darshan-core.c 31.1 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 51 52 53 54 55 56
static int darshan_log_write_header(
    MPI_File log_fh, int mod_count, int64_t rec_off, int64_t psx_off);
static int darshan_log_write_record_map(
    MPI_File log_fh, struct darshan_core_runtime *job,
    darshan_record_id *shared_recs, MPI_Offset *off);
static int darshan_log_coll_append(MPI_File log_fh, MPI_Offset *off,
    void *buf, int count);
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
    struct darshan_core_module *mod, *tmp;
190
    int internal_timing_flag = 0;
191 192
    char *envjobid;
    char *jobid_str;
193
    int jobid;
194
    struct tm *start_tm;
195
    time_t start_time_tmp;
196 197
    int ret = 0;
    int all_ret = 0;
198 199
    int64_t first_start_time;
    int64_t last_end_time;
200 201
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
202 203 204 205 206 207 208 209 210
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
    char *key;
    char *value;
    char *hints;
    char *tok_str;
    char *orig_tok_str;
    char *saveptr = NULL;
    char *mod_index;
    char *new_logfile_name;
211 212 213 214
    double start_log_time;
    double end_log_time;
    long offset;
    MPI_File log_fh;
Shane Snyder's avatar
Shane Snyder committed
215
    MPI_Offset my_off = 0;
216 217
    MPI_Info info;
    MPI_Status status;
218 219 220 221

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

222
    DARSHAN_CORE_LOCK();
223 224
    if(!darshan_core_job)
    {
225
        DARSHAN_CORE_UNLOCK();
226 227 228 229 230 231 232
        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;
233
    DARSHAN_CORE_UNLOCK();
234

235 236
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

237 238 239 240 241 242 243
    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
        darshan_core_cleanup(final_job);
        return;
    }

244
    /* set darshan job id/metadata and constuct log file name on rank 0 */
245
    if(my_rank == 0)
246 247 248
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
249
        if(!envjobid)
250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265
        {
            envjobid = CP_JOBID;
        }

        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();
        }

266
        final_job->log_job.jobid = (int64_t)jobid;
267

268 269 270 271 272
        /* if we are using any hints to write the log file, then record those
         * hints in the log file header
         */
        darshan_log_record_hints_and_ver(final_job);

273 274 275
        /* use human readable start time format in log filename */
        start_time_tmp = final_job->log_job.start_time;
        start_tm = localtime(&start_time_tmp);
276

277 278
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
279 280 281 282 283 284 285 286 287
    }

    /* 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 */
288
        free(logfile_name);
289 290 291 292 293 294
        darshan_core_cleanup(final_job);
        return;
    }

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

295 296 297 298 299 300 301 302 303 304
    /* 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;
    }
305

306 307 308 309 310 311 312 313 314 315
    /* 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);

316 317 318
    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
319 320
    MPI_Info_create(&info);

321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355
    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);
        }
    }

Shane Snyder's avatar
Shane Snyder committed
356 357 358
    /* TODO: ensuing error checking...does MPI ensure collective I/O functions return the same error
     * globally, or do I always need to allreduce????? */

359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376
    /* 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;

    }

377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393
    /* 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;
    }

Shane Snyder's avatar
Shane Snyder committed
402 403

    if(my_rank == 0)
404
    {
Shane Snyder's avatar
Shane Snyder committed
405 406 407 408 409 410
        my_off = sizeof(struct darshan_header);
        my_off += 2 * sizeof(int64_t); /* FIXME account for changes to index map */

        ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, my_off, &final_job->log_job,
                sizeof(struct darshan_job), MPI_BYTE, &status);
        if(ret != MPI_SUCCESS)
411 412 413 414 415
        {
            int msg_len;
            char msg[MPI_MAX_ERROR_STRING] = {0};

            MPI_Error_string(ret, msg, &msg_len);
Shane Snyder's avatar
Shane Snyder committed
416 417
            fprintf(stderr, "darshan library warning: unable to write job data to log file %s: %s\n",
                    logfile_name, msg);
418
        }
Shane Snyder's avatar
Shane Snyder committed
419 420

        my_off += sizeof(struct darshan_job);
421 422
    }

Shane Snyder's avatar
Shane Snyder committed
423
    int64_t rec_off = my_off; /* TODO: get rid of this hack */
424
    /* write the record name->id map to the log file */
Shane Snyder's avatar
Shane Snyder committed
425
    ret = darshan_log_write_record_map(log_fh, final_job, shared_recs, &my_off);
426 427 428 429 430 431 432 433 434 435 436 437 438 439

