darshan-core.c

/*
 *  (C) 2009 by Argonne National Laboratory.
 *      See COPYRIGHT 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 <sys/types.h>
#include <sys/stat.h>
#include <sys/vfs.h>
#include <zlib.h>
#include <mpi.h>
#include <assert.h>

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

/* TODO is __progname_full needed here */
extern char* __progname;

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

/* FS mount information */
#define DARSHAN_MAX_MNTS 64
#define DARSHAN_MAX_MNT_PATH 256
#define DARSHAN_MAX_MNT_TYPE 32
struct mnt_data
{
    int64_t 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_core_initialize(
    int *argc, char ***argv);
static void darshan_core_shutdown(
    void);
static void darshan_core_cleanup(
    struct darshan_core_runtime* core);
static void darshan_get_logfile_name(
    char* logfile_name, int jobid, struct tm* start_tm);
static void darshan_log_record_hints_and_ver(
    struct darshan_core_runtime* core);
static void darshan_get_exe_and_mounts_root(
    struct darshan_core_runtime *core, char* trailing_data,
    int space_left);
static char* darshan_get_exe_and_mounts(
    struct darshan_core_runtime *core);
static void darshan_get_shared_records(
    struct darshan_core_runtime *core, darshan_record_id *shared_recs);
static int darshan_log_coll_open(
    char *logfile_name, MPI_File *log_fh);
static int darshan_compress_buffer(void **pointers, int *lengths,
    int count, char *comp_buf, int *comp_length);
static int darshan_log_write_record_hash(
    MPI_File log_fh, struct darshan_core_runtime *core,
    struct darshan_log_map *map);
static int darshan_log_coll_write(
    MPI_File log_fh, void *buf, int count, struct darshan_log_map *map);

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

/* intercept MPI initialize and finalize to manage darshan core runtime */
int MPI_Init(int *argc, char ***argv)
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init)(argc, argv);
    if(ret != MPI_SUCCESS)
    {
        return(ret);
    }

    darshan_core_initialize(argc, argv);

    return(ret);
}

int MPI_Init_thread(int *argc, char ***argv, int required, int *provided)
{
    int ret;

    ret = DARSHAN_MPI_CALL(PMPI_Init_thread)(argc, argv, required, provided);
    if(ret != MPI_SUCCESS)
    {
        return(ret);
    }

    darshan_core_initialize(argc, argv);

    return(ret);
}

int MPI_Finalize(void)
{
    int ret;

    darshan_core_shutdown();

    ret = DARSHAN_MPI_CALL(PMPI_Finalize)();
    return(ret);
}

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

static void darshan_core_initialize(int *argc, char ***argv)
{
    int i;
    int internal_timing_flag = 0;
    double init_start, init_time, init_max;
    char* truncate_string = "<TRUNCATED>";
    int truncate_offset;
    int chars_left = 0;

    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)
    {
        /* TODO: darshan mem alignment code? */

        /* allocate structure to track darshan_core_runtime information */
        darshan_core = malloc(sizeof(*darshan_core));
        if(darshan_core)
        {
            memset(darshan_core, 0, sizeof(*darshan_core));

            darshan_core->log_job.uid = getuid();
            darshan_core->log_job.start_time = time(NULL);
            darshan_core->log_job.nprocs = nprocs;
            darshan_core->wtime_offset = DARSHAN_MPI_CALL(PMPI_Wtime)();

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

            /* if we don't see any arguments, then use glibc symbol to get
             * program name at least (this happens in fortran)
             */
            if(argc == 0)
            {
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
                strncat(darshan_core->exe, __progname, chars_left);
                chars_left = DARSHAN_EXE_LEN-strlen(darshan_core->exe);
                strncat(darshan_core->exe, " <unknown args>", chars_left);
            }

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

            /* collect information about command line and mounted file systems */
            darshan_core->trailing_data = darshan_get_exe_and_mounts(darshan_core);
        }
    }

