darshan-core.c 31.3 KB
Newer Older
1 2 3 4 5
/*
 *  (C) 2009 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

6
#define _GNU_SOURCE
7

8 9 10 11 12 13
#include "darshan-runtime-config.h"

#include <stdio.h>
#ifdef HAVE_MNTENT_H
#include <mntent.h>
#endif
14 15 16 17 18
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <limits.h>
#include <pthread.h>
19 20 21 22
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/vfs.h>
#include <mpi.h>
23
#include <assert.h>
24

25
#include "uthash.h"
26 27
#include "darshan-core.h"

28 29
/* TODO is __progname_full needed here */
extern char* __progname;
30

31
/* internal variable delcarations */
32 33
static struct darshan_core_runtime *darshan_core_job = NULL;
static pthread_mutex_t darshan_core_mutex = PTHREAD_MUTEX_INITIALIZER;
34
static int my_rank = -1;
35
static int nprocs = -1;
36

37 38 39 40 41 42 43 44 45 46 47 48 49
/* 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(
    struct darshan_core_runtime* job);
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
    struct darshan_core_runtime* job);
static int darshan_get_shared_record_ids(
    struct darshan_core_runtime *job, darshan_record_id *shared_recs);
Shane Snyder's avatar
Shane Snyder committed
50
static int darshan_log_write_record_map(
51 52 53 54
    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);
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
/* intercept MPI initialize and finalize to manage darshan core runtime */
60 61 62 63 64 65 66 67 68 69
int MPI_Init(int *argc, char ***argv)
{
    int ret;

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

70
    darshan_core_initialize(argc, argv);
71 72 73 74

    return(ret);
}

75
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
76 77 78 79
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
80
    if(ret != MPI_SUCCESS)
81 82 83 84
    {
        return(ret);
    }

85
    darshan_core_initialize(argc, argv);
86 87 88 89 90 91 92 93

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

94
    darshan_core_shutdown();
95 96 97 98 99

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

100 101
/* *********************************** */

102
static void darshan_core_initialize(int *argc, char ***argv)
103 104 105 106 107 108 109 110 111
{
    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);
112
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
113 114 115 116

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

117
    if(internal_timing_flag)
118 119 120
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
121
    if(!getenv("DARSHAN_DISABLE") && !darshan_core_job)
122
    {
123 124 125
        /* allocate structure to track darshan_core_job information */
        darshan_core_job = malloc(sizeof(*darshan_core_job));
        if(darshan_core_job)
126
        {
127
            memset(darshan_core_job, 0, sizeof(*darshan_core_job));
128

129 130 131 132
            darshan_core_job->log_job.uid = getuid();
            darshan_core_job->log_job.start_time = time(NULL);
            darshan_core_job->log_job.nprocs = nprocs;
            darshan_core_job->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();
133 134 135 136

            /* record exe and arguments */
            for(i=0; i<(*argc); i++)
            {
137 138
                chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
                strncat(darshan_core_job->exe, *(argv[i]), chars_left);
139 140
                if(i < ((*argc)-1))
                {
141 142
                    chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
                    strncat(darshan_core_job->exe, " ", chars_left);
143 144 145 146 147 148 149 150
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
151
                chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
152
                strncat(darshan_core_job->exe, __progname, chars_left);
153 154
                chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
                strncat(darshan_core_job->exe, " <unknown args>", chars_left);
155 156 157 158 159 160
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
                truncate_offset = CP_EXE_LEN - strlen(truncate_string);
161
                sprintf(&darshan_core_job->exe[truncate_offset], "%s",
162 163 164 165 166 167 168 169 170 171
                    truncate_string);
            }
        }
    }

    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);
172
        if(my_rank == 0)
173 174 175 176 177 178 179 180 181
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

182
static void darshan_core_shutdown()
183
{
184
    int i;
185
    char *logfile_name;
186
    struct darshan_core_runtime *final_job;
187
    int internal_timing_flag = 0;
188 189
    char *envjobid;
    char *jobid_str;
190
    int jobid;
191
    struct tm *start_tm;
192
    time_t start_time_tmp;
193 194
    int ret = 0;
    int all_ret = 0;
195 196
    int64_t first_start_time;
    int64_t last_end_time;
197 198
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
199 200
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
    char *hints;
201 202
    double start_log_time;
    long offset;
203
    struct darshan_header log_header;
204
    MPI_File log_fh;
205
    MPI_Offset tmp_off;
206 207
    MPI_Info info;
    MPI_Status status;
208 209 210 211

