darshan-core.c 51 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"
29 30
#include "darshan-core.h"

31
extern char* __progname;
32
extern char* __progname_full;
33

34
/* internal variable delcarations */
35
static struct darshan_core_runtime *darshan_core = NULL;
36
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
37
static int my_rank = -1;
38
static int nprocs = -1;
39

40 41 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/",
NULL
};

Shane Snyder's avatar
Shane Snyder committed
55 56 57
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

58 59 60 61 62 63 64 65 66 67 68 69 70
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
    int64_t block_size;
    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;

71 72
/* prototypes for internal helper functions */
static void darshan_core_initialize(
73
    int argc, char **argv);
74 75 76
static void darshan_core_shutdown(
    void);
static void darshan_core_cleanup(
77
    struct darshan_core_runtime* core);
78 79 80
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
81 82 83 84 85 86
    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);
87
static void darshan_get_shared_records(
88
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
89
static int darshan_log_open_all(
90
    char *logfile_name, MPI_File *log_fh);
91 92 93
static int darshan_deflate_buffer(
    void **pointers, int *lengths, int count, int nocomp_flag,
    char *comp_buf, int *comp_length);
94
static int darshan_log_write_record_hash(
95
    MPI_File log_fh, struct darshan_core_runtime *core,
96 97 98 99
    uint64_t *inout_off);
static int darshan_log_append_all(
    MPI_File log_fh, struct darshan_core_runtime *core, void *buf,
    int count, uint64_t *inout_off, uint64_t *agg_uncomp_sz);
100

101
/* intercept MPI initialize and finalize to manage darshan core runtime */
102 103 104 105 106 107 108 109 110 111
int MPI_Init(int *argc, char ***argv)
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init)(argc, argv);
    if(ret != MPI_SUCCESS)
    {
        return(ret);
    }

112 113 114 115 116 117 118 119 120
    if(argc && argv)
    {
        darshan_core_initialize(*argc, *argv);
    }
    else
    {
        /* we don't see argc and argv here in fortran */
        darshan_core_initialize(0, NULL);
    }
121 122 123 124

    return(ret);
}

125
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
126 127 128 129
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
130
    if(ret != MPI_SUCCESS)
131 132 133 134
    {
        return(ret);
    }

135 136 137 138 139 140 141 142 143
    if(argc && argv)
    {
        darshan_core_initialize(*argc, *argv);
    }
    else
    {
        /* we don't see argc and argv here in fortran */
        darshan_core_initialize(0, NULL);
    }
144 145 146 147 148 149 150 151

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

152
    darshan_core_shutdown();
153 154 155 156 157

    ret = DARSHAN_MPI_CALL(PMPI_Finalize)();
    return(ret);
}

158 159
/* *********************************** */

160
static void darshan_core_initialize(int argc, char **argv)
161 162 163 164 165 166 167 168 169
{
    int i;
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int chars_left = 0;

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
170
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
171 172 173 174

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

175
    if(internal_timing_flag)
176 177 178
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
179
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
180
    {
181 182 183 184 185
        /* TODO: darshan mem alignment code? */

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
186
        {
187
            memset(darshan_core, 0, sizeof(*darshan_core));
188

189 190 191 192
            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)();
193 194

            /* record exe and arguments */
195
            for(i=0; i<argc; i++)
196
            {
197
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
198 199
                strncat(darshan_core->exe, argv[i], chars_left);
                if(i < (argc-1))
200
                {
201
                    chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
202
                    strncat(darshan_core->exe, " ", chars_left);
203 204 205 206 207 208 209 210
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
211
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
212
                strncat(darshan_core->exe, __progname_full, chars_left);
213
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
214
                strncat(darshan_core->exe, " <unknown args>", chars_left);
215 216 217 218 219
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
220
                truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
221
                sprintf(&darshan_core->exe[truncate_offset], "%s",
222 223
                    truncate_string);
            }
224 225 226

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
227 228 229 230 231 232 233 234
        }
    }

    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);
235
        if(my_rank == 0)
236 237 238 239 240 241 242 243 244
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

