darshan-core.c 38.1 KB
Newer Older
1 2 3 4 5
/*
 *  (C) 2009 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

6
#define _XOPEN_SOURCE 500
7
#define _GNU_SOURCE
8

9 10 11 12 13 14
#include "darshan-runtime-config.h"

#include <stdio.h>
#ifdef HAVE_MNTENT_H
#include <mntent.h>
#endif
15 16 17 18 19
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <limits.h>
#include <pthread.h>
20 21 22
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/vfs.h>
23
#include <zlib.h>
24
#include <mpi.h>
25
#include <assert.h>
26

27
#include "uthash.h"
28 29
#include "darshan-core.h"

30 31
/* TODO is __progname_full needed here */
extern char* __progname;
32

33
/* internal variable delcarations */
34
static struct darshan_core_runtime *darshan_core = NULL;
35
static pthread_mutex_t darshan_core_mutex = PTHREAD_MUTEX_INITIALIZER;
36
static int my_rank = -1;
37
static int nprocs = -1;
38

39 40 41 42 43 44 45 46 47 48 49 50 51
/* 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;

52 53 54 55 56 57
/* prototypes for internal helper functions */
static void darshan_core_initialize(
    int *argc, char ***argv);
static void darshan_core_shutdown(
    void);
static void darshan_core_cleanup(
58
    struct darshan_core_runtime* core);
59 60 61
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
62 63 64 65 66 67
    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);
68
static void darshan_get_shared_record_ids(
69
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
70 71 72
static int darshan_log_coll_open(
    char *logfile_name, MPI_File *log_fh);
static int darshan_log_write_record_hash(
73 74 75 76
    MPI_File log_fh, struct darshan_core_record_ref *rec_hash,
    darshan_record_id *shared_recs, struct darshan_log_map *map);
static int darshan_log_coll_write(
    MPI_File log_fh, void *buf, int count, struct darshan_log_map *map);
77

78 79
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)
80

81
/* intercept MPI initialize and finalize to manage darshan core runtime */
82 83 84 85 86 87 88 89 90 91
int MPI_Init(int *argc, char ***argv)
{
    int ret;

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

92
    darshan_core_initialize(argc, argv);
93 94 95 96

    return(ret);
}

97
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
98 99 100 101
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
102
    if(ret != MPI_SUCCESS)
103 104 105 106
    {
        return(ret);
    }

107
    darshan_core_initialize(argc, argv);
108 109 110 111 112 113 114 115

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

116
    darshan_core_shutdown();
117 118 119 120 121

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

122 123
/* *********************************** */

124
static void darshan_core_initialize(int *argc, char ***argv)
125 126 127 128 129 130 131 132 133
{
    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);
134
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
135 136 137 138

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

139
    if(internal_timing_flag)
140 141 142
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
143
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
144
    {
145 146 147 148 149
        /* TODO: darshan mem alignment code? */

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
150
        {
151
            memset(darshan_core, 0, sizeof(*darshan_core));
152

153 154 155 156
            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)();
157 158 159 160

            /* record exe and arguments */
            for(i=0; i<(*argc); i++)
            {
161 162
                chars_left = CP_EXE_LEN-strlen(darshan_core->exe);
                strncat(darshan_core->exe, (*argv)[i], chars_left);
163 164
                if(i < ((*argc)-1))
                {
165 166
                    chars_left = CP_EXE_LEN-strlen(darshan_core->exe);
                    strncat(darshan_core->exe, " ", chars_left);
167 168 169 170 171 172 173 174
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
175 176 177 178
                chars_left = CP_EXE_LEN-strlen(darshan_core->exe);
                strncat(darshan_core->exe, __progname, chars_left);
                chars_left = CP_EXE_LEN-strlen(darshan_core->exe);
                strncat(darshan_core->exe, " <unknown args>", chars_left);
179 180 181 182 183 184
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
                truncate_offset = CP_EXE_LEN - strlen(truncate_string);
185
                sprintf(&darshan_core->exe[truncate_offset], "%s",
186 187
                    truncate_string);
            }
188 189 190

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
191 192 193 194 195 196 197 198
        }
    }

    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);
199
        if(my_rank == 0)
200 201 202 203 204 205 206 207 208
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

