darshan-core.c 51.2 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 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
static int darshan_mem_alignment = 1;
40

41 42 43 44 45 46 47 48 49 50 51 52 53 54 55
/* 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
56 57 58
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

59 60 61 62 63 64 65 66 67 68 69 70 71
/* 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;

72 73
/* prototypes for internal helper functions */
static void darshan_core_initialize(
74
    int argc, char **argv);
75 76 77
static void darshan_core_shutdown(
    void);
static void darshan_core_cleanup(
78
    struct darshan_core_runtime* core);
79 80 81
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
82 83 84 85 86 87
    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);
88
static void darshan_get_shared_records(
89
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
90
static int darshan_log_open_all(
91
    char *logfile_name, MPI_File *log_fh);
92 93 94
static int darshan_deflate_buffer(
    void **pointers, int *lengths, int count, int nocomp_flag,
    char *comp_buf, int *comp_length);
95
static int darshan_log_write_record_hash(
96
    MPI_File log_fh, struct darshan_core_runtime *core,
97 98 99 100
    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);
101

102
/* intercept MPI initialize and finalize to manage darshan core runtime */
103 104 105 106 107 108 109 110 111 112
int MPI_Init(int *argc, char ***argv)
{
    int ret;

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

113 114 115 116 117 118 119 120 121
    if(argc && argv)
    {
        darshan_core_initialize(*argc, *argv);
    }
    else
    {
        /* we don't see argc and argv here in fortran */
        darshan_core_initialize(0, NULL);
    }
122 123 124 125

    return(ret);
}

126
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
127 128 129 130
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
131
    if(ret != MPI_SUCCESS)
132 133 134 135
    {
        return(ret);
    }

136 137 138 139 140 141 142 143 144
    if(argc && argv)
    {
        darshan_core_initialize(*argc, *argv);
    }
    else
    {
        /* we don't see argc and argv here in fortran */
        darshan_core_initialize(0, NULL);
    }
145 146 147 148 149 150 151 152

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

153
    darshan_core_shutdown();
154 155 156 157 158

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

159 160
/* *********************************** */

161
static void darshan_core_initialize(int argc, char **argv)
162 163 164 165 166 167 168 169 170
{
    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);
171
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
172 173 174 175

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

176
    if(internal_timing_flag)
177 178 179
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
180
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
181
    {
182 183 184 185 186
        /* TODO: darshan mem alignment code? */

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
187
        {
188
            memset(darshan_core, 0, sizeof(*darshan_core));
189

190 191 192 193
            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)();
194 195

            /* record exe and arguments */
196
            for(i=0; i<argc; i++)
197
            {
198
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
199 200
                strncat(darshan_core->exe, argv[i], chars_left);
                if(i < (argc-1))
201
                {
202
                    chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
203
                    strncat(darshan_core->exe, " ", chars_left);
204 205 206 207 208 209 210 211
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
212
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
213
                strncat(darshan_core->exe, __progname_full, chars_left);
214
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
215
                strncat(darshan_core->exe, " <unknown args>", chars_left);
216 217 218 219 220
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
221
                truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
222
                sprintf(&darshan_core->exe[truncate_offset], "%s",
223 224
                    truncate_string);
            }
225 226 227

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
228 229 230 231 232 233 234 235
        }
    }

    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);
236
        if(my_rank == 0)
237 238 239 240 241 242 243 244 245
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

246
static void darshan_core_shutdown()
247
{
248
    int i;
249
    char *logfile_name;
250
    struct darshan_core_runtime *final_core;
251
    int internal_timing_flag = 0;
252 253
    char *envjobid;
    char *jobid_str;
254
    int jobid;
255
    struct tm *start_tm;
256
    time_t start_time_tmp;
257 258
    int ret = 0;
    int all_ret = 0;
259 260
    int64_t first_start_time;
    int64_t last_end_time;
261 262
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
263
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
264
    double start_log_time;
265 266 267 268 269 270 271
    double open1, open2;
    double job1, job2;
    double rec1, rec2;
    double mod1[DARSHAN_MAX_MODS] = {0};
    double mod2[DARSHAN_MAX_MODS] = {0};
    double header1, header2;
    double tm_end;
272 273
    uint64_t gz_fp = 0;
    uint64_t tmp_off = 0;
274 275
    MPI_File log_fh;
    MPI_Status status;
276 277 278 279

