darshan-core.c 47.3 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 34
#define DARSHAN_MPI_CALL(func) func

35
/* internal variable delcarations */
36
static struct darshan_core_runtime *darshan_core = NULL;
37
static pthread_mutex_t darshan_core_mutex = PTHREAD_MUTEX_INITIALIZER;
38
static int my_rank = -1;
39
static int nprocs = -1;
40

41 42 43 44 45 46 47 48 49 50 51 52 53
/* 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;

54 55 56 57 58 59
/* 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(
60
    struct darshan_core_runtime* core);
61 62 63
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
64 65 66 67 68 69
    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);
70
static void darshan_get_shared_records(
71
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
72 73
static int darshan_log_coll_open(
    char *logfile_name, MPI_File *log_fh);
74 75
static int darshan_compress_buffer(void **pointers, int *lengths,
    int count, char *comp_buf, int *comp_length);
76
static int darshan_log_write_record_hash(
77
    MPI_File log_fh, struct darshan_core_runtime *core,
78
    struct darshan_log_map *map);
79 80
static int darshan_log_coll_write(
    MPI_File log_fh, void *buf, int count, struct darshan_log_map *map);
81

82 83
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)
84

85
/* intercept MPI initialize and finalize to manage darshan core runtime */
86 87 88 89 90 91 92 93 94 95
int MPI_Init(int *argc, char ***argv)
{
    int ret;

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

96
    darshan_core_initialize(argc, argv);
97 98 99 100

    return(ret);
}

101
int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
102 103 104 105
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
106
    if(ret != MPI_SUCCESS)
107 108 109 110
    {
        return(ret);
    }

111
    darshan_core_initialize(argc, argv);
112 113 114 115 116 117 118 119

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

120
    darshan_core_shutdown();
121 122 123 124 125

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

126 127
/* *********************************** */

128
static void darshan_core_initialize(int *argc, char ***argv)
129 130 131 132 133 134 135 136 137
{
    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);
138
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
139 140 141 142

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

143
    if(internal_timing_flag)
144 145 146
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
147
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
148
    {
149 150 151 152 153
        /* TODO: darshan mem alignment code? */

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
154
        {
155
            memset(darshan_core, 0, sizeof(*darshan_core));
156

157 158 159 160
            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)();
161 162 163 164

            /* record exe and arguments */
            for(i=0; i<(*argc); i++)
            {
165
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
166
                strncat(darshan_core->exe, (*argv)[i], chars_left);
167 168
                if(i < ((*argc)-1))
                {
169
                    chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
170
                    strncat(darshan_core->exe, " ", chars_left);
171 172 173 174 175 176 177 178
                }
            }

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
179
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
180
                strncat(darshan_core->exe, __progname, chars_left);
181
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
182
                strncat(darshan_core->exe, " <unknown args>", chars_left);
183 184 185 186 187
            }

            if(chars_left == 0)
            {
                /* we ran out of room; mark that string was truncated */
188
                truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
189
                sprintf(&darshan_core->exe[truncate_offset], "%s",
190 191
                    truncate_string);
            }
192 193 194

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
195 196 197 198 199 200 201 202
        }
    }

    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);
203
        if(my_rank == 0)
204 205 206 207 208 209 210 211 212
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

213
static void darshan_core_shutdown()
214
{
215
    int i;
216
    char *logfile_name;
217
    struct darshan_core_runtime *final_core;
218
    int internal_timing_flag = 0;
219 220
    char *envjobid;
    char *jobid_str;
221
    int jobid;
222
    struct tm *start_tm;
223
    time_t start_time_tmp;
224 225
    int ret = 0;
    int all_ret = 0;
226 227
    int64_t first_start_time;
    int64_t last_end_time;
228 229
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
230
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
231
    double start_log_time;
232 233 234 235 236 237 238
    double open1, open2;
    double job1, job2;
    double rec1, rec2;
    double mod1[DARSHAN_MAX_MODS] = {0};
    double mod2[DARSHAN_MAX_MODS] = {0};
    double header1, header2;
    double tm_end;
239
    long offset;
240
    struct darshan_header log_header;
241
    MPI_File log_fh;
242
    MPI_Offset tmp_off = 0;
243
    MPI_Status status;
244 245 246 247

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

248 249
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

Shane Snyder's avatar
Shane Snyder committed
250
    /* disable darhan-core while we shutdown */
251
    DARSHAN_CORE_LOCK();
252
    if(!darshan_core)
253
    {
254
        DARSHAN_CORE_UNLOCK();
255 256
        return;
    }
257 258
    final_core = darshan_core;
    darshan_core = NULL;
Shane Snyder's avatar
Shane Snyder committed
259

260
    /* we also need to set which modules were registered on this process and
Shane Snyder's avatar
Shane Snyder committed
261 262 263 264 265 266 267 268 269 270
     * 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();
        }
    }
271
    DARSHAN_CORE_UNLOCK();
272 273 274 275

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
276
        darshan_core_cleanup(final_core);
277 278 279
        return;
    }

