darshan-core.c 63.4 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 <pwd.h>
29
#include <zlib.h>
30
#include <mpi.h>
31
#include <assert.h>
32

33
#include "uthash.h"
Shane Snyder's avatar
Shane Snyder committed
34
#include "darshan.h"
35
#include "darshan-core.h"
Shane Snyder's avatar
Shane Snyder committed
36
#include "darshan-dynamic.h"
37

38
#ifdef DARSHAN_LUSTRE
39
#include <lustre/lustre_user.h>
40
#endif
41

42
extern char* __progname;
43
extern char* __progname_full;
44

45
/* internal variable delcarations */
46
static struct darshan_core_runtime *darshan_core = NULL;
47
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
48
static int my_rank = -1;
49
static int nprocs = -1;
50
static int darshan_mem_alignment = 1;
51
static long darshan_mod_mem_quota = DARSHAN_MOD_MEM_MAX;
52

53 54 55 56 57 58 59 60 61 62 63 64
/* paths prefixed with the following directories are not traced by darshan */
char* darshan_path_exclusions[] = {
"/etc/",
"/dev/",
"/usr/",
"/bin/",
"/boot/",
"/lib/",
"/opt/",
"/sbin/",
"/sys/",
"/proc/",
65
"/var/",
66 67 68
NULL
};

69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
#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
};

84 85 86 87 88 89 90
#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
91 92 93
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

94 95 96 97 98 99 100 101
/* 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];
102
    struct darshan_fs_info fs_info;
103 104 105 106
};
static struct mnt_data mnt_data_array[DARSHAN_MAX_MNTS];
static int mnt_data_count = 0;

107
/* prototypes for internal helper functions */
108
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
109 110
static void *darshan_init_mmap_log(
    struct darshan_core_runtime* core, int jobid);
111
#endif
112
static void darshan_log_record_hints_and_ver(
113
    struct darshan_core_runtime* core);
114 115
static void darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core, int argc, char **argv);
116 117
static void darshan_fs_info_from_path(
    const char *path, struct darshan_fs_info *fs_info);
118
static int darshan_add_name_record_ref(
119
    struct darshan_core_runtime *core, darshan_record_id rec_id,
120
    const char *name, darshan_module_id mod_id);
121
static void darshan_get_username(
122
    char *user);
123 124
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
125
static void darshan_get_shared_records(
126 127
    struct darshan_core_runtime *core, darshan_record_id **shared_recs,
    int *shared_rec_cnt);
128
static int darshan_log_open_all(
129
    char *logfile_name, MPI_File *log_fh);
130
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
131 132
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
133
static int darshan_log_write_name_record_hash(
134
    MPI_File log_fh, struct darshan_core_runtime *core,
135 136 137
    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
138
    int count, uint64_t *inout_off);
Shane Snyder's avatar
Shane Snyder committed
139 140
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
141

142 143
/* *********************************** */

Shane Snyder's avatar
Shane Snyder committed
144
void darshan_core_initialize(int argc, char **argv)
145
{
146
    struct darshan_core_runtime *init_core = NULL;
147 148
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
149
    char *envstr;
150 151
    char *jobid_str;
    int jobid;
152 153
    int ret;
    int tmpval;
154
    int i;
155 156

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
157
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
158 159 160 161

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

162
    if(internal_timing_flag)
163 164 165
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
166
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
167
    {
168
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
169 170
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
171
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
172 173 174 175 176 177 178 179 180 181 182
        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
        {
183
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
184 185 186
        }

        /* avoid floating point errors on faulty input */
187
        if(darshan_mem_alignment < 1)
188 189 190
        {
            darshan_mem_alignment = 1;
        }
191

192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211
        /* 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();
        }

212 213 214 215 216 217 218 219 220 221 222 223
        /* set the memory quota for darshan modules' records */
        envstr = getenv(DARSHAN_MOD_MEM_OVERRIDE);
        if(envstr)
        {
            ret = sscanf(envstr, "%d", &tmpval);
            /* silently ignore if the env variable is set poorly */
            if(ret == 1 && tmpval > 0)
            {
                darshan_mod_mem_quota = tmpval * 1024 * 1024; /* convert from MiB */
            }
        }

224 225 226
        /* allocate structure to track darshan core runtime information */
        init_core = malloc(sizeof(*init_core));
        if(init_core)
227
        {
228 229 230
            memset(init_core, 0, sizeof(*init_core));
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

231 232
        /* TODO: do we alloc new memory as we go or just do everything up front? */

