darshan-core.c 49.1 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
55
56
57
/* 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/",
NULL
};

58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
#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
73
74
75
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

76
77
78
79
80
81
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
82
    int block_size;
83
84
85
86
87
88
    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;

89
90
91
92
/* 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(
93
94
95
96
97
98
    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);
99
100
static void darshan_block_size_from_path(
    const char *path, int *block_size);
101
static void darshan_get_shared_records(
102
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
103
static int darshan_log_open_all(
104
    char *logfile_name, MPI_File *log_fh);
105
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
106
107
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
108
static int darshan_log_write_record_hash(
109
    MPI_File log_fh, struct darshan_core_runtime *core,
110
111
112
    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
113
    int count, uint64_t *inout_off);
Shane Snyder's avatar
Shane Snyder committed
114
115
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
116

117
118
/* *********************************** */

Shane Snyder's avatar
Shane Snyder committed
119
void darshan_core_initialize(int argc, char **argv)
120
121
122
123
{
    int i;
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
124
    char *envstr;
125
126
127
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int chars_left = 0;
128
129
    int ret;
    int tmpval;
130
131

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
132
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
133
134
135
136

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

137
    if(internal_timing_flag)
138
139
140
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
141
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
142
    {
143
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
144
145
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
146
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
147
148
149
150
151
152
153
154
155
156
157
        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
        {
158
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
159
160
161
162
163
164
165
        }

        /* avoid floating point errors on faulty input */
        if (darshan_mem_alignment < 1)
        {
            darshan_mem_alignment = 1;
        }
166
167
168
169

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
170
        {
171
            memset(darshan_core, 0, sizeof(*darshan_core));
172

173
174
175
176
            darshan_core->log_job.uid = getuid();
            darshan_core->log_job.start_time = time(NULL);
            darshan_core->log_job.nprocs = nprocs;
            darshan_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();
177
178

            /* record exe and arguments */
179
            for(i=0; i<argc; i++)
180
            {
181
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
182
183
                strncat(darshan_core->exe, argv[i], chars_left);
                if(i < (argc-1))
184
                {
185
                    chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
186
                    strncat(darshan_core->exe, " ", chars_left);
187
188
189
190
191
192
193
194
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
195
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
196
                strncat(darshan_core->exe, __progname_full, chars_left);
197
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
198
                strncat(darshan_core->exe, " <unknown args>", chars_left);
199
200
201
202
203
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
204
                truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
205
                sprintf(&darshan_core->exe[truncate_offset], "%s",
206
207
                    truncate_string);
            }
208
209
210

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
211
212
        }

213
214
215
216
217
218
219
        /* maybe bootstrap modules with static initializers */
        i = 0;
        while(mod_static_init_fns[i])
        {
            (*mod_static_init_fns[i])();
            i++;
        }
220
221
    }

222
223
224
225
226
    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);
227
        if(my_rank == 0)
228
        {
229
230
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
231
232
233
234
235
236
        }
    }

    return;
}

Shane Snyder's avatar
Shane Snyder committed
237
void darshan_core_shutdown()
238
{
239
    int i;
240
    char *logfile_name;
241
    struct darshan_core_runtime *final_core;
242
    int internal_timing_flag = 0;
243
244
    char *envjobid;
    char *jobid_str;
245
    int jobid;
246
    struct tm *start_tm;
247
    time_t start_time_tmp;
248
249
    int ret = 0;
    int all_ret = 0;
250
251
    int64_t first_start_time;
    int64_t last_end_time;
252
253
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
254
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
255
    double start_log_time;
256
257
258
259
260
261
262
    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;
263
    uint64_t gz_fp = 0;
264
265
    MPI_File log_fh;
    MPI_Status status;
266
267
268
269

