darshan-core.c

/*
 * Copyright (C) 2015 University of Chicago.
 * See COPYRIGHT notice in top-level directory.
 *
 */

#define _XOPEN_SOURCE 500
#define _GNU_SOURCE

#include "darshan-runtime-config.h"

#include <stdio.h>
#ifdef HAVE_MNTENT_H
#include <mntent.h>
#endif
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <limits.h>
#include <pthread.h>
#include <fcntl.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/mman.h>
#include <sys/vfs.h>
#include <zlib.h>
#include <mpi.h>
#include <assert.h>

#include "uthash.h"
#include "darshan.h"
#include "darshan-core.h"
#include "darshan-dynamic.h"

extern char* __progname;
extern char* __progname_full;

/* internal variable delcarations */
static struct darshan_core_runtime *darshan_core = NULL;
static pthread_mutex_t darshan_core_mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
static int my_rank = -1;
static int nprocs = -1;
static int darshan_mem_alignment = 1;

/* 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
};

#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
};

#define DARSHAN_CORE_LOCK() pthread_mutex_lock(&darshan_core_mutex)
#define DARSHAN_CORE_UNLOCK() pthread_mutex_unlock(&darshan_core_mutex)

/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
    int block_size;
    char path[DARSHAN_MAX_MNT_PATH];
    char type[DARSHAN_MAX_MNT_TYPE];
};
static struct mnt_data mnt_data_array[DARSHAN_MAX_MNTS];
static int mnt_data_count = 0;

/* prototypes for internal helper functions */
static void darshan_log_record_hints_and_ver(
    struct darshan_core_runtime* core);
static void darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core, int argc, char **argv);
static void darshan_add_record_hashref(
    struct darshan_core_runtime *core, char *name,
    darshan_record_id id, struct darshan_core_record_ref **ref);
static int darshan_block_size_from_path(
    const char *path);
static void darshan_get_user_name(
    char *user);
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_get_shared_records(
    struct darshan_core_runtime *core, darshan_record_id **shared_recs,
    int *shared_rec_cnt);
static int darshan_log_open_all(
    char *logfile_name, MPI_File *log_fh);
static int darshan_deflate_buffer(
    void **pointers, int *lengths, int count, char *comp_buf,
    int *comp_buf_length);
static int darshan_log_write_record_hash(
    MPI_File log_fh, struct darshan_core_runtime *core,
    uint64_t *inout_off);
static int darshan_log_append_all(
    MPI_File log_fh, struct darshan_core_runtime *core, void *buf,
    int count, uint64_t *inout_off);
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);

/* *********************************** */

void darshan_core_initialize(int argc, char **argv)
{
    struct darshan_core_runtime *init_core = NULL;
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
    char *envstr;
    char *jobid_str;
    int jobid;
    int ret;
    int tmpval;
    int i;
#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

    DARSHAN_MPI_CALL(PMPI_Comm_size)(MPI_COMM_WORLD, &nprocs);
    DARSHAN_MPI_CALL(PMPI_Comm_rank)(MPI_COMM_WORLD, &my_rank);

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

    if(internal_timing_flag)
        init_start = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* setup darshan runtime if darshan is enabled and hasn't been initialized already */
    if(!getenv("DARSHAN_DISABLE") && !darshan_core)
    {
        #if (__DARSHAN_MEM_ALIGNMENT < 1)
            #error Darshan must be configured with a positive value for --with-mem-align
        #endif
        envstr = getenv(DARSHAN_MEM_ALIGNMENT_OVERRIDE);
        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
        {
            darshan_mem_alignment = __DARSHAN_MEM_ALIGNMENT;
        }

        /* avoid floating point errors on faulty input */
        if(darshan_mem_alignment < 1)
        {
            darshan_mem_alignment = 1;
        }

        /* 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();
        }

        /* allocate structure to track darshan core runtime information */
        init_core = malloc(sizeof(*init_core));
        if(init_core)
        {
            memset(init_core, 0, sizeof(*init_core));
            init_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