245
static void darshan_core_shutdown()
246
{
247
    int i;
248
    char *logfile_name;
249
    struct darshan_core_runtime *final_core;
250
    int internal_timing_flag = 0;
251 252
    char *envjobid;
    char *jobid_str;
253
    int jobid;
254
    struct tm *start_tm;
255
    time_t start_time_tmp;
256 257
    int ret = 0;
    int all_ret = 0;
258 259
    int64_t first_start_time;
    int64_t last_end_time;
260 261
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
262
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
263
    double start_log_time;
264 265 266 267 268 269 270
    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;
271 272
    uint64_t gz_fp = 0;
    uint64_t tmp_off = 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
    /* we also need to set which modules were registered on this process and
Shane Snyder's avatar
Shane Snyder committed
292 293 294 295 296 297 298 299 300 301
     * disable tracing within those modules while we shutdown
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
            local_mod_use[i] = 1;
            final_core->mod_array[i]->mod_funcs.disable_instrumentation();
        }
    }
302
    DARSHAN_CORE_UNLOCK();
303 304 305 306

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
307
        darshan_core_cleanup(final_core);
308 309 310
        return;
    }

311
    /* set darshan job id/metadata and constuct log file name on rank 0 */
312
    if(my_rank == 0)
313 314 315
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
316
        if(!envjobid)
317 318 319 320
        {
            envjobid = CP_JOBID;
        }

321
        /* find a job id */
322 323 324 325 326 327 328 329 330 331 332 333
        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();
        }

334
        final_core->log_job.jobid = (int64_t)jobid;
335

336
        /* if we are using any hints to write the log file, then record those
337
         * hints with the darshan job information
338
         */
339
        darshan_log_record_hints_and_ver(final_core);
340

341
        /* use human readable start time format in log filename */
342
        start_time_tmp = final_core->log_job.start_time;
343
        start_tm = localtime(&start_time_tmp);
344

345 346
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
347 348 349 350 351 352 353 354 355
    }

    /* 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 */
356
        free(logfile_name);
357
        darshan_core_cleanup(final_core);
358 359 360
        return;
    }

361
    final_core->log_job.end_time = time(NULL);
362

363 364 365
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
366 367
    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);
368 369
    if(my_rank == 0)
    {
370 371
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
372
    }
373

374 375 376
    /* 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);

377
    /* get a list of records which are shared across all processes */
378
    darshan_get_shared_records(final_core, shared_recs);
379

380 381
    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
382
    /* collectively open the darshan log file */
383
    ret = darshan_log_open_all(logfile_name, &log_fh);
384 385
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
386 387 388 389 390 391 392 393

    /* 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)
        {
394 395
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
396 397 398
            unlink(logfile_name);
        }
        free(logfile_name);
399
        darshan_core_cleanup(final_core);
400 401 402
        return;
    }

403 404
    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
405
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
406
    if(my_rank == 0)
407
    {
408 409 410
        void *pointers[2] = {&final_core->log_job, final_core->trailing_data};
        int lengths[2] = {sizeof(struct darshan_job), DARSHAN_EXE_LEN+1};
        int comp_buf_sz = 0;
411

412
        /* compress the job info and the trailing mount/exe data */
413
        all_ret = darshan_deflate_buffer(pointers, lengths, 2, 0,
414 415
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
416
        {
417
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
418
            unlink(logfile_name);
419
        }
420 421 422
        else
        {
            /* write the job information, preallocing space for the log header */
423 424 425
            gz_fp += sizeof(struct darshan_header) + 23; /* gzip headers/trailers ... */
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, gz_fp,
                final_core->comp_buf, comp_buf_sz, MPI_BYTE, &status);
426 427 428 429 430
            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
431
                
432
            }
Shane Snyder's avatar
Shane Snyder committed
433

434 435
            /* set the beginning offset of record hash, which follows job info just written */
            gz_fp += comp_buf_sz;
436
        }
437 438
    }

439 440 441 442 443
    /* 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);
444
        darshan_core_cleanup(final_core);
445 446
        return;
    }
447 448
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
449

450 451
    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
452
    /* write the record name->id hash to the log file */
453 454
    ret = darshan_log_write_record_hash(log_fh, final_core, &gz_fp);
    tmp_off = final_core->log_header.rec_map.off + final_core->log_header.rec_map.len;
455

