darshan-core.c 65.1 KB
Newer Older
1
/*
Shane Snyder's avatar
Shane Snyder committed
2 3 4
 * Copyright (C) 2015 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
5 6
 */

7
#define _XOPEN_SOURCE 500
8
#define _GNU_SOURCE
9

10 11 12 13 14 15
#include "darshan-runtime-config.h"

#include <stdio.h>
#ifdef HAVE_MNTENT_H
#include <mntent.h>
#endif
16 17 18 19 20
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <limits.h>
#include <pthread.h>
21
#include <fcntl.h>
22 23
#include <dirent.h>
#include <sys/ioctl.h>
24 25
#include <sys/types.h>
#include <sys/stat.h>
26
#include <sys/mman.h>
27
#include <sys/vfs.h>
28
#include <zlib.h>
29
#include <mpi.h>
30
#include <assert.h>
31

32
#include "uthash.h"
Shane Snyder's avatar
Shane Snyder committed
33
#include "darshan.h"
34
#include "darshan-core.h"
Shane Snyder's avatar
Shane Snyder committed
35
#include "darshan-dynamic.h"
36

37
#ifdef DARSHAN_LUSTRE
38
#include <lustre/lustre_user.h>
39
#endif
40

41
extern char* __progname;
42
extern char* __progname_full;
43

44
/* internal variable delcarations */
45
static struct darshan_core_runtime *darshan_core = NULL;
46
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
47
static int my_rank = -1;
48
static int nprocs = -1;
49
static int darshan_mem_alignment = 1;
50
static long darshan_mod_mem_quota = DARSHAN_MOD_MEM_MAX;
51

52
/* paths prefixed with the following directories are not tracked by darshan */
53
char* darshan_path_exclusions[] = {
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
    "/etc/",
    "/dev/",
    "/usr/",
    "/bin/",
    "/boot/",
    "/lib/",
    "/opt/",
    "/sbin/",
    "/sys/",
    "/proc/",
    "/var/",
    NULL
};
/* paths prefixed with the following directories are tracked by darshan even if
 * they share a root with a path listed in darshan_path_exclusions
 */
char* darshan_path_inclusions[] = {
    "/var/opt/cray/dws/mounts/",
    NULL
73 74
};

75 76 77 78 79 80 81 82 83 84 85 86 87 88 89
#ifdef DARSHAN_BGQ
extern void bgq_runtime_initialize();
#endif

/* array of init functions for modules which need to be statically
 * initialized by darshan at startup time
 */
void (*mod_static_init_fns[])(void) =
{
#ifdef DARSHAN_BGQ
    &bgq_runtime_initialize,
#endif
    NULL
};

90 91 92 93 94 95 96
#ifdef DARSHAN_LUSTRE
/* XXX need to use extern to get Lustre module's instrumentation function
 * since modules have no way of providing this to darshan-core
 */
extern void darshan_instrument_lustre_file(const char *filepath, int fd);
#endif

Shane Snyder's avatar
Shane Snyder committed
97 98 99
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

100 101 102 103 104 105 106 107
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
    char path[DARSHAN_MAX_MNT_PATH];
    char type[DARSHAN_MAX_MNT_TYPE];
108
    struct darshan_fs_info fs_info;
109 110 111 112
};
static struct mnt_data mnt_data_array[DARSHAN_MAX_MNTS];
static int mnt_data_count = 0;

113
/* prototypes for internal helper functions */
114
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
115 116
static void *darshan_init_mmap_log(
    struct darshan_core_runtime* core, int jobid);
117
#endif
118
static void darshan_log_record_hints_and_ver(
119
    struct darshan_core_runtime* core);
120 121
static void darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core, int argc, char **argv);
122 123
static void darshan_fs_info_from_path(
    const char *path, struct darshan_fs_info *fs_info);
124
static int darshan_add_name_record_ref(
125
    struct darshan_core_runtime *core, darshan_record_id rec_id,
126
    const char *name, darshan_module_id mod_id);
127 128
static void darshan_get_user_name(
    char *user);
129 130
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
131
static void darshan_get_shared_records(
132 133
    struct darshan_core_runtime *core, darshan_record_id **shared_recs,
    int *shared_rec_cnt);
134
static int darshan_log_open_all(
135
    char *logfile_name, MPI_File *log_fh);
136
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
137 138
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
139
static int darshan_log_write_name_record_hash(
140
    MPI_File log_fh, struct darshan_core_runtime *core,
141 142 143
    uint64_t *inout_off);
static int darshan_log_append_all(
    MPI_File log_fh, struct darshan_core_runtime *core, void *buf,
Shane Snyder's avatar
Shane Snyder committed
144
    int count, uint64_t *inout_off);
Shane Snyder's avatar
Shane Snyder committed
145 146
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
147

