storing I/O data
     *  2) reduce 'mod_ver' array to determine which log format version each
     *     module used for this output log
     */
577
578
    if(my_rank == 0)
    {
579
580
581
582
583
584
585
586
        DARSHAN_MPI_CALL(PMPI_Reduce)(MPI_IN_PLACE,
            &(final_core->log_header.partial_flag), 1, MPI_UINT32_T,
            MPI_BOR, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(MPI_IN_PLACE,
            final_core->log_header.mod_ver, DARSHAN_MAX_MODS, MPI_UINT32_T,
            MPI_MAX, 0, MPI_COMM_WORLD);

        /* rank 0 is responsible for writing the log header */
587
588
589
        /* initialize the remaining header fields */
        strcpy(final_core->log_header.version_string, DARSHAN_LOG_VERSION);
        final_core->log_header.magic_nr = DARSHAN_MAGIC_NR;
590
        final_core->log_header.comp_type = DARSHAN_ZLIB_COMP;
591

Shane Snyder's avatar
Shane Snyder committed
592
593
594
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &(final_core->log_header),
            sizeof(struct darshan_header), MPI_BYTE, &status);
        if(all_ret != MPI_SUCCESS)
595
        {
Shane Snyder's avatar
Shane Snyder committed
596
597
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
598
            unlink(logfile_name);
599
        }
600
    }
601
602
603
604
605
606
607
608
609
    else
    {
        DARSHAN_MPI_CALL(PMPI_Reduce)(&(final_core->log_header.partial_flag),
            &(final_core->log_header.partial_flag), 1, MPI_UINT32_T,
            MPI_BOR, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(final_core->log_header.mod_ver,
            final_core->log_header.mod_ver, DARSHAN_MAX_MODS, MPI_UINT32_T,
            MPI_MAX, 0, MPI_COMM_WORLD);
    }
610

611
612
613
614
615
    /* error out if unable to write log header */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&all_ret, 1, MPI_INT, 0, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        free(logfile_name);
616
        darshan_core_cleanup(final_core);
617
618
        return;
    }
619
620
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
621

622
623
624
    DARSHAN_MPI_CALL(PMPI_File_close)(&log_fh);

    /* if we got this far, there are no errors, so rename from *.darshan_partial
625
     * to *-<logwritetime>.darshan, which indicates that this log file is
626
627
     * complete and ready for analysis
     */
628
629
    if(my_rank == 0)
    {
Shane Snyder's avatar
Shane Snyder committed
630
        if(getenv("DARSHAN_LOGFILE"))
631
        {
632
#ifdef __DARSHAN_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
633
            chmod(logfile_name, (S_IRUSR|S_IRGRP));
634
#else
Shane Snyder's avatar
Shane Snyder committed
635
            chmod(logfile_name, (S_IRUSR));
636
#endif
Shane Snyder's avatar
Shane Snyder committed
637
638
639
640
641
642
643
644
645
646
647
648
649
650
        }
        else
        {
            char* tmp_index;
            double end_log_time;
            char* new_logfile_name;

            new_logfile_name = malloc(PATH_MAX);
            if(new_logfile_name)
            {
                new_logfile_name[0] = '\0';
                end_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
                strcat(new_logfile_name, logfile_name);
                tmp_index = strstr(new_logfile_name, ".darshan_partial");
651
                sprintf(tmp_index, "_%d.darshan", (int)(end_log_time-start_log_time+1));
Shane Snyder's avatar
Shane Snyder committed
652
653
                rename(logfile_name, new_logfile_name);
                /* set permissions on log file */
654
#ifdef __DARSHAN_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
655
656
657
658
659
660
                chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
#else
                chmod(new_logfile_name, (S_IRUSR));
#endif
                free(new_logfile_name);
            }
661
        }
662
    }
663

