darshan-core.c 53.2 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
#include <fcntl.h>
22
23
#include <sys/types.h>
#include <sys/stat.h>
24
#include <sys/mman.h>
25
#include <sys/vfs.h>
26
#include <zlib.h>
27
#include <mpi.h>
28
#include <assert.h>
29

30
#include "uthash.h"
Shane Snyder's avatar
Shane Snyder committed
31
#include "darshan.h"
32
#include "darshan-core.h"
Shane Snyder's avatar
Shane Snyder committed
33
#include "darshan-dynamic.h"
34

35
extern char* __progname;
36
extern char* __progname_full;
37

38
/* internal variable delcarations */
39
static struct darshan_core_runtime *darshan_core = NULL;
40
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
41
static int my_rank = -1;
42
static int nprocs = -1;
43
static int darshan_mem_alignment = 1;
44

45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
/* 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
};

60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
#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
75
76
77
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

78
79
80
81
82
83
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
84
    int block_size;
85
86
87
88
89
90
    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;

91
92
93
94
/* 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(
95
96
    struct darshan_core_runtime* core);
static void darshan_get_exe_and_mounts_root(
97
98
99
    struct darshan_core_runtime *core, int argc, char **argv);
static void darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core, int argc, char **argv);
100
101
static void darshan_block_size_from_path(
    const char *path, int *block_size);
102
static void darshan_get_shared_records(
103
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
104
static int darshan_log_open_all(
105
    char *logfile_name, MPI_File *log_fh);
106
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
107
108
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
109
static int darshan_log_write_record_hash(
110
    MPI_File log_fh, struct darshan_core_runtime *core,
111
112
113
    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
114
    int count, uint64_t *inout_off);
Shane Snyder's avatar
Shane Snyder committed
115
116
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
117

118
119
/* *********************************** */

Shane Snyder's avatar
Shane Snyder committed
120
void darshan_core_initialize(int argc, char **argv)
121
{
122
    struct darshan_core_runtime *init_core = NULL;
123
124
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
125
126
127
128
    char *mmap_log_name = "darshan-log.out";
    int mmap_fd;
    int mmap_size;
    int sys_page_size;
129
    char *envstr;
130
131
    char *jobid_str;
    int jobid;
132
133
    int ret;
    int tmpval;
134
    int i;
135
136

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
137
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
138
139
140
141

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

142
    if(internal_timing_flag)
143
144
145
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
146
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
147
    {
148
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
149
150
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
151
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
152
153
154
155
156
157
158
159
160
161
162
        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
        {
163
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
164
165
166
167
168
169
170
        }

        /* avoid floating point errors on faulty input */
        if (darshan_mem_alignment < 1)
        {
            darshan_mem_alignment = 1;
        }
171

172
173
174
        /* allocate structure to track darshan core runtime information */
        init_core = malloc(sizeof(*init_core));
        if(init_core)
175
        {
176
177
178
179
180
            memset(init_core, 0, sizeof(*init_core));
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

            sys_page_size = sysconf(_SC_PAGESIZE);
            assert(sys_page_size > 0);
181

182
183
184
185
186
187
188
            /* set the size of the mmap, making sure to round up to the
             * nearest page size. One mmap chunk is used for the job-level
             * metadata, and the rest are statically assigned to modules
             */
            mmap_size = (1 + DARSHAN_MAX_MODS) * DARSHAN_MMAP_CHUNK_SIZE;
            if(mmap_size % sys_page_size)
                mmap_size = ((mmap_size / sys_page_size) + 1) * sys_page_size;
189

190
191
192
            /* TODO: logfile name should have process rank in it for uniqueness */
            mmap_fd = open(mmap_log_name, O_CREAT|O_RDWR|O_EXCL , 0644);
            if(mmap_fd < 0)
193
            {
194
195
196
197
                fprintf(stderr, "darshan library warning: "
                    "unable to create darshan log file %s\n", mmap_log_name);
                free(init_core);
                return;
198
199
            }