#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
             */

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

            mmap_size = sizeof(struct darshan_header) + DARSHAN_JOB_RECORD_SIZE +
                + DARSHAN_RECORD_BUF_SIZE + DARSHAN_MOD_MEM_MAX;
            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_dir = envstr;
            else
                mmap_log_dir = 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(init_core->mmap_log_name, PATH_MAX,
                "/%s/%s_%s_id%d_mmap-log-%d.darshan",
                mmap_log_dir, cuser, __progname, jobid, my_rank);

            /* create the temporary mmapped darshan log */
            mmap_fd = open(init_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", init_core->mmap_log_name);
                free(init_core);
                return;
            }

            /* TODO: what's more expensive? truncate or write zeros? perf test this call and later accesses */
            /* 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", init_core->mmap_log_name);
                free(init_core);
                close(mmap_fd);
                unlink(init_core->mmap_log_name);
                return;
            }

            /* memory map buffers for getting at least some summary i/o data
             * into a log file if darshan does not shut down properly
             */
            void *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", init_core->mmap_log_name);
                free(init_core);
                close(mmap_fd);
                unlink(init_core->mmap_log_name);
                return;
            }

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

            /* set the memory pointers for each log file region */
            init_core->log_hdr_p = (struct darshan_header *)mmap_p;
            init_core->log_job_p = (struct darshan_job *)
                ((char *)init_core->log_hdr_p + sizeof(struct darshan_header));
            init_core->log_exemnt_p = (char *)
                ((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);

            /* set header fields needed for the mmap log mechanism */
            init_core->log_hdr_p->comp_type = DARSHAN_NO_COMP;
            init_core->log_hdr_p->rec_map.off =
                sizeof(struct darshan_header) + DARSHAN_JOB_RECORD_SIZE;
#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;

            /* 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;

            /* 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);

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

            /* bootstrap any modules with static initialization routines */
            i = 0;
            while(mod_static_init_fns[i])
            {
                (*mod_static_init_fns[i])();
                i++;
            }

            darshan_core = init_core;
        }
    }

    /* 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);

    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);
        if(my_rank == 0)
        {
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            fprintf(stderr, "darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

void darshan_core_shutdown()
{
    struct darshan_core_runtime *final_core;
    struct darshan_header out_header;
    struct darshan_job out_job;
    int internal_timing_flag = 0;
    struct tm *start_tm;
    time_t start_time_tmp;
    int64_t last_end_time;
    double start_log_time;
    double tm_end;
    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;

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

    if(internal_timing_flag)
        start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

    /* disable darhan-core while we shutdown */
    DARSHAN_CORE_LOCK();
    if(!darshan_core)
    {
        DARSHAN_CORE_UNLOCK();
        return;
    }
    final_core = darshan_core;
    darshan_core = NULL;
    DARSHAN_CORE_UNLOCK();

    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?? */
    /* XXX just copy mmap files somewhere else to avoid corruption */
    DARSHAN_MPI_CALL(PMPI_Barrier)(MPI_COMM_WORLD);
    char cp_cmd[500] = {0};
    sprintf(cp_cmd, "cp %s ~/Desktop", final_core->mmap_log_name);
    system(cp_cmd);
    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;
    /* TODO: do we ever check for darshan_shutdown? */
    snprintf(m, meta_remain, "darshan_shutdown=yes\n");
#endif

    /* we also need to set which modules were registered on this process and
     * call into those modules and give them a chance to perform any necessary
     * pre-shutdown steps.
     */
    for(i = 0; i < DARSHAN_MAX_MODS; i++)
    {
        if(final_core->mod_array[i])
        {
            local_mod_use[i] = 1;
            final_core->mod_array[i]->mod_funcs.begin_shutdown();
        }
    }

    final_core->comp_buf = malloc(DARSHAN_COMP_BUF_SIZE);
    if(!(final_core->comp_buf))
    {
        darshan_core_cleanup(final_core);
        return;
    }

    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;
    }

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

    /* reduce to report last end time across all ranks at rank 0 */
    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)
    {
        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);
    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);
    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)
    {
        mode_t chmod_mode = S_IRUSR;
#ifdef __DARSHAN_GROUP_READABLE_LOGS
        chmod_mode |= S_IRGRP;
#endif

        if(getenv("DARSHAN_LOGFILE"))
        {
            chmod(logfile_name, chmod_mode);
        }
        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 */
                chmod(new_logfile_name, chmod_mode);
                free(new_logfile_name);
            }
        }
    }