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

212
    DARSHAN_CORE_LOCK();
213 214
    if(!darshan_core_job)
    {
215
        DARSHAN_CORE_UNLOCK();
216 217 218 219 220 221 222
        return;
    }
    /* disable further tracing while hanging onto the data so that we can
     * write it out
     */
    final_job = darshan_core_job;
    darshan_core_job = NULL;
223
    DARSHAN_CORE_UNLOCK();
224

225 226
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

227 228 229 230 231 232 233
    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
        darshan_core_cleanup(final_job);
        return;
    }

234
    /* set darshan job id/metadata and constuct log file name on rank 0 */
235
    if(my_rank == 0)
236 237 238
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
239
        if(!envjobid)
240 241 242 243
        {
            envjobid = CP_JOBID;
        }

244
        /* find a job id */
245 246 247 248 249 250 251 252 253 254 255 256
        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();
        }

257
        final_job->log_job.jobid = (int64_t)jobid;
258

259 260
/* TODO */
#if 0
261
        /* if we are using any hints to write the log file, then record those
262
         * hints with the darshan job information
263 264
         */
        darshan_log_record_hints_and_ver(final_job);
265
#endif
266

267 268 269
        /* use human readable start time format in log filename */
        start_time_tmp = final_job->log_job.start_time;
        start_tm = localtime(&start_time_tmp);
270

271 272
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
273 274 275 276 277 278 279 280 281
    }

    /* 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 */
282
        free(logfile_name);
283 284 285 286 287 288
        darshan_core_cleanup(final_job);
        return;
    }

    final_job->log_job.end_time = time(NULL);

289 290 291 292 293 294 295 296 297 298
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_job->log_job.start_time, &first_start_time, 1, MPI_LONG_LONG, MPI_MIN, 0, MPI_COMM_WORLD);
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_job->log_job.end_time, &last_end_time, 1, MPI_LONG_LONG, MPI_MAX, 0, MPI_COMM_WORLD);
    if(my_rank == 0)
    {
        final_job->log_job.start_time = first_start_time;
        final_job->log_job.end_time = last_end_time;
    }
299

300 301 302 303
    /* XXX */
    /* TODO: ensuing error checking...does MPI ensure collective I/O functions return the same error
     * globally, or do I always need to allreduce????? */

304 305 306 307 308 309 310 311 312 313
    /* set which local modules were actually used */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_job->mod_array[i])
            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);

314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330
    /* get a list of records which are shared across all processes */
    ret = darshan_get_shared_record_ids(final_job, shared_recs);

    /* error out if unable to determine shared file records */
    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 determine shared file records\n");
        }
        free(logfile_name);
        darshan_core_cleanup(final_job);
        return;
    }

331 332 333
    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
334 335
    MPI_Info_create(&info);

336 337 338 339 340 341 342 343
    hints = getenv(CP_LOG_HINTS_OVERRIDE);
    if(!hints)
    {
        hints = __CP_LOG_HINTS;
    }

    if(hints && strlen(hints) > 0)
    {
344 345 346 347 348 349
        char *tok_str;
        char *orig_tok_str;
        char *key;
        char *value;
        char *saveptr = NULL;

350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
        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);
    MPI_Info_free(&info);

    /* 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)
        {
            int msg_len;
            char msg[MPI_MAX_ERROR_STRING] = {0};

            MPI_Error_string(ret, msg, &msg_len);
            fprintf(stderr, "darshan library warning: unable to open log file %s: %s\n",
394
                logfile_name, msg);
395 396 397 398 399 400 401
            unlink(logfile_name);
        }
        free(logfile_name);
        darshan_core_cleanup(final_job);
        return;
    }

402
    /* rank 0 is responsible for writing the darshan job information */
Shane Snyder's avatar
Shane Snyder committed
403
    if(my_rank == 0)
404
    {
405 406 407
        /* write the job information, making sure to prealloc space for the log header */
        ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, sizeof(struct darshan_header),
                &final_job->log_job, sizeof(struct darshan_job), MPI_BYTE, &status);
Shane Snyder's avatar
Shane Snyder committed
408
        if(ret != MPI_SUCCESS)
409 410 411 412 413
        {
            int msg_len;
            char msg[MPI_MAX_ERROR_STRING] = {0};

            MPI_Error_string(ret, msg, &msg_len);
Shane Snyder's avatar
Shane Snyder committed
414 415
            fprintf(stderr, "darshan library warning: unable to write job data to log file %s: %s\n",
                    logfile_name, msg);
416
        }
Shane Snyder's avatar
Shane Snyder committed
417