200
201
202
203
204
205
206
207
208
209
210
211
212
213
            /* allocate the necessary space in the log file */
            ret = ftruncate(mmap_fd, mmap_size);
            if(ret < 0)
            {
                fprintf(stderr, "darshan library warning: "
                    "unable to allocate darshan log file %s\n", mmap_log_name);
                free(init_core);
                close(mmap_fd);
                unlink(mmap_log_name);
                return;
            }

            /* memory map buffers for getting at least some summary i/o data
             * into a log file if darshan does not shut down properly
214
             */
215
216
217
            init_core->mmap_p = mmap(NULL, mmap_size, PROT_WRITE, MAP_SHARED,
                mmap_fd, 0);
            if(init_core->mmap_p == MAP_FAILED)
218
            {
219
220
221
222
223
224
                fprintf(stderr, "darshan library warning: "
                    "unable to mmap darshan log file %s\n", mmap_log_name);
                free(init_core);
                close(mmap_fd);
                unlink(mmap_log_name);
                return;
225
226
            }

227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
            /* close darshan log file (this does *not* unmap the log file) */
            close(mmap_fd);

            /* set the pointers for each log file region */
            init_core->mmap_job_p = (struct darshan_job *)(init_core->mmap_p);
            init_core->mmap_exe_mnt_p =
                (char *)(((char *)init_core->mmap_p) + sizeof(struct darshan_job));
            init_core->mmap_mod_p =
                (void *)(((char *)init_core->mmap_p) + DARSHAN_MMAP_CHUNK_SIZE);

            /* set known job-level metadata files for the log file */
            init_core->mmap_job_p->uid = getuid();
            init_core->mmap_job_p->start_time = time(NULL);
            init_core->mmap_job_p->nprocs = nprocs;

            /* Use DARSHAN_JOBID_OVERRIDE for the env var for __DARSHAN_JOBID */
            envstr = getenv(DARSHAN_JOBID_OVERRIDE);
            if(!envstr)
245
            {
246
                envstr = __DARSHAN_JOBID;
247
            }
248

249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
            /* find a job id */
            jobid_str = getenv(envstr);
            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();
            }
            init_core->mmap_job_p->jobid = (int64_t)jobid;

            /* if we are using any hints to write the log file, then record those
             * hints with the darshan job information
             */
            darshan_log_record_hints_and_ver(init_core);

268
            /* collect information about command line and mounted file systems */
269
            darshan_get_exe_and_mounts(init_core, argc, argv);
270

271
272
273
274
275
276
277
278
279
280
281
            /* TODO: what would be needed in a termination routine? set job end time? */

            /* maybe bootstrap modules with static initializers */
            i = 0;
            while(mod_static_init_fns[i])
            {
                (*mod_static_init_fns[i])();
                i++;
            }

            darshan_core = init_core;
282
        }
283
284
    }

285
286
287
288
289
    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);
290
        if(my_rank == 0)
291
        {
292
293
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
294
295
296
297
298
299
        }
    }

    return;
}

Shane Snyder's avatar
Shane Snyder committed
300
void darshan_core_shutdown()
301
{
302
303
304

    return;
#if 0
305
    int i;
306
    char *logfile_name;
307
    struct darshan_core_runtime *final_core;
308
    int internal_timing_flag = 0;
309
310
    char *envjobid;
    char *jobid_str;
311
    int jobid;
312
    struct tm *start_tm;
313
    time_t start_time_tmp;
314
315
    int ret = 0;
    int all_ret = 0;
316
317
    int64_t first_start_time;
    int64_t last_end_time;
318
319
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
320
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
321
    double start_log_time;
322
323
324
325
326
327
328
    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;
329
    uint64_t gz_fp = 0;
330
    unsigned char tmp_partial_flag;
331
332
    MPI_File log_fh;
    MPI_Status status;
333
334
335
336