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

270
271
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

Shane Snyder's avatar
Shane Snyder committed
272
    /* disable darhan-core while we shutdown */
273
    DARSHAN_CORE_LOCK();
274
    if(!darshan_core)
275
    {
276
        DARSHAN_CORE_UNLOCK();
277
278
        return;
    }
279
280
    final_core = darshan_core;
    darshan_core = NULL;
Shane Snyder's avatar
Shane Snyder committed
281

282
    /* we also need to set which modules were registered on this process and
283
284
     * call into those modules and give them a chance to perform any necessary
     * pre-shutdown steps.
Shane Snyder's avatar
Shane Snyder committed
285
286
287
288
289
290
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
            local_mod_use[i] = 1;
291
            final_core->mod_array[i]->mod_funcs.begin_shutdown();
Shane Snyder's avatar
Shane Snyder committed
292
293
        }
    }
294
    DARSHAN_CORE_UNLOCK();
295
296
297
298

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
299
        darshan_core_cleanup(final_core);
300
301
302
        return;
    }

303
    /* set darshan job id/metadata and constuct log file name on rank 0 */
304
    if(my_rank == 0)
305
    {
306
        /* Use DARSHAN_JOBID_OVERRIDE for the env var for __DARSHAN_JOBID */
307
        envjobid = getenv(DARSHAN_JOBID_OVERRIDE);
308
        if(!envjobid)
309
        {
310
            envjobid = __DARSHAN_JOBID;
311
312
        }

313
        /* find a job id */
314
315
316
317
318
319
320
321
322
323
324
325
        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();
        }

326
        final_core->log_job.jobid = (int64_t)jobid;
327

328
        /* if we are using any hints to write the log file, then record those
329
         * hints with the darshan job information
330
         */
331
        darshan_log_record_hints_and_ver(final_core);
332

333
        /* use human readable start time format in log filename */
334
        start_time_tmp = final_core->log_job.start_time;
335
        start_tm = localtime(&start_time_tmp);
336

337
338
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
339
340
341
342
343
344
345
346
347
    }

    /* 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 */
348
        free(logfile_name);
349
        darshan_core_cleanup(final_core);
350
351
352
        return;
    }

353
    final_core->log_job.end_time = time(NULL);
354

355
356
357
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
358
359
    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);
360
361
    if(my_rank == 0)
    {
362
363
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
364
    }
365

366
367
368
    /* 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);

369
    /* get a list of records which are shared across all processes */
370
    darshan_get_shared_records(final_core, shared_recs);
371

372
373
    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
374
    /* collectively open the darshan log file */
375
    ret = darshan_log_open_all(logfile_name, &log_fh);
376
377
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
378
379
380
381
382
383
384
385

    /* 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)
        {
386
387
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
388
389
390
            unlink(logfile_name);
        }
        free(logfile_name);
391
        darshan_core_cleanup(final_core);
392
393
394
        return;
    }

395
396
    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
397
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
398
    if(my_rank == 0)
399
    {
400
        void *pointers[2] = {&final_core->log_job, final_core->trailing_data};
401
        int lengths[2] = {sizeof(struct darshan_job), strlen(final_core->trailing_data)};
402
        int comp_buf_sz = 0;
403

404
        /* compress the job info and the trailing mount/exe data */
Shane Snyder's avatar
Shane Snyder committed
405
        all_ret = darshan_deflate_buffer(pointers, lengths, 2,
406
407
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
408
        {
409
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
410
            unlink(logfile_name);
411
        }
412
413
414
        else
        {
            /* write the job information, preallocing space for the log header */
Shane Snyder's avatar
Shane Snyder committed
415
            gz_fp += sizeof(struct darshan_header);
416
417
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, gz_fp,
                final_core->comp_buf, comp_buf_sz, MPI_BYTE, &status);
418
419
420
421
422
            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
423
                
424
            }
425
            gz_fp += comp_buf_sz;
426
        }
427
428
    }

429
430
431
432
433
    /* 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);
434
        darshan_core_cleanup(final_core);
435
436
        return;
    }
437
438
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
439