233 234 235 236 237
#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);
238
            init_core->log_name_p = malloc(DARSHAN_NAME_RECORD_BUF_SIZE);
239
            init_core->log_mod_p = malloc(darshan_mod_mem_quota);
240 241

            if(!(init_core->log_hdr_p) || !(init_core->log_job_p) ||
242
               !(init_core->log_exemnt_p) || !(init_core->log_name_p) ||
243 244 245 246 247 248 249 250 251
               !(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);
252
            memset(init_core->log_name_p, 0, DARSHAN_NAME_RECORD_BUF_SIZE);
253
            memset(init_core->log_mod_p, 0, darshan_mod_mem_quota);
254 255 256 257
#else
            /* if mmap logs are enabled, we need to initialize the mmap region
             * before setting the corresponding log file region pointers
             */
258 259
            void *mmap_p = darshan_init_mmap_log(init_core, jobid);
            if(!mmap_p)
260
            {
261 262
                free(init_core);
                return;
263 264
            }

265
            /* set the memory pointers for each log file region */
266
            init_core->log_hdr_p = (struct darshan_header *)mmap_p;
267
            init_core->log_job_p = (struct darshan_job *)
268
                ((char *)init_core->log_hdr_p + sizeof(struct darshan_header));
269
            init_core->log_exemnt_p = (char *)
270
                ((char *)init_core->log_job_p + sizeof(struct darshan_job));
271
            init_core->log_name_p = (void *)
272 273
                ((char *)init_core->log_exemnt_p + DARSHAN_EXE_LEN + 1);
            init_core->log_mod_p = (void *)
274
                ((char *)init_core->log_name_p + DARSHAN_NAME_RECORD_BUF_SIZE);
275

276
            /* set header fields needed for the mmap log mechanism */
277
            init_core->log_hdr_p->comp_type = DARSHAN_NO_COMP;
278
            init_core->log_hdr_p->name_map.off =
279
                ((char *)init_core->log_name_p - (char *)init_core->log_hdr_p);
280 281 282 283 284
#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;
285

286 287 288 289 290
            /* 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;
291 292 293 294 295 296

            /* 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);

297
            /* collect information about command line and mounted file systems */
298
            darshan_get_exe_and_mounts(init_core, argc, argv);
299

300 301 302 303
            /* if darshan was successfully initialized, set the global pointer
             * and bootstrap any modules with static initialization routines
             */
            DARSHAN_CORE_LOCK();
304
            darshan_core = init_core;
305 306 307 308 309 310 311 312
            DARSHAN_CORE_UNLOCK();

            i = 0;
            while(mod_static_init_fns[i])
            {
                (*mod_static_init_fns[i])();
                i++;
            }
313
        }
314 315
    }

316 317 318 319 320
    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);
321
        if(my_rank == 0)
322
        {
323 324
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
325 326 327 328 329 330
        }
    }

    return;
}

Shane Snyder's avatar
Shane Snyder committed
331
void darshan_core_shutdown()
332
{
333
    struct darshan_core_runtime *final_core;
334
    int internal_timing_flag = 0;
335 336
    struct tm *start_tm;
    time_t start_time_tmp;
337
    int64_t first_start_time;
338
    int64_t last_end_time;
339
    double start_log_time;
340
    double tm_end;
341 342 343
    double open1 = 0, open2 = 0;
    double job1 = 0, job2 = 0;
    double rec1 = 0, rec2 = 0;
344 345
    double mod1[DARSHAN_MAX_MODS] = {0};
    double mod2[DARSHAN_MAX_MODS] = {0};
346
    double header1 = 0, header2 = 0;
347 348 349 350 351 352 353 354 355 356 357 358
    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;
359 360 361 362

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

363 364
    /* synchronize before getting start time */
    DARSHAN_MPI_CALL(PMPI_Barrier)(MPI_COMM_WORLD);
Shane Snyder's avatar
Shane Snyder committed
365
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
366

Shane Snyder's avatar
Shane Snyder committed
367
    /* disable darhan-core while we shutdown */
368
    DARSHAN_CORE_LOCK();
369
    if(!darshan_core)
370
    {
371
        DARSHAN_CORE_UNLOCK();
372 373
        return;
    }
374 375
    final_core = darshan_core;
    darshan_core = NULL;
376 377
    DARSHAN_CORE_UNLOCK();

378
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
379 380 381 382 383
    /* 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.
384
     */
385
    unlink(final_core->mmap_log_name);
386
#endif
Shane Snyder's avatar
Shane Snyder committed
387

388 389 390 391 392 393 394 395
    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
396
    {
397 398
        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
399
    }
400