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

337
338
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

Shane Snyder's avatar
Shane Snyder committed
339
    /* disable darhan-core while we shutdown */
340
    DARSHAN_CORE_LOCK();
341
    if(!darshan_core)
342
    {
343
        DARSHAN_CORE_UNLOCK();
344
345
        return;
    }
346
347
    final_core = darshan_core;
    darshan_core = NULL;
Shane Snyder's avatar
Shane Snyder committed
348

349
    /* we also need to set which modules were registered on this process and
350
351
     * call into those modules and give them a chance to perform any necessary
     * pre-shutdown steps.
Shane Snyder's avatar
Shane Snyder committed
352
353
354
355
356
357
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
            local_mod_use[i] = 1;
358
            final_core->mod_array[i]->mod_funcs.begin_shutdown();
Shane Snyder's avatar
Shane Snyder committed
359
360
        }
    }
361
    DARSHAN_CORE_UNLOCK();
362
363
364
365

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
366
        darshan_core_cleanup(final_core);
367
368
369
        return;
    }

370
    /* set darshan job id/metadata and constuct log file name on rank 0 */
371
    if(my_rank == 0)
372
    {
373
        /* Use DARSHAN_JOBID_OVERRIDE for the env var for __DARSHAN_JOBID */
374
        envjobid = getenv(DARSHAN_JOBID_OVERRIDE);
375
        if(!envjobid)
376
        {
377
            envjobid = __DARSHAN_JOBID;
378
379
        }

380
        /* find a job id */
381
382
383
384
385
386
387
388
389
390
391
392
        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();
        }

393
        final_core->log_job.jobid = (int64_t)jobid;
394

395
        /* if we are using any hints to write the log file, then record those
396
         * hints with the darshan job information
397
         */
398
        darshan_log_record_hints_and_ver(final_core);
399

400
        /* use human readable start time format in log filename */
401
        start_time_tmp = final_core->log_job.start_time;
402
        start_tm = localtime(&start_time_tmp);
403

404
405
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
406
407
408
409
410
411
412
413
414
    }

    /* 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 */
415
        free(logfile_name);
416
        darshan_core_cleanup(final_core);
417
418
419
        return;
    }

420
    final_core->log_job.end_time = time(NULL);
421

422
423
424
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
425
426
    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);
427
428
    if(my_rank == 0)
    {
429
430
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
431
    }
432

433
434
435
    /* 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);

436
    /* get a list of records which are shared across all processes */
437
    darshan_get_shared_records(final_core, shared_recs);
438

439
440
    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
441
    /* collectively open the darshan log file */
442
    ret = darshan_log_open_all(logfile_name, &log_fh);
443
444
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
445
446
447
448
449
450
451
452

    /* 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)
        {
453
454
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
455
456
457
            unlink(logfile_name);
        }
        free(logfile_name);
458
        darshan_core_cleanup(final_core);
459
460
461
        return;
    }

462
463
    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
464
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
465
    if(my_rank == 0)
466
    {
467
        void *pointers[2] = {&final_core->log_job, final_core->trailing_data};
468
        int lengths[2] = {sizeof(struct darshan_job), strlen(final_core->trailing_data)};
469
        int comp_buf_sz = 0;
470

471
        /* compress the job info and the trailing mount/exe data */
Shane Snyder's avatar
Shane Snyder committed
472
        all_ret = darshan_deflate_buffer(pointers, lengths, 2,
473
474
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
475
        {
476
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
477
            unlink(logfile_name);
478
        }
479
480
481
        else
        {
            /* write the job information, preallocing space for the log header */
Shane Snyder's avatar
Shane Snyder committed
482
            gz_fp += sizeof(struct darshan_header);
483
484
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, gz_fp,
                final_core->comp_buf, comp_buf_sz, MPI_BYTE, &status);
485
486
487
488
489
            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
490
                
491
            }
492
            gz_fp += comp_buf_sz;
493
        }
494
495
    }

496
497
498
499
500
    /* 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);
501
        darshan_core_cleanup(final_core);
502
503
        return;
    }
504
505
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
506