456
    /* error out if unable to write record hash */
457 458 459 460 461 462
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
463
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
464
                logfile_name);
465
            unlink(logfile_name);
466 467
        }
        free(logfile_name);
468
        darshan_core_cleanup(final_core);
469 470
        return;
    }
Shane Snyder's avatar
Shane Snyder committed
471 472
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
473 474

    /* loop over globally used darshan modules and:
475
     *      - perform shared file reductions, if possible
476
     *      - get final output buffer
477
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
478
     *      - append compressed buffer to log file
479 480
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
481
     */
482
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
483
    {
484
        struct darshan_core_module* this_mod = final_core->mod_array[i];
485 486
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        struct darshan_core_record_ref *ref = NULL;
487
        void* mod_buf = NULL;
488
        int mod_buf_sz = 0;
489
        int j;
490

491
        if(global_mod_use_count[i] == 0)
492 493
        {
            if(my_rank == 0)
494 495 496 497
            {
                final_core->log_header.mod_map[i].off = 0;
                final_core->log_header.mod_map[i].len = 0;
            }
498
            continue;
499
        }
500 501 502 503
 
        if(internal_timing_flag)
            mod1[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();   
        /* if all processes used this module, prepare to do a shared file reduction */
Shane Snyder's avatar
Shane Snyder committed
504
        if(global_mod_use_count[i] == nprocs)
505
        {
506 507 508 509 510 511 512 513 514 515 516 517 518
            int shared_rec_count = 0;
            int rec_sz = 0;
            void *red_send_buf = NULL, *red_recv_buf = NULL;
            MPI_Datatype red_type;
            MPI_Op red_op;

            /* 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);
Shane Snyder's avatar
Shane Snyder committed
519
                if(DARSHAN_CORE_MOD_ISSET(ref->global_mod_flags, i))
520 521 522 523 524 525
                {
                    mod_shared_recs[shared_rec_count++] = shared_recs[j];
                }
            }

            /* if there are globally shared files, do a shared file reduction */
526 527
            if(shared_rec_count && this_mod->mod_funcs.prepare_for_reduction &&
               this_mod->mod_funcs.record_reduction_op)
528 529 530 531 532 533 534 535 536 537 538 539 540
            {
                this_mod->mod_funcs.prepare_for_reduction(mod_shared_recs, &shared_rec_count,
                    &red_send_buf, &red_recv_buf, &rec_sz);

                if(shared_rec_count)
                {
                    /* construct a datatype for a file record.  This is serving no purpose
                     * except to make sure we can do a reduction on proper boundaries
                     */
                    DARSHAN_MPI_CALL(PMPI_Type_contiguous)(rec_sz, MPI_BYTE, &red_type);
                    DARSHAN_MPI_CALL(PMPI_Type_commit)(&red_type);

                    /* register a reduction operator for this module */
541
                    DARSHAN_MPI_CALL(PMPI_Op_create)(this_mod->mod_funcs.record_reduction_op,
542 543 544 545 546 547 548 549 550 551
                        1, &red_op);

                    /* reduce shared file records for this module */
                    DARSHAN_MPI_CALL(PMPI_Reduce)(red_send_buf, red_recv_buf,
                        shared_rec_count, red_type, red_op, 0, MPI_COMM_WORLD);

                    DARSHAN_MPI_CALL(PMPI_Type_free)(&red_type);
                    DARSHAN_MPI_CALL(PMPI_Op_free)(&red_op);
                }
            }
552
        }
553

554
        /* if module is registered locally, get the corresponding output buffer */
555
        if(this_mod)
556 557
        {
            /* get output buffer from module */
558
            this_mod->mod_funcs.get_output_data(&mod_buf, &mod_buf_sz);
559 560
        }

561
        final_core->log_header.mod_map[i].off = tmp_off;
562

563 564 565 566
        /* append this module's data to the darshan log */
        ret = darshan_log_append_all(log_fh, final_core, mod_buf, mod_buf_sz,
            &gz_fp, &(final_core->log_header.mod_map[i].len));
        tmp_off += final_core->log_header.mod_map[i].len;
567

568
        /* error out if the log append failed */
569 570 571
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
572
        {
573 574 575 576 577 578 579 580
            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);
581
            darshan_core_cleanup(final_core);
582
            return;
583 584 585
        }