148 149
/* *********************************** */

Shane Snyder's avatar
Shane Snyder committed
150
void darshan_core_initialize(int argc, char **argv)
151
{
152
    struct darshan_core_runtime *init_core = NULL;
153 154
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
155
    char *envstr;
156 157
    char *jobid_str;
    int jobid;
158
    int ret;
159
    int i;
160 161
    int tmpval;
    double tmpfloat;
162 163

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
164
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
165 166 167 168

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

169
    if(internal_timing_flag)
170 171 172
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
173
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
174
    {
175
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
176 177
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
178
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
179 180 181 182 183 184 185 186 187 188 189
        if(envstr)
        {
            ret = sscanf(envstr, "%d", &tmpval);
            /* silently ignore if the env variable is set poorly */
            if(ret == 1 && tmpval > 0)
            {
                darshan_mem_alignment = tmpval;
            }
        }
        else
        {
190
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
191 192 193
        }

        /* avoid floating point errors on faulty input */
194
        if(darshan_mem_alignment < 1)
195 196 197
        {
            darshan_mem_alignment = 1;
        }
198

199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
        /* Use DARSHAN_JOBID_OVERRIDE for the env var for __DARSHAN_JOBID */
        envstr = getenv(DARSHAN_JOBID_OVERRIDE);
        if(!envstr)
        {
            envstr = __DARSHAN_JOBID;
        }

        /* find a job id */
        jobid_str = getenv(envstr);
        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();
        }

219 220 221 222
        /* set the memory quota for darshan modules' records */
        envstr = getenv(DARSHAN_MOD_MEM_OVERRIDE);
        if(envstr)
        {
223
            ret = sscanf(envstr, "%lf", &tmpfloat);
224
            /* silently ignore if the env variable is set poorly */
225
            if(ret == 1 && tmpfloat > 0)
226
            {
227
                darshan_mod_mem_quota = tmpfloat * 1024 * 1024; /* convert from MiB */
228 229 230
            }
        }

231 232 233
        /* allocate structure to track darshan core runtime information */
        init_core = malloc(sizeof(*init_core));
        if(init_core)
234
        {
235 236 237
            memset(init_core, 0, sizeof(*init_core));
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

238 239
        /* TODO: do we alloc new memory as we go or just do everything up front? */

240 241 242 243 244
#ifndef __DARSHAN_ENABLE_MMAP_LOGS
            /* just allocate memory for each log file region */
            init_core->log_hdr_p = malloc(sizeof(struct darshan_header));
            init_core->log_job_p = malloc(sizeof(struct darshan_job));
            init_core->log_exemnt_p = malloc(DARSHAN_EXE_LEN+1);
245
            init_core->log_name_p = malloc(DARSHAN_NAME_RECORD_BUF_SIZE);
246
            init_core->log_mod_p = malloc(darshan_mod_mem_quota);
247 248

            if(!(init_core->log_hdr_p) || !(init_core->log_job_p) ||
249
               !(init_core->log_exemnt_p) || !(init_core->log_name_p) ||
250 251 252 253 254 255 256 257 258
               !(init_core->log_mod_p))
            {
                free(init_core);
                return;
            }
            /* if allocation succeeds, zero fill memory regions */
            memset(init_core->log_hdr_p, 0, sizeof(struct darshan_header));
            memset(init_core->log_job_p, 0, sizeof(struct darshan_job));
            memset(init_core->log_exemnt_p, 0, DARSHAN_EXE_LEN+1);
259
            memset(init_core->log_name_p, 0, DARSHAN_NAME_RECORD_BUF_SIZE);
260
            memset(init_core->log_mod_p, 0, darshan_mod_mem_quota);
261 262 263 264
#else
            /* if mmap logs are enabled, we need to initialize the mmap region
             * before setting the corresponding log file region pointers
             */
265 266
            void *mmap_p = darshan_init_mmap_log(init_core, jobid);
            if(!mmap_p)
267
            {
268 269
                free(init_core);
                return;
270 271
            }

