darshan-core.c 56 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"
31
#include "darshan.h"
32
#include "darshan-core.h"
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

45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
/* 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/",
NULL
};

60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
#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
75 76 77
#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

78 79 80 81 82 83
/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
84
    int block_size;
85 86 87 88 89 90
    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;

91 92
/* prototypes for internal helper functions */
static void darshan_log_record_hints_and_ver(
93
    struct darshan_core_runtime* core);
94 95
static void darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core, int argc, char **argv);
96 97 98
static void darshan_add_record_hashref(
    struct darshan_core_runtime *core, char *name,
    darshan_record_id id, struct darshan_core_record_ref **ref);
99 100
static int darshan_block_size_from_path(
    const char *path);
101 102
static void darshan_get_user_name(
    char *user);
103 104
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
105
static void darshan_get_shared_records(
106 107
    struct darshan_core_runtime *core, darshan_record_id **shared_recs,
    int *shared_rec_cnt);
108
static int darshan_log_open_all(
109
    char *logfile_name, MPI_File *log_fh);
110
static int darshan_deflate_buffer(
Shane Snyder's avatar
Shane Snyder committed
111 112
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
113
static int darshan_log_write_record_hash(
114
    MPI_File log_fh, struct darshan_core_runtime *core,
115 116 117
    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
118
    int count, uint64_t *inout_off);
119 120
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
121

122 123
/* *********************************** */

124
void darshan_core_initialize(int argc, char **argv)
125
{
126
    struct darshan_core_runtime *init_core = NULL;
127 128
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
129
    char *envstr;
130 131
    char *jobid_str;
    int jobid;
132 133
    int ret;
    int tmpval;
134
    int i;
135 136 137 138 139 140 141
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
    int mmap_fd;
    int mmap_size;
    int sys_page_size;
    char cuser[L_cuserid] = {0};
    char *mmap_log_dir;
#endif
142 143

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
144
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);
145 146 147 148

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

149
    if(internal_timing_flag)
150 151 152
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
153
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
154
    {
155
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
156 157
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
158
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
159 160 161 162 163 164 165 166 167 168 169
        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
        {
170
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
171 172 173
        }

        /* avoid floating point errors on faulty input */
174
        if(darshan_mem_alignment < 1)
175 176 177
        {
            darshan_mem_alignment = 1;
        }
178

179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198
        /* 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();
        }

199 200 201
        /* allocate structure to track darshan core runtime information */
        init_core = malloc(sizeof(*init_core));
        if(init_core)
202
        {
203 204 205
            memset(init_core, 0, sizeof(*init_core));
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
#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);
            init_core->log_rec_p = malloc(DARSHAN_RECORD_BUF_SIZE);
            init_core->log_mod_p = malloc(DARSHAN_MOD_MEM_MAX);

            if(!(init_core->log_hdr_p) || !(init_core->log_job_p) ||
               !(init_core->log_exemnt_p) || !(init_core->log_rec_p) ||
               !(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);
            memset(init_core->log_rec_p, 0, DARSHAN_RECORD_BUF_SIZE);
            memset(init_core->log_mod_p, 0, DARSHAN_MOD_MEM_MAX);
#else
            /* if mmap logs are enabled, we need to initialize the mmap region
             * before setting the corresponding log file region pointers
             */

232 233
            sys_page_size = sysconf(_SC_PAGESIZE);
            assert(sys_page_size > 0);
234

235
            mmap_size = sizeof(struct darshan_header) + DARSHAN_JOB_RECORD_SIZE +
236
                + DARSHAN_RECORD_BUF_SIZE + DARSHAN_MOD_MEM_MAX;
237 238
            if(mmap_size % sys_page_size)
                mmap_size = ((mmap_size / sys_page_size) + 1) * sys_page_size;
239

240 241 242 243 244 245 246 247
            envstr = getenv(DARSHAN_MMAP_LOG_PATH_OVERRIDE);
            if(envstr)
                mmap_log_dir = envstr;
            else
                mmap_log_dir = DARSHAN_DEF_MMAP_LOG_PATH;

            darshan_get_user_name(cuser);

248 249 250
            /* construct a unique temporary log file name for this process
             * to write mmap log data to
             */
251 252 253
            snprintf(init_core->mmap_log_name, PATH_MAX,
                "/%s/%s_%s_id%d_mmap-log-%d.darshan",
                mmap_log_dir, cuser, __progname, jobid, my_rank);
254 255

            /* create the temporary mmapped darshan log */
256
            mmap_fd = open(init_core->mmap_log_name, O_CREAT|O_RDWR|O_EXCL , 0644);
257
            if(mmap_fd < 0)
258
            {
259
                fprintf(stderr, "darshan library warning: "
260
                    "unable to create darshan log file %s\n", init_core->mmap_log_name);
261 262
                free(init_core);
                return;
263 264
            }

