darshan-core.c 59.9 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 <sys/types.h>
#include <sys/stat.h>
24
#include <sys/mman.h>
25
#include <sys/vfs.h>
26
#include <zlib.h>
27
#include <mpi.h>
28
#include <assert.h>
29

30
#include "uthash.h"
Shane Snyder's avatar
Shane Snyder committed
31
#include "darshan.h"
32
#include "darshan-core.h"
Shane Snyder's avatar
Shane Snyder committed
33
#include "darshan-dynamic.h"
34

35
extern char* __progname;
36
extern char* __progname_full;
37

38
/* internal variable delcarations */
39
static struct darshan_core_runtime *darshan_core = NULL;
40
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
41
static int my_rank = -1;
42
static int nprocs = -1;
43
static int darshan_mem_alignment = 1;
44
static long darshan_mod_mem_quota = DARSHAN_MOD_MEM_MAX;
45

46 47 48 49 50 51 52 53 54 55 56 57
/* 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/",
58
"/var/",
59 60 61
NULL
};

62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
#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
};

Shane Snyder's avatar
Shane Snyder committed
77 78 79
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

80 81 82 83 84 85
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
86
    int block_size;
87 88 89 90 91 92
    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;

93
/* prototypes for internal helper functions */
94
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
95 96
static void *darshan_init_mmap_log(
    struct darshan_core_runtime* core, int jobid);
97
#endif
98
static void darshan_log_record_hints_and_ver(
99
    struct darshan_core_runtime* core);
100 101
static void darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core, int argc, char **argv);
102
static int darshan_add_name_record_ref(
103
    struct darshan_core_runtime *core, darshan_record_id rec_id,
104
    const char *name, darshan_module_id mod_id);
105 106
static int darshan_block_size_from_path(
    const char *path);
107 108
static void darshan_get_user_name(
    char *user);
109 110
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
111
static void darshan_get_shared_records(
112 113
    struct darshan_core_runtime *core, darshan_record_id **shared_recs,
    int *shared_rec_cnt);
114
static int darshan_log_open_all(
115
    char *logfile_name, MPI_File *log_fh);
116
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
117 118
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
119
static int darshan_log_write_name_record_hash(
120
    MPI_File log_fh, struct darshan_core_runtime *core,
121 122 123
    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
124
    int count, uint64_t *inout_off);
Shane Snyder's avatar
Shane Snyder committed
125 126
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
127

128 129
/* *********************************** */

Shane Snyder's avatar
Shane Snyder committed
130
void darshan_core_initialize(int argc, char **argv)
131
{
132
    struct darshan_core_runtime *init_core = NULL;
133 134
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
135
    char *envstr;
136 137
    char *jobid_str;
    int jobid;
138 139
    int ret;
    int tmpval;
140
    int i;
141 142

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
143
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
144 145 146 147

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

148
    if(internal_timing_flag)
149 150 151
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
152
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
153
    {
154
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
155 156
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
157
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
158 159 160 161 162 163 164 165 166 167 168
        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
        {
169
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
170 171 172
        }

        /* avoid floating point errors on faulty input */
173
        if(darshan_mem_alignment < 1)
174 175 176
        {
            darshan_mem_alignment = 1;
        }
177

178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197
        /* 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();
        }

198 199 200
        /* allocate structure to track darshan core runtime information */
        init_core = malloc(sizeof(*init_core));
        if(init_core)
201
        {
202 203 204
            memset(init_core, 0, sizeof(*init_core));
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

205 206
        /* TODO: do we alloc new memory as we go or just do everything up front? */

207 208 209 210 211
#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);
212
            init_core->log_name_p = malloc(DARSHAN_NAME_RECORD_BUF_SIZE);
213
            init_core->log_mod_p = malloc(darshan_mod_mem_quota);
214 215

            if(!(init_core->log_hdr_p) || !(init_core->log_job_p) ||
216
               !(init_core->log_exemnt_p) || !(init_core->log_name_p) ||
217 218 219 220 221 222 223 224 225
               !(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);
226
            memset(init_core->log_name_p, 0, DARSHAN_NAME_RECORD_BUF_SIZE);
227
            memset(init_core->log_mod_p, 0, darshan_mod_mem_quota);
228 229 230 231
#else
            /* if mmap logs are enabled, we need to initialize the mmap region
             * before setting the corresponding log file region pointers
             */
232 233
            void *mmap_p = darshan_init_mmap_log(init_core, jobid);
            if(!mmap_p)
234
            {
235 236
                free(init_core);
                return;
237 238
            }