272
            /* set the memory pointers for each log file region */
273
            init_core->log_hdr_p = (struct darshan_header *)mmap_p;
274
            init_core->log_job_p = (struct darshan_job *)
275
                ((char *)init_core->log_hdr_p + sizeof(struct darshan_header));
276
            init_core->log_exemnt_p = (char *)
277
                ((char *)init_core->log_job_p + sizeof(struct darshan_job));
278
            init_core->log_name_p = (void *)
279 280
                ((char *)init_core->log_exemnt_p + DARSHAN_EXE_LEN + 1);
            init_core->log_mod_p = (void *)
281
                ((char *)init_core->log_name_p + DARSHAN_NAME_RECORD_BUF_SIZE);
282

283
            /* set header fields needed for the mmap log mechanism */
284
            init_core->log_hdr_p->comp_type = DARSHAN_NO_COMP;
285
            init_core->log_hdr_p->name_map.off =
286
                ((char *)init_core->log_name_p - (char *)init_core->log_hdr_p);
287 288 289 290 291
#endif

            /* set known header fields for the log file */
            strcpy(init_core->log_hdr_p->version_string, DARSHAN_LOG_VERSION);
            init_core->log_hdr_p->magic_nr = DARSHAN_MAGIC_NR;
292

293 294 295 296 297
            /* set known job-level metadata fields for the log file */
            init_core->log_job_p->uid = getuid();
            init_core->log_job_p->start_time = time(NULL);
            init_core->log_job_p->nprocs = nprocs;
            init_core->log_job_p->jobid = (int64_t)jobid;
298 299 300 301 302 303

            /* if we are using any hints to write the log file, then record those
             * hints with the darshan job information
             */
            darshan_log_record_hints_and_ver(init_core);

304
            /* collect information about command line and mounted file systems */
305
            darshan_get_exe_and_mounts(init_core, argc, argv);
306

307 308 309 310
            /* if darshan was successfully initialized, set the global pointer
             * and bootstrap any modules with static initialization routines
             */
            DARSHAN_CORE_LOCK();
311
            darshan_core = init_core;
312 313 314 315 316 317 318 319
            DARSHAN_CORE_UNLOCK();

            i = 0;
            while(mod_static_init_fns[i])
            {
                (*mod_static_init_fns[i])();
                i++;
            }
320
        }
321 322
    }

323 324 325 326 327
    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);
328
        if(my_rank == 0)
329
        {
330 331
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
332 333 334 335 336 337
        }
    }

    return;
}

Shane Snyder's avatar
Shane Snyder committed
338
void darshan_core_shutdown()
339
{
340
    struct darshan_core_runtime *final_core;
341
    int internal_timing_flag = 0;
342 343
    struct tm *start_tm;
    time_t start_time_tmp;
344
    int64_t first_start_time;
345
    int64_t last_end_time;
346
    double start_log_time;
347
    double tm_end;
348 349 350
    double open1 = 0, open2 = 0;
    double job1 = 0, job2 = 0;
    double rec1 = 0, rec2 = 0;
351 352
    double mod1[DARSHAN_MAX_MODS] = {0};
    double mod2[DARSHAN_MAX_MODS] = {0};
353
    double header1 = 0, header2 = 0;
354 355 356 357 358 359 360 361 362 363 364 365
    char *logfile_name;
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
    darshan_record_id *shared_recs;
    darshan_record_id *mod_shared_recs;
    int shared_rec_cnt = 0;
    int ret = 0;
    int all_ret = 0;
    int i;
    uint64_t gz_fp = 0;
    MPI_File log_fh;
    MPI_Status status;
366 367 368 369

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

370 371
    /* synchronize before getting start time */
    DARSHAN_MPI_CALL(PMPI_Barrier)(MPI_COMM_WORLD);
Shane Snyder's avatar
Shane Snyder committed
372
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
373

Shane Snyder's avatar
Shane Snyder committed
374
    /* disable darhan-core while we shutdown */
375
    DARSHAN_CORE_LOCK();
376
    if(!darshan_core)
377
    {
378
        DARSHAN_CORE_UNLOCK();
379 380
        return;
    }
381 382
    final_core = darshan_core;
    darshan_core = NULL;
383 384
    DARSHAN_CORE_UNLOCK();