265
            /* TODO: what's more expensive? truncate or write zeros? perf test this call and later accesses */
266 267 268 269 270
            /* allocate the necessary space in the log file */
            ret = ftruncate(mmap_fd, mmap_size);
            if(ret < 0)
            {
                fprintf(stderr, "darshan library warning: "
271
                    "unable to allocate darshan log file %s\n", init_core->mmap_log_name);
272 273
                free(init_core);
                close(mmap_fd);
274
                unlink(init_core->mmap_log_name);
275 276 277 278 279
                return;
            }

            /* memory map buffers for getting at least some summary i/o data
             * into a log file if darshan does not shut down properly
280
             */
281 282
            void *mmap_p = mmap(NULL, mmap_size, PROT_WRITE, MAP_SHARED, mmap_fd, 0);
            if(mmap_p == MAP_FAILED)
283
            {
284
                fprintf(stderr, "darshan library warning: "
285
                    "unable to mmap darshan log file %s\n", init_core->mmap_log_name);
286 287
                free(init_core);
                close(mmap_fd);
288
                unlink(init_core->mmap_log_name);
289
                return;
290 291
            }

292 293 294
            /* close darshan log file (this does *not* unmap the log file) */
            close(mmap_fd);

295
            /* set the memory pointers for each log file region */
296
            init_core->log_hdr_p = (struct darshan_header *)mmap_p;
297
            init_core->log_job_p = (struct darshan_job *)
298
                ((char *)init_core->log_hdr_p + sizeof(struct darshan_header));
299
            init_core->log_exemnt_p = (char *)
300 301 302 303 304
                ((char *)init_core->log_job_p + sizeof(struct darshan_job));
            init_core->log_rec_p = (void *)
                ((char *)init_core->log_exemnt_p + DARSHAN_EXE_LEN + 1);
            init_core->log_mod_p = (void *)
                ((char *)init_core->log_rec_p + DARSHAN_RECORD_BUF_SIZE);
305

306
            /* set header fields needed for the mmap log mechanism */
307
            init_core->log_hdr_p->comp_type = DARSHAN_NO_COMP;
308 309
            init_core->log_hdr_p->rec_map.off =
                sizeof(struct darshan_header) + DARSHAN_JOB_RECORD_SIZE;
310 311 312 313 314
#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;
315

316 317 318 319 320
            /* 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;
321 322 323 324 325 326

            /* 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);

327
            /* collect information about command line and mounted file systems */
328
            darshan_get_exe_and_mounts(init_core, argc, argv);
329

330
            /* bootstrap any modules with static initialization routines */
331 332 333 334 335 336 337 338
            i = 0;
            while(mod_static_init_fns[i])
            {
                (*mod_static_init_fns[i])();
                i++;
            }

            darshan_core = init_core;
339
        }
340 341
    }

342 343 344 345
    /* reduce so all ranks agree on the job start time */
    DARSHAN_MPI_CALL(PMPI_Allreduce)(MPI_IN_PLACE, &init_core->log_job_p->start_time,
        1, MPI_INT64_T, MPI_MIN, MPI_COMM_WORLD);

346 347 348 349 350
    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);
351
        if(my_rank == 0)
352
        {
353 354
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
355 356 357 358 359 360
        }
    }

    return;
}

361
void darshan_core_shutdown()
362
{
363
    struct darshan_core_runtime *final_core;
364 365
    struct darshan_header out_header;
    struct darshan_job out_job;
366
    int internal_timing_flag = 0;
367 368 369
    struct tm *start_tm;
    time_t start_time_tmp;
    int64_t last_end_time;
370
    double start_log_time;
371
    double tm_end;
372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
    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;
390 391 392 393

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

394
    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
395

Shane Snyder's avatar
Shane Snyder committed
396
    /* disable darhan-core while we shutdown */
397
    DARSHAN_CORE_LOCK();
398
    if(!darshan_core)
399
    {
400
        DARSHAN_CORE_UNLOCK();
401 402
        return;
    }
403 404
    final_core = darshan_core;
    darshan_core = NULL;
405 406
    DARSHAN_CORE_UNLOCK();

407 408 409 410
    memcpy(&out_job, final_core->log_job_p, sizeof(struct darshan_job));

#ifdef __DARSHAN_ENABLE_MMAP_LOGS
    /* TODO: can we get rid of out_ header and job?? */
411 412
    /* XXX just copy mmap files somewhere else to avoid corruption */
    DARSHAN_MPI_CALL(PMPI_Barrier)(MPI_COMM_WORLD);
413 414 415
    char cp_cmd[500] = {0};
    sprintf(cp_cmd, "cp %s ~/Desktop", final_core->mmap_log_name);
    system(cp_cmd);
416 417 418 419 420 421 422 423 424 425
    DARSHAN_MPI_CALL(PMPI_Barrier)(MPI_COMM_WORLD);

    /* indicate in the metadata field of the temporary darshan log file that
     * the darshan shutdown process was invoked on the data in the log. since
     * we have no way of knowing how far the shutdown process got, the data
     * could potentially be in inconsistent or corrupt state. in this case,
     * darshan currently does not generate a log file at all.
     */
    char *m = final_core->log_job_p->metadata + strlen(final_core->log_job_p->metadata);
    int meta_remain = DARSHAN_JOB_METADATA_LEN - strlen(final_core->log_job_p->metadata) - 1;
426
    /* TODO: do we ever check for darshan_shutdown? */
427
    snprintf(m, meta_remain, "darshan_shutdown=yes\n");
428
#endif
Shane Snyder's avatar
Shane Snyder committed
429

