darshan-core.c 44.3 KB
Newer Older
1 2 3 4 5
/*
 *  (C) 2009 by Argonne National Laboratory.
 *      See COPYRIGHT in top-level directory.
 */

6
#define _XOPEN_SOURCE 500
7
#define _GNU_SOURCE
8

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

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

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

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

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

39 40 41 42 43 44 45 46 47 48 49 50 51
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
    int64_t block_size;
    char path[DARSHAN_MAX_MNT_PATH];
    char type[DARSHAN_MAX_MNT_TYPE];
};
static struct mnt_data mnt_data_array[DARSHAN_MAX_MNTS];
static int mnt_data_count = 0;

52 53 54 55 56 57
/* prototypes for internal helper functions */
static void darshan_core_initialize(
    int *argc, char ***argv);
static void darshan_core_shutdown(
    void);
static void darshan_core_cleanup(
58
    struct darshan_core_runtime* core);
59 60 61
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
62 63 64 65 66 67
    struct darshan_core_runtime* core);
static void darshan_get_exe_and_mounts_root(
    struct darshan_core_runtime *core, char* trailing_data,
    int space_left);
static char* darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core);
68
static void darshan_get_shared_records(
69
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
70 71
static int darshan_log_coll_open(
    char *logfile_name, MPI_File *log_fh);
72 73
static int darshan_compress_buffer(void **pointers, int *lengths,
    int count, char *comp_buf, int *comp_length);
74
static int darshan_log_write_record_hash(
75
    MPI_File log_fh, struct darshan_core_runtime *core,
76
    struct darshan_log_map *map);
77 78
static int darshan_log_coll_write(
    MPI_File log_fh, void *buf, int count, struct darshan_log_map *map);
79

80 81
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)
82

83
/* intercept MPI initialize and finalize to manage darshan core runtime */
84 85 86 87 88 89 90 91 92 93
int MPI_Init(int *argc, char ***argv)
{
    int ret;

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

94
    darshan_core_initialize(argc, argv);
95 96 97 98

    return(ret);
}

99
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
100 101 102 103
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
104
    if(ret != MPI_SUCCESS)
105 106 107 108
    {
        return(ret);
    }

109
    darshan_core_initialize(argc, argv);
110 111 112 113 114 115 116 117

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

118
    darshan_core_shutdown();
119 120 121 122 123

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

124 125
/* *********************************** */

126
static void darshan_core_initialize(int *argc, char ***argv)
127 128 129 130 131 132 133 134 135
{
    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);
136
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
137 138 139 140

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

141
    if(internal_timing_flag)
142 143 144
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
145
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
146
    {
147 148 149 150 151
        /* TODO: darshan mem alignment code? */

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
152
        {
153
            memset(darshan_core, 0, sizeof(*darshan_core));
154

155 156 157 158
            darshan_core->log_job.uid = getuid();
            darshan_core->log_job.start_time = time(NULL);
            darshan_core->log_job.nprocs = nprocs;
            darshan_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();
159 160 161 162

            /* record exe and arguments */
            for(i=0; i<(*argc); i++)
            {
163
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
164
                strncat(darshan_core->exe, (*argv)[i], chars_left);
165 166
                if(i < ((*argc)-1))
                {
167
                    chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
168
                    strncat(darshan_core->exe, " ", chars_left);
169 170 171 172 173 174 175 176
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
177
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
178
                strncat(darshan_core->exe, __progname, chars_left);
179
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
180
                strncat(darshan_core->exe, " <unknown args>", chars_left);
181 182 183 184 185
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
186
                truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
187
                sprintf(&darshan_core->exe[truncate_offset], "%s",
188 189
                    truncate_string);
            }
190 191 192

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
193 194 195 196 197 198 199 200
        }
    }

    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);
201
        if(my_rank == 0)
202 203 204 205 206 207 208 209 210
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

