darshan-core.c 32 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
/* 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);
48
static void darshan_get_shared_record_ids(
49
    struct darshan_core_runtime *job, darshan_record_id *shared_recs);
50 51 52
static int darshan_log_coll_open(
    char *logfile_name, MPI_File *log_fh);
static int darshan_log_write_record_hash(
53 54 55 56
    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);
57

58 59
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)
60

61
/* intercept MPI initialize and finalize to manage darshan core runtime */
62 63 64 65 66 67 68 69 70 71
int MPI_Init(int *argc, char ***argv)
{
    int ret;

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

72
    darshan_core_initialize(argc, argv);
73 74 75 76

    return(ret);
}

77
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
78 79 80 81
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
82
    if(ret != MPI_SUCCESS)
83 84 85 86
    {
        return(ret);
    }

87
    darshan_core_initialize(argc, argv);
88 89 90 91 92 93 94 95

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

96
    darshan_core_shutdown();
97 98 99 100 101

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

102 103
/* *********************************** */

104
static void darshan_core_initialize(int *argc, char ***argv)
105 106 107 108 109 110 111 112 113
{
    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);
114
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
115 116 117 118

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

119
    if(internal_timing_flag)
120 121 122
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
123
    if(!getenv("DARSHAN_DISABLE") && !darshan_core_job)
124
    {
125 126 127
        /* allocate structure to track darshan_core_job information */
        darshan_core_job = malloc(sizeof(*darshan_core_job));
        if(darshan_core_job)
128
        {
129
            memset(darshan_core_job, 0, sizeof(*darshan_core_job));
130

131 132 133 134
            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)();
135 136 137 138

            /* record exe and arguments */
            for(i=0; i<(*argc); i++)
            {
139 140
                chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
                strncat(darshan_core_job->exe, *(argv[i]), chars_left);
141 142
                if(i < ((*argc)-1))
                {
143 144
                    chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
                    strncat(darshan_core_job->exe, " ", chars_left);
145 146 147 148 149 150 151 152
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
153
                chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
154
                strncat(darshan_core_job->exe, __progname, chars_left);
155 156
                chars_left = CP_EXE_LEN-strlen(darshan_core_job->exe);
                strncat(darshan_core_job->exe, " <unknown args>", chars_left);
157 158 159 160 161 162
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
                truncate_offset = CP_EXE_LEN - strlen(truncate_string);
163
                sprintf(&darshan_core_job->exe[truncate_offset], "%s",
164 165 166 167 168 169 170 171 172 173
                    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);
174
        if(my_rank == 0)
175 176 177 178 179 180 181 182 183
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

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

300 301 302 303 304 305 306 307 308 309
    /* 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);

310
    /* get a list of records which are shared across all processes */
311
    darshan_get_shared_record_ids(final_job, shared_recs);
312

313 314
    /* collectively open the darshan log file */
    ret = darshan_log_coll_open(logfile_name, &log_fh);
315 316 317 318 319 320 321 322

    /* 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)
        {
323 324
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
325 326 327 328 329 330 331
            unlink(logfile_name);
        }
        free(logfile_name);
        darshan_core_cleanup(final_job);
        return;
    }

332
    /* rank 0 is responsible for writing the darshan job information */
Shane Snyder's avatar
Shane Snyder committed
333
    if(my_rank == 0)
334
    {
335
        /* write the job information, making sure to prealloc space for the log header */
336
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, sizeof(struct darshan_header),
337
                &final_job->log_job, sizeof(struct darshan_job), MPI_BYTE, &status);
338
        if(all_ret != MPI_SUCCESS)
339
        {
340 341 342
            fprintf(stderr, "darshan library warning: unable to write job data to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
343
        }
Shane Snyder's avatar
Shane Snyder committed
344

345
        /* TODO: after compression is added, this should be fixed */
346
        log_header.rec_map.off = sizeof(struct darshan_header) + sizeof(struct darshan_job);
347 348
    }

349 350 351 352 353 354 355 356 357 358 359
    /* 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);
        darshan_core_cleanup(final_job);
        return;
    }

    /* write the record name->id hash to the log file */
    ret = darshan_log_write_record_hash(log_fh, final_job->rec_hash,
360
        shared_recs, &log_header.rec_map);
361

362
    /* error out if unable to write record hash */
363 364 365 366 367 368 369 370
    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);
371
            unlink(logfile_name);
372 373 374 375 376
        }
        free(logfile_name);
        darshan_core_cleanup(final_job);
        return;
    }
377 378 379

    /* loop over globally used darshan modules and:
     *      - get final output buffer
380
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
381
     *      - append compressed buffer to log file
382 383
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
384
     */
385
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
386
    {
387 388 389 390 391 392
        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])
393 394 395 396
        {
            if(my_rank == 0)
                log_header.mod_map[i].off = log_header.mod_map[i].len = 0;

397
            continue;
398
        }
399 400 401 402 403 404 405 406 407 408

        /* 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);
        }
409

410 411 412 413 414 415 416
        /* 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);
        }

417 418 419 420 421 422 423 424
        /* 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]);
425 426 427 428 429