430
    /* we also need to set which modules were registered on this process and
431 432
     * call into those modules and give them a chance to perform any necessary
     * pre-shutdown steps.
Shane Snyder's avatar
Shane Snyder committed
433 434 435 436 437
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
438
            local_mod_use[i] = 1;
439
            final_core->mod_array[i]->mod_funcs.begin_shutdown();
Shane Snyder's avatar
Shane Snyder committed
440 441
        }
    }
442

443 444 445 446 447 448
    final_core->comp_buf = malloc(DARSHAN_COMP_BUF_SIZE);
    if(!(final_core->comp_buf))
    {
        darshan_core_cleanup(final_core);
        return;
    }
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
    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;
    }
478

479
    final_core->log_job_p->end_time = time(NULL);
480

481
    /* reduce to report last end time across all ranks at rank 0 */
482
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job_p->end_time, &last_end_time,
483
        1, MPI_INT64_T, MPI_MAX, 0, MPI_COMM_WORLD);
484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687
    if(my_rank == 0)
    {
        out_job.end_time = last_end_time;
    }

    /* 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)
    {
        void *pointers[2] = {&out_job, final_core->log_exemnt_p};
        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 */
    out_header.rec_map.off = gz_fp;
    ret = darshan_log_write_record_hash(log_fh, final_core, &gz_fp);
    out_header.rec_map.len = gz_fp - out_header.rec_map.off;

    /* error out if unable to write record hash */
    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:
     *      - perform shared file reductions, if possible
     *      - get final output buffer
     *      - compress (zlib) provided output buffer
     *      - append compressed buffer to log file
     *      - add module index info (file offset/length) to log header
     *      - shutdown the module
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        struct darshan_core_module* this_mod = final_core->mod_array[i];
        struct darshan_core_record_ref *ref = NULL;
        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)
            {
                out_header.mod_map[i].off = 0;
                out_header.mod_map[i].len = 0;
            }
            continue;
        }

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

        /* set the shared file list for this module */
        memset(mod_shared_recs, 0, shared_rec_cnt * sizeof(darshan_record_id));
        for(j = 0; j < shared_rec_cnt; j++)
        {
            HASH_FIND(hlink, final_core->rec_hash, &shared_recs[j],
                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
         * modules, if there are file records shared globally.
         */
        if(this_mod)
        {
            this_mod->mod_funcs.get_output_data(MPI_COMM_WORLD, mod_shared_recs,
                mod_shared_rec_cnt, &mod_buf, &mod_buf_sz);
        }

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

        /* 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;
        }

        /* shutdown module if registered locally */
        if(this_mod)
        {
            this_mod->mod_funcs.shutdown();
        }
        if(internal_timing_flag)
            mod2[i] = DARSHAN_MPI_CALL(PMPI_Wtime)();
    }

    if(internal_timing_flag)
        header1 = DARSHAN_MPI_CALL(PMPI_Wtime)();
    /* write out log header, after running 2 reduction on header variables:
     *  1) reduce 'partial_flag' variable to determine which modules ran out
     *     of memory for storing I/O data
     *  2) reduce 'mod_ver' array to determine which log format version each
     *     module used for this output log
     */
    DARSHAN_MPI_CALL(PMPI_Reduce)(&(final_core->log_hdr_p->partial_flag),
        &(out_header.partial_flag), 1, MPI_UINT32_T, MPI_BOR, 0, MPI_COMM_WORLD);
688 689 690
    DARSHAN_MPI_CALL(PMPI_Reduce)(&(final_core->log_hdr_p->mod_ver),
        &(out_header.mod_ver), DARSHAN_MAX_MODS, MPI_UINT32_T, MPI_MAX,
        0, MPI_COMM_WORLD);
691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727
    if(my_rank == 0)
    {
        /* rank 0 is responsible for writing the log header */
        /* initialize the remaining header fields */
        strcpy(out_header.version_string, DARSHAN_LOG_VERSION);
        out_header.magic_nr = DARSHAN_MAGIC_NR;
        out_header.comp_type = DARSHAN_ZLIB_COMP;

        all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(log_fh, 0, &out_header,
            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);
        }
    }

    /* 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)
    {
728
        mode_t chmod_mode = S_IRUSR;
729
#ifdef __DARSHAN_GROUP_READABLE_LOGS
730
        chmod_mode |= S_IRGRP;
731
#endif
732 733 734 735

        if(getenv("DARSHAN_LOGFILE"))
        {
            chmod(logfile_name, chmod_mode);
736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752
        }
        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 */
753
                chmod(new_logfile_name, chmod_mode);
754 755 756 757 758
                free(new_logfile_name);
            }
        }
    }

759
#ifdef __DARSHAN_ENABLE_MMAP_LOGS
760 761
    /* remove the temporary mmap log files */
    unlink(final_core->mmap_log_name);
762
#endif
763 764

    free(logfile_name);
765 766
    free(shared_recs);
    free(mod_shared_recs);
767
    darshan_core_cleanup(final_core);
768