239
            /* set the memory pointers for each log file region */
240
            init_core->log_hdr_p = (struct darshan_header *)mmap_p;
241
            init_core->log_job_p = (struct darshan_job *)
242
                ((char *)init_core->log_hdr_p + sizeof(struct darshan_header));
243
            init_core->log_exemnt_p = (char *)
244
                ((char *)init_core->log_job_p + sizeof(struct darshan_job));
245
            init_core->log_name_p = (void *)
246 247
                ((char *)init_core->log_exemnt_p + DARSHAN_EXE_LEN + 1);
            init_core->log_mod_p = (void *)
248
                ((char *)init_core->log_name_p + DARSHAN_NAME_RECORD_BUF_SIZE);
249

250
            /* set header fields needed for the mmap log mechanism */
251
            init_core->log_hdr_p->comp_type = DARSHAN_NO_COMP;
252
            init_core->log_hdr_p->name_map.off =
253
                ((char *)init_core->log_name_p - (char *)init_core->log_hdr_p);
254 255 256 257 258
#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;
259

260 261 262 263 264
            /* 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;
265 266 267 268 269 270

            /* 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);

271
            /* collect information about command line and mounted file systems */
272
            darshan_get_exe_and_mounts(init_core, argc, argv);
273

274
            darshan_core = init_core;
275
        }
276 277
    }

278 279 280 281 282
    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);
283
        if(my_rank == 0)
284
        {
285 286
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
287 288 289
        }
    }

290 291 292
    /* if darshan was successfully initialized, set the global pointer and
     * bootstrap any modules with static initialization routines
     */
293
    if(init_core)
294
    {
295 296
        darshan_core = init_core;

297 298 299 300 301 302 303 304
        i = 0;
        while(mod_static_init_fns[i])
        {
            (*mod_static_init_fns[i])();
            i++;
        }
    }

305 306 307
    return;
}

Shane Snyder's avatar
Shane Snyder committed
308
void darshan_core_shutdown()
309
{
310
    struct darshan_core_runtime *final_core;
311
    int internal_timing_flag = 0;
312 313
    struct tm *start_tm;
    time_t start_time_tmp;
314
    int64_t first_start_time;
315
    int64_t last_end_time;
316
    double start_log_time;
317
    double tm_end;
318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
    double open1, open2;
    double job1, job2;
    double rec1, rec2;
    double mod1[DARSHAN_MAX_MODS] = {0};
    double mod2[DARSHAN_MAX_MODS] = {0};
    double header1, header2;
    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;
336 337 338 339

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

Shane Snyder's avatar
Shane Snyder committed
340
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
341

Shane Snyder's avatar
Shane Snyder committed
342
    /* disable darhan-core while we shutdown */
343
    DARSHAN_CORE_LOCK();
344
    if(!darshan_core)
345
    {
346
        DARSHAN_CORE_UNLOCK();
347 348
        return;
    }
349 350
    final_core = darshan_core;
    darshan_core = NULL;
351 352
    DARSHAN_CORE_UNLOCK();

353
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
354 355 356 357 358
    /* 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.
359
     */
360
    unlink(final_core->mmap_log_name);
361
#endif
Shane Snyder's avatar
Shane Snyder committed
362

363 364 365 366 367 368 369 370
    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
371
    {
372 373
        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
374
    }
375

376 377 378 379 380 381
    final_core->comp_buf = malloc(DARSHAN_COMP_BUF_SIZE);
    if(!(final_core->comp_buf))
    {
        darshan_core_cleanup(final_core);
        return;
    }
382

383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410
    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;
    }
411

412 413 414 415 416
    /* 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++)
417
    {
418 419
        if(final_core->mod_array[i])
            local_mod_use[i] = 1;
420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455
    }

    /* 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)
    {
456
        void *pointers[2] = {final_core->log_job_p, final_core->log_exemnt_p};
457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499
        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 */
500
    final_core->log_hdr_p->name_map.off = gz_fp;
501
    ret = darshan_log_write_name_record_hash(log_fh, final_core, &gz_fp);
502
    final_core->log_hdr_p->name_map.len = gz_fp - final_core->log_hdr_p->name_map.off;
503