385
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
386 387 388 389 390
    /* remove the temporary mmap log files */
    /* NOTE: this unlink is not immediate as it must wait for the mapping
     * to no longer be referenced, which in our case happens when the
     * executable exits. If the application terminates mid-shutdown, then
     * there will be no mmap files and no final log file.
391
     */
392
    unlink(final_core->mmap_log_name);
393
#endif
Shane Snyder's avatar
Shane Snyder committed
394

395 396 397 398 399 400 401 402
    final_core->log_job_p->end_time = time(NULL);

    /* reduce to report first start and last end time across all ranks at rank 0 */
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job_p->start_time, &first_start_time,
        1, MPI_INT64_T, MPI_MIN, 0, MPI_COMM_WORLD);
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job_p->end_time, &last_end_time,
        1, MPI_INT64_T, MPI_MAX, 0, MPI_COMM_WORLD);
    if(my_rank == 0)
Shane Snyder's avatar
Shane Snyder committed
403
    {
404 405
        final_core->log_job_p->start_time = first_start_time;
        final_core->log_job_p->end_time = last_end_time;
Shane Snyder's avatar
Shane Snyder committed
406
    }
407

408 409 410 411 412 413
    final_core->comp_buf = malloc(DARSHAN_COMP_BUF_SIZE);
    if(!(final_core->comp_buf))
    {
        darshan_core_cleanup(final_core);
        return;
    }
414

415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442
    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
        darshan_core_cleanup(final_core);
        return;
    }

    /* set the log file name on rank 0 */
    if(my_rank == 0)
    {
        /* use human readable start time format in log filename */
        start_time_tmp = final_core->log_job_p->start_time;
        start_tm = localtime(&start_time_tmp);

        darshan_get_logfile_name(logfile_name, final_core->log_job_p->jobid, start_tm);
    }

    /* 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 */
        free(logfile_name);
        darshan_core_cleanup(final_core);
        return;
    }
443

444 445 446 447 448
    /* set which modules were registered locally, and call into
     * them to disable further instrumentation and to perform any
     * other pre-shutdown steps
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
449
    {
450 451
        if(final_core->mod_array[i])
            local_mod_use[i] = 1;
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
    }

    /* 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);

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

    if(internal_timing_flag)
        open1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
    /* collectively open the darshan log file */
    ret = darshan_log_open_all(logfile_name, &log_fh);
    if(internal_timing_flag)
        open2 = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* 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)
        {
            fprintf(stderr, "darshan library warning: unable to open log file %s\n",
                logfile_name);
        }
        free(logfile_name);
        darshan_core_cleanup(final_core);
        return;
    }

    if(internal_timing_flag)
        job1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
    /* rank 0 is responsible for writing the compressed darshan job information */
    if(my_rank == 0)
    {
488
        void *pointers[2] = {final_core->log_job_p, final_core->log_exemnt_p};
489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531
        int lengths[2] = {sizeof(struct darshan_job), strlen(final_core->log_exemnt_p)};
        int comp_buf_sz = 0;

        /* compress the job info and the trailing mount/exe data */
        all_ret = darshan_deflate_buffer(pointers, lengths, 2,
            final_core->comp_buf, &comp_buf_sz);
        if(all_ret)
        {
            fprintf(stderr, "darshan library warning: unable to compress job data\n");
            unlink(logfile_name);
        }
        else
        {
            /* write the job information, preallocing space for the log header */
            gz_fp += sizeof(struct darshan_header);
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, gz_fp,
                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);

            }
            gz_fp += comp_buf_sz;
        }
    }

    /* error out if unable to write job information */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&all_ret, 1, MPI_INT, 0, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        free(logfile_name);
        darshan_core_cleanup(final_core);
        return;
    }
    if(internal_timing_flag)
        job2 = DARSHAN_MPI_CALL(PMPI_Wtime)();

    if(internal_timing_flag)
        rec1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
    /* write the record name->id hash to the log file */
532
    final_core->log_hdr_p->name_map.off = gz_fp;
533
    ret = darshan_log_write_name_record_hash(log_fh, final_core, &gz_fp);
534
    final_core->log_hdr_p->name_map.len = gz_fp - final_core->log_hdr_p->name_map.off;
535