        /* 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)
430
        {
431 432 433 434 435 436 437 438 439 440
            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);
            darshan_core_cleanup(final_job);
            return;
441 442
        }

443 444
        tmp_off += log_header.mod_map[i].len;

445 446 447 448 449 450 451 452 453 454
        /* shutdown module if registered locally */
        if(local_mod_use[i])
        {
            this_mod->mod_funcs.shutdown();
            this_mod = NULL;
        }

        MPI_Comm_free(&mod_comm);
    }

455
    /* rank 0 is responsible for writing the log header */
456 457
    if(my_rank == 0)
    {
458 459 460 461 462
        /* 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;

463
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &log_header,
464
            sizeof(struct darshan_header), MPI_BYTE, &status);
465
        if(all_ret != MPI_SUCCESS)
466
        {
467 468 469
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
470
        }
471 472
    }

473 474 475 476 477 478 479 480 481
    /* 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);
        darshan_core_cleanup(final_job);
        return;
    }

482 483 484 485 486 487
    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
     */
488
    /* TODO: support user given logfile path/name */
489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504
    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 */
505
#ifdef __CP_GROUP_READABLE_LOGS
506
            chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
507
#else
508
            chmod(new_logfile_name, (S_IRUSR));
509
#endif
510 511
            free(new_logfile_name);
        }
512
    }
513

514
    free(logfile_name);
515
    darshan_core_cleanup(final_job);
516

517
    if(internal_timing_flag)
518
    {
519 520 521 522 523
        /* TODO: what do we want to time in new darshan version? */
    }
    
    return;
}
524

525
/* free darshan core data structures to shutdown */
526
static void darshan_core_cleanup(struct darshan_core_runtime* job)
527
{
528
    int i;
529

530
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
531
    {
532 533 534 535 536
        if(job->mod_array[i])
        {        
            free(job->mod_array[i]);
            job->mod_array[i] = NULL;
        }
537
    }
538

539
    free(job);
540 541 542 543

    return;
}

544
/* construct the darshan log file name */
545
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
546
{
547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676
    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;
677 678
}

679
/* record any hints used to write the darshan log in the log header */
680
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* job)
681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
{
    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;
}

725
static void darshan_get_shared_record_ids(struct darshan_core_runtime *job,
726
    darshan_record_id *shared_recs)
727 728 729 730 731 732 733 734 735 736 737 738
{
    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;
739
        HASH_ITER(hlink, job->rec_hash, ref, tmp)
740
        {
741
            id_array[ndx++] = ref->rec.id;           
742 743 744 745
        }
    }

    /* broadcast root's list of records to all other processes */
746
    DARSHAN_MPI_CALL(PMPI_Bcast)(id_array,
747 748 749 750 751 752
        (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++)
    {
753
        HASH_ITER(hlink, job->rec_hash, ref, tmp)
754
        {
755
            if(id_array[i] == ref->rec.id)
756 757 758 759 760 761 762 763 764
            {
                /* we opened that record too */
                mask_array[i] = 1;
                break;
            }
        }
    }

    /* now allreduce so everyone agrees which files are shared */
765
    DARSHAN_MPI_CALL(PMPI_Allreduce)(mask_array, all_mask_array,
766 767 768 769 770 771 772 773 774 775 776
        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];
        }
    }

777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837
    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);
Shane Snyder's avatar
Shane Snyder committed
838 839 840
    return(0);
}

841 842 843
/* NOTE: the map written to file may contain duplicate id->name entries if a
 *       record is opened by multiple ranks, but not all ranks
 */
844
static int darshan_log_write_record_hash(MPI_File log_fh, struct darshan_core_record_ref *rec_hash,
845
    darshan_record_id *shared_recs, struct darshan_log_map *map)
846
{
847 848 849 850 851
    int i;
    int ret;
    struct darshan_core_record_ref *ref, *tmp;
    uint32_t name_len;
    size_t record_sz;
852 853 854
    size_t hash_buf_sz = 0;
    unsigned char *hash_buf;
    unsigned char *hash_buf_off;
855 856
    MPI_Status status;

857
    /* non-root ranks (rank > 0) remove shared records from their map --
858 859 860 861 862 863
     * these records will be written by rank 0
     */
    if(my_rank > 0)
    {
        for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && shared_recs[i]); i++)
        {
864
            HASH_FIND(hlink, rec_hash, &shared_recs[i], sizeof(darshan_record_id), ref);
865
            assert(ref); /* this id had better be in the hash ... */
866
            HASH_DELETE(hlink, rec_hash, ref);
867 868 869 870 871 872 873
            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 */
874 875 876
    hash_buf_sz = DARSHAN_CORE_MAX_RECORDS * 64;
    hash_buf = malloc(hash_buf_sz);
    if(!hash_buf)
877
    {
Shane Snyder's avatar
Shane Snyder committed
878
        return(-1);
879 880
    }

Shane Snyder's avatar
Shane Snyder committed
881
    /* serialize the record map into a buffer for writing */
882 883
    hash_buf_off = hash_buf;
    HASH_ITER(hlink, rec_hash, ref, tmp)
884 885 886 887
    {
        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 */
888
        if((hash_buf_off + record_sz) > (hash_buf + hash_buf_sz))
889 890 891 892 893
        {
            unsigned char *tmp_buf;
            size_t old_buf_sz;