211
static void darshan_core_shutdown()
212
{
213
    int i;
214
    char *logfile_name;
215
    struct darshan_core_runtime *final_core;
216
    int internal_timing_flag = 0;
217 218
    char *envjobid;
    char *jobid_str;
219
    int jobid;
220
    struct tm *start_tm;
221
    time_t start_time_tmp;
222 223
    int ret = 0;
    int all_ret = 0;
224 225
    int64_t first_start_time;
    int64_t last_end_time;
226 227
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
228
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
229 230
    double start_log_time;
    long offset;
231
    struct darshan_header log_header;
232
    MPI_File log_fh;
233
    MPI_Offset tmp_off = 0;
234
    MPI_Status status;
235 236 237 238

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

Shane Snyder's avatar
Shane Snyder committed
239
    /* disable darhan-core while we shutdown */
240
    DARSHAN_CORE_LOCK();
241
    if(!darshan_core)
242
    {
243
        DARSHAN_CORE_UNLOCK();
244 245
        return;
    }
246 247
    final_core = darshan_core;
    darshan_core = NULL;
Shane Snyder's avatar
Shane Snyder committed
248 249 250 251 252 253 254 255 256 257 258 259

    /* we also need to set which modules were registerd on this process and
     * disable tracing within those modules while we shutdown
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
            local_mod_use[i] = 1;
            final_core->mod_array[i]->mod_funcs.disable_instrumentation();
        }
    }
260
    DARSHAN_CORE_UNLOCK();
261

262 263
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

264 265 266
    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
267
        darshan_core_cleanup(final_core);
268 269 270
        return;
    }

271
    /* set darshan job id/metadata and constuct log file name on rank 0 */
272
    if(my_rank == 0)
273 274 275
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
276
        if(!envjobid)
277 278 279 280
        {
            envjobid = CP_JOBID;
        }

281
        /* find a job id */
282 283 284 285 286 287 288 289 290 291 292 293
        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();
        }

294
        final_core->log_job.jobid = (int64_t)jobid;
295

296
        /* if we are using any hints to write the log file, then record those
297
         * hints with the darshan job information
298
         */
299
        darshan_log_record_hints_and_ver(final_core);
300

301
        /* use human readable start time format in log filename */
302
        start_time_tmp = final_core->log_job.start_time;
303
        start_tm = localtime(&start_time_tmp);
304

305 306
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
307 308 309 310 311 312 313 314 315
    }

    /* 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 */
316
        free(logfile_name);
317
        darshan_core_cleanup(final_core);
318 319 320
        return;
    }

321
    final_core->log_job.end_time = time(NULL);
322

323 324 325
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
326 327
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job.start_time, &first_start_time, 1, MPI_LONG_LONG, MPI_MIN, 0, MPI_COMM_WORLD);
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job.end_time, &last_end_time, 1, MPI_LONG_LONG, MPI_MAX, 0, MPI_COMM_WORLD);
328 329
    if(my_rank == 0)
    {
330 331
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
332
    }
333

334 335 336
    /* 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);

337
    /* get a list of records which are shared across all processes */
338
    darshan_get_shared_records(final_core, shared_recs);
339

340 341
    /* collectively open the darshan log file */
    ret = darshan_log_coll_open(logfile_name, &log_fh);
342 343 344 345 346 347 348 349

    /* 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)
        {
350 351
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
352 353 354
            unlink(logfile_name);
        }
        free(logfile_name);
355
        darshan_core_cleanup(final_core);
356 357 358
        return;
    }

359
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
360
    if(my_rank == 0)
361
    {
362 363 364 365 366 367 368 369
        void *pointers[2] = {&final_core->log_job, final_core->trailing_data};
        int lengths[2] = {sizeof(struct darshan_job), DARSHAN_EXE_LEN+1};
        int comp_buf_sz = 0;
        