664
    free(logfile_name);
665
    darshan_core_cleanup(final_core);
666

667
    if(internal_timing_flag)
668
    {
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
        double open_tm, open_slowest;
        double header_tm, header_slowest;
        double job_tm, job_slowest;
        double rec_tm, rec_slowest;
        double mod_tm[DARSHAN_MAX_MODS], mod_slowest[DARSHAN_MAX_MODS];
        double all_tm, all_slowest;

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

        open_tm = open2 - open1;
        header_tm = header2 - header1;
        job_tm = job2 - job1;
        rec_tm = rec2 - rec1;
        all_tm = tm_end - start_log_time;
        for(i = 0;i < DARSHAN_MAX_MODS; i++)
        {
            mod_tm[i] = mod2[i] - mod1[i];
        }

        DARSHAN_MPI_CALL(PMPI_Reduce)(&open_tm, &open_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&header_tm, &header_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&job_tm, &job_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&rec_tm, &rec_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&all_tm, &all_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(mod_tm, mod_slowest, DARSHAN_MAX_MODS,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);

        if(my_rank == 0)
        {
703
704
705
706
707
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:log_open\t%d\t%f\n", nprocs, open_slowest);
            fprintf(stderr, "darshan:job_write\t%d\t%f\n", nprocs, job_slowest);
            fprintf(stderr, "darshan:hash_write\t%d\t%f\n", nprocs, rec_slowest);
            fprintf(stderr, "darshan:header_write\t%d\t%f\n", nprocs, header_slowest);
708
709
710
            for(i = 0; i < DARSHAN_MAX_MODS; i++)
            {
                if(global_mod_use_count[i])
711
                    fprintf(stderr, "darshan:%s_shutdown\t%d\t%f\n", darshan_module_names[i],
Shane Snyder's avatar
Shane Snyder committed
712
                        nprocs, mod_slowest[i]);
713
            }
714
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
715
        }
716
717
718
719
    }
    
    return;
}
720

Shane Snyder's avatar
Shane Snyder committed
721
/* *********************************** */
722

723
/* construct the darshan log file name */
724
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
725
{
Shane Snyder's avatar
Shane Snyder committed
726
    char* user_logfile_name;
727
728
729
    char* logpath;
    char* logname_string;
    char* logpath_override = NULL;
730
#ifdef __DARSHAN_LOG_ENV
731
732
733
734
735
736
737
738
739
    char env_check[256];
    char* env_tok;
#endif
    uint64_t hlevel;
    char hname[HOST_NAME_MAX];
    uint64_t logmod;
    char cuser[L_cuserid] = {0};
    int ret;

Shane Snyder's avatar
Shane Snyder committed
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
    /* first, check if user specifies a complete logpath to use */
    user_logfile_name = getenv("DARSHAN_LOGFILE");
    if(user_logfile_name)
    {
        if(strlen(user_logfile_name) >= (PATH_MAX-1))
        {
            fprintf(stderr, "darshan library warning: user log file name too long.\n");
            logfile_name[0] = '\0';
        }
        else
        {
            strcpy(logfile_name, user_logfile_name);
        }
    }
    else
755
    {
Shane Snyder's avatar
Shane Snyder committed
756
757
        /* otherwise, generate the log path automatically */

758
759
        /* Use DARSHAN_LOG_PATH_OVERRIDE for the value or __DARSHAN_LOG_PATH */
        logpath = getenv(DARSHAN_LOG_PATH_OVERRIDE);
Shane Snyder's avatar
Shane Snyder committed
760
761
        if(!logpath)
        {
762
763
#ifdef __DARSHAN_LOG_PATH
            logpath = __DARSHAN_LOG_PATH;
764
#endif
Shane Snyder's avatar
Shane Snyder committed
765
        }
766

Shane Snyder's avatar
Shane Snyder committed
767
768
769
770
771
772
773
774
775
776
        /* get the username for this job.  In order we will try each of the
         * following until one of them succeeds:
         *
         * - cuserid()
         * - getenv("LOGNAME")
         * - snprintf(..., geteuid());
         *
         * Note that we do not use getpwuid() because it generally will not
         * work in statically compiled binaries.
         */
777
778