401 402 403 404 405 406
    final_core->comp_buf = malloc(DARSHAN_COMP_BUF_SIZE);
    if(!(final_core->comp_buf))
    {
        darshan_core_cleanup(final_core);
        return;
    }
407

408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435
    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;
    }
436

437 438 439 440 441
    /* 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++)
442
    {
443 444
        if(final_core->mod_array[i])
            local_mod_use[i] = 1;
445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480
    }

    /* 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)
    {
481
        void *pointers[2] = {final_core->log_job_p, final_core->log_exemnt_p};
482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524
        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 */
525
    final_core->log_hdr_p->name_map.off = gz_fp;
526
    ret = darshan_log_write_name_record_hash(log_fh, final_core, &gz_fp);
527
    final_core->log_hdr_p->name_map.len = gz_fp - final_core->log_hdr_p->name_map.off;
528

529
    /* error out if unable to write the name record hash */
530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554
    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
555
     *      - add module map info (file offset/length) to log header
556 557 558 559 560
     *      - shutdown the module
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        struct darshan_core_module* this_mod = final_core->mod_array[i];
561
        struct darshan_core_name_record_ref *ref = NULL;
562 563 564 565 566 567 568 569 570
        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)
            {
571 572
                final_core->log_hdr_p->mod_map[i].off = 0;
                final_core->log_hdr_p->mod_map[i].len = 0;
573 574 575 576 577 578 579
            }
            continue;
        }

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

580
        /* set the shared record list for this module */
581 582
        for(j = 0; j < shared_rec_cnt; j++)
        {
583
            HASH_FIND(hlink, final_core->name_hash, &shared_recs[j],
584 585 586 587 588 589 590 591 592 593 594
                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
595
         * modules, if there are records shared globally.
596 597 598
         */
        if(this_mod)
        {
599 600
            mod_buf = final_core->mod_array[i]->rec_buf_start;
            mod_buf_sz = final_core->mod_array[i]->rec_buf_p - mod_buf;
601
            this_mod->mod_shutdown_func(MPI_COMM_WORLD, mod_shared_recs,
602 603 604 605
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
        }

        /* append this module's data to the darshan log */
606
        final_core->log_hdr_p->mod_map[i].off = gz_fp;
607
        ret = darshan_log_append_all(log_fh, final_core, mod_buf, mod_buf_sz, &gz_fp);
608 609
        final_core->log_hdr_p->mod_map[i].len =
            gz_fp - final_core->log_hdr_p->mod_map[i].off;
610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633

        /* 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)();
634
    /* write out log header, after running 2 reductions on header variables:
635
     *  1) reduce 'partial_flag' variable to determine which modules ran out
636
     *     of memory for storing data
637 638 639 640 641 642
     *  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 */
643 644 645 646 647 648 649 650
        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);
651

652
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, final_core->log_hdr_p,
653 654 655 656 657 658 659 660
            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);
        }
    }
661 662 663 664 665 666 667 668 669
    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);
    }
670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689

    /* 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)
    {
690
        mode_t chmod_mode = S_IRUSR;
691
#ifdef __DARSHAN_GROUP_READABLE_LOGS
692
        chmod_mode |= S_IRGRP;
693
#endif
694 695 696 697

        if(getenv("DARSHAN_LOGFILE"))
        {
            chmod(logfile_name, chmod_mode);
698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714
        }
        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 */
715
                chmod(new_logfile_name, chmod_mode);
716 717 718 719 720 721
                free(new_logfile_name);
            }
        }
    }

    free(logfile_name);
722 723
    free(shared_recs);
    free(mod_shared_recs);
724
    darshan_core_cleanup(final_core);
725

726
    if(internal_timing_flag)
727
    {
728 729 730 731 732
        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];
733 734 735 736
        double all_tm, all_slowest;

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

737 738 739 740
        open_tm = open2 - open1;
        header_tm = header2 - header1;
        job_tm = job2 - job1;
        rec_tm = rec2 - rec1;
741
        all_tm = tm_end - start_log_time;
742 743 744 745
        for(i = 0;i < DARSHAN_MAX_MODS; i++)
        {
            mod_tm[i] = mod2[i] - mod1[i];
        }
746

747 748 749 750 751 752 753 754
        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);
755 756
        DARSHAN_MPI_CALL(PMPI_Reduce)(&all_tm, &all_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
757 758
        DARSHAN_MPI_CALL(PMPI_Reduce)(mod_tm, mod_slowest, DARSHAN_MAX_MODS,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
759 760 761

        if(my_rank == 0)
        {
762
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
763 764 765 766 767 768 769 770 771 772
            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]);
            }
773
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
774
        }
775
    }
776