        /* compress the job info and the trailing mount/exe data */
        all_ret = darshan_compress_buffer(pointers, lengths, 2,
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
370
        {
371
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
372
            unlink(logfile_name);
373
        }
374 375 376 377 378 379 380 381 382 383 384 385
        else
        {
            /* write the job information, preallocing space for the log header */
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(
                log_fh, sizeof(struct darshan_header), final_core->comp_buf,
                comp_buf_sz, MPI_BYTE, &status);
            if(all_ret != MPI_SUCCESS)
            {
                fprintf(stderr, "darshan library warning: unable to write job data to log file %s\n",
                        logfile_name);
                unlink(logfile_name);
            }
Shane Snyder's avatar
Shane Snyder committed
386

387 388 389
            /* set the beginning offset of record hash, which precedes job info just written */
            log_header.rec_map.off = sizeof(struct darshan_header) + comp_buf_sz;
        }
390 391
    }

392 393 394 395 396
    /* 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);
397
        darshan_core_cleanup(final_core);
398 399 400 401
        return;
    }

    /* write the record name->id hash to the log file */
402
    ret = darshan_log_write_record_hash(log_fh, final_core, &log_header.rec_map);
403

404
    /* error out if unable to write record hash */
405 406 407 408 409 410
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
411
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
412
                logfile_name);
413
            unlink(logfile_name);
414 415
        }
        free(logfile_name);
416
        darshan_core_cleanup(final_core);
417 418
        return;
    }
419 420

    /* loop over globally used darshan modules and:
421
     *      - perform shared file reductions, if possible
422
     *      - get final output buffer
423
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
424
     *      - append compressed buffer to log file
425 426
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
427
     */
428
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
429
    {
430
        struct darshan_core_module* this_mod = final_core->mod_array[i];
431 432 433
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        struct darshan_core_record_ref *ref = NULL;
        int shared_rec_cnt = 0;
434
        void* mod_buf = NULL;
435 436
        int mod_buf_sz = 0;
        int comp_buf_sz = 0;
437
        int j;
438

439
        if(global_mod_use_count[i] == 0)
440 441 442 443
        {
            if(my_rank == 0)
                log_header.mod_map[i].off = log_header.mod_map[i].len = 0;

444
            continue;
445
        }
446
        else if(global_mod_use_count[j] == nprocs)
447
        {
448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494
            int shared_rec_count = 0;
            int rec_sz = 0;
            void *red_send_buf = NULL, *red_recv_buf = NULL;
            MPI_Datatype red_type;
            MPI_Op red_op;

            /* if all processes used this module, prepare to do a shared file reduction */

            /* set the shared file list for this module */
            memset(mod_shared_recs, 0, DARSHAN_CORE_MAX_RECORDS * sizeof(darshan_record_id));
            for(j = 0; j < DARSHAN_CORE_MAX_RECORDS && shared_recs[j] != 0; j++)
            {
                HASH_FIND(hlink, final_core->rec_hash, &shared_recs[j],
                    sizeof(darshan_record_id), ref);
                assert(ref);
                if(ref->global_mod_flags & (1 << i)) /* TODO: MACRO? */
                {
                    mod_shared_recs[shared_rec_count++] = shared_recs[j];
                }
            }

            /* if there are globally shared files, do a shared file reduction */
            if(shared_rec_count)
            {
                this_mod->mod_funcs.prepare_for_reduction(mod_shared_recs, &shared_rec_count,
                    &red_send_buf, &red_recv_buf, &rec_sz);

                if(shared_rec_count)
                {
                    /* construct a datatype for a file record.  This is serving no purpose
                     * except to make sure we can do a reduction on proper boundaries
                     */
                    DARSHAN_MPI_CALL(PMPI_Type_contiguous)(rec_sz, MPI_BYTE, &red_type);
                    DARSHAN_MPI_CALL(PMPI_Type_commit)(&red_type);

                    /* register a reduction operator for this module */
                    DARSHAN_MPI_CALL(PMPI_Op_create)(this_mod->mod_funcs.reduce_record,
                        1, &red_op);