536
    /* error out if unable to write the name record hash */
537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561
    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 hash to log file %s\n",
                logfile_name);
            unlink(logfile_name);
        }
        free(logfile_name);
        darshan_core_cleanup(final_core);
        return;
    }
    if(internal_timing_flag)
        rec2 = DARSHAN_MPI_CALL(PMPI_Wtime)();

    mod_shared_recs = malloc(shared_rec_cnt * sizeof(darshan_record_id));
    assert(mod_shared_recs);

    /* loop over globally used darshan modules and:
     *      - get final output buffer
     *      - compress (zlib) provided output buffer
     *      - append compressed buffer to log file
562
     *      - add module map info (file offset/length) to log header
563 564 565 566 567
     *      - shutdown the module
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        struct darshan_core_module* this_mod = final_core->mod_array[i];
568
        struct darshan_core_name_record_ref *ref = NULL;
569 570 571 572 573 574 575 576 577
        int mod_shared_rec_cnt = 0;
        void* mod_buf = NULL;
        int mod_buf_sz = 0;
        int j;

        if(global_mod_use_count[i] == 0)
        {
            if(my_rank == 0)
            {
578 579
                final_core->log_hdr_p->mod_map[i].off = 0;
                final_core->log_hdr_p->mod_map[i].len = 0;
580 581 582 583 584 585 586
            }
            continue;
        }

        if(internal_timing_flag)
            mod1[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();

587
        /* set the shared record list for this module */
588 589
        for(j = 0; j < shared_rec_cnt; j++)
        {
590
            HASH_FIND(hlink, final_core->name_hash, &shared_recs[j],
591 592 593 594 595 596 597 598 599 600 601
                sizeof(darshan_record_id), ref);
            assert(ref);
            if(DARSHAN_MOD_FLAG_ISSET(ref->global_mod_flags, i))
            {
                mod_shared_recs[mod_shared_rec_cnt++] = shared_recs[j];
            }
        }

        /* if module is registered locally, get the corresponding output buffer
         * 
         * NOTE: this function can be used to run collective operations across
602
         * modules, if there are records shared globally.
603 604 605
         */
        if(this_mod)
        {
606 607
            mod_buf = final_core->mod_array[i]->rec_buf_start;
            mod_buf_sz = final_core->mod_array[i]->rec_buf_p - mod_buf;
608
            this_mod->mod_shutdown_func(MPI_COMM_WORLD, mod_shared_recs,
609 610 611 612
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
        }

        /* append this module's data to the darshan log */
613
        final_core->log_hdr_p->mod_map[i].off = gz_fp;
614
        ret = darshan_log_append_all(log_fh, final_core, mod_buf, mod_buf_sz, &gz_fp);
615 616
        final_core->log_hdr_p->mod_map[i].len =
            gz_fp - final_core->log_hdr_p->mod_map[i].off;
617

618 619 620 621 622 623
        /* XXX: DXT manages its own module memory buffers, so we need to
         * explicitly free them
         */
        if(i == DXT_POSIX_MOD || i == DXT_MPIIO_MOD)
            free(mod_buf);

624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646
        /* error out if the log append failed */
        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 %s module data to log file %s\n",
                    darshan_module_names[i], logfile_name);
                unlink(logfile_name);
            }
            free(logfile_name);
            darshan_core_cleanup(final_core);
            return;
        }

        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
    }

    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
647
    /* write out log header, after running 2 reductions on header variables:
648
     *  1) reduce 'partial_flag' variable to determine which modules ran out
649
     *     of memory for storing data
650 651 652 653 654 655
     *  2) reduce 'mod_ver' array to determine which log format version each
     *     module used for this output log
     */
    if(my_rank == 0)
    {
        /* rank 0 is responsible for writing the log header */
656 657 658 659 660 661 662 663
        final_core->log_hdr_p->comp_type = DARSHAN_ZLIB_COMP;

        DARSHAN_MPI_CALL(PMPI_Reduce)(
            MPI_IN_PLACE, &(final_core->log_hdr_p->partial_flag),
            1, MPI_UINT32_T, MPI_BOR, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(
            MPI_IN_PLACE, &(final_core->log_hdr_p->mod_ver),
            DARSHAN_MAX_MODS, MPI_UINT32_T, MPI_MAX, 0, MPI_COMM_WORLD);
664