#ifdef __DARSHAN_ENABLE_MMAP_LOGS
    /* remove the temporary mmap log files */
    unlink(final_core->mmap_log_name);
#endif

    free(logfile_name);
    free(shared_recs);
    free(mod_shared_recs);
    darshan_core_cleanup(final_core);

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

        tm_end = DARSHAN_MPI_CALL(PMPI_Wtime)();

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

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

        if(my_rank == 0)
        {
            fprintf(stderr, "#darshan:<op>\t<nprocs>\t<time>\n");
            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]);
            }
            fprintf(stderr, "darshan:core_shutdown\t%d\t%f\n", nprocs, all_slowest);
        }
    }

    return;
}

/* *********************************** */

/* record any hints used to write the darshan log in the job data */
static void darshan_log_record_hints_and_ver(struct darshan_core_runtime* core)
{
    char* hints;
    char* job_hints;
    int meta_remain = 0;
    char* m;

    /* check environment variable to see if the default MPI file hints have
     * been overridden
     */
    hints = getenv(DARSHAN_LOG_HINTS_OVERRIDE);
    if(!hints)
    {
        hints = __DARSHAN_LOG_HINTS;
    }

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

    job_hints = strdup(hints);
    if(!job_hints)
        return;

    meta_remain = DARSHAN_JOB_METADATA_LEN -
        strlen(core->log_job_p->metadata) - 1;
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
        sprintf(core->log_job_p->metadata, "lib_ver=%s\n", PACKAGE_VERSION);
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(job_hints)))
    {
        m = core->log_job_p->metadata + strlen(core->log_job_p->metadata);
        /* We have room to store the hints in the metadata portion of
         * the job structure.  We just prepend an h= to the hints list.  The
         * metadata parser will ignore = characters that appear in the value
         * portion of the metadata key/value pair.
         */
        sprintf(m, "h=%s\n", job_hints);
    }
    free(job_hints);

    return;
}

static int mnt_data_cmp(const void* a, const void* b)
{
    const struct mnt_data *d_a = (const struct mnt_data*)a;
    const struct mnt_data *d_b = (const struct mnt_data*)b;

    if(strlen(d_a->path) > strlen(d_b->path))
        return(-1);
    else if(strlen(d_a->path) < strlen(d_b->path))
        return(1);
    else
        return(0);
}

/* adds an entry to table of mounted file systems */
static void add_entry(char* buf, int* space_left, struct mntent *entry)
{
    int ret;
    char tmp_mnt[256];
    struct statfs statfsbuf;

    strncpy(mnt_data_array[mnt_data_count].path, entry->mnt_dir,
        DARSHAN_MAX_MNT_PATH-1);
    strncpy(mnt_data_array[mnt_data_count].type, entry->mnt_type,
        DARSHAN_MAX_MNT_TYPE-1);
    /* NOTE: we now try to detect the preferred block size for each file 
     * system using fstatfs().  On Lustre we assume a size of 1 MiB 
     * because fstatfs() reports 4 KiB. 
     */
#ifndef LL_SUPER_MAGIC
#define LL_SUPER_MAGIC 0x0BD00BD0
#endif
    ret = statfs(entry->mnt_dir, &statfsbuf);
    if(ret == 0 && statfsbuf.f_type != LL_SUPER_MAGIC)
        mnt_data_array[mnt_data_count].block_size = statfsbuf.f_bsize;
    else if(ret == 0 && statfsbuf.f_type == LL_SUPER_MAGIC)
        mnt_data_array[mnt_data_count].block_size = 1024*1024;
    else
        mnt_data_array[mnt_data_count].block_size = 4096;

    /* store mount information with the job-level metadata in darshan log */
    ret = snprintf(tmp_mnt, 256, "\n%s\t%s",
        entry->mnt_type, entry->mnt_dir);
    if(ret < 256 && strlen(tmp_mnt) <= (*space_left))
    {
        strcat(buf, tmp_mnt);
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