209
static void darshan_core_shutdown()
210
{
211
    int i;
212
    char *logfile_name;
213
    struct darshan_core_runtime *final_core;
214
    int internal_timing_flag = 0;
215 216
    char *envjobid;
    char *jobid_str;
217
    int jobid;
218
    struct tm *start_tm;
219
    time_t start_time_tmp;
220 221
    int ret = 0;
    int all_ret = 0;
222 223
    int64_t first_start_time;
    int64_t last_end_time;
224 225
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
226
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
227 228
    double start_log_time;
    long offset;
229
    struct darshan_header log_header;
230
    MPI_File log_fh;
231
    MPI_Offset tmp_off;
232
    MPI_Status status;
233 234 235 236

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

237
    DARSHAN_CORE_LOCK();
238
    if(!darshan_core)
239
    {
240
        DARSHAN_CORE_UNLOCK();
241 242 243 244 245
        return;
    }
    /* disable further tracing while hanging onto the data so that we can
     * write it out
     */
246 247
    final_core = darshan_core;
    darshan_core = NULL;
248
    DARSHAN_CORE_UNLOCK();
249

250 251
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

252 253 254
    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
255
        darshan_core_cleanup(final_core);
256 257 258
        return;
    }

259
    /* set darshan job id/metadata and constuct log file name on rank 0 */
260
    if(my_rank == 0)
261 262 263
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
264
        if(!envjobid)
265 266 267 268
        {
            envjobid = CP_JOBID;
        }

269
        /* find a job id */
270 271 272 273 274 275 276 277 278 279 280 281
        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();
        }

282
        final_core->log_job.jobid = (int64_t)jobid;
283

284
        /* if we are using any hints to write the log file, then record those
285
         * hints with the darshan job information
286
         */
287
        darshan_log_record_hints_and_ver(final_core);
288

289
        /* use human readable start time format in log filename */
290
        start_time_tmp = final_core->log_job.start_time;
291
        start_tm = localtime(&start_time_tmp);
292

293 294
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
295 296 297 298 299 300 301 302 303
    }

    /* 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 */
304
        free(logfile_name);
305
        darshan_core_cleanup(final_core);
306 307 308
        return;
    }

309
    final_core->log_job.end_time = time(NULL);
310

311 312 313
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
314 315
    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);
316 317
    if(my_rank == 0)
    {
318 319
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
320
    }
321

322 323 324
    /* set which local modules were actually used */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
325
        if(final_core->mod_array[i])
326 327 328 329 330 331
            local_mod_use[i] = 1;
    }

    /* 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);

332
    /* get a list of records which are shared across all processes */
333
    darshan_get_shared_record_ids(final_core, shared_recs);
334

335 336
    /* collectively open the darshan log file */
    ret = darshan_log_coll_open(logfile_name, &log_fh);
337 338 339 340 341 342 343 344

    /* 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)
        {
345 346
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
347 348 349
            unlink(logfile_name);
        }
        free(logfile_name);
350
        darshan_core_cleanup(final_core);
351 352 353
        return;
    }

354
    /* rank 0 is responsible for writing the darshan job information */
355
    if(my_rank == 0)
356
    {
357 358 359 360 361 362 363 364 365
        unsigned char tmp_buf[CP_JOB_RECORD_SIZE];
        unsigned char *tmp_ptr;

        /* pack the job info and exe/mount info into a buffer for writing */
        tmp_ptr = tmp_buf;
        memcpy(tmp_ptr, &final_core->log_job, sizeof(struct darshan_job));
        tmp_ptr += sizeof(struct darshan_job);
        memcpy(tmp_ptr, final_core->trailing_data, CP_EXE_LEN+1);

366
        /* write the job information, making sure to prealloc space for the log header */
367
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, sizeof(struct darshan_header),
368
                tmp_buf, CP_JOB_RECORD_SIZE, MPI_BYTE, &status);
369
        if(all_ret != MPI_SUCCESS)
370
        {
371 372 373
            fprintf(stderr, "darshan library warning: unable to write job data to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
374
        }
375

376
        /* TODO: after compression is added, this should be fixed */
377
        log_header.rec_map.off = sizeof(struct darshan_header) + CP_JOB_RECORD_SIZE;
378 379
    }

380 381 382 383 384
    /* 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);
385
        darshan_core_cleanup(final_core);
386 387 388 389
        return;
    }