507
508
    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
509
    /* write the record name->id hash to the log file */
Shane Snyder's avatar
Shane Snyder committed
510
    final_core->log_header.rec_map.off = gz_fp;
511
    ret = darshan_log_write_record_hash(log_fh, final_core, &gz_fp);
Shane Snyder's avatar
Shane Snyder committed
512
    final_core->log_header.rec_map.len = gz_fp - final_core->log_header.rec_map.off;
513

514
    /* error out if unable to write record hash */
515
516
517
518
519
520
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
521
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
522
                logfile_name);
523
            unlink(logfile_name);
524
525
        }
        free(logfile_name);
526
        darshan_core_cleanup(final_core);
527
528
        return;
    }
Shane Snyder's avatar
Shane Snyder committed
529
530
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
531
532

    /* loop over globally used darshan modules and:
533
     *      - perform shared file reductions, if possible
534
     *      - get final output buffer
535
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
536
     *      - append compressed buffer to log file
537
538
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
539
     */
540
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
541
    {
542
        struct darshan_core_module* this_mod = final_core->mod_array[i];
543
        struct darshan_core_record_ref *ref = NULL;
544
545
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        int mod_shared_rec_cnt = 0;
546
        void* mod_buf = NULL;
547
        int mod_buf_sz = 0;
548
        int j;
549

550
        if(global_mod_use_count[i] == 0)
551
552
        {
            if(my_rank == 0)
553
554
555
556
            {
                final_core->log_header.mod_map[i].off = 0;
                final_core->log_header.mod_map[i].len = 0;
            }
557
            continue;
558
        }
559
560
 
        if(internal_timing_flag)
561
            mod1[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
562

563
564
565
566
567
568
569
570
        /* 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))
571
            {
572
                mod_shared_recs[mod_shared_rec_cnt++] = shared_recs[j];
573
            }
574
        }
575

576
577
578
579
580
        /* 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.
         */
581
        if(this_mod)
582
        {
583
584
            this_mod->mod_funcs.get_output_data(MPI_COMM_WORLD, mod_shared_recs,
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
585
586
        }

587
        /* append this module's data to the darshan log */
Shane Snyder's avatar
Shane Snyder committed
588
589
590
591
        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;
592

593
        /* error out if the log append failed */
594
595
596
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
597
        {
598
599
600
601
602
603
604
605
            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);
606
            darshan_core_cleanup(final_core);
607
            return;
608
609
610
        }

        /* shutdown module if registered locally */
611
        if(this_mod)
612
613
614
        {
            this_mod->mod_funcs.shutdown();
        }
615
616
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
617
618
    }

619
620
621
622
623
624
625
    /* run a reduction to determine if any application processes had to set the
     * partial flag. this happens when a process has tracked too many records
     * at once and cannot track new records
     */
    DARSHAN_MPI_CALL(PMPI_Reduce)(&(final_core->log_header.partial_flag),
        &tmp_partial_flag, 1, MPI_UNSIGNED_CHAR, MPI_MAX, 0, MPI_COMM_WORLD);

626
627
    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
628
    /* rank 0 is responsible for writing the log header */
629
630
    if(my_rank == 0)
    {
631
632
633
        /* initialize the remaining header fields */
        strcpy(final_core->log_header.version_string, DARSHAN_LOG_VERSION);
        final_core->log_header.magic_nr = DARSHAN_MAGIC_NR;
634
        final_core->log_header.comp_type = DARSHAN_ZLIB_COMP;
635
        final_core->log_header.partial_flag = tmp_partial_flag;
636

Shane Snyder's avatar
Shane Snyder committed
637
638
639
        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)
640
        {
Shane Snyder's avatar
Shane Snyder committed
641
642
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
643
            unlink(logfile_name);
644
        }
645
646
    }

647
648
649
650
651
    /* 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);
652
        darshan_core_cleanup(final_core);
653
654
        return;
    }
655
656
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
657

658
659
660
    DARSHAN_MPI_CALL(PMPI_File_close)(&log_fh);