            /* if no room, reallocate the map buffer at twice the current size */
894 895 896
            old_buf_sz = hash_buf_off - hash_buf;
            hash_buf_sz *= 2;
            tmp_buf = malloc(hash_buf_sz);
897 898
            if(!tmp_buf)
            {
899
                free(hash_buf);
Shane Snyder's avatar
Shane Snyder committed
900
                return(-1);
901 902
            }

903 904 905 906
            memcpy(tmp_buf, hash_buf, old_buf_sz);
            free(hash_buf);
            hash_buf = tmp_buf;
            hash_buf_off = hash_buf + old_buf_sz;
907 908 909 910 911 912
        }

        /* now serialize the record into the map buffer.
         * NOTE: darshan record map serialization method: 
         *          ... darshan_record_id | (uint32_t) path_len | path ...
         */
913 914 915 916 917 918
        *((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;
919 920 921
    }

    /* collectively write out the record map to the darshan log */
922
    if(hash_buf_off > hash_buf)
923 924
    {
        /* we have records to contribute to the collective write of the record map */
925
        ret = darshan_log_coll_write(log_fh, hash_buf, (hash_buf_off-hash_buf), map);
926 927 928 929
    }
    else
    {
        /* we have no data to write, but participate in the collective anyway */
930
        ret = darshan_log_coll_write(log_fh, NULL, 0, map);
931 932
    }

933
    free(hash_buf);
934

Shane Snyder's avatar
Shane Snyder committed
935 936 937 938 939 940
    if(ret < 0)
        return(-1);

    return(0);
}

941 942 943 944 945 946
/* 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
947
 */
948 949
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
950 951 952 953 954
{
    MPI_Offset send_off, my_off;
    MPI_Status status;
    int ret;

955
    /* figure out where everyone is writing using scan */
Shane Snyder's avatar
Shane Snyder committed
956 957
    send_off = count;
    if(my_rank == 0)
958 959 960
    {
        send_off += map->off; /* rank 0 knows the beginning offset */
    }
Shane Snyder's avatar
Shane Snyder committed
961 962 963 964 965 966 967 968 969 970 971 972

    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);

973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993
    /* 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
994 995

    return(0);
996 997
}

998
/* ********************************************************* */
999 1000

void darshan_core_register_module(
1001
    darshan_module_id id,
1002 1003 1004
    struct darshan_module_funcs *funcs,
    int *runtime_mem_limit)
{
1005
    struct darshan_core_module* mod;
1006

1007
    DARSHAN_CORE_LOCK();
1008

1009
    *runtime_mem_limit = 0;
1010
    if(!darshan_core_job || (id >= DARSHAN_MAX_MODS))
1011
    {
1012
        DARSHAN_CORE_UNLOCK();
1013
        return;
1014
    }
1015 1016

    /* see if this module is already registered */
1017
    if(darshan_core_job->mod_array[id])
1018
    {
1019
        /* if module is already registered just return */
1020
        /* NOTE: we do not recalculate memory limit here, just set to 0 */
1021
        DARSHAN_CORE_UNLOCK();
1022
        return;
1023 1024
    }

1025
    /* this module has not been registered yet, allocate and initialize it */
1026 1027
    mod = malloc(sizeof(*mod));
    if(!mod)
1028
    {
1029
        DARSHAN_CORE_UNLOCK();
1030 1031
        return;
    }
1032
    memset(mod, 0, sizeof(*mod));
1033

1034
    mod->id = id;
1035
    mod->mod_funcs = *funcs;
1036 1037 1038

    /* register module with darshan */
    darshan_core_job->mod_array[id] = mod;
1039

1040 1041
    /* TODO: something smarter than just 2 MiB per module */
    *runtime_mem_limit = 2 * 1024 * 1024;
1042

1043
    DARSHAN_CORE_UNLOCK();
1044

1045 1046 1047
    return;
}

1048
void darshan_core_lookup_record_id(
1049 1050 1051
    void *name,
    int len,
    int printable_flag,
1052
    darshan_record_id *id)
1053
{
1054 1055
    darshan_record_id tmp_id;
    struct darshan_core_record_ref* ref;
1056

1057
    if(!darshan_core_job || !name)
1058 1059 1060 1061 1062 1063
        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);
1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074
 
    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)
        {
1075 1076 1077 1078
            ref->rec.id = tmp_id;
            ref->rec.name = malloc(strlen(name) + 1);
            if(ref->rec.name)
                strcpy(ref->rec.name, name);
1079

1080
            HASH_ADD(hlink, darshan_core_job->rec_hash, rec.id, sizeof(darshan_record_id), ref);
1081 1082 1083 1084
        }
    }   

    DARSHAN_CORE_UNLOCK();
1085

Shane Snyder's avatar
Shane Snyder committed
1086 1087
    printf("New Darshan record: %s (%"PRIu64")\n", (char *)name, tmp_id);

1088
    *id = tmp_id;
1089 1090 1091
    return;
}

1092 1093
double