504
    /* error out if unable to write the name record hash */
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
    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
530
     *      - add module map info (file offset/length) to log header
531 532 533 534 535
     *      - shutdown the module
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        struct darshan_core_module* this_mod = final_core->mod_array[i];
536
        struct darshan_core_name_record_ref *ref = NULL;
537 538 539 540 541 542 543 544 545
        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)
            {
546 547
                final_core->log_hdr_p->mod_map[i].off = 0;
                final_core->log_hdr_p->mod_map[i].len = 0;
548 549 550 551 552 553 554
            }
            continue;
        }

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

555
        /* set the shared record list for this module */
556 557
        for(j = 0; j < shared_rec_cnt; j++)
        {
558
            HASH_FIND(hlink, final_core->name_hash, &shared_recs[j],
559 560 561 562 563 564 565 566 567 568 569
                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
570
         * modules, if there are records shared globally.
571 572 573
         */
        if(this_mod)
        {
574 575
            mod_buf = final_core->mod_array[i]->rec_buf_start;
            mod_buf_sz = final_core->mod_array[i]->rec_buf_p - mod_buf;
576
            this_mod->mod_shutdown_func(MPI_COMM_WORLD, mod_shared_recs,
577 578 579 580
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
        }

        /* append this module's data to the darshan log */
581
        final_core->log_hdr_p->mod_map[i].off = gz_fp;
582
        ret = darshan_log_append_all(log_fh, final_core, mod_buf, mod_buf_sz, &gz_fp);
583 584
        final_core->log_hdr_p->mod_map[i].len =
            gz_fp - final_core->log_hdr_p->mod_map[i].off;
585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608

        /* 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)();
609
    /* write out log header, after running 2 reductions on header variables:
610
     *  1) reduce 'partial_flag' variable to determine which modules ran out
611
     *     of memory for storing data
612 613 614 615 616 617
     *  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 */
618 619 620 621 622 623 624 625
        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);
626

627
        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, final_core->log_hdr_p,
628 629 630 631 632 633 634 635
            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);
        }
    }
636 637 638 639 640 641 642 643 644
    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);
    }
645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664

    /* 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)
    {
665
        mode_t chmod_mode = S_IRUSR;
666
#ifdef __DARSHAN_GROUP_READABLE_LOGS
667
        chmod_mode |= S_IRGRP;
668
#endif
669 670 671 672

        if(getenv("DARSHAN_LOGFILE"))
        {
            chmod(logfile_name, chmod_mode);
673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689
        }
        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 */
690
                chmod(new_logfile_name, chmod_mode);
691 692 693 694 695 696
                free(new_logfile_name);
            }
        }
    }

    free(logfile_name);
697 698
    free(shared_recs);
    free(mod_shared_recs);
699
    darshan_core_cleanup(final_core);
700

701
    if(internal_timing_flag)
702
    {
703 704 705 706 707
        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];
708 709 710 711
        double all_tm, all_slowest;

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

712 713 714 715
        open_tm = open2 - open1;
        header_tm = header2 - header1;
        job_tm = job2 - job1;
        rec_tm = rec2 - rec1;
716
        all_tm = tm_end - start_log_time;
717 718 719 720
        for(i = 0;i < DARSHAN_MAX_MODS; i++)
        {
            mod_tm[i] = mod2[i] - mod1[i];
        }
721

722 723 724 725 726 727 728 729
        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);
730 731
        DARSHAN_MPI_CALL(PMPI_Reduce)(&all_tm, &all_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
732 733
        DARSHAN_MPI_CALL(PMPI_Reduce)(mod_tm, mod_slowest, DARSHAN_MAX_MODS,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
734 735 736

        if(my_rank == 0)
        {
737
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
738 739 740 741 742 743 744 745 746 747
            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]);
            }
748
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
749
        }
750
    }
751

752 753
    return;
}
754

Shane Snyder's avatar
Shane Snyder committed
755
/* *********************************** */
756

757
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
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
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};
    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);