    /* if we got this far, there are no errors, so rename from *.darshan_partial
661
     * to *-<logwritetime>.darshan, which indicates that this log file is
662
663
     * complete and ready for analysis
     */
664
665
    if(my_rank == 0)
    {
Shane Snyder's avatar
Shane Snyder committed
666
        if(getenv("DARSHAN_LOGFILE"))
667
        {
668
#ifdef __DARSHAN_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
669
            chmod(logfile_name, (S_IRUSR|S_IRGRP));
670
#else
Shane Snyder's avatar
Shane Snyder committed
671
            chmod(logfile_name, (S_IRUSR));
672
#endif
Shane Snyder's avatar
Shane Snyder committed
673
674
675
676
677
678
679
680
681
682
683
684
685
686
        }
        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");
687
                sprintf(tmp_index, "_%d.darshan", (int)(end_log_time-start_log_time+1));
Shane Snyder's avatar
Shane Snyder committed
688
689
                rename(logfile_name, new_logfile_name);
                /* set permissions on log file */
690
#ifdef __DARSHAN_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
691
692
693
694
695
696
                chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
#else
                chmod(new_logfile_name, (S_IRUSR));
#endif
                free(new_logfile_name);
            }
697
        }
698
    }
699

700
    free(logfile_name);
701
    darshan_core_cleanup(final_core);
702

703
    if(internal_timing_flag)
704
    {
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
        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)
        {
739
740
741
742
743
            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);
744
745
746
            for(i = 0; i < DARSHAN_MAX_MODS; i++)
            {
                if(global_mod_use_count[i])
747
                    fprintf(stderr, "darshan:%s_shutdown\t%d\t%f\n", darshan_module_names[i],
Shane Snyder's avatar
Shane Snyder committed
748
                        nprocs, mod_slowest[i]);
749
            }
750
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
751
        }
752
    }
753

754
    return;
755
#endif
756
}
757

Shane Snyder's avatar
Shane Snyder committed
758
/* *********************************** */
759

760
/* construct the darshan log file name */
761
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
762
{
Shane Snyder's avatar
Shane Snyder committed
763
    char* user_logfile_name;
764
765
766
    char* logpath;
    char* logname_string;
    char* logpath_override = NULL;
767
#ifdef __DARSHAN_LOG_ENV
768
769
770
771
772
773
774
775
776
    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
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
    /* 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
792
    {
Shane Snyder's avatar
Shane Snyder committed
793
794
        /* otherwise, generate the log path automatically */

795
796
        /* 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
797
798
        if(!logpath)
        {
799
800
#ifdef __DARSHAN_LOG_PATH
            logpath = __DARSHAN_LOG_PATH;
801
#endif
Shane Snyder's avatar
Shane Snyder committed
802
        }
803

Shane Snyder's avatar
Shane Snyder committed
804
805
806
807
808
809
810
811
812
813
        /* 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.
         */
814
815

#ifndef DARSHAN_DISABLE_CUSERID
Shane Snyder's avatar
Shane Snyder committed
816
        cuserid(cuser);
817
818
#endif

Shane Snyder's avatar
Shane Snyder committed
819
820
        /* if cuserid() didn't work, then check the environment */
        if(strcmp(cuser, "") == 0)
821
        {
Shane Snyder's avatar
Shane Snyder committed
822
823
824
825
826
            logname_string = getenv("LOGNAME");
            if(logname_string)
            {
                strncpy(cuser, logname_string, (L_cuserid-1));
            }
827
828
        }

Shane Snyder's avatar
Shane Snyder committed
829
830
831
832
833
834
        /* 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);
        }
835

Shane Snyder's avatar
Shane Snyder committed
836
837
838
839
        /* 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);
840