769
    if(internal_timing_flag)
770
    {
771 772 773 774 775
        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];
776 777 778 779
        double all_tm, all_slowest;

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

780 781 782 783
        open_tm = open2 - open1;
        header_tm = header2 - header1;
        job_tm = job2 - job1;
        rec_tm = rec2 - rec1;
784
        all_tm = tm_end - start_log_time;
785 786 787 788
        for(i = 0;i < DARSHAN_MAX_MODS; i++)
        {
            mod_tm[i] = mod2[i] - mod1[i];
        }
789

790 791 792 793 794 795 796 797
        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);
798 799
        DARSHAN_MPI_CALL(PMPI_Reduce)(&all_tm, &all_slowest, 1,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
800 801
        DARSHAN_MPI_CALL(PMPI_Reduce)(mod_tm, mod_slowest, DARSHAN_MAX_MODS,
            MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
802 803 804

        if(my_rank == 0)
        {
805
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
806 807 808 809 810 811 812 813 814 815
            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]);
            }
816
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
817
        }
818
    }
819

820 821
    return;
}
822

823
/* *********************************** */
824

825
/* record any hints used to write the darshan log in the job data */
826
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
827 828
{
    char* hints;
829
    char* job_hints;
830 831 832 833 834 835
    int meta_remain = 0;
    char* m;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
836
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
837 838
    if(!hints)
    {
839
        hints = __DARSHAN_LOG_HINTS;
840 841 842 843 844
    }

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

845 846
    job_hints = strdup(hints);
    if(!job_hints)
847 848 849
        return;

    meta_remain = DARSHAN_JOB_METADATA_LEN -
850
        strlen(core->log_job_p->metadata) - 1;
851 852
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
853
        sprintf(core->log_job_p->metadata, "lib_ver=%s\n", PACKAGE_VERSION);
854 855
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
856
    if(meta_remain >= (3 + strlen(job_hints)))
857
    {
858
        m = core->log_job_p->metadata + strlen(core->log_job_p->metadata);
859
        /* We have room to store the hints in the metadata portion of
860
         * the job structure.  We just prepend an h= to the hints list.  The
861 862 863
         * metadata parser will ignore = characters that appear in the value
         * portion of the metadata key/value pair.
         */
864
        sprintf(m, "h=%s\n", job_hints);
865
    }
866
    free(job_hints);
867 868 869 870

    return;
}

871 872 873 874 875 876 877 878 879 880 881 882 883 884
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 */
885
static void add_entry(char* buf, int* space_left, struct mntent *entry)
886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909
{
    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;

910
    /* store mount information with the job-level metadata in darshan log */
911
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
912 913 914
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
915
        strcat(buf, tmp_mnt);
916 917 918 919 920 921 922
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

923
/* darshan_get_exe_and_mounts()
924 925
 *
 * collects command line and list of mounted file systems into a string that
926
 * will be stored with the job-level metadata
927
 */
928
static void darshan_get_exe_and_mounts(struct darshan_core_runtime *core,
929
    int argc, char **argv)
930 931 932 933
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
934 935 936 937
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int space_left = DARSHAN_EXE_LEN;
    int i;
938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958
    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
    };

959 960 961
    /* record exe and arguments */
    for(i=0; i<argc; i++)
    {
962 963
        strncat(core->log_exemnt_p, argv[i], space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
964 965
        if(i < (argc-1))
        {
966 967
            strncat(core->log_exemnt_p, " ", space_left);
            space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
968 969 970 971 972 973 974 975
        }
    }

    /* if we don't see any arguments, then use glibc symbol to get
     * program name at least (this happens in fortran)
     */
    if(argc == 0)
    {
976 977 978 979
        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);
980 981 982 983 984 985
    }

    if(space_left == 0)
    {
        /* we ran out of room; mark that string was truncated */
        truncate_offset = DARSHAN_EXE_LEN - strlen(truncate_string);
986
        sprintf(&(core->log_exemnt_p[truncate_offset]), "%s",
987 988 989
            truncate_string);
    }

990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016
    /* we make two passes through mounted file systems; in the first pass we
     * grab any non-nfs mount points, then on the second pass we grab nfs
     * mount points
     */

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

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

1017
        add_entry(core->log_exemnt_p, &space_left, entry);
1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029
    }
    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;

1030
        add_entry(core->log_exemnt_p, &space_left, entry);
1031 1032 1033
    }
    endmntent(tab);

1034
    /* sort mount points in order of longest path to shortest path.  This is
1035 1036 1037 1038 1039 1040 1041
     * 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;
}

1042
static int darshan_block_size_from_path(const char *path)
1043
{
1044
    int i;
1045
    int block_size = -1;
1046

1047
    for(i=0; i<mnt_data_count; i++)
1048
    {
1049
        if(!(strncmp(mnt_data_array[i].path, path, strlen(mnt_data_array[i].path))))
1050
        {
1051 1052
            block_size = mnt_data_array[i].block_size;
            break;;
1053 1054 1055
        }
    }