280
    /* set darshan job id/metadata and constuct log file name on rank 0 */
281
    if(my_rank == 0)
282 283 284
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
285
        if(!envjobid)
286 287 288 289
        {
            envjobid = CP_JOBID;
        }

290
        /* find a job id */
291 292 293 294 295 296 297 298 299 300 301 302
        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();
        }

303
        final_core->log_job.jobid = (int64_t)jobid;
304

305
        /* if we are using any hints to write the log file, then record those
306
         * hints with the darshan job information
307
         */
308
        darshan_log_record_hints_and_ver(final_core);
309

310
        /* use human readable start time format in log filename */
311
        start_time_tmp = final_core->log_job.start_time;
312
        start_tm = localtime(&start_time_tmp);
313

314 315
        /* construct log file name */
        darshan_get_logfile_name(logfile_name, jobid, start_tm);
316 317 318 319 320 321 322 323 324
    }

    /* 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 */
325
        free(logfile_name);
326
        darshan_core_cleanup(final_core);
327 328 329
        return;
    }

330
    final_core->log_job.end_time = time(NULL);
331

332 333 334
    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
335 336
    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);
337 338
    if(my_rank == 0)
    {
339 340
        final_core->log_job.start_time = first_start_time;
        final_core->log_job.end_time = last_end_time;
341
    }
342

343 344 345
    /* 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);

346
    /* get a list of records which are shared across all processes */
347
    darshan_get_shared_records(final_core, shared_recs);
348

349 350
    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
351 352
    /* collectively open the darshan log file */
    ret = darshan_log_coll_open(logfile_name, &log_fh);
353 354
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
355 356 357 358 359 360 361 362

    /* 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)
        {
363 364
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
365 366 367
            unlink(logfile_name);
        }
        free(logfile_name);
368
        darshan_core_cleanup(final_core);
369 370 371
        return;
    }

372 373
    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
374
    /* rank 0 is responsible for writing the compressed darshan job information */
Shane Snyder's avatar
Shane Snyder committed
375
    if(my_rank == 0)
376
    {
377 378 379
        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;
380

381 382 383 384
        /* 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)
385
        {
386
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
387
            unlink(logfile_name);
388
        }
389 390 391 392 393 394 395 396 397 398 399 400
        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
401

402 403 404
            /* 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;
        }
405 406
    }

407 408 409 410 411
    /* 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);
412
        darshan_core_cleanup(final_core);
413 414
        return;
    }
415 416
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
417

418 419
    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
420
    /* write the record name->id hash to the log file */
421
    ret = darshan_log_write_record_hash(log_fh, final_core, &log_header.rec_map);
422

423
    /* error out if unable to write record hash */
424 425 426 427 428 429
    DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
        MPI_LOR, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        if(my_rank == 0)
        {
430
            fprintf(stderr, "darshan library warning: unable to write record hash to log file %s\n",
431
                logfile_name);
432
            unlink(logfile_name);
433 434
        }
        free(logfile_name);
435
        darshan_core_cleanup(final_core);
436 437
        return;
    }
Shane Snyder's avatar
Shane Snyder committed
438 439
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
440

441
    /* TODO: would be nice to factor this out somehow ... a lot to look at */
442
    /* loop over globally used darshan modules and:
443
     *      - perform shared file reductions, if possible
444
     *      - get final output buffer
445
     *      - compress (zlib) provided output buffer
Shane Snyder's avatar
Shane Snyder committed
446
     *      - append compressed buffer to log file
447 448
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
449
     */
450
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
451
    {
452
        struct darshan_core_module* this_mod = final_core->mod_array[i];
453 454 455
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        struct darshan_core_record_ref *ref = NULL;
        int shared_rec_cnt = 0;
456
        void* mod_buf = NULL;
457 458
        int mod_buf_sz = 0;
        int comp_buf_sz = 0;
459
        int j;
460

461
        if(global_mod_use_count[i] == 0)
462 463 464 465
        {
            if(my_rank == 0)
                log_header.mod_map[i].off = log_header.mod_map[i].len = 0;

466
            continue;
467
        }
468 469 470 471 472
 
        if(internal_timing_flag)
            mod1[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();   
        /* if all processes used this module, prepare to do a shared file reduction */
        if(global_mod_use_count[j] == nprocs)
473
        {
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518
            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;

            /* 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);
                }
            }
519
        }
520

521
        /* if module is registered locally, get the corresponding output buffer */
522
        if(this_mod)
523 524
        {
            /* get output buffer from module */
525
            this_mod->mod_funcs.get_output_data(&mod_buf, &mod_buf_sz);
526 527
        }

528 529 530 531 532 533 534 535
        /* 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;
536

537
            log_header.mod_map[i].off = tmp_off;
538

539 540 541 542 543 544 545 546 547 548 549
            /* 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]);
        }
550

551
        /* error out if the compression or collective write failed */
552 553 554
        DARSHAN_MPI_CALL(PMPI_Allreduce)(&ret, &all_ret, 1, MPI_INT,
            MPI_LOR, MPI_COMM_WORLD);
        if(all_ret != 0)
555
        {
556 557 558 559 560 561 562 563
            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);
564
            darshan_core_cleanup(final_core);
565
            return;
566 567
        }