440
441
    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
442
    /* write the record name->id hash to the log file */
Shane Snyder's avatar
Shane Snyder committed
443
    final_core->log_header.rec_map.off = gz_fp;
444
    ret = darshan_log_write_record_hash(log_fh, final_core, &gz_fp);
Shane Snyder's avatar
Shane Snyder committed
445
    final_core->log_header.rec_map.len = gz_fp - final_core->log_header.rec_map.off;
446

447
    /* error out if unable to write record hash */
448
449
450
451
452
453
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
454
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
455
                logfile_name);
456
            unlink(logfile_name);
457
458
        }
        free(logfile_name);
459
        darshan_core_cleanup(final_core);
460
461
        return;
    }
Shane Snyder's avatar
Shane Snyder committed
462
463
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
464
465

    /* loop over globally used darshan modules and:
466
     *      - perform shared file reductions, if possible
467
     *      - get final output buffer
468
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
469
     *      - append compressed buffer to log file
470
471
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
472
     */
473
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
474
    {
475
        struct darshan_core_module* this_mod = final_core->mod_array[i];
476
        struct darshan_core_record_ref *ref = NULL;
477
478
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        int mod_shared_rec_cnt = 0;
479
        void* mod_buf = NULL;
480
        int mod_buf_sz = 0;
481
        int j;
482

483
        if(global_mod_use_count[i] == 0)
484
485
        {
            if(my_rank == 0)
486
487
488
489
            {
                final_core->log_header.mod_map[i].off = 0;
                final_core->log_header.mod_map[i].len = 0;
            }
490
            continue;
491
        }
492
493
 
        if(internal_timing_flag)
494
            mod1[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
495

496
497
498
499
500
501
502
503
        /* set the shared file list for this module */
        memset(mod_shared_recs, 0, DARSHAN_CORE_MAX_RECORDS * sizeof(darshan_record_id));
        for(j = 0; j < DARSHAN_CORE_MAX_RECORDS && shared_recs[j] != 0; j++)
        {
            HASH_FIND(hlink, final_core->rec_hash, &shared_recs[j],
                sizeof(darshan_record_id), ref);
            assert(ref);
            if(DARSHAN_CORE_MOD_ISSET(ref->global_mod_flags, i))
504
            {
505
                mod_shared_recs[mod_shared_rec_cnt++] = shared_recs[j];
506
            }
507
        }
508

509
510
511
512
513
        /* 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.
         */
514
        if(this_mod)
515
        {
516
517
            this_mod->mod_funcs.get_output_data(MPI_COMM_WORLD, mod_shared_recs,
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
518
519
        }

520
        /* append this module's data to the darshan log */
Shane Snyder's avatar
Shane Snyder committed
521
522
523
524
        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;
525

526
        /* error out if the log append failed */
527
528
529
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
530
        {
531
532
533
534
535
536
537
538
            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);
539
            darshan_core_cleanup(final_core);
540
            return;
541
542
543
        }

        /* shutdown module if registered locally */
544
        if(this_mod)
545
546
547
        {
            this_mod->mod_funcs.shutdown();
        }
548
549
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
550
551
    }

552
553
    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
554
    /* rank 0 is responsible for writing the log header */
555
556
    if(my_rank == 0)
    {
557
558
559
        /* initialize the remaining header fields */
        strcpy(final_core->log_header.version_string, DARSHAN_LOG_VERSION);
        final_core->log_header.magic_nr = DARSHAN_MAGIC_NR;
560
        final_core->log_header.comp_type = DARSHAN_ZLIB_COMP;
561

Shane Snyder's avatar
Shane Snyder committed
562
563
564
        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)
565
        {
Shane Snyder's avatar
Shane Snyder committed
566
567
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
568
            unlink(logfile_name);
569
        }
570
571
    }