    /* write the record name->id hash to the log file */
390
    ret = darshan_log_write_record_hash(log_fh, final_core->rec_hash,
391
        shared_recs, &log_header.rec_map);
392

393
    /* error out if unable to write record hash */
394 395 396 397 398 399 400 401
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
            fprintf(stderr, "darshan library warning: unable to write record map to log file %s\n",
                logfile_name);
402
            unlink(logfile_name);
403 404
        }
        free(logfile_name);
405
        darshan_core_cleanup(final_core);
406 407
        return;
    }
408 409 410

    /* loop over globally used darshan modules and:
     *      - get final output buffer
411
     *      - compress (zlib) provided output buffer
412
     *      - append compressed buffer to log file
413 414
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
415
     */
416
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
417
    {
418
        struct darshan_core_module* this_mod = final_core->mod_array[i];
419 420 421 422 423
        MPI_Comm mod_comm;
        void* mod_buf = NULL;
        int mod_buf_size = 0;

        if(!global_mod_use_count[i])
424 425 426 427
        {
            if(my_rank == 0)
                log_header.mod_map[i].off = log_header.mod_map[i].len = 0;

428
            continue;
429
        }
430 431 432 433

        /* create a communicator to use for shutting down the module */
        if(global_mod_use_count[i] == nprocs)
        {
434
            DARSHAN_MPI_CALL(PMPI_Comm_dup)(MPI_COMM_WORLD, &mod_comm);
435 436 437
        }
        else
        {
438
            DARSHAN_MPI_CALL(PMPI_Comm_split)(MPI_COMM_WORLD, local_mod_use[i], 0, &mod_comm);
439
        }
440

441 442 443 444 445 446 447
        /* if module is registered locally, get the corresponding output buffer */
        if(local_mod_use[i])
        {
            /* get output buffer from module */
            this_mod->mod_funcs.get_output_data(mod_comm, &mod_buf, &mod_buf_size);
        }

448 449 450 451 452 453 454 455
        /* set the starting offset of this module */
        if(tmp_off == 0)
            tmp_off = log_header.rec_map.off + log_header.rec_map.len;

        log_header.mod_map[i].off = tmp_off;

        /* write module data buffer to the darshan log file */
        ret = darshan_log_coll_write(log_fh, mod_buf, mod_buf_size, &log_header.mod_map[i]);
456 457 458 459 460

        /* error out if unable to write this module's data */
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
461
        {
462 463 464 465 466 467 468 469
            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);
470
            darshan_core_cleanup(final_core);
471
            return;
472 473
        }

474 475
        tmp_off += log_header.mod_map[i].len;

476 477 478 479 480 481 482
        /* shutdown module if registered locally */
        if(local_mod_use[i])
        {
            this_mod->mod_funcs.shutdown();
            this_mod = NULL;
        }

483
        DARSHAN_MPI_CALL(PMPI_Comm_free)(&mod_comm);
484 485
    }

486
    /* rank 0 is responsible for writing the log header */
487 488
    if(my_rank == 0)
    {
489 490 491 492 493
        /* initialize the remaining header fields */
        strcpy(log_header.version_string, CP_VERSION);
        log_header.magic_nr = CP_MAGIC_NR;
        log_header.comp_type = DARSHAN_GZ_COMP;

494
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &log_header,
495
            sizeof(struct darshan_header), MPI_BYTE, &status);
496
        if(all_ret != MPI_SUCCESS)
497
        {
498 499 500
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
501
        }
502 503
    }

504 505 506 507 508
    /* 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);
509
        darshan_core_cleanup(final_core);
510 511 512
        return;
    }

513 514 515 516 517 518
    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
     */
519
    /* TODO: support user given logfile path/name */
520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535
    if(my_rank == 0)
    {
        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 */
536
#ifdef __CP_GROUP_READABLE_LOGS
537
            chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
538
#else
539
            chmod(new_logfile_name, (S_IRUSR));
540
#endif
541 542
            free(new_logfile_name);
        }
543
    }
544

545
    free(logfile_name);
546
    darshan_core_cleanup(final_core);
547

548
    if(internal_timing_flag)
549
    {
550 551 552 553 554
        /* TODO: what do we want to time in new darshan version? */
    }
    
    return;
}
555

556
/* free darshan core data structures to shutdown */
557
static void darshan_core_cleanup(struct darshan_core_runtime* core)
558
{
559
    int i;
560