    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;
    }
Shane Snyder's avatar
Shane Snyder committed
440
    int64_t psx_off = my_off; /* TODO: get rid of this hack */
441 442 443

    /* loop over globally used darshan modules and:
     *      - get final output buffer
444
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
445
     *      - append compressed buffer to log file
446 447
     *      - shutdown the module
     */
448
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
449
    {
450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
        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])
            continue;

        /* 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);
        }
467

468 469 470 471 472 473 474
        /* 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);
        }

Shane Snyder's avatar
Shane Snyder committed
475 476 477
        /* append module data buffer to the darshan log file */
        ret = darshan_log_coll_append(log_fh, &my_off, mod_buf, mod_buf_size);
        if(ret < 0)
478
        {
Shane Snyder's avatar
Shane Snyder committed
479
            /* TODO: */
480 481 482 483 484 485 486 487 488 489 490 491
        }

        /* shutdown module if registered locally */
        if(local_mod_use[i])
        {
            this_mod->mod_funcs.shutdown();
            this_mod = NULL;
        }

        MPI_Comm_free(&mod_comm);
    }

Shane Snyder's avatar
Shane Snyder committed
492
    /* rank 0 is responsible for writing the log header and index map */
493 494
    if(my_rank == 0)
    {
Shane Snyder's avatar
Shane Snyder committed
495
        ret = darshan_log_write_header(log_fh, 1, rec_off, psx_off);
496 497
    }

498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519
    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
     */
    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);
        mod_index = strstr(new_logfile_name, ".darshan_partial");
        sprintf(mod_index, "_%d.darshan.gz", (int)(end_log_time-start_log_time+1));
        rename(logfile_name, new_logfile_name);
        /* set permissions on log file */
#ifdef __CP_GROUP_READABLE_LOGS
        chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
#else
        chmod(new_logfile_name, (S_IRUSR));
#endif
        free(new_logfile_name);
520
    }
521 522

    darshan_core_cleanup(final_job);
523
    free(logfile_name);
524

525
    if(internal_timing_flag)
526
    {
527 528 529 530 531
        /* TODO: what do we want to time in new darshan version? */
    }
    
    return;
}
532

533
/* free darshan core data structures to shutdown */
534
static void darshan_core_cleanup(struct darshan_core_runtime* job)
535
{
536
    int i;
537

538
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
539
    {
540 541 542 543 544
        if(job->mod_array[i])
        {        
            free(job->mod_array[i]);
            job->mod_array[i] = NULL;
        }
545
    }
546

547
    free(job);
548 549 550 551

    return;
}

552
/* construct the darshan log file name */
553
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
554
{
555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685
    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;
686 687
}

688
/* record any hints used to write the darshan log in the log header */
689
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* job)
690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733
{
    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;
}

734 735
static int darshan_get_shared_record_ids(struct darshan_core_runtime *job,
    darshan_record_id *shared_recs)
736 737 738 739 740 741 742 743 744 745 746 747 748
{
    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;
749
        HASH_ITER(hlink, job->rec_hash, ref, tmp)
750
        {
751
            id_array[ndx++] = ref->rec.id;           
752 753 754 755 756 757 758 759 760
        }
    }

    /* 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
761
        return(-1);
762 763 764 765 766
    }

    /* 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++)
    {
767
        HASH_ITER(hlink, job->rec_hash, ref, tmp)
768
        {
769
            if(id_array[i] == ref->rec.id)
770 771 772 773 774 775 776 777 778 779 780 781 782
            {
                /* 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
783
        return(-1);
784 785 786 787 788 789 790 791 792 793 794
    }

    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
795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839
    return(0);
}

static int darshan_log_write_header(MPI_File log_fh, int mod_count,
    int64_t rec_off, int64_t psx_off)
{
    struct darshan_header base_hdr;
    unsigned char *hdr_buf;
    unsigned char *tmp_p;
    int hdr_size;
    MPI_Status status;
    int i;
    int ret;