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

280 281
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

Shane Snyder's avatar
Shane Snyder committed
282
    /* disable darhan-core while we shutdown */
283
    DARSHAN_CORE_LOCK();
284
    if(!darshan_core)
285
    {
286
        DARSHAN_CORE_UNLOCK();
287 288
        return;
    }
289 290
    final_core = darshan_core;
    darshan_core = NULL;
Shane Snyder's avatar
Shane Snyder committed
291

292
    /* we also need to set which modules were registered on this process and
Shane Snyder's avatar
Shane Snyder committed
293 294 295 296 297 298 299 300 301 302
     * 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();
        }
    }
303
    DARSHAN_CORE_UNLOCK();
304 305 306 307

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
308
        darshan_core_cleanup(final_core);
309 310 311
        return;
    }

312
    /* set darshan job id/metadata and constuct log file name on rank 0 */
313
    if(my_rank == 0)
314 315 316
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
317
        if(!envjobid)
318 319 320 321
        {
            envjobid = CP_JOBID;
        }

322
        /* find a job id */
323 324 325 326 327 328 329 330 331 332 333 334
        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();
        }

335
        final_core->log_job.jobid = (int64_t)jobid;
336

337
        /* if we are using any hints to write the log file, then record those
338
         * hints with the darshan job information
339
         */
340
        darshan_log_record_hints_and_ver(final_core);
341

342
        /* use human readable start time format in log filename */
343
        start_time_tmp = final_core->log_job.start_time;
344
        start_tm = localtime(&start_time_tmp);
345

346 347
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
348 349 350 351 352 353 354 355 356
    }

    /* 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 */
357
        free(logfile_name);
358
        darshan_core_cleanup(final_core);
359 360 361
        return;
    }

362
    final_core->log_job.end_time = time(NULL);
363

364 365 366
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
367 368
    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);
369 370
    if(my_rank == 0)
    {
371 372
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
373
    }
374

375 376 377
    /* 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);

378
    /* get a list of records which are shared across all processes */
379
    darshan_get_shared_records(final_core, shared_recs);
380

381 382
    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
383
    /* collectively open the darshan log file */
384
    ret = darshan_log_open_all(logfile_name, &log_fh);
385 386
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
387 388 389 390 391 392 393 394

    /* 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)
        {
395 396
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
397 398 399
            unlink(logfile_name);
        }
        free(logfile_name);
400
        darshan_core_cleanup(final_core);
401 402 403
        return;
    }

404 405
    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
406
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
407
    if(my_rank == 0)
408
    {
409 410 411
        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;
412

413
        /* compress the job info and the trailing mount/exe data */
414
        all_ret = darshan_deflate_buffer(pointers, lengths, 2, 0,
415 416
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
417
        {
418
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
419
            unlink(logfile_name);
420
        }
421 422 423
        else
        {
            /* write the job information, preallocing space for the log header */
424 425 426
            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);
427 428 429 430 431
            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
432
                
433
            }
Shane Snyder's avatar
Shane Snyder committed
434

435 436
            /* set the beginning offset of record hash, which follows job info just written */
            gz_fp += comp_buf_sz;
437
        }
438 439
    }

440 441 442 443 444
    /* 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);
445
        darshan_core_cleanup(final_core);
446 447
        return;
    }
448 449
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
450

451 452
    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
453
    /* write the record name->id hash to the log file */
454 455
    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;
456

457
    /* error out if unable to write record hash */
458 459 460 461 462 463
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
464
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
465
                logfile_name);
466
            unlink(logfile_name);
467 468
        }
        free(logfile_name);
469
        darshan_core_cleanup(final_core);
470 471
        return;
    }
Shane Snyder's avatar
Shane Snyder committed
472 473
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
474 475