561
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
562
    {
563
        if(core->mod_array[i])
564
        {        
565 566
            free(core->mod_array[i]);
            core->mod_array[i] = NULL;
567
        }
568
    }
569

570
    free(core);
571 572 573 574

    return;
}

575
/* construct the darshan log file name */
576
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
577
{
578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707
    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;

    /* Use CP_LOG_PATH_OVERRIDE for the value or __CP_LOG_PATH */
    logpath = getenv(CP_LOG_PATH_OVERRIDE);
    if(!logpath)
    {
#ifdef __CP_LOG_PATH
        logpath = __CP_LOG_PATH;
#endif
    }

    /* 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.
     */

#ifndef DARSHAN_DISABLE_CUSERID
    cuserid(cuser);
#endif

    /* if cuserid() didn't work, then check the environment */
    if(strcmp(cuser, "") == 0)
    {
        logname_string = getenv("LOGNAME");
        if(logname_string)
        {
            strncpy(cuser, logname_string, (L_cuserid-1));
        }
    }

    /* 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);
    }

    /* 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);

    /* 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.
     */
#ifdef __CP_LOG_ENV
    /* just silently skip if the environment variable list is too big */
    if(strlen(__CP_LOG_ENV) < 256)
    {
        /* 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)
        {
            do
            {
                /* 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, ",")));
        }
    }
#endif

    if(logpath_override)
    {
        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);
        }
    }
    else if(logpath)
    {
        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';
    }

    return;
708 709
}

710
/* record any hints used to write the darshan log in the log header */
711
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734
{
    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 -
735
        strlen(core->log_job.metadata) - 1;
736 737
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
738
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
739 740 741 742
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
743
        m = core->log_job.metadata + strlen(core->log_job.metadata);
744 745 746 747 748 749 750 751 752 753 754 755
        /* 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;
}

756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795
static int mnt_data_cmp(const void* a, const void* b)
{
    const struct mnt_data *d_a = (const struct mnt_data*)a;
    const struct mnt_data *d_b = (const struct mnt_data*)b;

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

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

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

    /* store mount information for use in header of darshan log */
796
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
        strcat(trailing_data, tmp_mnt);
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

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

    /* skip these fs types */
    static char* fs_exclusions[] = {
        "tmpfs",
        "proc",
        "sysfs",
        "devpts",
        "binfmt_misc",
        "fusectl",
        "debugfs",
        "securityfs",
        "nfsd",
        "none",
        "rpc_pipefs",
        "hugetlbfs",
        "cgroup",
        NULL
    };

    /* length of exe has already been safety checked in darshan-posix.c */
    strcat(trailing_data, core->exe);
    space_left = CP_EXE_LEN - strlen(trailing_data);

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

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

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

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

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

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

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

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

    space_left = CP_EXE_LEN + 1;
    trailing_data = malloc(space_left);
    if(!trailing_data)
    {
        return(NULL);
    }
    memset(trailing_data, 0, space_left);

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

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

    return(trailing_data);
}

static void darshan_get_shared_record_ids(struct darshan_core_runtime *core,
933
    darshan_record_id *shared_recs)
934 935 936 937 938 939 940 941 942 943 944 945
{
    int i;
    int ndx;
    struct darshan_core_record_ref *ref, *tmp;
    darshan_record_id id_array[DARSHAN_CORE_MAX_RECORDS] = {0};
    darshan_record_id mask_array[DARSHAN_CORE_MAX_RECORDS] = {0};
    darshan_record_id all_mask_array[DARSHAN_CORE_MAX_RECORDS] = {0};

    /* first, determine list of records root process has opened */
    if(my_rank == 0)
    {
        ndx = 0;
946
        HASH_ITER(hlink, core->rec_hash, ref, tmp)
947
        {
948
            id_array[ndx++] = ref->rec.id;           
949 950 951 952
        }
    }

    /* broadcast root's list of records to all other processes */
953
    DARSHAN_MPI_CALL(PMPI_Bcast)(id_array,
954 955 956 957 958 959
        (DARSHAN_CORE_MAX_RECORDS * sizeof(darshan_record_id)),
        MPI_BYTE, 0, MPI_COMM_WORLD);