/* darshan_get_exe_and_mounts()
 *
 * collects command line and list of mounted file systems into a string that
 * will be stored with the job-level metadata
 */
static void darshan_get_exe_and_mounts(struct darshan_core_runtime *core,
    int argc, char **argv)
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int space_left = DARSHAN_EXE_LEN;
    int i;
    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
    };

    /* record exe and arguments */
    for(i=0; i<argc; i++)
    {
        strncat(core->log_exemnt_p, argv[i], space_left);
        space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
        if(i < (argc-1))
        {
            strncat(core->log_exemnt_p, " ", space_left);
            space_left = DARSHAN_EXE_LEN-strlen(core->log_exemnt_p);
        }
    }

    /* if we don't see any arguments, then use glibc symbol to get
     * program name at least (this happens in fortran)
     */
    if(argc == 0)
    {
        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);
    }

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

    /* we make two passes through mounted file systems; in the first pass we
     * grab any non-nfs mount points, then on the second pass we grab nfs
     * mount points
     */

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

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

        add_entry(core->log_exemnt_p, &space_left, entry);
    }
    endmntent(tab);

    tab = setmntent("/etc/mtab", "r");
    if(!tab)
        return;
    /* loop through list of mounted file systems */
    while(mnt_data_count<DARSHAN_MAX_MNTS && (entry = getmntent(tab)) != NULL)
    {
        if(strcmp(entry->mnt_type, "nfs") != 0)
            continue;

        add_entry(core->log_exemnt_p, &space_left, entry);
    }
    endmntent(tab);

    /* sort mount points in order of longest path to shortest path.  This is
     * necessary so that if we try to match file paths to mount points later
     * we don't match on "/" every time.
     */
    qsort(mnt_data_array, mnt_data_count, sizeof(mnt_data_array[0]), mnt_data_cmp);
    return;
}

static int darshan_block_size_from_path(const char *path)
{
    int i;
    int block_size = -1;

    for(i=0; i<mnt_data_count; i++)
    {
        if(!(strncmp(mnt_data_array[i].path, path, strlen(mnt_data_array[i].path))))
        {
            block_size = mnt_data_array[i].block_size;
            break;;
        }
    }

    return block_size;
}

static void darshan_get_user_name(char *cuser)
{
    char* logname_string;

    /* get the username for this job.  In order we will try each of the
     * following until one of them succeeds:
     *
     * - cuserid()
     * - getenv("LOGNAME")
     * - snprintf(..., geteuid());
     *
     * Note that we do not use getpwuid() because it generally will not
     * work in statically compiled binaries.
     */

#ifndef __DARSHAN_DISABLE_CUSERID
    cuserid(cuser);
#endif

    /* if cuserid() didn't work, then check the environment */
    if(strcmp(cuser, "") == 0)
    {
        logname_string = getenv("LOGNAME");
        if(logname_string)
        {
            strncpy(cuser, logname_string, (L_cuserid-1));
        }
    }

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

    return;
}

/* 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
        }

        darshan_get_user_name(cuser);

        /* 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;
}

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;
    darshan_record_id *id_p;
    char *name_p;

    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));

        /* 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);

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

        /* add the record to the hash table */
        HASH_ADD_KEYPTR(hlink, core->rec_hash, id_p, sizeof(darshan_record_id), (*ref));
        core->rec_hash_cnt++;
        core->log_hdr_p->rec_map.len += record_size;
    }

    return;
}

static void darshan_get_shared_records(struct darshan_core_runtime *core,
    darshan_record_id **shared_recs, int *shared_rec_cnt)
{
    int i, j;
    int tmp_cnt = core->rec_hash_cnt;
    struct darshan_core_record_ref *tmp, *ref;
    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));

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

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

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

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

    j = 0;
    for(i=0; i<tmp_cnt; i++)
    {
        if(global_mod_flags[i] != 0)
        {
            (*shared_recs)[j++] = id_array[i];

            /* 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];
        }
    }
    *shared_rec_cnt = j;

    free(id_array);
    free(mod_flags);
    free(global_mod_flags);
    return;
}

static int darshan_log_open_all(char *logfile_name, MPI_File *log_fh)
{
    char *hints;
    char *tok_str;