572
573
574
575
576
    /* 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);
577
        darshan_core_cleanup(final_core);
578
579
        return;
    }
580
581
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
582

583
584
585
    DARSHAN_MPI_CALL(PMPI_File_close)(&log_fh);

    /* if we got this far, there are no errors, so rename from *.darshan_partial
586
     * to *-<logwritetime>.darshan, which indicates that this log file is
587
588
     * complete and ready for analysis
     */
589
590
    if(my_rank == 0)
    {
Shane Snyder's avatar
Shane Snyder committed
591
        if(getenv("DARSHAN_LOGFILE"))
592
        {
593
#ifdef __DARSHAN_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
594
            chmod(logfile_name, (S_IRUSR|S_IRGRP));
595
#else
Shane Snyder's avatar
Shane Snyder committed
596
            chmod(logfile_name, (S_IRUSR));
597
#endif
Shane Snyder's avatar
Shane Snyder committed
598
599
600
601
602
603
604
605
606
607
608
609
610
611
        }
        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");
612
                sprintf(tmp_index, "_%d.darshan", (int)(end_log_time-start_log_time+1));
Shane Snyder's avatar
Shane Snyder committed
613
614
                rename(logfile_name, new_logfile_name);
                /* set permissions on log file */
615
#ifdef __DARSHAN_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
616
617
618
619
620
621
                chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
#else
                chmod(new_logfile_name, (S_IRUSR));
#endif
                free(new_logfile_name);
            }
622
        }
623
    }
624

625
    free(logfile_name);
626
    darshan_core_cleanup(final_core);
627

628
    if(internal_timing_flag)
629
    {
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
        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)
        {
664
665
666
667
668
            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);
669
670
671
            for(i = 0; i < DARSHAN_MAX_MODS; i++)
            {
                if(global_mod_use_count[i])
672
                    fprintf(stderr, "darshan:%s_shutdown\t%d\t%f\n", darshan_module_names[i],
Shane Snyder's avatar
Shane Snyder committed
673
                        nprocs, mod_slowest[i]);
674
            }
675
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
676
        }
677
678
679
680
    }
    
    return;
}
681

Shane Snyder's avatar
Shane Snyder committed
682
/* *********************************** */
683

684
/* construct the darshan log file name */
685
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
686
{
Shane Snyder's avatar
Shane Snyder committed
687
    char* user_logfile_name;
688
689
690
    char* logpath;
    char* logname_string;
    char* logpath_override = NULL;
691
#ifdef __DARSHAN_LOG_ENV
692
693
694
695
696
697
698
699
700
    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
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
    /* 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
716
    {
Shane Snyder's avatar
Shane Snyder committed
717
718
        /* otherwise, generate the log path automatically */

719
720
        /* 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
721
722
        if(!logpath)
        {
723
724
#ifdef __DARSHAN_LOG_PATH
            logpath = __DARSHAN_LOG_PATH;
725
#endif
Shane Snyder's avatar
Shane Snyder committed
726
        }
727

Shane Snyder's avatar
Shane Snyder committed
728
729
730
731
732
733
734
735
736
737
        /* 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.
         */
738
739

#ifndef DARSHAN_DISABLE_CUSERID
Shane Snyder's avatar
Shane Snyder committed
740
        cuserid(cuser);
741
742
#endif

Shane Snyder's avatar
Shane Snyder committed
743
744
        /* if cuserid() didn't work, then check the environment */
        if(strcmp(cuser, "") == 0)
745
        {
Shane Snyder's avatar
Shane Snyder committed
746
747
748
749
750
            logname_string = getenv("LOGNAME");
            if(logname_string)
            {
                strncpy(cuser, logname_string, (L_cuserid-1));
            }
751
752
        }

Shane Snyder's avatar
Shane Snyder committed
753
754
755
756
757
758
        /* 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);
        }
759

Shane Snyder's avatar
Shane Snyder committed
760
761
762
763
        /* 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);
764