    /* everyone looks to see if they opened the same records as root */
    for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && id_array[i] != 0); i++)
    {
960
        HASH_ITER(hlink, core->rec_hash, ref, tmp)
961
        {
962
            if(id_array[i] == ref->rec.id)
963 964 965 966 967 968 969 970 971
            {
                /* we opened that record too */
                mask_array[i] = 1;
                break;
            }
        }
    }

    /* now allreduce so everyone agrees which files are shared */
972
    DARSHAN_MPI_CALL(PMPI_Allreduce)(mask_array, all_mask_array,
973 974 975 976 977 978 979 980 981 982 983
        DARSHAN_CORE_MAX_RECORDS, MPI_INT, MPI_LAND, MPI_COMM_WORLD);

    ndx = 0;
    for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && id_array[i] != 0); i++)
    {
        if(all_mask_array[i] != 0)
        {
            shared_recs[ndx++] = id_array[i];
        }
    }

984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044
    return;
}

static int darshan_log_coll_open(char *logfile_name, MPI_File *log_fh)
{
    char *hints;
    char *tok_str;
    char *orig_tok_str;
    char *key;
    char *value;
    char *saveptr = NULL;
    int ret;
    MPI_Info info;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
    MPI_Info_create(&info);

    hints = getenv(CP_LOG_HINTS_OVERRIDE);
    if(!hints)
    {
        hints = __CP_LOG_HINTS;
    }

    if(hints && strlen(hints) > 0)
    {
        tok_str = strdup(hints);
        if(tok_str)
        {
            orig_tok_str = tok_str;
            do
            {
                /* split string on semicolon */
                key = strtok_r(tok_str, ";", &saveptr);
                if(key)
                {
                    tok_str = NULL;
                    /* look for = sign splitting key/value pairs */
                    value = index(key, '=');
                    if(value)
                    {
                        /* break key and value into separate null terminated strings */
                        value[0] = '\0';
                        value++;
                        if(strlen(key) > 0)
                            MPI_Info_set(info, key, value);
                    }
                }
            }while(key != NULL);
            free(orig_tok_str);
        }
    }

    /* open the darshan log file for writing */
    ret = DARSHAN_MPI_CALL(PMPI_File_open)(MPI_COMM_WORLD, logfile_name,
        MPI_MODE_CREATE | MPI_MODE_WRONLY | MPI_MODE_EXCL, info, log_fh);
    if(ret < 0)
        return(-1);

    MPI_Info_free(&info);
1045 1046 1047
    return(0);
}

1048 1049 1050
/* NOTE: the map written to file may contain duplicate id->name entries if a
 *       record is opened by multiple ranks, but not all ranks
 */
1051
static int darshan_log_write_record_hash(MPI_File log_fh, struct darshan_core_record_ref *rec_hash,
1052
    darshan_record_id *shared_recs, struct darshan_log_map *map)
1053
{
1054 1055 1056 1057 1058
    int i;
    int ret;
    struct darshan_core_record_ref *ref, *tmp;
    uint32_t name_len;
    size_t record_sz;
1059 1060 1061
    size_t hash_buf_sz = 0;
    unsigned char *hash_buf;
    unsigned char *hash_buf_off;
1062 1063
    MPI_Status status;

1064
    /* non-root ranks (rank > 0) remove shared records from their map --
1065 1066 1067 1068 1069 1070
     * these records will be written by rank 0
     */
    if(my_rank > 0)
    {
        for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && shared_recs[i]); i++)
        {
1071
            HASH_FIND(hlink, rec_hash, &shared_recs[i], sizeof(darshan_record_id), ref);
1072
            assert(ref); /* this id had better be in the hash ... */
1073
            HASH_DELETE(hlink, rec_hash, ref);
1074 1075 1076 1077 1078 1079 1080
            if(ref->rec.name) free(ref->rec.name);
            free(ref);
        }
    }

    /* allocate a buffer to store at most 64 bytes for each of a max number of records */
    /* NOTE: this buffer may be reallocated if estimate is too small */
1081 1082 1083
    hash_buf_sz = DARSHAN_CORE_MAX_RECORDS * 64;
    hash_buf = malloc(hash_buf_sz);
    if(!hash_buf)
1084
    {
1085
        return(-1);
1086 1087
    }