418 419
        /* TODO */
        log_header.rec_map.off = sizeof(struct darshan_header) + sizeof(struct darshan_job);
420 421
    }

422
    /* write the record name->id map to the log file */
423 424
    ret = darshan_log_write_record_map(log_fh, final_job->rec_hash,
        shared_recs, &log_header.rec_map);
425 426 427 428 429 430 431 432 433 434 435 436 437 438

    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);
        }
        free(logfile_name);
        darshan_core_cleanup(final_job);
        return;
    }
439 440 441

    /* loop over globally used darshan modules and:
     *      - get final output buffer
442
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
443
     *      - append compressed buffer to log file
444
     *      - shutdown the module TODO
445
     */
446
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
447
    {
448 449 450 451 452 453
        struct darshan_core_module* this_mod = final_job->mod_array[i];
        MPI_Comm mod_comm;
        void* mod_buf = NULL;
        int mod_buf_size = 0;

        if(!global_mod_use_count[i])
454 455 456 457
        {
            if(my_rank == 0)
                log_header.mod_map[i].off = log_header.mod_map[i].len = 0;

458
            continue;
459
        }
460 461 462 463 464 465 466 467 468 469

        /* create a communicator to use for shutting down the module */
        if(global_mod_use_count[i] == nprocs)
        {
            MPI_Comm_dup(MPI_COMM_WORLD, &mod_comm);
        }
        else
        {
            MPI_Comm_split(MPI_COMM_WORLD, local_mod_use[i], 0, &mod_comm);
        }
470

471 472 473 474 475 476 477
        /* 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);
        }

478 479 480 481 482 483 484 485
        /* 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]);
Shane Snyder's avatar
Shane Snyder committed
486
        if(ret < 0)
487
        {
Shane Snyder's avatar
Shane Snyder committed
488
            /* TODO: */
489 490
        }

491 492
        tmp_off += log_header.mod_map[i].len;

493 494 495 496 497 498 499 500 501 502
        /* shutdown module if registered locally */
        if(local_mod_use[i])
        {
            this_mod->mod_funcs.shutdown();
            this_mod = NULL;
        }

        MPI_Comm_free(&mod_comm);
    }

503
    /* rank 0 is responsible for writing the log header */
504 505
    if(my_rank == 0)
    {
506 507 508 509 510 511 512 513 514 515 516
        /* 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;

        ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &log_header,
            sizeof(struct darshan_header), MPI_BYTE, &status);
        if(ret != MPI_SUCCESS)
        {
            /* TODO */
        }
517 518
    }

519 520 521 522 523 524
    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
     */
525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540
    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 */
541
#ifdef __CP_GROUP_READABLE_LOGS
542
            chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
543
#else
544
            chmod(new_logfile_name, (S_IRUSR));
545
#endif
546 547
            free(new_logfile_name);
        }
548
    }
549

550
    free(logfile_name);
551
    darshan_core_cleanup(final_job);
552

553
    if(internal_timing_flag)
554
    {
555 556 557 558 559
        /* TODO: what do we want to time in new darshan version? */
    }
    
    return;
}
560

561
/* free darshan core data structures to shutdown */
562
static void darshan_core_cleanup(struct darshan_core_runtime* job)
563
{
564
    int i;
565

566
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
567
    {
568 569 570 571 572
        if(job->mod_array[i])
        {        
            free(job->mod_array[i]);
            job->mod_array[i] = NULL;
        }
573
    }
574

575
    free(job);
576 577 578 579

    return;
}

580
/* construct the darshan log file name */
581
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
582
{
583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
    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;
713 714
}

715
/* record any hints used to write the darshan log in the log header */
716
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* job)
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760
{
    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 -
        strlen(job->log_job.metadata) - 1;
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
        sprintf(job->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
        m = job->log_job.metadata + strlen(job->log_job.metadata);
        /* 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;
}

761 762
static int darshan_get_shared_record_ids(struct darshan_core_runtime *job,
    darshan_record_id *shared_recs)
763 764 765 766 767 768 769 770 771 772 773 774 775
{
    int i;
    int ndx;
    int ret;
    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;
776
        HASH_ITER(hlink, job->rec_hash, ref, tmp)
777
        {
778
            id_array[ndx++] = ref->rec.id;           
779 780 781 782 783 784 785 786 787
        }
    }

    /* broadcast root's list of records to all other processes */
    ret = DARSHAN_MPI_CALL(PMPI_Bcast)(id_array,
        (DARSHAN_CORE_MAX_RECORDS * sizeof(darshan_record_id)),
        MPI_BYTE, 0, MPI_COMM_WORLD);
    if(ret != 0)
    {
Shane Snyder's avatar
Shane Snyder committed
788
        return(-1);
789 790 791 792 793
    }