    /* loop over globally used darshan modules and:
476
     *      - perform shared file reductions, if possible
477
     *      - get final output buffer
478
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
479
     *      - append compressed buffer to log file
480 481
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
482
     */
483
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
484
    {
485
        struct darshan_core_module* this_mod = final_core->mod_array[i];
486 487
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        struct darshan_core_record_ref *ref = NULL;
488
        void* mod_buf = NULL;
489
        int mod_buf_sz = 0;
490
        int j;
491

492
        if(global_mod_use_count[i] == 0)
493 494
        {
            if(my_rank == 0)
495 496 497 498
            {
                final_core->log_header.mod_map[i].off = 0;
                final_core->log_header.mod_map[i].len = 0;
            }
499
            continue;
500
        }
501 502 503 504
 
        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
505
        if(global_mod_use_count[i] == nprocs)
506
        {
507 508 509 510 511 512 513 514 515 516 517 518 519
            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
520
                if(DARSHAN_CORE_MOD_ISSET(ref->global_mod_flags, i))
521 522 523 524 525 526
                {
                    mod_shared_recs[shared_rec_count++] = shared_recs[j];
                }
            }

            /* if there are globally shared files, do a shared file reduction */
527 528
            if(shared_rec_count && this_mod->mod_funcs.prepare_for_reduction &&
               this_mod->mod_funcs.record_reduction_op)
529 530 531 532 533 534 535 536 537 538 539 540 541
            {
                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 */
542
                    DARSHAN_MPI_CALL(PMPI_Op_create)(this_mod->mod_funcs.record_reduction_op,
543 544 545 546 547 548 549 550 551 552
                        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);
                }
            }
553
        }
554

555
        /* if module is registered locally, get the corresponding output buffer */
556
        if(this_mod)
557 558
        {
            /* get output buffer from module */
559
            this_mod->mod_funcs.get_output_data(&mod_buf, &mod_buf_sz);
560 561
        }

562
        final_core->log_header.mod_map[i].off = tmp_off;
563

564 565 566 567
        /* 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;
568

569
        /* error out if the log append failed */
570 571 572
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
573
        {
574 575 576 577 578 579 580 581
            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);
582
            darshan_core_cleanup(final_core);
583
            return;
584 585 586
        }

        /* shutdown module if registered locally */
587
        if(this_mod)
588 589 590
        {
            this_mod->mod_funcs.shutdown();
        }
591 592
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
593 594
    }

595 596
    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
597
    /* rank 0 is responsible for writing the log header */
598 599
    if(my_rank == 0)
    {
600 601 602
        void *header_buf = &(final_core->log_header);
        int header_buf_sz = sizeof(struct darshan_header);
        int comp_buf_sz = 0;
603

604 605 606 607 608 609 610 611 612 613 614 615 616 617
        /* 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)
618
        {
619
            fprintf(stderr, "darshan library warning: unable to compress header\n");
620
            unlink(logfile_name);
621
        }
622 623 624 625 626 627 628 629 630 631 632
        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);
            }
        }
633 634
    }

635 636 637 638 639
    /* 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);
640
        darshan_core_cleanup(final_core);
641 642
        return;
    }
643 644
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
645

646 647 648 649 650 651
    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
     */
652 653
    if(my_rank == 0)
    {
Shane Snyder's avatar
Shane Snyder committed
654
        if(getenv("DARSHAN_LOGFILE"))
655
        {
656
#ifdef __CP_GROUP_READABLE_LOGS
Shane Snyder's avatar
Shane Snyder committed
657
            chmod(logfile_name, (S_IRUSR|S_IRGRP));
658
#else
Shane Snyder's avatar
Shane Snyder committed
659
            chmod(logfile_name, (S_IRUSR));
660
#endif
Shane Snyder's avatar
Shane Snyder committed
661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684
        }
        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);
            }
685
        }
686
    }
687