1088
    /* serialize the record map into a buffer for writing */
1089 1090
    hash_buf_off = hash_buf;
    HASH_ITER(hlink, rec_hash, ref, tmp)
1091 1092 1093 1094
    {
        name_len = strlen(ref->rec.name);
        record_sz = sizeof(darshan_record_id) + sizeof(int) + name_len;
        /* make sure there is room in the buffer for this record */
1095
        if((hash_buf_off + record_sz) > (hash_buf + hash_buf_sz))
1096 1097 1098 1099 1100
        {
            unsigned char *tmp_buf;
            size_t old_buf_sz;

            /* if no room, reallocate the map buffer at twice the current size */
1101 1102 1103
            old_buf_sz = hash_buf_off - hash_buf;
            hash_buf_sz *= 2;
            tmp_buf = malloc(hash_buf_sz);
1104 1105
            if(!tmp_buf)
            {
1106
                free(hash_buf);
1107
                return(-1);
1108 1109
            }

1110 1111 1112 1113
            memcpy(tmp_buf, hash_buf, old_buf_sz);
            free(hash_buf);
            hash_buf = tmp_buf;
            hash_buf_off = hash_buf + old_buf_sz;
1114 1115 1116 1117 1118 1119
        }

        /* now serialize the record into the map buffer.
         * NOTE: darshan record map serialization method: 
         *          ... darshan_record_id | (uint32_t) path_len | path ...
         */
1120 1121 1122 1123 1124 1125
        *((darshan_record_id *)hash_buf_off) = ref->rec.id;
        hash_buf_off += sizeof(darshan_record_id);
        *((uint32_t *)hash_buf_off) = name_len;
        hash_buf_off += sizeof(uint32_t);
        memcpy(hash_buf_off, ref->rec.name, name_len);
        hash_buf_off += name_len;
1126 1127 1128
    }

    /* collectively write out the record map to the darshan log */
1129
    if(hash_buf_off > hash_buf)
1130 1131
    {
        /* we have records to contribute to the collective write of the record map */
1132
        ret = darshan_log_coll_write(log_fh, hash_buf, (hash_buf_off-hash_buf), map);
1133 1134 1135 1136
    }
    else
    {
        /* we have no data to write, but participate in the collective anyway */
1137
        ret = darshan_log_coll_write(log_fh, NULL, 0, map);
1138 1139
    }

1140
    free(hash_buf);
1141

1142 1143 1144 1145 1146 1147
    if(ret < 0)
        return(-1);

    return(0);
}

1148 1149 1150 1151 1152 1153
/* NOTE: The in/out param 'map' is only valid on rank 0 and is used
 *       to provide the starting offset of this collective write and
 *       to store the aggregate size of this write upon completion.
 *       This implies ONLY rank 0 can specify the starting offset
 *       and that only rank 0 knows the ending log file offset upon
 *       return from this function (starting off + aggregate size).
1154
 */
1155 1156
static int darshan_log_coll_write(MPI_File log_fh, void *buf, int count,
    struct darshan_log_map *map)
1157 1158 1159 1160 1161
{
    MPI_Offset send_off, my_off;
    MPI_Status status;
    int ret;

1162
    /* figure out where everyone is writing using scan */
1163 1164
    send_off = count;
    if(my_rank == 0)
1165 1166 1167
    {
        send_off += map->off; /* rank 0 knows the beginning offset */
    }
1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179

    DARSHAN_MPI_CALL(PMPI_Scan)(&send_off, &my_off, 1, MPI_OFFSET,
        MPI_SUM, MPI_COMM_WORLD);
    /* scan in inclusive; subtract local size back out */
    my_off -= count;

    /* perform the collective write */
    ret = DARSHAN_MPI_CALL(PMPI_File_write_at_all)(log_fh, my_off, buf,
        count, MPI_BYTE, &status);
    if(ret < 0)
        return(-1);