665
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, final_core->log_hdr_p,
666 667 668 669 670 671 672 673
            sizeof(struct darshan_header), MPI_BYTE, &status);
        if(all_ret != MPI_SUCCESS)
        {
            fprintf(stderr, "darshan library warning: unable to write header to log file %s\n",
                    logfile_name);
            unlink(logfile_name);
        }
    }
674 675 676 677 678 679 680 681 682
    else
    {
        DARSHAN_MPI_CALL(PMPI_Reduce)(
            &(final_core->log_hdr_p->partial_flag), &(final_core->log_hdr_p->partial_flag),
            1, MPI_UINT32_T, MPI_BOR, 0, MPI_COMM_WORLD);
        DARSHAN_MPI_CALL(PMPI_Reduce)(
            &(final_core->log_hdr_p->mod_ver), &(final_core->log_hdr_p->mod_ver),
            DARSHAN_MAX_MODS, MPI_UINT32_T, MPI_MAX, 0, MPI_COMM_WORLD);
    }
683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702

    /* error out if unable to write log header */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&all_ret, 1, MPI_INT, 0, MPI_COMM_WORLD);
    if(all_ret != 0)
    {
        free(logfile_name);
        darshan_core_cleanup(final_core);
        return;
    }
    if(internal_timing_flag)
        header2 = DARSHAN_MPI_CALL(PMPI_Wtime)();

    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, which indicates that this log file is
     * complete and ready for analysis
     */
    if(my_rank == 0)
    {
703
        mode_t chmod_mode = S_IRUSR;
704
#ifdef __DARSHAN_GROUP_READABLE_LOGS
705
        chmod_mode |= S_IRGRP;
706
#endif
707 708 709 710

        if(getenv("DARSHAN_LOGFILE"))
        {
            chmod(logfile_name, chmod_mode);
711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727
        }
        else
        {
            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", (int)(end_log_time-start_log_time+1));
                rename(logfile_name, new_logfile_name);
                /* set permissions on log file */
728
                chmod(new_logfile_name, chmod_mode);
729 730 731 732 733 734
                free(new_logfile_name);
            }
        }
    }

    free(logfile_name);
735 736
    free(shared_recs);
    free(mod_shared_recs);
737
    darshan_core_cleanup(final_core);
738

739
    if(internal_timing_flag)
740
    {
741 742 743 744 745
        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];
746 747 748 749
        double all_tm, all_slowest;

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

750 751 752 753
        open_tm = open2 - open1;
        header_tm = header2 - header1;
        job_tm = job2 - job1;
        rec_tm = rec2 - rec1;
754
        all_tm = tm_end - start_log_time;
755 756 757 758
        for(i = 0;i < DARSHAN_MAX_MODS; i++)
        {
            mod_tm[i] = mod2[i] - mod1[i];
        }
759

760 761 762 763 764 765 766 767
        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);
768 769
        DARSHAN_MPI_CALL(PMPI_Reduce)(&all_tm, &all_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
770 771
        DARSHAN_MPI_CALL(PMPI_Reduce)(mod_tm, mod_slowest, DARSHAN_MAX_MODS,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
772 773 774

        if(my_rank == 0)
        {
775
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
776 777 778 779 780 781 782 783 784 785
            fprintf(stderr, "darshan:log_open\t%d\t%f\n", nprocs, open_slowest);
            fprintf(stderr, "darshan:job_write\t%d\t%f\n", nprocs, job_slowest);
            fprintf(stderr, "darshan:hash_write\t%d\t%f\n", nprocs, rec_slowest);
            fprintf(stderr, "darshan:header_write\t%d\t%f\n", nprocs, header_slowest);
            for(i = 0; i < DARSHAN_MAX_MODS; i++)
            {
                if(global_mod_use_count[i])
                    fprintf(stderr, "darshan:%s_shutdown\t%d\t%f\n", darshan_module_names[i],
                        nprocs, mod_slowest[i]);
            }
786
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
787
        }
788
    }
789

790 791
    return;
}
792

Shane Snyder's avatar
Shane Snyder committed
793
/* *********************************** */
794