Shane Snyder's avatar
Shane Snyder committed
841
842
843
844
        /* 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.
         */
845
#ifdef __DARSHAN_LOG_ENV
Shane Snyder's avatar
Shane Snyder committed
846
        /* just silently skip if the environment variable list is too big */
847
        if(strlen(__DARSHAN_LOG_ENV) < 256)
848
        {
Shane Snyder's avatar
Shane Snyder committed
849
            /* copy env variable list to a temporary buffer */
850
            strcpy(env_check, __DARSHAN_LOG_ENV);
Shane Snyder's avatar
Shane Snyder committed
851
852
853
            /* tokenize the comma-separated list */
            env_tok = strtok(env_check, ",");
            if(env_tok)
854
            {
Shane Snyder's avatar
Shane Snyder committed
855
                do
856
                {
Shane Snyder's avatar
Shane Snyder committed
857
858
859
860
861
862
863
864
865
                    /* 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, ",")));
            }
866
867
868
        }
#endif

Shane Snyder's avatar
Shane Snyder committed
869
        if(logpath_override)
870
        {
Shane Snyder's avatar
Shane Snyder committed
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
            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);
            }
886
        }
Shane Snyder's avatar
Shane Snyder committed
887
        else if(logpath)
888
        {
Shane Snyder's avatar
Shane Snyder committed
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
            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';
909
910
911
912
        }
    }

    return;
913
914
}

915
/* record any hints used to write the darshan log in the log header */
916
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
917
918
919
920
921
922
923
924
925
{
    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
     */
926
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
927
928
    if(!hints)
    {
929
        hints = __DARSHAN_LOG_HINTS;
930
931
932
933
934
935
936
937
938
939
    }

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

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

    meta_remain = DARSHAN_JOB_METADATA_LEN -
940
        strlen(core->mmap_job_p->metadata) - 1;
941
942
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
943
        sprintf(core->mmap_job_p->metadata, "lib_ver=%s\n", PACKAGE_VERSION);
944
945
946
947
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
948
        m = core->mmap_job_p->metadata + strlen(core->mmap_job_p->metadata);
949
950
951
952
953
954
955
956
957
958
959
960
        /* 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;
}

961
962
963
964
965
966
967
968
969
970
971
972
973
974
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 */
975
static void add_entry(char* buf, int* space_left, struct mntent *entry)
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
{
    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 */
1001
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
1002
1003
1004
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
1005
        strcat(buf, tmp_mnt);
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
        (*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,
1019
    int argc, char **argv)
1020
1021
1022
1023
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
1024
1025
1026
1027
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int space_left = DARSHAN_EXE_LEN;
    int i;
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
    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
    };

1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
    /* record exe and arguments */
    for(i=0; i<argc; i++)
    {
        strncat(core->mmap_exe_mnt_p, argv[i], space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->mmap_exe_mnt_p);
        if(i < (argc-1))
        {
            strncat(core->mmap_exe_mnt_p, " ", space_left);
            space_left = DARSHAN_EXE_LEN-strlen(core->mmap_exe_mnt_p);
        }
    }

    /* if we don't see any arguments, then use glibc symbol to get
     * program name at least (this happens in fortran)
     */
    if(argc == 0)
    {
        strncat(core->mmap_exe_mnt_p, __progname_full, space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->mmap_exe_mnt_p);
        strncat(core->mmap_exe_mnt_p, " <unknown args>", space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->mmap_exe_mnt_p);
    }

    if(space_left == 0)
    {
        /* we ran out of room; mark that string was truncated */
        truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
        sprintf(&core->mmap_exe_mnt_p[truncate_offset], "%s",
            truncate_string);
    }

1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107

    /* 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;

1108
        add_entry(core->mmap_exe_mnt_p, &space_left, entry);
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
    }
    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;

1121
        add_entry(core->mmap_exe_mnt_p, &space_left, entry);
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
    }
    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
 */
1138
1139
static void darshan_get_exe_and_mounts(struct darshan_core_runtime *core,
    int argc, char **argv)
1140
1141
1142
{
    if(my_rank == 0)
    {
1143
        darshan_get_exe_and_mounts_root(core, argc, argv);
1144
1145
1146
1147
1148
1149
1150
1151
1152
    }

    /* broadcast mount count to all nodes */