    /* construct a unique temporary log file name for this process
     * to write mmap log data to
     */
    snprintf(core->mmap_log_name, PATH_MAX,
        "/%s/%s_%s_id%d_mmap-log-%d.darshan",
        mmap_log_path, cuser, __progname, jobid, my_rank);

    /* 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);
}
831
#endif
832

833
/* record any hints used to write the darshan log in the job data */
834
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
835 836
{
    char* hints;
837
    char* job_hints;
838 839 840 841 842 843
    int meta_remain = 0;
    char* m;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
844
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
845 846
    if(!hints)
    {
847
        hints = __DARSHAN_LOG_HINTS;
848 849 850 851 852
    }

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

853 854
    job_hints = strdup(hints);
    if(!job_hints)
855 856 857
        return;

    meta_remain = DARSHAN_JOB_METADATA_LEN -
858
        strlen(core->log_job_p->metadata) - 1;
859 860
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
861
        sprintf(core->log_job_p->metadata, "lib_ver=%s\n", PACKAGE_VERSION);
862 863
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
864
    if(meta_remain >= (3 + strlen(job_hints)))
865
    {
866
        m = core->log_job_p->metadata + strlen(core->log_job_p->metadata);
867
        /* We have room to store the hints in the metadata portion of
868
         * the job structure.  We just prepend an h= to the hints list.  The
869 870 871
         * metadata parser will ignore = characters that appear in the value
         * portion of the metadata key/value pair.
         */
872
        sprintf(m, "h=%s\n", job_hints);
873
    }
874
    free(job_hints);
875 876 877 878

    return;
}

879 880 881 882 883 884 885 886 887 888 889 890 891 892
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 */
893
static void add_entry(char* buf, int* space_left, struct mntent* entry)
894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917
{
    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;

918
    /* store mount information with the job-level metadata in darshan log */
919
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
920 921 922
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
923
        strcat(buf, tmp_mnt);
924 925 926 927 928 929 930
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

931
/* darshan_get_exe_and_mounts()
932 933
 *
 * collects command line and list of mounted file systems into a string that
934
 * will be stored with the job-level metadata
935
 */
936
static void darshan_get_exe_and_mounts(struct darshan_core_runtime *core,
937
    int argc, char **argv)
938 939 940 941
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
942 943 944 945
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int space_left = DARSHAN_EXE_LEN;
    int i;
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966
    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
    };

967 968 969
    /* record exe and arguments */
    for(i=0; i<argc; i++)
    {
970 971
        strncat(core->log_exemnt_p, argv[i], space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
972 973
        if(i < (argc-1))
        {
974 975
            strncat(core->log_exemnt_p, " ", space_left);
            space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
976 977 978 979 980 981 982 983
        }
    }

    /* if we don't see any arguments, then use glibc symbol to get
     * program name at least (this happens in fortran)
     */
    if(argc == 0)
    {
984 985 986 987
        strncat(core->log_exemnt_p, __progname_full, space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
        strncat(core->log_exemnt_p, " <unknown args>", space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
988 989 990 991 992 993
    }

    if(space_left == 0)
    {
        /* we ran out of room; mark that string was truncated */
        truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
994
        sprintf(&(core->log_exemnt_p[truncate_offset]), "%s",
995 996 997
            truncate_string);
    }

998 999 1000 1001
    /* 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
     */
1002
    mnt_data_count = 0;
1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025

    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;

1026
        add_entry(core->log_exemnt_p, &space_left, entry);
1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
    }
    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;

1039
        add_entry(core->log_exemnt_p, &space_left, entry);
1040 1041 1042
    }
    endmntent(tab);

1043
    /* sort mount points in order of longest path to shortest path.  This is
1044 1045 1046 1047 1048 1049 1050
     * 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;
}

1051
static int darshan_block_size_from_path(const char *path)
1052
{
1053
    int i;
1054
    int block_size = -1;
1055

1056
    for(i=0; i<mnt_data_count; i++)
1057
    {
1058
        if(!(strncmp(mnt_data_array[i].path, path, strlen(mnt_data_array[i].path))))
1059
        {
1060 1061
            block_size = mnt_data_array[i].block_size;
            break;;
1062 1063 1064
        }
    }

1065
    return block_size;
1066 1067
}

1068
static void darshan_get_user_name(char *cuser)
1069
{
1070
    char* logname_string;
1071