    /* 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++)
    {
794
        HASH_ITER(hlink, job->rec_hash, ref, tmp)
795
        {
796
            if(id_array[i] == ref->rec.id)
797 798 799 800 801 802 803 804 805 806 807 808 809
            {
                /* we opened that record too */
                mask_array[i] = 1;
                break;
            }
        }
    }

    /* now allreduce so everyone agrees which files are shared */
    ret = DARSHAN_MPI_CALL(PMPI_Allreduce)(mask_array, all_mask_array,
        DARSHAN_CORE_MAX_RECORDS, MPI_INT, MPI_LAND, MPI_COMM_WORLD);
    if(ret != 0)
    {
Shane Snyder's avatar
Shane Snyder committed
810
        return(-1);
811 812 813 814 815 816 817 818 819 820 821
    }

    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];
        }
    }

Shane Snyder's avatar
Shane Snyder committed
822 823 824
    return(0);
}

825 826 827
/* NOTE: the map written to file may contain duplicate id->name entries if a
 *       record is opened by multiple ranks, but not all ranks
 */
828 829
static int darshan_log_write_record_map(MPI_File log_fh, struct darshan_core_record_ref *rec_hash,
    darshan_record_id *shared_recs, struct darshan_log_map *map)
830
{
831 832 833 834 835
    int i;
    int ret;
    struct darshan_core_record_ref *ref, *tmp;
    uint32_t name_len;
    size_t record_sz;
836 837 838
    size_t hash_buf_sz = 0;
    unsigned char *hash_buf;
    unsigned char *hash_buf_off;
839 840
    MPI_Status status;

841
    /* non-root ranks (rank > 0) remove shared records from their map --
842 843 844 845 846 847
     * these records will be written by rank 0
     */
    if(my_rank > 0)
    {
        for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && shared_recs[i]); i++)
        {
848
            HASH_FIND(hlink, rec_hash, &shared_recs[i], sizeof(darshan_record_id), ref);
849
            assert(ref); /* this id had better be in the hash ... */
850
            HASH_DELETE(hlink, rec_hash, ref);
851 852 853 854 855 856 857
            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 */
858 859 860
    hash_buf_sz = DARSHAN_CORE_MAX_RECORDS * 64;
    hash_buf = malloc(hash_buf_sz);
    if(!hash_buf)
861
    {
Shane Snyder's avatar
Shane Snyder committed
862
        return(-1);
863 864
    }

Shane Snyder's avatar
Shane Snyder committed
865
    /* serialize the record map into a buffer for writing */
866 867
    hash_buf_off = hash_buf;
    HASH_ITER(hlink, rec_hash, ref, tmp)
868 869 870 871
    {
        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 */
872
        if((hash_buf_off + record_sz) > (hash_buf + hash_buf_sz))
873 874 875 876 877
        {
            unsigned char *tmp_buf;
            size_t old_buf_sz;

            /* if no room, reallocate the map buffer at twice the current size */
878 879 880
            old_buf_sz = hash_buf_off - hash_buf;
            hash_buf_sz *= 2;
            tmp_buf = malloc(hash_buf_sz);
881 882
            if(!tmp_buf)
            {
883
                free(hash_buf);
Shane Snyder's avatar
Shane Snyder committed
884
                return(-1);
885 886
            }

887 888 889 890
            memcpy(tmp_buf, hash_buf, old_buf_sz);
            free(hash_buf);
            hash_buf = tmp_buf;
            hash_buf_off = hash_buf + old_buf_sz;
891 892 893 894 895 896
        }

        /* now serialize the record into the map buffer.
         * NOTE: darshan record map serialization method: 
         *          ... darshan_record_id | (uint32_t) path_len | path ...
         */
897 898 899 900 901 902
        *((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;
903 904 905
    }

    /* collectively write out the record map to the darshan log */
906
    if(hash_buf_off > hash_buf)
907 908
    {
        /* we have records to contribute to the collective write of the record map */
909
        ret = darshan_log_coll_write(log_fh, hash_buf, (hash_buf_off-hash_buf), map);
910 911 912 913
    }
    else
    {
        /* we have no data to write, but participate in the collective anyway */
914
        ret = darshan_log_coll_write(log_fh, NULL, 0, map);
915 916
    }