        /* shutdown module if registered locally */
586
        if(this_mod)
587 588 589
        {
            this_mod->mod_funcs.shutdown();
        }
590 591
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
592 593
    }

594 595
    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
596
    /* rank 0 is responsible for writing the log header */
597 598
    if(my_rank == 0)
    {
599 600 601
        void *header_buf = &(final_core->log_header);
        int header_buf_sz = sizeof(struct darshan_header);
        int comp_buf_sz = 0;
602

603 604 605 606 607 608 609 610 611 612 613 614 615 616
        /* initialize the remaining header fields */
        strcpy(final_core->log_header.version_string, DARSHAN_LOG_VERSION);
        final_core->log_header.magic_nr = DARSHAN_MAGIC_NR;

        /* deflate the header */
        /* NOTE: the header is not actually compressed because space for it must
         *       be preallocated before writing. i.e., the "compressed" header
         *       must be constant sized, sizeof(struct darshan_header) + 23.
         *       it is still necessary to deflate the header or the resulting log
         *       file will not be a valid gzip file.
         */
        all_ret = darshan_deflate_buffer((void **)&header_buf, &header_buf_sz, 1, 1,
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
617
        {
618
            fprintf(stderr, "darshan library warning: unable to compress header\n");
619
            unlink(logfile_name);
620
        }
621 622 623 624 625 626 627 628 629 630 631
        else
        {
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, final_core->comp_buf,
                comp_buf_sz, MPI_BYTE, &status);
            if(all_ret != MPI_SUCCESS)
            {
                fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                        logfile_name);
                unlink(logfile_name);
            }
        }
632 633
    }

634 635 636 637 638
    /* 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);
639
        darshan_core_cleanup(final_core);
640 641
        return;
    }
642 643
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
644

645 646 647 648 649 650
    DARSHAN_MPI_CALL(PMPI_File_close)(&log_fh);

    /* if we got this far, there are no errors, so rename from *.darshan_partial
     * to *-<logwritetime>.darshan.gz, which indicates that this log file is
     * complete and ready for analysis
     */
651 652
    if(my_rank == 0)
    {
Shane Snyder's avatar
Shane Snyder committed
653
        if(getenv("DARSHAN_LOGFILE"))
654
        {
655
#ifdef __CP_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
656
            chmod(logfile_name, (S_IRUSR|S_IRGRP));
657
#else
Shane Snyder's avatar
Shane Snyder committed
658
            chmod(logfile_name, (S_IRUSR));
659
#endif
Shane Snyder's avatar
Shane Snyder committed
660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683
        }
        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");
                sprintf(tmp_index, "_%d.darshan.gz", (int)(end_log_time-start_log_time+1));
                rename(logfile_name, new_logfile_name);
                /* set permissions on log file */
#ifdef __CP_GROUP_READABLE_LOGS
                chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
#else
                chmod(new_logfile_name, (S_IRUSR));
#endif
                free(new_logfile_name);
            }
684
        }
685
    }
686

687
    free(logfile_name);
688
    darshan_core_cleanup(final_core);
689

690
    if(internal_timing_flag)
691
    {
692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729
        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)
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:log_open\t%d\t%f\n", nprocs, open_slowest);
            printf("darshan:job_write\t%d\t%f\n", nprocs, job_slowest);
            printf("darshan:hash_write\t%d\t%f\n", nprocs, rec_slowest);
Shane Snyder's avatar
Shane Snyder committed
730
            printf("darshan:header_write\t%d\t%f\n", nprocs, header_slowest);
731 732 733 734
            for(i = 0; i < DARSHAN_MAX_MODS; i++)
            {
                if(global_mod_use_count[i])
                    printf("darshan:%s_shutdown\t%d\t%f\n", darshan_module_names[i],
Shane Snyder's avatar
Shane Snyder committed
735
                        nprocs, mod_slowest[i]);
736 737 738
            }
            printf("darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
        }
739 740 741 742
    }
    
    return;
}
743

744
/* free darshan core data structures to shutdown */
745
static void darshan_core_cleanup(struct darshan_core_runtime* core)
746
{
747
    struct darshan_core_record_ref *tmp, *ref;
748
    int i;
749

750 751 752 753 754 755
    HASH_ITER(hlink, core->rec_hash, ref, tmp)
    {
        HASH_DELETE(hlink, core->rec_hash, ref);
        free(ref->rec.name);
        free(ref);
    }
756