568 569
        tmp_off += log_header.mod_map[i].len;

570
        /* shutdown module if registered locally */
571
        if(this_mod)
572 573 574
        {
            this_mod->mod_funcs.shutdown();
        }
575 576
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
577 578
    }

579 580
    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
581
    /* rank 0 is responsible for writing the log header */
582 583
    if(my_rank == 0)
    {
584
        /* initialize the remaining header fields */
585 586
        strcpy(log_header.version_string, DARSHAN_LOG_VERSION);
        log_header.magic_nr = DARSHAN_MAGIC_NR;
587 588
        log_header.comp_type = DARSHAN_GZ_COMP;

589
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &log_header,
590
            sizeof(struct darshan_header), MPI_BYTE, &status);
591
        if(all_ret != MPI_SUCCESS)
592
        {
593 594 595
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
596
        }
597 598
    }

599 600 601 602 603
    /* 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);
604
        darshan_core_cleanup(final_core);
605 606
        return;
    }
607 608
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();
609

610 611 612 613 614 615
    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
     */
616
    /* TODO: support user given logfile path/name */
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632
    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 */
633
#ifdef __CP_GROUP_READABLE_LOGS
634
            chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
635
#else
636
            chmod(new_logfile_name, (S_IRUSR));
637
#endif
638 639
            free(new_logfile_name);
        }
640
    }
641

642
    free(logfile_name);
643
    darshan_core_cleanup(final_core);
644

645
    if(internal_timing_flag)
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
        double open_tm, open_slowest;
        double header_tm, header_slowest;
        double job_tm, job_slowest;
        double rec_tm, rec_slowest;
        double mod_tm[DARSHAN_MAX_MODS], mod_slowest[DARSHAN_MAX_MODS];
        double all_tm, all_slowest;

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

        open_tm = open2 - open1;
        header_tm = header2 - header1;
        job_tm = job2 - job1;
        rec_tm = rec2 - rec1;
        all_tm = tm_end - start_log_time;
        for(i = 0;i < DARSHAN_MAX_MODS; i++)
        {
            mod_tm[i] = mod2[i] - mod1[i];
        }

        DARSHAN_MPI_CALL(PMPI_Reduce)(&open_tm, &open_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&header_tm, &header_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&job_tm, &job_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&rec_tm, &rec_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(&all_tm, &all_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(mod_tm, mod_slowest, DARSHAN_MAX_MODS,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);

        if(my_rank == 0)
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:log_open\t%d\t%f\n", nprocs, open_slowest);
            printf("darshan:job_write\t%d\t%f\n", nprocs, job_slowest);
            printf("darshan:hash_write\t%d\t%f\n", nprocs, rec_slowest);
Shane Snyder's avatar
Shane Snyder committed
685
            printf("darshan:header_write\t%d\t%f\n", nprocs, header_slowest);
686 687 688 689
            for(i = 0; i < DARSHAN_MAX_MODS; i++)
            {
                if(global_mod_use_count[i])
                    printf("darshan:%s_shutdown\t%d\t%f\n", darshan_module_names[i],
Shane Snyder's avatar
Shane Snyder committed
690
                        nprocs, mod_slowest[i]);
691 692 693
            }
            printf("darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
        }
694 695 696 697
    }
    
    return;
}
698

699
/* free darshan core data structures to shutdown */
700
static void darshan_core_cleanup(struct darshan_core_runtime* core)
701
{
702
    struct darshan_core_record_ref *tmp, *ref;
703
    int i;
704

705 706 707 708 709 710
    HASH_ITER(hlink, core->rec_hash, ref, tmp)
    {
        HASH_DELETE(hlink, core->rec_hash, ref);
        free(ref->rec.name);
        free(ref);
    }
711

712
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
713
    {
714
        if(core->mod_array[i])
715
        {        
716 717
            free(core->mod_array[i]);
            core->mod_array[i] = NULL;
718
        }
719
    }
720

721
    free(core->trailing_data);
722
    free(core);
723 724 725 726

    return;
}

727
/* construct the darshan log file name */
728
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
729
{
730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859
    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;
860 861
}

862
/* record any hints used to write the darshan log in the log header */
863
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886
{
    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 -
887
        strlen(core->log_job.metadata) - 1;
888 889
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
890
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
891 892 893 894
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
895
        m = core->log_job.metadata + strlen(core->log_job.metadata);
896 897 898 899 900 901 902 903 904 905 906 907
        /* 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;
}

908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947
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 */
948
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993
        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);
994
    space_left = DARSHAN_EXE_LEN - strlen(trailing_data);
995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057

    /* 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;

1058
    space_left = DARSHAN_EXE_LEN + 1;
1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
    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);
}

1084
static void darshan_get_shared_records(struct darshan_core_runtime *core,
1085
    darshan_record_id *shared_recs)
1086 1087 1088
{
    int i;
    int ndx;
1089
    struct darshan_core_record_ref *tmp, *ref;
1090
    darshan_record_id id_array[DARSHAN_CORE_MAX_RECORDS] = {0};