                    /* reduce shared file records for this module */
                    DARSHAN_MPI_CALL(PMPI_Reduce)(red_send_buf, red_recv_buf,
                        shared_rec_count, red_type, red_op, 0, MPI_COMM_WORLD);

                    DARSHAN_MPI_CALL(PMPI_Type_free)(&red_type);
                    DARSHAN_MPI_CALL(PMPI_Op_free)(&red_op);
                }
            }
495
        }
496

497
        /* if module is registered locally, get the corresponding output buffer */
498
        if(this_mod)
499 500
        {
            /* get output buffer from module */
501
            this_mod->mod_funcs.get_output_data(&mod_buf, &mod_buf_sz);
502 503
        }

504 505 506 507 508 509 510 511
        /* compress the module buffer */
        ret = darshan_compress_buffer((void**)&mod_buf, &mod_buf_sz, 1,
            final_core->comp_buf, &comp_buf_sz);
        if(ret == 0)
        {
            /* set the starting offset of this module */
            if(tmp_off == 0)
                tmp_off = log_header.rec_map.off + log_header.rec_map.len;
512

513
            log_header.mod_map[i].off = tmp_off;
514

515 516 517 518 519 520 521 522 523 524 525
            /* write (compressed) module data buffer to the darshan log file */
            ret = darshan_log_coll_write(log_fh, final_core->comp_buf, comp_buf_sz,
                &log_header.mod_map[i]);
        }
        else
        {
            /* error in compression, participate in collective write to avoid
             * deadlock, but preserve return value to terminate darshan_shutdown.
             */
            (void)darshan_log_coll_write(log_fh, NULL, 0, &log_header.mod_map[i]);
        }
526

527
        /* error out if the compression or collective write failed */
528 529 530
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
531
        {
532 533 534 535 536 537 538 539
            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);
540
            darshan_core_cleanup(final_core);
541
            return;
542 543
        }

544 545
        tmp_off += log_header.mod_map[i].len;

546
        /* shutdown module if registered locally */
547
        if(this_mod)
548 549 550 551 552
        {
            this_mod->mod_funcs.shutdown();
        }
    }

553
    /* rank 0 is responsible for writing the log header */
554 555
    if(my_rank == 0)
    {
556
        /* initialize the remaining header fields */
557 558
        strcpy(log_header.version_string, DARSHAN_LOG_VERSION);
        log_header.magic_nr = DARSHAN_MAGIC_NR;
559 560
        log_header.comp_type = DARSHAN_GZ_COMP;

561
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &log_header,
562
            sizeof(struct darshan_header), MPI_BYTE, &status);
563
        if(all_ret != MPI_SUCCESS)
564
        {
565 566 567
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
568
        }
569 570
    }

571 572 573 574 575
    /* 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);
576
        darshan_core_cleanup(final_core);
577 578 579
        return;
    }

580 581 582 583 584 585
    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
     */
586
    /* TODO: support user given logfile path/name */
587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602
    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 */
603
#ifdef __CP_GROUP_READABLE_LOGS
604
            chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
605
#else
606
            chmod(new_logfile_name, (S_IRUSR));
607
#endif
608 609
            free(new_logfile_name);
        }
610
    }
611

612
    free(logfile_name);
613
    darshan_core_cleanup(final_core);
614

615
    if(internal_timing_flag)
616
    {
617 618 619 620 621
        /* TODO: what do we want to time in new darshan version? */
    }
    
    return;
}
622

623
/* free darshan core data structures to shutdown */
624
static void darshan_core_cleanup(struct darshan_core_runtime* core)
625
{
626
    struct darshan_core_record_ref *tmp, *ref;
627
    int i;
628

629 630 631 632 633 634
    HASH_ITER(hlink, core->rec_hash, ref, tmp)
    {
        HASH_DELETE(hlink, core->rec_hash, ref);
        free(ref->rec.name);
        free(ref);
    }
635

636
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
637
    {
638
        if(core->mod_array[i])
639
        {        
640 641
            free(core->mod_array[i]);
            core->mod_array[i] = NULL;
642
        }
643
    }
644