795
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
796 797 798 799 800 801 802
static void *darshan_init_mmap_log(struct darshan_core_runtime* core, int jobid)
{
    int ret;
    int mmap_fd;
    int mmap_size;
    int sys_page_size;
    char cuser[L_cuserid] = {0};
803 804 805
    uint64_t hlevel;
    char hname[HOST_NAME_MAX];
    uint64_t logmod;
806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825
    char *envstr;
    char *mmap_log_path;
    void *mmap_p;

    sys_page_size = sysconf(_SC_PAGESIZE);
    assert(sys_page_size > 0);

    mmap_size = sizeof(struct darshan_header) + DARSHAN_JOB_RECORD_SIZE +
        + DARSHAN_NAME_RECORD_BUF_SIZE + darshan_mod_mem_quota;
    if(mmap_size % sys_page_size)
        mmap_size = ((mmap_size / sys_page_size) + 1) * sys_page_size;

    envstr = getenv(DARSHAN_MMAP_LOG_PATH_OVERRIDE);
    if(envstr)
        mmap_log_path = envstr;
    else
        mmap_log_path = DARSHAN_DEF_MMAP_LOG_PATH;

    darshan_get_user_name(cuser);

826 827 828 829 830 831 832 833 834 835 836 837 838
    /* generate a random number to help differentiate the temporary log */
    /* NOTE: job id is not sufficient for constructing a unique log file name,
     * since a job could be composed of multiple application runs, so we also
     * add a random number component to the log name
     */
    if(my_rank == 0)
    {
        hlevel=DARSHAN_MPI_CALL(PMPI_Wtime)() * 1000000;
        (void)gethostname(hname, sizeof(hname));
        logmod = darshan_hash((void*)hname,strlen(hname),hlevel);
    }
    DARSHAN_MPI_CALL(PMPI_Bcast)(&logmod, 1, MPI_UINT64_T, 0, MPI_COMM_WORLD);

839 840 841 842
    /* construct a unique temporary log file name for this process
     * to write mmap log data to
     */
    snprintf(core->mmap_log_name, PATH_MAX,
843 844
        "/%s/%s_%s_id%d_mmap-log-%" PRIu64 "-%d.darshan",
        mmap_log_path, cuser, __progname, jobid, logmod, my_rank);
845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884

    /* create the temporary mmapped darshan log */
    mmap_fd = open(core->mmap_log_name, O_CREAT|O_RDWR|O_EXCL , 0644);
    if(mmap_fd < 0)
    {
        fprintf(stderr, "darshan library warning: "
            "unable to create darshan log file %s\n", core->mmap_log_name);
        return(NULL);
    }

    /* TODO: ftruncate or just zero fill? */
    /* allocate the necessary space in the log file */
    ret = ftruncate(mmap_fd, mmap_size);
    if(ret < 0)
    {
        fprintf(stderr, "darshan library warning: "
            "unable to allocate darshan log file %s\n", core->mmap_log_name);
        close(mmap_fd);
        unlink(core->mmap_log_name);
        return(NULL);
    }

    /* create the memory map for darshan's data structures so they are
     * persisted to file as the application executes
     */
    mmap_p = mmap(NULL, mmap_size, PROT_WRITE, MAP_SHARED, mmap_fd, 0);
    if(mmap_p == MAP_FAILED)
    {
        fprintf(stderr, "darshan library warning: "
            "unable to mmap darshan log file %s\n", core->mmap_log_name);
        close(mmap_fd);
        unlink(core->mmap_log_name);
        return(NULL);
    }

    /* close darshan log file (this does *not* unmap the log file) */
    close(mmap_fd);

    return(mmap_p);
}
885
#endif
886

887
/* record any hints used to write the darshan log in the job data */
888
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
889 890
{
    char* hints;
891
    char* job_hints;
892 893 894 895 896 897
    int meta_remain = 0;
    char* m;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
898
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
899 900
    if(!hints)
    {
901
        hints = __DARSHAN_LOG_HINTS;
902 903 904 905 906
    }

    if(!hints || strlen(hints) < 1)
        return;