    /* set the fields of the darshan header */
    strcpy(base_hdr.version_string, CP_VERSION);
    base_hdr.magic_nr = CP_MAGIC_NR;
    base_hdr.comp_type = DARSHAN_GZ_COMP;
    base_hdr.mod_count = mod_count;

    hdr_size = sizeof(struct darshan_header) + (2 * sizeof(int64_t)); /* TODO: */
    hdr_buf = malloc(hdr_size);
    if(!hdr_buf)
    {
        return(-1);
    }

    /* pack the header in buffer for writing */
    tmp_p = hdr_buf;
    *((struct darshan_header *)tmp_p) = base_hdr;
    tmp_p += sizeof(struct darshan_header);

    /* TODO: we need to have a way of identifying different modules in index map */
    *((int64_t *)tmp_p) = rec_off;
    tmp_p += sizeof(int64_t);
    *((int64_t *)tmp_p) = psx_off;

    ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, hdr_buf, hdr_size,
        MPI_BYTE, &status);
    if(ret < 0)
    {
        return(-1);
    }

    return(0);
840 841
}

842 843 844
/* NOTE: the map written to file may contain duplicate id->name entries if a
 *       record is opened by multiple ranks, but not all ranks
 */
Shane Snyder's avatar
Shane Snyder committed
845 846
static int darshan_log_write_record_map(MPI_File log_fh, struct darshan_core_runtime *job,
    darshan_record_id *shared_recs, MPI_Offset *off)
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
    int i;
    int ret;
    struct darshan_core_record_ref *ref, *tmp;
    uint32_t name_len;
    size_t record_sz;
    size_t map_buf_sz = 0;
    unsigned char *map_buf;
    unsigned char *map_buf_off;
    MPI_Status status;

    /* non-root ranks (rank 0) remove shared records from their map --
     * these records will be written by rank 0
     */
    if(my_rank > 0)
    {
        for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && shared_recs[i]); i++)
        {
            HASH_FIND(hlink, job->rec_hash, &shared_recs[i], sizeof(darshan_record_id), ref);
            assert(ref); /* this id had better be in the hash ... */
            HASH_DELETE(hlink, job->rec_hash, ref);
            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 */
    map_buf_sz = DARSHAN_CORE_MAX_RECORDS * 64;
    map_buf = malloc(map_buf_sz);
    if(!map_buf)
    {
Shane Snyder's avatar
Shane Snyder committed
879
        return(-1);
880 881
    }

Shane Snyder's avatar
Shane Snyder committed
882
    /* serialize the record map into a buffer for writing */
883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900
    map_buf_off = map_buf;
    HASH_ITER(hlink, job->rec_hash, ref, tmp)
    {
        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 */
        if((map_buf_off + record_sz) > (map_buf + map_buf_sz))
        {
            unsigned char *tmp_buf;
            size_t old_buf_sz;

            /* if no room, reallocate the map buffer at twice the current size */
            old_buf_sz = map_buf_off - map_buf;
            map_buf_sz *= 2;
            tmp_buf = malloc(map_buf_sz);
            if(!tmp_buf)
            {
                free(map_buf);
Shane Snyder's avatar
Shane Snyder committed
901
                return(-1);
902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925
            }

            memcpy(tmp_buf, map_buf, old_buf_sz);
            free(map_buf);
            map_buf = tmp_buf;
            map_buf_off = map_buf + old_buf_sz;
        }

        /* now serialize the record into the map buffer.
         * NOTE: darshan record map serialization method: 
         *          ... darshan_record_id | (uint32_t) path_len | path ...
         */
        *((darshan_record_id *)map_buf_off) = ref->rec.id;
        map_buf_off += sizeof(darshan_record_id);
        *((uint32_t *)map_buf_off) = name_len;
        map_buf_off += sizeof(uint32_t);
        memcpy(map_buf_off, ref->rec.name, name_len);
        map_buf_off += name_len;
    }

    /* collectively write out the record map to the darshan log */
    if(map_buf_off > map_buf)
    {
        /* we have records to contribute to the collective write of the record map */
Shane Snyder's avatar
Shane Snyder committed
926
        ret = darshan_log_coll_append(log_fh, off, map_buf, (map_buf_off-map_buf));
927 928 929 930
    }
    else
    {
        /* we have no data to write, but participate in the collective anyway */
Shane Snyder's avatar
Shane Snyder committed
931
        ret = darshan_log_coll_append(log_fh, off, NULL, 0);
932 933 934 935
    }

    free(map_buf);