917
    free(hash_buf);
918

Shane Snyder's avatar
Shane Snyder committed
919 920 921 922 923 924
    if(ret < 0)
        return(-1);

    return(0);
}

925 926 927 928 929 930
/* 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).
Shane Snyder's avatar
Shane Snyder committed
931
 */
932 933
static int darshan_log_coll_write(MPI_File log_fh, void *buf, int count,
    struct darshan_log_map *map)
Shane Snyder's avatar
Shane Snyder committed
934 935 936 937 938
{
    MPI_Offset send_off, my_off;
    MPI_Status status;
    int ret;

939
    /* figure out where everyone is writing using scan */
Shane Snyder's avatar
Shane Snyder committed
940 941
    send_off = count;
    if(my_rank == 0)
942 943 944
    {
        send_off += map->off; /* rank 0 knows the beginning offset */
    }
Shane Snyder's avatar
Shane Snyder committed
945 946 947 948 949 950 951 952 953 954 955 956

    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);

957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977
    /* 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;
    }
Shane Snyder's avatar
Shane Snyder committed
978 979

    return(0);
980 981
}

982
/* ********************************************************* */
983 984

void darshan_core_register_module(
985
    darshan_module_id id,
986 987 988 989
    char *name,
    struct darshan_module_funcs *funcs,
    int *runtime_mem_limit)
{
990
    struct darshan_core_module* mod;
991

992
    DARSHAN_CORE_LOCK();
993

994
    *runtime_mem_limit = 0;
995
    if(!darshan_core_job || (id >= DARSHAN_MAX_MODS))
996
    {
997
        DARSHAN_CORE_UNLOCK();
998
        return;
999
    }
1000 1001

    /* see if this module is already registered */
1002
    if(darshan_core_job->mod_array[id])
1003
    {
1004
        /* if module is already registered just return */
1005
        /* NOTE: we do not recalculate memory limit here, just set to 0 */
1006
        DARSHAN_CORE_UNLOCK();
1007
        return;
1008 1009
    }

1010
    /* this module has not been registered yet, allocate and initialize it */
1011 1012
    mod = malloc(sizeof(*mod));
    if(!mod)
1013
    {
1014
        DARSHAN_CORE_UNLOCK();
1015 1016
        return;
    }
1017
    memset(mod, 0, sizeof(*mod));
1018

1019
    mod->id = id;
1020 1021
    strncpy(mod->name, name, DARSHAN_MOD_NAME_LEN);
    mod->mod_funcs = *funcs;
1022 1023 1024

    /* register module with darshan */
    darshan_core_job->mod_array[id] = mod;
1025

1026 1027
    /* TODO: something smarter than just 2 MiB per module */
    *runtime_mem_limit = 2 * 1024 * 1024;
1028

1029
    DARSHAN_CORE_UNLOCK();
1030

1031 1032 1033
    return;
}

1034
void darshan_core_lookup_record_id(
1035 1036 1037
    void *name,
    int len,
    int printable_flag,
1038
    darshan_record_id *id)
1039
{
1040 1041
    darshan_record_id tmp_id;
    struct darshan_core_record_ref* ref;
1042

1043
    if(!darshan_core_job || !name)
1044 1045 1046 1047 1048 1049
        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);
1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060
 
    DARSHAN_CORE_LOCK();

    /* check to see if we've already stored the id->name mapping for this record */
    HASH_FIND(hlink, darshan_core_job->rec_hash, &tmp_id, sizeof(darshan_record_id), ref);
    if(!ref)
    {
        /* if not, add this record to the hash */
        ref = malloc(sizeof(struct darshan_core_record_ref));
        if(ref)
        {
1061 1062 1063 1064
            ref->rec.id = tmp_id;
            ref->rec.name = malloc(strlen(name) + 1);
            if(ref->rec.name)
                strcpy(ref->rec.name, name);
1065

1066
            HASH_ADD(hlink, darshan_core_job->rec_hash, rec.id, sizeof(darshan_record_id), ref);
1067 1068 1069 1070
        }
    }   

    DARSHAN_CORE_UNLOCK();
1071

Shane Snyder's avatar
Shane Snyder committed
1072 1073
    printf("New Darshan record: %s (%"PRIu64")\n", (char *)name, tmp_id);

1074
    *id = tmp_id;
1075 1076 1077
    return;
}

1078 1079
double darshan_core_wtime()
{
1080
    if(!darshan_core_job)
1081 1082 1083 1084
    {
        return(0);
    }

1085
    return(DARSHAN_MPI_CALL(PMPI_Wtime)() - darshan_core_job->wtime_offset);
1086
}
1087 1088 1089 1090 1091 1092 1093 1094 1095

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