1056
    return block_size;
1057 1058
}

1059
static void darshan_get_user_name(char *cuser)
1060
{
1061
    char* logname_string;
1062

1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079
    /* get the username for this job.  In order we will try each of the
     * following until one of them succeeds:
     *
     * - cuserid()
     * - getenv("LOGNAME")
     * - snprintf(..., geteuid());
     *
     * Note that we do not use getpwuid() because it generally will not
     * work in statically compiled binaries.
     */

#ifndef __DARSHAN_DISABLE_CUSERID
    cuserid(cuser);
#endif

    /* if cuserid() didn't work, then check the environment */
    if(strcmp(cuser, "") == 0)
1080
    {
1081 1082
        logname_string = getenv("LOGNAME");
        if(logname_string)
1083
        {
1084
            strncpy(cuser, logname_string, (L_cuserid-1));
1085 1086 1087
        }
    }

1088 1089 1090 1091 1092 1093 1094
    /* if cuserid() and environment both fail, then fall back to uid */
    if(strcmp(cuser, "") == 0)
    {
        uid_t uid = geteuid();
        snprintf(cuser, L_cuserid, "%u", uid);
    }

1095 1096 1097
    return;
}

1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140
/* construct the darshan log file name */
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
{
    char* user_logfile_name;
    char* logpath;
    char* logpath_override = NULL;
#ifdef __DARSHAN_LOG_ENV
    char env_check[256];
    char* env_tok;
#endif
    uint64_t hlevel;
    char hname[HOST_NAME_MAX];
    uint64_t logmod;
    char cuser[L_cuserid] = {0};
    int ret;

    /* first, check if user specifies a complete logpath to use */
    user_logfile_name = getenv("DARSHAN_LOGFILE");
    if(user_logfile_name)
    {
        if(strlen(user_logfile_name) >= (PATH_MAX-1))
        {
            fprintf(stderr, "darshan library warning: user log file name too long.\n");
            logfile_name[0] = '\0';
        }
        else
        {
            strcpy(logfile_name, user_logfile_name);
        }
    }
    else
    {
        /* otherwise, generate the log path automatically */

        /* Use DARSHAN_LOG_PATH_OVERRIDE for the value or __DARSHAN_LOG_PATH */
        logpath = getenv(DARSHAN_LOG_PATH_OVERRIDE);
        if(!logpath)
        {
#ifdef __DARSHAN_LOG_PATH
            logpath = __DARSHAN_LOG_PATH;
#endif
        }

1141
        darshan_get_user_name(cuser);
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221

        /* generate a random number to help differentiate the log */
        hlevel=DARSHAN_MPI_CALL(PMPI_Wtime)() * 1000000;
        (void)gethostname(hname, sizeof(hname));
        logmod = darshan_hash((void*)hname,strlen(hname),hlevel);

        /* see if darshan was configured using the --with-logpath-by-env
         * argument, which allows the user to specify an absolute path to
         * place logs via an env variable.
         */
#ifdef __DARSHAN_LOG_ENV
        /* just silently skip if the environment variable list is too big */
        if(strlen(__DARSHAN_LOG_ENV) < 256)
        {
            /* copy env variable list to a temporary buffer */
            strcpy(env_check, __DARSHAN_LOG_ENV);
            /* tokenize the comma-separated list */
            env_tok = strtok(env_check, ",");
            if(env_tok)
            {
                do
                {
                    /* check each env variable in order */
                    logpath_override = getenv(env_tok);
                    if(logpath_override)
                    {
                        /* stop as soon as we find a match */
                        break;
                    }
                }while((env_tok = strtok(NULL, ",")));
            }
        }
#endif

        if(logpath_override)
        {
            ret = snprintf(logfile_name, PATH_MAX,
                "%s/%s_%s_id%d_%d-%d-%d-%" PRIu64 ".darshan_partial",
                logpath_override,
                cuser, __progname, jobid,
                (start_tm->tm_mon+1),
                start_tm->tm_mday,
                (start_tm->tm_hour*60*60 + start_tm->tm_min*60 + start_tm->tm_sec),
                logmod);
            if(ret == (PATH_MAX-1))
            {
                /* file name was too big; squish it down */
                snprintf(logfile_name, PATH_MAX,
                    "%s/id%d.darshan_partial",
                    logpath_override, jobid);
            }
        }
        else if(logpath)
        {
            ret = snprintf(logfile_name, PATH_MAX,
                "%s/%d/%d/%d/%s_%s_id%d_%d-%d-%d-%" PRIu64 ".darshan_partial",
                logpath, (start_tm->tm_year+1900),
                (start_tm->tm_mon+1), start_tm->tm_mday,
                cuser, __progname, jobid,
                (start_tm->tm_mon+1),
                start_tm->tm_mday,
                (start_tm->tm_hour*60*60 + start_tm->tm_min*60 + start_tm->tm_sec),
                logmod);
            if(ret == (PATH_MAX-1))
            {
                /* file name was too big; squish it down */
                snprintf(logfile_name, PATH_MAX,
                    "%s/id%d.darshan_partial",
                    logpath, jobid);
            }
        }
        else
        {
            logfile_name[0] = '\0';
        }
    }

    return;
}

1222 1223 1224 1225
static void darshan_add_record_hashref(struct darshan_core_runtime *core,
    char *name, darshan_record_id id, struct darshan_core_record_ref **ref)
{
    int record_size = sizeof(darshan_record_id) + strlen(name) + 1;
1226 1227
    darshan_record_id *id_p;
    char *name_p;
1228 1229 1230 1231 1232 1233 1234 1235 1236

    if((record_size + core->log_hdr_p->rec_map.len) > DARSHAN_RECORD_BUF_SIZE)
        return;