Shane Snyder's avatar
Shane Snyder committed
765
766
767
768
        /* 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.
         */
769
#ifdef __DARSHAN_LOG_ENV
Shane Snyder's avatar
Shane Snyder committed
770
        /* just silently skip if the environment variable list is too big */
771
        if(strlen(__DARSHAN_LOG_ENV) < 256)
772
        {
Shane Snyder's avatar
Shane Snyder committed
773
            /* copy env variable list to a temporary buffer */
774
            strcpy(env_check, __DARSHAN_LOG_ENV);
Shane Snyder's avatar
Shane Snyder committed
775
776
777
            /* tokenize the comma-separated list */
            env_tok = strtok(env_check, ",");
            if(env_tok)
778
            {
Shane Snyder's avatar
Shane Snyder committed
779
                do
780
                {
Shane Snyder's avatar
Shane Snyder committed
781
782
783
784
785
786
787
788
789
                    /* 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, ",")));
            }
790
791
792
        }
#endif

Shane Snyder's avatar
Shane Snyder committed
793
        if(logpath_override)
794
        {
Shane Snyder's avatar
Shane Snyder committed
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
            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);
            }
810
        }
Shane Snyder's avatar
Shane Snyder committed
811
        else if(logpath)
812
        {
Shane Snyder's avatar
Shane Snyder committed
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
            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';
833
834
835
836
        }
    }

    return;
837
838
}

839
/* record any hints used to write the darshan log in the log header */
840
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
841
842
843
844
845
846
847
848
849
{
    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
     */
850
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
851
852
    if(!hints)
    {
853
        hints = __DARSHAN_LOG_HINTS;
854
855
856
857
858
859
860
861
862
863
    }

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

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

    meta_remain = DARSHAN_JOB_METADATA_LEN -
864
        strlen(core->log_job.metadata) - 1;
865
866
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
867
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
868
869
870
871
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
872
        m = core->log_job.metadata + strlen(core->log_job.metadata);
873
874
875
876
877
878
879
880
881
882
883
884
        /* 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;
}

885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
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 */
925
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
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
964
965
966
967
968
969
970
        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",
        "none",
        "rpc_pipefs",
        "hugetlbfs",
        "cgroup",
        NULL
    };

    /* length of exe has already been safety checked in darshan-posix.c */
    strcat(trailing_data, core->exe);
971
    space_left = DARSHAN_EXE_LEN - strlen(trailing_data);
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034

    /* we make two passes through mounted file systems; in the first pass we
     * grab any non-nfs mount points, then on the second pass we grab nfs
     * mount points
     */

    tab = setmntent("/etc/mtab", "r");
    if(!tab)
        return;
    /* loop through list of mounted file systems */
    while(mnt_data_count<DARSHAN_MAX_MNTS && (entry = getmntent(tab)) != NULL)
    {
        /* filter out excluded fs types */
        tmp_index = 0;
        skip = 0;
        while((exclude = fs_exclusions[tmp_index]))
        {
            if(!(strcmp(exclude, entry->mnt_type)))
            {
                skip =1;
                break;
            }
            tmp_index++;
        }

        if(skip || (strcmp(entry->mnt_type, "nfs") == 0))
            continue;

        add_entry(trailing_data, &space_left, entry);
    }
    endmntent(tab);

    tab = setmntent("/etc/mtab", "r");
    if(!tab)
        return;
    /* loop through list of mounted file systems */
    while(mnt_data_count<DARSHAN_MAX_MNTS && (entry = getmntent(tab)) != NULL)
    {
        if(strcmp(entry->mnt_type, "nfs") != 0)
            continue;

        add_entry(trailing_data, &space_left, entry);
    }
    endmntent(tab);

    /* Sort mount points in order of longest path to shortest path.  This is
     * necessary so that if we try to match file paths to mount points later
     * we don't match on "/" every time.
     */
    qsort(mnt_data_array, mnt_data_count, sizeof(mnt_data_array[0]), mnt_data_cmp);
    return;
}