Shane Snyder's avatar
Shane Snyder committed
936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974
    if(ret < 0)
        return(-1);

    return(0);
}

/* NOTE: This function assumes that rank 0 passes in the log file offset
 *       to start appending at, so the caller must make sure rank 0
 *       has this info. After returning from this function, it is
 *       guaranteed all ranks will have the ending log file offset.
 */
static int darshan_log_coll_append(MPI_File log_fh, MPI_Offset *off,
    void *buf, int count)
{
    MPI_Offset send_off, my_off;
    MPI_Status status;
    int ret;

    /* figure out where everyone is writing */
    send_off = count;
    if(my_rank == 0)
        send_off += *off; /* rank 0 knows the beginning offset */

    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);

    *off = my_off + count;
    /* broadcast the final offset so everyone knows */
    DARSHAN_MPI_CALL(PMPI_Bcast)(off, 1, MPI_OFFSET, (nprocs-1), MPI_COMM_WORLD);

    return(0);
975 976
}

977
/* ********************************************************* */
978 979

void darshan_core_register_module(
980
    darshan_module_id id,
981 982 983 984
    char *name,
    struct darshan_module_funcs *funcs,
    int *runtime_mem_limit)
{
985
    struct darshan_core_module* mod;
986

987
    DARSHAN_CORE_LOCK();
988

989
    *runtime_mem_limit = 0;
990
    if(!darshan_core_job || (id >= DARSHAN_MAX_MODS))
991
    {
992
        DARSHAN_CORE_UNLOCK();
993
        return;
994
    }
995 996

    /* see if this module is already registered */
997
    if(darshan_core_job->mod_array[id])
998
    {
999
        /* if module is already registered just return */
1000
        /* NOTE: we do not recalculate memory limit here, just set to 0 */
1001
        DARSHAN_CORE_UNLOCK();
1002
        return;
1003 1004
    }

1005
    /* this module has not been registered yet, allocate and initialize it */
1006 1007
    mod = malloc(sizeof(*mod));
    if(!mod)
1008
    {
1009
        DARSHAN_CORE_UNLOCK();
1010 1011
        return;
    }
1012
    memset(mod, 0, sizeof(*mod));
1013

1014
    mod->id = id;
1015 1016
    strncpy(mod->name, name, DARSHAN_MOD_NAME_LEN);
    mod->mod_funcs = *funcs;
1017 1018 1019

    /* register module with darshan */
    darshan_core_job->mod_array[id] = mod;
1020

1021 1022
    /* TODO: something smarter than just 2 MiB per module */
    *runtime_mem_limit = 2 * 1024 * 1024;
1023

1024
    DARSHAN_CORE_UNLOCK();
1025

1026 1027 1028
    return;
}

1029
void darshan_core_lookup_record_id(
1030 1031 1032
    void *name,
    int len,
    int printable_flag,
1033
    darshan_record_id *id)
1034
{
1035 1036
    darshan_record_id tmp_id;
    struct darshan_core_record_ref* ref;
1037

1038
    if(!darshan_core_job || !name)
1039 1040 1041 1042 1043 1044
        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);
1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055
 
    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)
        {
1056 1057 1058 1059
            ref->rec.id = tmp_id;
            ref->rec.name = malloc(strlen(name) + 1);
            if(ref->rec.name)
                strcpy(ref->rec.name, name);
1060

1061
            HASH_ADD(hlink, darshan_core_job->rec_hash, rec.id, sizeof(darshan_record_id), ref);
1062 1063 1064 1065
        }
    }   

    DARSHAN_CORE_UNLOCK();
1066

Shane Snyder's avatar
Shane Snyder committed
1067 1068
    printf("New Darshan record: %s (%"PRIu64")\n", (char *)name, tmp_id);

1069
    *id = tmp_id;
1070 1071 1072
    return;
}

1073 1074
double darshan_core_wtime()
{
1075
    if(!darshan_core_job)
1076 1077 1078 1079
    {
        return(0);
    }

1080
    return(DARSHAN_MPI_CALL(PMPI_Wtime)() - darshan_core_job->wtime_offset);
1081
}
1082 1083 1084 1085 1086 1087 1088 1089 1090

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