#ifndef DARSHAN_DISABLE_CUSERID
Shane Snyder's avatar
Shane Snyder committed
779
        cuserid(cuser);
780
781
#endif

Shane Snyder's avatar
Shane Snyder committed
782
783
        /* if cuserid() didn't work, then check the environment */
        if(strcmp(cuser, "") == 0)
784
        {
Shane Snyder's avatar
Shane Snyder committed
785
786
787
788
789
            logname_string = getenv("LOGNAME");
            if(logname_string)
            {
                strncpy(cuser, logname_string, (L_cuserid-1));
            }
790
791
        }

Shane Snyder's avatar
Shane Snyder committed
792
793
794
795
796
797
        /* if cuserid() and environment both fail, then fall back to uid */
        if(strcmp(cuser, "") == 0)
        {
            uid_t uid = geteuid();
            snprintf(cuser, sizeof(cuser), "%u", uid);
        }
798

Shane Snyder's avatar
Shane Snyder committed
799
800
801
802
        /* generate a random number to help differentiate the log */
        hlevel=DARSHAN_MPI_CALL(PMPI_Wtime)() * 1000000;
        (void)gethostname(hname, sizeof(hname));
        logmod = darshan_hash((void*)hname,strlen(hname),hlevel);
803

Shane Snyder's avatar
Shane Snyder committed
804
805
806
807
        /* see if darshan was configured using the --with-logpath-by-env
         * argument, which allows the user to specify an absolute path to
         * place logs via an env variable.
         */
808
#ifdef __DARSHAN_LOG_ENV
Shane Snyder's avatar
Shane Snyder committed
809
        /* just silently skip if the environment variable list is too big */
810
        if(strlen(__DARSHAN_LOG_ENV) < 256)
811
        {
Shane Snyder's avatar
Shane Snyder committed
812
            /* copy env variable list to a temporary buffer */
813
            strcpy(env_check, __DARSHAN_LOG_ENV);
Shane Snyder's avatar
Shane Snyder committed
814
815
816
            /* tokenize the comma-separated list */
            env_tok = strtok(env_check, ",");
            if(env_tok)
817
            {
Shane Snyder's avatar
Shane Snyder committed
818
                do
819
                {
Shane Snyder's avatar
Shane Snyder committed
820
821
822
823
824
825
826
827
828
                    /* check each env variable in order */
                    logpath_override = getenv(env_tok);
                    if(logpath_override)
                    {
                        /* stop as soon as we find a match */
                        break;
                    }
                }while((env_tok = strtok(NULL, ",")));
            }
829
830
831
        }
#endif

Shane Snyder's avatar
Shane Snyder committed
832
        if(logpath_override)
833
        {
Shane Snyder's avatar
Shane Snyder committed
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
            ret = snprintf(logfile_name, PATH_MAX,
                "%s/%s_%s_id%d_%d-%d-%d-%" PRIu64 ".darshan_partial",
                logpath_override,
                cuser, __progname, jobid,
                (start_tm->tm_mon+1),
                start_tm->tm_mday,
                (start_tm->tm_hour*60*60 + start_tm->tm_min*60 + start_tm->tm_sec),
                logmod);
            if(ret == (PATH_MAX-1))
            {
                /* file name was too big; squish it down */
                snprintf(logfile_name, PATH_MAX,
                    "%s/id%d.darshan_partial",
                    logpath_override, jobid);
            }
849
        }
Shane Snyder's avatar
Shane Snyder committed
850
        else if(logpath)
851
        {
Shane Snyder's avatar
Shane Snyder committed
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
            ret = snprintf(logfile_name, PATH_MAX,
                "%s/%d/%d/%d/%s_%s_id%d_%d-%d-%d-%" PRIu64 ".darshan_partial",
                logpath, (start_tm->tm_year+1900),
                (start_tm->tm_mon+1), start_tm->tm_mday,
                cuser, __progname, jobid,
                (start_tm->tm_mon+1),
                start_tm->tm_mday,
                (start_tm->tm_hour*60*60 + start_tm->tm_min*60 + start_tm->tm_sec),
                logmod);
            if(ret == (PATH_MAX-1))
            {
                /* file name was too big; squish it down */
                snprintf(logfile_name, PATH_MAX,
                    "%s/id%d.darshan_partial",
                    logpath, jobid);
            }
        }
        else
        {
            logfile_name[0] = '\0';
872
873
874
875
        }
    }