688
    free(logfile_name);
689
    darshan_core_cleanup(final_core);
690

691
    if(internal_timing_flag)
692
    {
693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730
        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
731
            printf("darshan:header_write\t%d\t%f\n", nprocs, header_slowest);
732 733 734 735
            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
736
                        nprocs, mod_slowest[i]);
737 738 739
            }
            printf("darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
        }
740 741 742 743
    }
    
    return;
}
744

745
/* free darshan core data structures to shutdown */
746
static void darshan_core_cleanup(struct darshan_core_runtime* core)
747
{
748
    struct darshan_core_record_ref *tmp, *ref;
749
    int i;
750

751 752 753 754 755 756
    HASH_ITER(hlink, core->rec_hash, ref, tmp)
    {
        HASH_DELETE(hlink, core->rec_hash, ref);
        free(ref->rec.name);
        free(ref);
    }
757

758
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
759
    {
760
        if(core->mod_array[i])
761
        {        
762 763
            free(core->mod_array[i]);
            core->mod_array[i] = NULL;
764
        }
765
    }
766

767
    free(core->trailing_data);
768
    free(core);
769 770 771 772

    return;
}

773
/* construct the darshan log file name */
774
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
775
{
Shane Snyder's avatar
Shane Snyder committed
776
    char* user_logfile_name;
777 778 779 780 781 782 783 784 785 786 787 788 789
    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
790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
    /* 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
805
    {
Shane Snyder's avatar
Shane Snyder committed
806 807 808 809 810 811
        /* 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)
        {
812
#ifdef __CP_LOG_PATH
Shane Snyder's avatar
Shane Snyder committed
813
            logpath = __CP_LOG_PATH;
814
#endif
Shane Snyder's avatar
Shane Snyder committed
815
        }
816

Shane Snyder's avatar
Shane Snyder committed
817 818 819 820 821 822 823 824 825 826
        /* 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.
         */
827 828

#ifndef DARSHAN_DISABLE_CUSERID
Shane Snyder's avatar
Shane Snyder committed
829
        cuserid(cuser);
830 831
#endif

Shane Snyder's avatar
Shane Snyder committed
832 833
        /* if cuserid() didn't work, then check the environment */
        if(strcmp(cuser, "") == 0)
834
        {
Shane Snyder's avatar
Shane Snyder committed
835 836 837 838 839
            logname_string = getenv("LOGNAME");
            if(logname_string)
            {
                strncpy(cuser, logname_string, (L_cuserid-1));
            }
840 841
        }

Shane Snyder's avatar
Shane Snyder committed
842 843 844 845 846 847
        /* 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);
        }
848

Shane Snyder's avatar
Shane Snyder committed
849 850 851 852
        /* 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);
853

Shane Snyder's avatar
Shane Snyder committed
854 855 856 857
        /* 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.
         */
858
#ifdef __CP_LOG_ENV
Shane Snyder's avatar
Shane Snyder committed
859 860
        /* just silently skip if the environment variable list is too big */
        if(strlen(__CP_LOG_ENV) < 256)
861
        {
Shane Snyder's avatar
Shane Snyder committed
862 863 864 865 866
            /* 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)
867
            {
Shane Snyder's avatar
Shane Snyder committed
868
                do
869
                {
Shane Snyder's avatar
Shane Snyder committed
870 871 872 873 874 875 876 877 878
                    /* 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, ",")));
            }
879 880 881
        }
#endif

Shane Snyder's avatar
Shane Snyder committed
882
        if(logpath_override)
883
        {
Shane Snyder's avatar
Shane Snyder committed
884 885 886 887 888 889 890 891 892 893 894 895 896 897 898
            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);
            }
899
        }
Shane Snyder's avatar
Shane Snyder committed
900
        else if(logpath)
901
        {
Shane Snyder's avatar
Shane Snyder committed
902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921
            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';
922 923 924 925
        }
    }

    return;
926 927
}

928
/* record any hints used to write the darshan log in the log header */
929
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952
{
    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 -
953
        strlen(core->log_job.metadata) - 1;
954 955
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
956
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
957 958 959 960
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
961
        m = core->log_job.metadata + strlen(core->log_job.metadata);