907 908
    job_hints = strdup(hints);
    if(!job_hints)
909 910 911
        return;

    meta_remain = DARSHAN_JOB_METADATA_LEN -
912
        strlen(core->log_job_p->metadata) - 1;
913 914
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
915
        sprintf(core->log_job_p->metadata, "lib_ver=%s\n", PACKAGE_VERSION);
916 917
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
918
    if(meta_remain >= (3 + strlen(job_hints)))
919
    {
920
        m = core->log_job_p->metadata + strlen(core->log_job_p->metadata);
921
        /* We have room to store the hints in the metadata portion of
922
         * the job structure.  We just prepend an h= to the hints list.  The
923 924 925
         * metadata parser will ignore = characters that appear in the value
         * portion of the metadata key/value pair.
         */
926
        sprintf(m, "h=%s\n", job_hints);
927
    }
928
    free(job_hints);
929 930 931 932

    return;
}

933 934 935 936 937 938 939 940 941 942 943 944 945 946
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 */
947
static void add_entry(char* buf, int* space_left, struct mntent* entry)
948
{
949
    int i;
950 951 952 953
    int ret;
    char tmp_mnt[256];
    struct statfs statfsbuf;

954 955 956 957 958 959 960 961 962 963
    /* avoid adding the same mount points multiple times -- to limit
     * storage space and potential statfs, ioctl, etc calls
     */
    for(i = 0; i < mnt_data_count; i++)
    {
        if((strncmp(mnt_data_array[i].path, entry->mnt_dir, DARSHAN_MAX_MNT_PATH) == 0) &&
           (strncmp(mnt_data_array[i].type, entry->mnt_type, DARSHAN_MAX_MNT_PATH) == 0))
            return;
    }

964 965 966 967 968 969 970 971 972 973 974 975
    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);
976
    mnt_data_array[mnt_data_count].fs_info.fs_type = statfsbuf.f_type;
977
    if(ret == 0 && statfsbuf.f_type != LL_SUPER_MAGIC)
978
        mnt_data_array[mnt_data_count].fs_info.block_size = statfsbuf.f_bsize;
979
    else if(ret == 0 && statfsbuf.f_type == LL_SUPER_MAGIC)
980
        mnt_data_array[mnt_data_count].fs_info.block_size = 1024*1024;
981
    else
982
        mnt_data_array[mnt_data_count].fs_info.block_size = 4096;
983

984
#ifdef DARSHAN_LUSTRE
985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003
    /* attempt to retrieve OST and MDS counts from Lustre */
    mnt_data_array[mnt_data_count].fs_info.ost_count = -1;
    mnt_data_array[mnt_data_count].fs_info.mdt_count = -1;
    if ( statfsbuf.f_type == LL_SUPER_MAGIC )
    {
        int n_ost, n_mdt;
        int ret_ost, ret_mdt;
        DIR *mount_dir;

        mount_dir = opendir( entry->mnt_dir );
        if ( mount_dir  ) 
        {
            /* n_ost and n_mdt are used for both input and output to ioctl */
            n_ost = 0;
            n_mdt = 1;

            ret_ost = ioctl( dirfd(mount_dir), LL_IOC_GETOBDCOUNT, &n_ost );
            ret_mdt = ioctl( dirfd(mount_dir), LL_IOC_GETOBDCOUNT, &n_mdt );

1004
            if ( !(ret_ost < 0 || ret_mdt < 0) )
1005 1006 1007 1008 1009 1010 1011
            {
                mnt_data_array[mnt_data_count].fs_info.ost_count = n_ost;
                mnt_data_array[mnt_data_count].fs_info.mdt_count = n_mdt;
            }
            closedir( mount_dir );
        }
    }
1012
#endif
1013

1014
    /* store mount information with the job-level metadata in darshan log */
1015
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
1016 1017 1018
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
1019
        strcat(buf, tmp_mnt);
1020 1021 1022 1023 1024 1025 1026
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

1027
/* darshan_get_exe_and_mounts()
1028 1029
 *
 * collects command line and list of mounted file systems into a string that
1030
 * will be stored with the job-level metadata
1031
 */
1032
static void darshan_get_exe_and_mounts(struct darshan_core_runtime *core,
1033
    int argc, char **argv)
1034 1035 1036 1037
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
1038 1039 1040
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int space_left = DARSHAN_EXE_LEN;
1041 1042 1043
    FILE *fh;
    int i, ii;
    char cmdl[DARSHAN_EXE_LEN];
1044 1045