777 778
    return;
}
779

Shane Snyder's avatar
Shane Snyder committed
780
/* *********************************** */
781

782
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
783 784 785 786 787 788
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;
789
    char username[DARSHAN_USERNAME_LEN_MAX] = {0};
790 791 792
    uint64_t hlevel;
    char hname[HOST_NAME_MAX];
    uint64_t logmod;
793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810
    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;

811
    darshan_get_username(username);
812

813 814 815 816 817 818 819 820 821 822 823 824 825
    /* 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);

826 827 828 829
    /* construct a unique temporary log file name for this process
     * to write mmap log data to
     */
    snprintf(core->mmap_log_name, PATH_MAX,
830
        "/%s/%s_%s_id%d_mmap-log-%" PRIu64 "-%d.darshan",
831
        mmap_log_path, username, __progname, jobid, logmod, my_rank);
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

    /* 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);
}
872
#endif
873

874
/* record any hints used to write the darshan log in the job data */
875
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
876 877
{
    char* hints;
878
    char* job_hints;
879 880 881 882 883 884
    int meta_remain = 0;
    char* m;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
885
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
886 887
    if(!hints)
    {
888
        hints = __DARSHAN_LOG_HINTS;
889 890 891 892 893
    }

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

894 895
    job_hints = strdup(hints);
    if(!job_hints)
896 897 898
        return;

    meta_remain = DARSHAN_JOB_METADATA_LEN -
899
        strlen(core->log_job_p->metadata) - 1;
900 901
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
902
        sprintf(core->log_job_p->metadata, "lib_ver=%s\n", PACKAGE_VERSION);
903 904
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
905
    if(meta_remain >= (3 + strlen(job_hints)))
906
    {
907
        m = core->log_job_p->metadata + strlen(core->log_job_p->metadata);
908
        /* We have room to store the hints in the metadata portion of
909
         * the job structure.  We just prepend an h= to the hints list.  The
910 911 912
         * metadata parser will ignore = characters that appear in the value
         * portion of the metadata key/value pair.
         */
913
        sprintf(m, "h=%s\n", job_hints);
914
    }
915
    free(job_hints);
916 917 918 919

    return;
}

920 921 922 923 924 925 926 927 928 929 930 931 932 933
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 */
934
static void add_entry(char* buf, int* space_left, struct mntent* entry)
935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951
{
    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);
952
    mnt_data_array[mnt_data_count].fs_info.fs_type = statfsbuf.f_type;
953
    if(ret == 0 && statfsbuf.f_type != LL_SUPER_MAGIC)
954
        mnt_data_array[mnt_data_count].fs_info.block_size = statfsbuf.f_bsize;
955
    else if(ret == 0 && statfsbuf.f_type == LL_SUPER_MAGIC)
956
        mnt_data_array[mnt_data_count].fs_info.block_size = 1024*1024;
957
    else
958
        mnt_data_array[mnt_data_count].fs_info.block_size = 4096;
959

960
#ifdef DARSHAN_LUSTRE
961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979
    /* 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 );

980
            if ( !(ret_ost < 0 || ret_mdt < 0) )
981 982 983 984 985 986 987
            {
                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 );
        }
    }
988
#endif
989

990
    /* store mount information with the job-level metadata in darshan log */
991
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
992 993 994
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
995
        strcat(buf, tmp_mnt);
996 997 998 999 1000 1001 1002
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

1003
/* darshan_get_exe_and_mounts()
1004 1005
 *
 * collects command line and list of mounted file systems into a string that
1006
 * will be stored with the job-level metadata
1007
 */
1008
static void darshan_get_exe_and_mounts(struct darshan_core_runtime *core,
1009
    int argc, char **argv)
1010 1011 1012 1013
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
1014 1015 1016 1017
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int space_left = DARSHAN_EXE_LEN;
    int i;
1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
    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
    };

1039 1040 1041
    /* record exe and arguments */
    for(i=0; i<argc; i++)
    {
1042 1043
        strncat(core->log_exemnt_p, argv[i], space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
1044 1045
        if(i < (argc-1))
        {
1046 1047
            strncat(core->log_exemnt_p, " ", space_left);
            space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
1048 1049 1050 1051 1052 1053 1054 1055
        }
    }

    /* if we don't see any arguments, then use glibc symbol to get
     * program name at least (this happens in fortran)
     */
    if(argc == 0)
    {
1056 1057 1058 1059
        strncat(core->log_exemnt_p, __progname_full, space_left);
        space_left = DARSHAN_EXE_LEN-