    return;
876
877
}

878
/* record any hints used to write the darshan log in the log header */
879
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
880
881
882
883
884
885
886
887
888
{
    char* hints;
    char* header_hints;
    int meta_remain = 0;
    char* m;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
889
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
890
891
    if(!hints)
    {
892
        hints = __DARSHAN_LOG_HINTS;
893
894
895
896
897
898
899
900
901
902
    }

    if(!hints || strlen(hints) < 1)
        return;

    header_hints = strdup(hints);
    if(!header_hints)
        return;

    meta_remain = DARSHAN_JOB_METADATA_LEN -
903
        strlen(core->log_job.metadata) - 1;
904
905
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
906
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
907
908
909
910
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
911
        m = core->log_job.metadata + strlen(core->log_job.metadata);
912
913
914
915
916
917
918
919
920
921
922
923
        /* We have room to store the hints in the metadata portion of
         * the job header.  We just prepend an h= to the hints list.  The
         * metadata parser will ignore = characters that appear in the value
         * portion of the metadata key/value pair.
         */
        sprintf(m, "h=%s\n", header_hints);
    }
    free(header_hints);

    return;
}

924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
static int mnt_data_cmp(const void* a, const void* b)
{
    const struct mnt_data *d_a = (const struct mnt_data*)a;
    const struct mnt_data *d_b = (const struct mnt_data*)b;

    if(strlen(d_a->path) > strlen(d_b->path))
        return(-1);
    else if(strlen(d_a->path) < strlen(d_b->path))
        return(1);
    else
        return(0);
}

/* adds an entry to table of mounted file systems */
static void add_entry(char* trailing_data, int* space_left, struct mntent *entry)
{
    int ret;
    char tmp_mnt[256];
    struct statfs statfsbuf;

    strncpy(mnt_data_array[mnt_data_count].path, entry->mnt_dir,
        DARSHAN_MAX_MNT_PATH-1);
    strncpy(mnt_data_array[mnt_data_count].type, entry->mnt_type,
        DARSHAN_MAX_MNT_TYPE-1);
    /* NOTE: we now try to detect the preferred block size for each file 
     * system using fstatfs().  On Lustre we assume a size of 1 MiB 
     * because fstatfs() reports 4 KiB. 
     */
#ifndef LL_SUPER_MAGIC
#define LL_SUPER_MAGIC 0x0BD00BD0
#endif
    ret = statfs(entry->mnt_dir, &statfsbuf);
    if(ret == 0 && statfsbuf.f_type != LL_SUPER_MAGIC)
        mnt_data_array[mnt_data_count].block_size = statfsbuf.f_bsize;
    else if(ret == 0 && statfsbuf.f_type == LL_SUPER_MAGIC)
        mnt_data_array[mnt_data_count].block_size = 1024*1024;
    else
        mnt_data_array[mnt_data_count].block_size = 4096;

    /* store mount information for use in header of darshan log */
964
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
        strcat(trailing_data, tmp_mnt);
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

/* darshan_get_exe_and_mounts_root()
 *
 * collects command line and list of mounted file systems into a string that
 * will be stored with the job header
 */
static void darshan_get_exe_and_mounts_root(struct darshan_core_runtime *core,
    char* trailing_data, int space_left)
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
    int tmp_index = 0;
    int skip = 0;

    /* skip these fs types */
    static char* fs_exclusions[] = {
        "tmpfs",
        "proc",
        "sysfs",
        "devpts",
        "binfmt_misc",
        "fusectl",
        "debugfs",
        "securityfs",
        "nfsd",
For faster browsing, not all history is shown. View entire blame