645
    free(core->trailing_data);
646
    free(core);
647 648 649 650

    return;
}

651
/* construct the darshan log file name */
652
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
653
{
654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783
    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;
784 785
}

786
/* record any hints used to write the darshan log in the log header */
787
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
{
    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 -
811
        strlen(core->log_job.metadata) - 1;
812 813
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
814
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
815 816 817 818
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
819
        m = core->log_job.metadata + strlen(core->log_job.metadata);
820 821 822 823 824 825 826 827 828 829 830 831
        /* 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;
}

832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871
static int mnt_data_cmp(const void* a, const void* b)
{
    const struct mnt_data *d_a = (const struct mnt_data*)a;
    const struct mnt_data *d_b = (const struct mnt_data*)b;

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

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

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

    /* store mount information for use in header of darshan log */
872
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
        strcat(trailing_data, tmp_mnt);
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

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

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

    /* length of exe has already been safety checked in darshan-posix.c */
    strcat(trailing_data, core->exe);
918
    space_left = DARSHAN_EXE_LEN - strlen(trailing_data);
919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981

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

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

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

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

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

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

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

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

982
    space_left = DARSHAN_EXE_LEN + 1;
983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
    trailing_data = malloc(space_left);
    if(!trailing_data)
    {
        return(NULL);
    }
    memset(trailing_data, 0, space_left);

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

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

    return(trailing_data);
}

1008
static void darshan_get_shared_records(struct darshan_core_runtime *core,
1009
    darshan_record_id *shared_recs)
1010 1011 1012
{
    int i;
    int ndx;
1013
    struct darshan_core_record_ref *tmp, *ref;
1014
    darshan_record_id id_array[DARSHAN_CORE_MAX_RECORDS] = {0};
1015 1016
    uint64_t mod_flags[DARSHAN_CORE_MAX_RECORDS] = {0};
    uint64_t global_mod_flags[DARSHAN_CORE_MAX_RECORDS] = {0};
1017 1018 1019 1020 1021

    /* first, determine list of records root process has opened */
    if(my_rank == 0)
    {
        ndx = 0;
1022
        HASH_ITER(hlink, core->rec_hash, ref, tmp)
1023
        {
1024
            id_array[ndx++] = ref->rec.id;           
1025 1026 1027 1028
        }
    }

    /* broadcast root's list of records to all other processes */
1029
    DARSHAN_MPI_CALL(PMPI_Bcast)(id_array,
1030 1031 1032 1033 1034 1035
        (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++)
    {
1036 1037
        HASH_FIND(hlink, core->rec_hash, &id_array[i], sizeof(darshan_record_id), ref);
        if(ref)
1038
        {
1039 1040 1041
            /* we opened that record too, save the mod_flags */
            mod_flags[i] = ref->mod_flags;
            break;
1042 1043 1044
        }
    }

1045 1046 1047 1048 1049
    /* now allreduce so everyone agrees which files are shared and
     * which modules accessed them collectively
     */
    DARSHAN_MPI_CALL(PMPI_Allreduce)(mod_flags, global_mod_flags,
        DARSHAN_CORE_MAX_RECORDS, MPI_UINT64_T, MPI_LAND, MPI_COMM_WORLD);
1050 1051 1052 1053

    ndx = 0;
    for(i=0; (i<DARSHAN_CORE_MAX_RECORDS && id_array[i] != 0); i++)
    {
1054
        if(global_mod_flags[i] != 0)
1055 1056
        {
            shared_recs[ndx++] = id_array[i];
1057 1058 1059 1060 1061 1062 1063 1064

            /* set global_mod_flags so we know which modules collectively
             * accessed this module. we need this info to support shared
             * file reductions
             */
            HASH_FIND(hlink, core->rec_hash, &id_array[i], sizeof(darshan_record_id), ref);
            assert(ref);
            ref->global_mod_flags = global_mod_flags[i];
1065 1066 1067
        }
    }

1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128
    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)