    *ref = malloc(sizeof(**ref));
    if(*ref)
    {
        memset(*ref, 0, sizeof(**ref));

1237 1238 1239 1240 1241 1242
        /* serialize the record id and name into the record map buffer */
        id_p = (darshan_record_id *)
            ((char *)core->log_rec_p + core->log_hdr_p->rec_map.len);
        *id_p = id;
        name_p = (char *)id_p + sizeof(darshan_record_id);
        strcpy(name_p, name);
1243

1244 1245
        /* save pointer to this record mapping buffer */
        (*ref)->rec_p = id_p;
1246

1247 1248
        /* add the record to the hash table */
        HASH_ADD_KEYPTR(hlink, core->rec_hash, id_p, sizeof(darshan_record_id), (*ref));
1249 1250 1251 1252 1253 1254 1255
        core->rec_hash_cnt++;
        core->log_hdr_p->rec_map.len += record_size;
    }

    return;
}

1256
static void darshan_get_shared_records(struct darshan_core_runtime *core,
1257
    darshan_record_id **shared_recs, int *shared_rec_cnt)
1258
{
1259
    int i, j;
1260
    int tmp_cnt = core->rec_hash_cnt;
1261
    struct darshan_core_record_ref *tmp, *ref;
1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278
    darshan_record_id *id_array;
    uint64_t *mod_flags;
    uint64_t *global_mod_flags;

    /* broadcast root's number of records to all other processes */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&tmp_cnt, 1, MPI_INT, 0, MPI_COMM_WORLD);

    /* use root record count to allocate data structures */
    id_array = malloc(tmp_cnt * sizeof(darshan_record_id));
    mod_flags = malloc(tmp_cnt * sizeof(uint64_t));
    global_mod_flags = malloc(tmp_cnt * sizeof(uint64_t));
    *shared_recs = malloc(tmp_cnt * sizeof(darshan_record_id));
    assert(id_array && mod_flags && global_mod_flags && *shared_recs);

    memset(mod_flags, 0, tmp_cnt * sizeof(uint64_t));
    memset(global_mod_flags, 0, tmp_cnt * sizeof(uint64_t));
    memset(*shared_recs, 0, tmp_cnt * sizeof(darshan_record_id));
1279 1280 1281 1282

    /* first, determine list of records root process has opened */
    if(my_rank == 0)
    {
1283
        i = 0;
1284
        HASH_ITER(hlink, core->rec_hash, ref, tmp)
1285
        {
1286 1287
            /* dereference the record pointer to get corresponding id */
            id_array[i++] = *(darshan_record_id *)ref->rec_p;
1288 1289 1290 1291
        }
    }

    /* broadcast root's list of records to all other processes */
1292
    DARSHAN_MPI_CALL(PMPI_Bcast)(id_array, (tmp_cnt * sizeof(darshan_record_id)),
1293 1294 1295
        MPI_BYTE, 0, MPI_COMM_WORLD);

    /* everyone looks to see if they opened the same records as root */
1296
    for(i=0; i<tmp_cnt; i++)
1297
    {
1298 1299
        HASH_FIND(hlink, core->rec_hash, &id_array[i], sizeof(darshan_record_id), ref);
        if(ref)
1300
        {
1301 1302
            /* we opened that record too, save the mod_flags */
            mod_flags[i] = ref->mod_flags;
1303 1304 1305
        }
    }

1306 1307 1308
    /* now allreduce so everyone agrees which files are shared and
     * which modules accessed them collectively
     */
1309 1310
    DARSHAN_MPI_CALL(PMPI_Allreduce)(mod_flags, global_mod_flags, tmp_cnt,
        MPI_UINT64_T, MPI_BAND, MPI_COMM_WORLD);
1311

1312 1313
    j = 0;
    for(i=0; i<tmp_cnt; i++)
1314
    {
1315
        if(global_mod_flags[i] != 0)
1316
        {
1317
            (*shared_recs)[j++] = id_array[i];
1318 1319 1320 1321 1322 1323 1324 1325

            /* set global_mod_flags so we know which modules collectively
             * accessed this module. we need this info to support shared
             * file reductions
             */
            HASH_FIND(hlink, core->rec_hash, &id_array[i], sizeof(darshan_record_id), ref);
            assert(ref);
            ref->global_mod_flags = global_mod_flags[i];
1326 1327
        }
    }
1328
    *shared_rec_cnt = j;
1329

1330 1331 1332
    free(id_array);
    free(mod_flags);
    free(global_mod_flags);
1333 1334 1335
    return;
}