1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200
    /* send the ending offset from rank (n-1) to rank 0 */
    if(nprocs > 1)
    {
        if(my_rank == (nprocs-1))
        {
            my_off += count;
            DARSHAN_MPI_CALL(PMPI_Send)(&my_off, 1, MPI_OFFSET, 0, 0,
                MPI_COMM_WORLD);
        }
        else if(my_rank == 0)
        {
            DARSHAN_MPI_CALL(PMPI_Recv)(&my_off, 1, MPI_OFFSET, (nprocs-1), 0,
                MPI_COMM_WORLD, &status);

            map->len = my_off - map->off;
        }
    }
    else
    {
        map->len = my_off + count - map->off;
    }
1201 1202

    return(0);
1203 1204
}

1205
/* ********************************************************* */
1206 1207

void darshan_core_register_module(
1208
    darshan_module_id id,
1209 1210 1211
    struct darshan_module_funcs *funcs,
    int *runtime_mem_limit)
{
1212
    struct darshan_core_module* mod;
1213

1214
    DARSHAN_CORE_LOCK();
1215

1216
    *runtime_mem_limit = 0;
1217
    if(!darshan_core || (id >= DARSHAN_MAX_MODS))
1218
    {
1219
        DARSHAN_CORE_UNLOCK();
1220
        return;
1221
    }
1222 1223

    /* see if this module is already registered */
1224
    if(darshan_core->mod_array[id])
1225
    {
1226
        /* if module is already registered just return */
1227
        /* NOTE: we do not recalculate memory limit here, just set to 0 */
1228
        DARSHAN_CORE_UNLOCK();
1229
        return;
1230 1231
    }

1232
    /* this module has not been registered yet, allocate and initialize it */
1233 1234
    mod = malloc(sizeof(*mod));
    if(!mod)
1235
    {
1236
        DARSHAN_CORE_UNLOCK();
1237 1238
        return;
    }
1239
    memset(mod, 0, sizeof(*mod));
1240

1241
    mod->id = id;
1242
    mod->mod_funcs = *funcs;
1243 1244

    /* register module with darshan */
1245
    darshan_core->mod_array[id] = mod;
1246

1247 1248
    /* TODO: something smarter than just 2 MiB per module */
    *runtime_mem_limit = 2 * 1024 * 1024;
1249

1250
    DARSHAN_CORE_UNLOCK();
1251

1252 1253 1254
    return;
}

1255
void darshan_core_lookup_record_id(
1256 1257 1258
    void *name,
    int len,
    int printable_flag,
1259
    darshan_record_id *id)
1260
{
1261 1262
    darshan_record_id tmp_id;
    struct darshan_core_record_ref* ref;
1263

1264
    if(!darshan_core || !name)
1265 1266 1267 1268 1269 1270
        return;

    /* TODO: what do you do with printable flag? */

    /* hash the input name to get a unique id for this record */
    tmp_id = darshan_hash(name, len, 0);
1271 1272 1273 1274
 
    DARSHAN_CORE_LOCK();

    /* check to see if we've already stored the id->name mapping for this record */
1275
    HASH_FIND(hlink, darshan_core->rec_hash, &tmp_id, sizeof(darshan_record_id), ref);
1276 1277 1278 1279 1280 1281
    if(!ref)
    {
        /* if not, add this record to the hash */
        ref = malloc(sizeof(struct darshan_core_record_ref));
        if(ref)
        {
1282 1283 1284 1285
            ref->rec.id = tmp_id;
            ref->rec.name = malloc(strlen(name) + 1);
            if(ref->rec.name)
                strcpy(ref->rec.name, name);
1286

1287
            HASH_ADD(hlink, darshan_core->rec_hash, rec.id, sizeof(darshan_record_id), ref);
1288 1289 1290 1291
        }
    }   

    DARSHAN_CORE_UNLOCK();
1292

1293 1294
    printf("New Darshan record: %s (%"PRIu64")\n", (char *)name, tmp_id);

1295
    *id = tmp_id;
1296 1297 1298
    return;
}

1299 1300
double darshan_core_wtime()
{
1301
    if(!darshan_core)
1302 1303 1304 1305
    {
        return(0);
    }

1306
    return(DARSHAN_MPI_CALL(PMPI_Wtime)() - darshan_core->wtime_offset);
1307
}
1308 1309 1310 1311 1312 1313 1314 1315 1316

/*
 * Local variables:
 *  c-indent-level: 4
 *  c-basic-offset: 4
 * End:
 *
 * vim: ts=8 sts=4 sw=4 expandtab
 */