757
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
758
    {
759
        if(core->mod_array[i])
760
        {        
761 762
            free(core->mod_array[i]);
            core->mod_array[i] = NULL;
763
        }
764
    }
765

766
    free(core->trailing_data);
767
    free(core);
768 769 770 771

    return;
}

772
/* construct the darshan log file name */
773
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
774
{
Shane Snyder's avatar
Shane Snyder committed
775
    char* user_logfile_name;
776 777 778 779 780 781 782 783 784 785 786 787 788
    char* logpath;
    char* logname_string;
    char* logpath_override = NULL;
#ifdef __CP_LOG_ENV
    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
789 790 791 792 793 794 795 796 797 798 799 800 801 802 803
    /* 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
804
    {
Shane Snyder's avatar
Shane Snyder committed
805 806 807 808 809 810
        /* otherwise, generate the log path automatically */

        /* Use CP_LOG_PATH_OVERRIDE for the value or __CP_LOG_PATH */
        logpath = getenv(CP_LOG_PATH_OVERRIDE);
        if(!logpath)
        {
811
#ifdef __CP_LOG_PATH
Shane Snyder's avatar
Shane Snyder committed
812
            logpath = __CP_LOG_PATH;
813
#endif
Shane Snyder's avatar
Shane Snyder committed
814
        }
815

Shane Snyder's avatar
Shane Snyder committed
816 817 818 819 820 821 822 823 824 825
        /* 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.
         */
826 827

#ifndef DARSHAN_DISABLE_CUSERID
Shane Snyder's avatar
Shane Snyder committed
828
        cuserid(cuser);
829 830
#endif

Shane Snyder's avatar
Shane Snyder committed
831 832
        /* if cuserid() didn't work, then check the environment */
        if(strcmp(cuser, "") == 0)
833
        {
Shane Snyder's avatar
Shane Snyder committed
834 835 836 837 838
            logname_string = getenv("LOGNAME");
            if(logname_string)
            {
                strncpy(cuser, logname_string, (L_cuserid-1));
            }
839 840
        }

Shane Snyder's avatar
Shane Snyder committed
841 842 843 844 845 846
        /* 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);
        }
847

Shane Snyder's avatar
Shane Snyder committed
848 849 850 851
        /* 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);
852

Shane Snyder's avatar
Shane Snyder committed
853 854 855 856
        /* 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.
         */
857
#ifdef __CP_LOG_ENV
Shane Snyder's avatar
Shane Snyder committed
858 859
        /* just silently skip if the environment variable list is too big */
        if(strlen(__CP_LOG_ENV) < 256)
860
        {
Shane Snyder's avatar
Shane Snyder committed
861 862 863 864 865
            /* copy env variable list to a temporary buffer */
            strcpy(env_check, __CP_LOG_ENV);
            /* tokenize the comma-separated list */
            env_tok = strtok(env_check, ",");
            if(env_tok)
866
            {
Shane Snyder's avatar
Shane Snyder committed
867
                do
868
                {
Shane Snyder's avatar
Shane Snyder committed
869 870 871 872 873 874 875 876 877
                    /* 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, ",")));
            }
878 879 880
        }
#endif

Shane Snyder's avatar
Shane Snyder committed
881
        if(logpath_override)
882
        {
Shane Snyder's avatar
Shane Snyder committed
883 884 885 886 887 888 889 890 891 892 893 894 895 896 897
            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);
            }
898
        }
Shane Snyder's avatar
Shane Snyder committed
899
        else if(logpath)
900
        {
Shane Snyder's avatar
Shane Snyder committed
901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920
            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';
921 922 923 924
        }
    }

    return;
925 926
}

927
/* record any hints used to write the darshan log in the log header */
928
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951
{
    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
     */
    hints = getenv(CP_LOG_HINTS_OVERRIDE);
    if(!hints)
    {
        hints = __CP_LOG_HINTS;
    }

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

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

    meta_remain = DARSHAN_JOB_METADATA_LEN -
952
        strlen(core->log_job.metadata) - 1;
953 954
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
955
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
956 957 958 959
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
960
        m = core->log_job.metadata + strlen(core->log_job.metadata);
961 962 963 964 965 966