1336
static int darshan_log_open_all(char *logfile_name, MPI_File *log_fh)
1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351
{
    char *hints;
    char *tok_str;
    char *orig_tok_str;
    char *key;
    char *value;
    char *saveptr = NULL;
    int ret;
    MPI_Info info;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
    MPI_Info_create(&info);

1352
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
1353 1354
    if(!hints)
    {
1355
        hints = __DARSHAN_LOG_HINTS;
1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389
    }

    if(hints && strlen(hints) > 0)
    {
        tok_str = strdup(hints);
        if(tok_str)
        {
            orig_tok_str = tok_str;
            do
            {
                /* split string on semicolon */
                key = strtok_r(tok_str, ";", &saveptr);
                if(key)
                {
                    tok_str = NULL;
                    /* look for = sign splitting key/value pairs */
                    value = index(key, '=');
                    if(value)
                    {
                        /* break key and value into separate null terminated strings */
                        value[0] = '\0';
                        value++;
                        if(strlen(key) > 0)
                            MPI_Info_set(info, key, value);
                    }
                }
            }while(key != NULL);
            free(orig_tok_str);
        }
    }

    /* open the darshan log file for writing */
    ret = DARSHAN_MPI_CALL(PMPI_File_open)(MPI_COMM_WORLD, logfile_name,
        MPI_MODE_CREATE | MPI_MODE_WRONLY | MPI_MODE_EXCL, info, log_fh);
Shane Snyder's avatar
Shane Snyder committed
1390
    if(ret != MPI_SUCCESS)
1391 1392 1393
        return(-1);

    MPI_Info_free(&info);
1394 1395 1396
    return(0);
}

1397
static int darshan_deflate_buffer(void **pointers, int *lengths, int count,
Shane Snyder's avatar
Shane Snyder committed
1398
    char *comp_buf, int *comp_buf_length)
1399 1400 1401 1402 1403 1404
{
    int ret = 0;
    int i;
    int total_target = 0;
    z_stream tmp_stream;

1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415
    /* just return if there is no data */
    for(i = 0; i < count; i++)
    {
        total_target += lengths[i];
    }
    if(total_target)
    {
        total_target = 0;
    }
    else
    {
Shane Snyder's avatar
Shane Snyder committed
1416
        *comp_buf_length = 0;
1417 1418 1419
        return(0);
    }

1420 1421 1422 1423 1424 1425 1426
    memset(&tmp_stream, 0, sizeof(tmp_stream));
    tmp_stream.zalloc = Z_NULL;
    tmp_stream.zfree = Z_NULL;
    tmp_stream.opaque = Z_NULL;

    /* initialize the zlib compression parameters */
    /* TODO: check these parameters? */
Shane Snyder's avatar
Shane Snyder committed
1427 1428 1429
//    ret = deflateInit2(&tmp_stream, Z_DEFAULT_COMPRESSION, Z_DEFLATED,
//        15 + 16, 8, Z_DEFAULT_STRATEGY);
    ret = deflateInit(&tmp_stream, Z_DEFAULT_COMPRESSION);
1430 1431 1432 1433 1434
    if(ret != Z_OK)
    {
        return(-1);
    }

Shane Snyder's avatar
Shane Snyder committed
1435
    tmp_stream.next_out = (unsigned char *)comp_buf;
1436
    tmp_stream.avail_out = DARSHAN_COMP_BUF_SIZE;
1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478

    /* loop over the input pointers */
    for(i = 0; i < count; i++)
    {
        total_target += lengths[i];
        tmp_stream.next_in = pointers[i];
        tmp_stream.avail_in = lengths[i];
        /* while we have not finished consuming all of the data available to
         * this point in the loop
         */
        while(tmp_stream.total_in < total_target)
        {
            if(tmp_stream.avail_out == 0)
            {
                /* We ran out of buffer space for compression.  In theory,
                 * we could start using some of the file_array buffer space
                 * without having to malloc again.  In practice, this case 
                 * is going to be practically impossible to hit.
                 */
                deflateEnd(&tmp_stream);
                return(-1);
            }

            /* compress data */
            ret = deflate(&tmp_stream, Z_NO_FLUSH);
            if(ret != Z_OK)
            {
                deflateEnd(&tmp_stream);
                return(-1);
            }
        }
    }

    /* flush compression and end */
    ret = deflate(&tmp_stream, Z_FINISH);
    if(ret != Z_STREAM_END)
    {
        deflateEnd(&tmp_stream);
        return(-1);
    }
    deflateEnd(&tmp_stream);

Shane Snyder's avatar
Shane Snyder committed
1479
    *comp_buf_length = tmp_stream.total_out;
1480 1481 1482
    return(0);
}

1483 1484 1485
/* NOTE: the map written to file may contain duplicate id->name entries if a
 *       record is opened by multiple ranks, but not all ranks
 */
1486
static int darshan_log_write_record_hash(MPI_File log_fh, struct darshan_core_runtime *core,
1487
    uint64_t *inout_off)
1488
{
1489
    struct darshan_core_record_ref *ref, *tmp;
1490
    int ret;
1491