/* darshan_get_exe_and_mounts()
 *
 * collects command line and list of mounted file systems into a string that
 * will be stored with the job header
 */
static char* darshan_get_exe_and_mounts(struct darshan_core_runtime *core)
{
    char* trailing_data;
    int space_left;

1035
    space_left = DARSHAN_EXE_LEN + 1;
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
    trailing_data = malloc(space_left);
    if(!trailing_data)
    {
        return(NULL);
    }
    memset(trailing_data, 0, space_left);

    if(my_rank == 0)
    {
        darshan_get_exe_and_mounts_root(core, trailing_data, space_left);
    }

    /* broadcast trailing data to all nodes */
    DARSHAN_MPI_CALL(PMPI_Bcast)(trailing_data, space_left, MPI_CHAR, 0,
        MPI_COMM_WORLD);
    /* broadcast mount count to all nodes */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&mnt_data_count, 1, MPI_INT, 0,
        MPI_COMM_WORLD);
    /* broadcast mount data to all nodes */
    DARSHAN_MPI_CALL(PMPI_Bcast)(mnt_data_array,
        mnt_data_count*sizeof(mnt_data_array[0]), MPI_BYTE, 0, MPI_COMM_WORLD);

    return(trailing_data);
}

1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
static void darshan_block_size_from_path(const char *path, int *block_size)
{
    int i;
    *block_size = -1;

    for(i=0; i<mnt_data_count; i++)
    {
        if(!(strncmp(mnt_data_array[i].path, path, strlen(mnt_data_array[i].path))))
        {
            *block_size = mnt_data_array[i].block_size;
            return;
        }
    }

    return;
}

1078
static void darshan_get_shared_records(struct darshan_core_runtime *core,
1079
    darshan_record_id *shared_recs)
1080
1081
1082
{
    int i;
    int ndx;
1083
    struct darshan_core_record_ref *tmp, *ref;
1084
    darshan_record_id id_array[DARSHAN_CORE_MAX_RECORDS] = {0};
1085
1086
    uint64_t mod_flags[DARSHAN_CORE_MAX_RECORDS] = {0};
    uint64_t global_mod_flags[DARSHAN_CORE_MAX_RECORDS] = {0};
1087
1088
1089
1090
1091

    /* first, determine list of records root process has opened */
    if(my_rank == 0)
    {
        ndx = 0;
1092
        HASH_ITER(hlink, core->rec_hash, ref, tmp)
1093
        {
1094
            id_array[ndx++] = ref->rec.id;           
1095
1096
1097
1098
        }
    }

    /* broadcast root's list of records to all other processes */
1099
    DARSHAN_MPI_CALL(PMPI_Bcast)(id_array,
1100
1101
1102
1103
1104
1105
        (DARSHAN_CORE_MAX_RECORDS * sizeof(darshan_record_id)),
        MPI_BYTE, 0, MPI_COMM_WORLD);

    /* everyone looks to see if they opened the same records as root */
    for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && id_array[i] != 0); i++)
    {
1106
1107
        HASH_FIND(hlink, core->rec_hash, &id_array[i], sizeof(darshan_record_id), ref);
        if(ref)
1108
        {
1109
1110
            /* we opened that record too, save the mod_flags */
            mod_flags[i] = ref->mod_flags;
1111
1112
1113
        }
    }

1114
1115
1116
1117
    /* now allreduce so everyone agrees which files are shared and
     * which modules accessed them collectively
     */
    DARSHAN_MPI_CALL(PMPI_Allreduce)(mod_flags, global_mod_flags,
Shane Snyder's avatar
Shane Snyder committed
1118
        DARSHAN_CORE_MAX_RECORDS, MPI_UINT64_T, MPI_BAND, MPI_COMM_WORLD);
1119
1120