1492 1493
    /* serialize the record hash into a buffer for writing */
    HASH_ITER(hlink, core->rec_hash, ref, tmp)
1494
    {
1495 1496
        /* to avoid duplicate records, only rank 0 will write shared records */
        if(my_rank > 0 && ref->global_mod_flags)
1497
        {
1498
            /* TODO: remove ref */
1499 1500
        }
    }
1501

1502
    /* collectively write out the record hash to the darshan log */
1503
    ret = darshan_log_append_all(log_fh, core, core->log_rec_p,
1504
        core->log_hdr_p->rec_map.len, inout_off);
1505

1506
    return(ret);
1507 1508
}

1509 1510
/* NOTE: inout_off contains the starting offset of this append at the beginning
 *       of the call, and contains the ending offset at the end of the call.
Shane Snyder's avatar
Shane Snyder committed
1511
 *       This variable is only valid on the root rank (rank 0).
1512
 */
1513
static int darshan_log_append_all(MPI_File log_fh, struct darshan_core_runtime *core,
Shane Snyder's avatar
Shane Snyder committed
1514
    void *buf, int count, uint64_t *inout_off)
1515 1516 1517
{
    MPI_Offset send_off, my_off;
    MPI_Status status;
1518
    int comp_buf_sz = 0;
1519 1520
    int ret;

1521
    /* compress the input buffer */
Shane Snyder's avatar
Shane Snyder committed
1522
    ret = darshan_deflate_buffer((void **)&buf, &count, 1,
1523 1524 1525 1526
        core->comp_buf, &comp_buf_sz);
    if(ret < 0)
        comp_buf_sz = 0;

1527
    /* figure out where everyone is writing using scan */
1528
    send_off = comp_buf_sz;
1529
    if(my_rank == 0)
1530
    {
1531
        send_off += *inout_off; /* rank 0 knows the beginning offset */
1532
    }
1533 1534 1535

    DARSHAN_MPI_CALL(PMPI_Scan)(&send_off, &my_off, 1, MPI_OFFSET,
        MPI_SUM, MPI_COMM_WORLD);
1536
    /* scan is inclusive; subtract local size back out */
1537
    my_off -= comp_buf_sz;
1538

1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552
    if(ret == 0)
    {
        /* no compression errors, proceed with the collective write */
        ret = DARSHAN_MPI_CALL(PMPI_File_write_at_all)(log_fh, my_off,
            core->comp_buf, comp_buf_sz, MPI_BYTE, &status);
    }
    else
    {
        /* error during compression. preserve and return error to caller,
         * but participate in collective write to avoid deadlock.
         */
        (void)DARSHAN_MPI_CALL(PMPI_File_write_at_all)(log_fh, my_off,
            core->comp_buf, comp_buf_sz, MPI_BYTE, &status);
    }
1553

1554 1555
    if(nprocs > 1)
    {
Shane Snyder's avatar
Shane Snyder committed
1556
        /* send the ending offset from rank (n-1) to rank 0 */
1557 1558
        if(my_rank == (nprocs-1))
        {
1559
            my_off += comp_buf_sz;
1560 1561 1562
            DARSHAN_MPI_CALL(PMPI_Send)(&my_off, 1, MPI_OFFSET, 0, 0,
                MPI_COMM_WORLD);
        }
Shane Snyder's avatar
Shane Snyder committed
1563
        if(my_rank == 0)
1564 1565 1566 1567
        {
            DARSHAN_MPI_CALL(PMPI_Recv)(&my_off, 1, MPI_OFFSET, (nprocs-1), 0,
                MPI_COMM_WORLD, &status);

1568
            *inout_off = my_off;
1569 1570 1571 1572
        }
    }
    else
    {
1573
        *inout_off = my_off + comp_buf_sz;
1574
    }
1575

1576 1577
    if(ret != 0)
        return(-1);
1578
    return(0);
1579 1580
}

1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601
/* free darshan core data structures to shutdown */
static void darshan_core_cleanup(struct darshan_core_runtime* core)
{
    struct darshan_core_record_ref *tmp, *ref;
    int i;

    HASH_ITER(hlink, core->rec_hash, ref, tmp)
    {
        HASH_DELETE(hlink, core->rec_hash, ref);
        free(ref);
    }

    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(core->mod_array[i])
        {        
            free(core->mod_array[i]);
            core->mod_array[i] = NULL;
        }
    }

1602 1603 1604 1605 1606 1607 1608 1609
#ifndef __DARSHAN_ENABLE_MMAP_LOGS
    free(core->log_hdr_p);
    free(core->log_job_p);
    free(core->log_exemnt_p);
    free(core->log_rec_p);
    free(core->log_mod_p);
#endif

1610 1611
    if(core->comp_buf)
        free(core->comp_buf);
1612 1613 1614 1615 1616
    free(core);

    return;
}

1617
/* ********************************************************* */
1618

1619
/* TODO: do we alloc new space as we go or just do everything up front? */
1620
void darshan_core_register_module(
1621
    darshan_module_id mod_id,
1622
    struct darshan_module_funcs *funcs,