    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)
        {
            printf("#darshan:<op>\t<nprocs>\t<time>\n");
            printf("darshan:init\t%d\t%f\n", nprocs, init_max);
        }
    }

    return;
}

static void darshan_core_shutdown()
{
    int i;
    char *logfile_name;
    struct darshan_core_runtime *final_core;
    int internal_timing_flag = 0;
    char *envjobid;
    char *jobid_str;
    int jobid;
    struct tm *start_tm;
    time_t start_time_tmp;
    int ret = 0;
    int all_ret = 0;
    int64_t first_start_time;
    int64_t last_end_time;
    int local_mod_use[DARSHAN_MAX_MODS] = {0};
    int global_mod_use_count[DARSHAN_MAX_MODS] = {0};
    darshan_record_id shared_recs[DARSHAN_CORE_MAX_RECORDS] = {0};
    double start_log_time;
    long offset;
    struct darshan_header log_header;
    MPI_File log_fh;
    MPI_Offset tmp_off = 0;
    MPI_Status status;

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

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

    /* we also need to set which modules were registerd on this process and
     * disable tracing within those modules while we shutdown
     */
    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.disable_instrumentation();
        }
    }
    DARSHAN_CORE_UNLOCK();

    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();

    logfile_name = malloc(PATH_MAX);
    if(!logfile_name)
    {
        darshan_core_cleanup(final_core);
        return;
    }

    /* set darshan job id/metadata and constuct log file name on rank 0 */
    if(my_rank == 0)
    {
        /* Use CP_JOBID_OVERRIDE for the env var or CP_JOBID */
        envjobid = getenv(CP_JOBID_OVERRIDE);
        if(!envjobid)
        {
            envjobid = CP_JOBID;
        }

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

        final_core->log_job.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(final_core);

        /* use human readable start time format in log filename */
        start_time_tmp = final_core->log_job.start_time;
        start_tm = localtime(&start_time_tmp);

        /* construct log file name */
        darshan_get_logfile_name(logfile_name, 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.end_time = time(NULL);

    /* reduce to report first start time and last end time across all ranks
     * at rank 0
     */
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job.start_time, &first_start_time, 1, MPI_LONG_LONG, MPI_MIN, 0, MPI_COMM_WORLD);
    DARSHAN_MPI_CALL(PMPI_Reduce)(&final_core->log_job.end_time, &last_end_time, 1, MPI_LONG_LONG, MPI_MAX, 0, MPI_COMM_WORLD);
    if(my_rank == 0)
    {
        final_core->log_job.start_time = first_start_time;
        final_core->log_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);

    /* collectively open the darshan log file */
    ret = darshan_log_coll_open(logfile_name, &log_fh);

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

    /* rank 0 is responsible for writing the compressed darshan job information */
    if(my_rank == 0)
    {
        void *pointers[2] = {&final_core->log_job, final_core->trailing_data};
        int lengths[2] = {sizeof(struct darshan_job), DARSHAN_EXE_LEN+1};
        int comp_buf_sz = 0;
        
        /* compress the job info and the trailing mount/exe data */
        all_ret = darshan_compress_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 */
            all_ret = DARSHAN_MPI_CALL(PMPI_File_write_at)(
                log_fh, sizeof(struct darshan_header), 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);
            }

            /* set the beginning offset of record hash, which precedes job info just written */
            log_header.rec_map.off = sizeof(struct darshan_header) + 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;
    }

    /* write the record name->id hash to the log file */
    ret = darshan_log_write_record_hash(log_fh, final_core, &log_header.rec_map);

    /* 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;
    }

    /* 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];
        darshan_record_id mod_shared_recs[DARSHAN_CORE_MAX_RECORDS];
        struct darshan_core_record_ref *ref = NULL;
        int shared_rec_cnt = 0;
        void* mod_buf = NULL;
        int mod_buf_sz = 0;
        int comp_buf_sz = 0;
        int j;

        if(global_mod_use_count[i] == 0)
        {
            if(my_rank == 0)
                log_header.mod_map[i].off = log_header.mod_map[i].len = 0;

            continue;
        }
        else if(global_mod_use_count[j] == nprocs)
        {
            int shared_rec_count = 0;
            int rec_sz = 0;
            void *red_send_buf = NULL, *red_recv_buf = NULL;
            MPI_Datatype red_type;
            MPI_Op red_op;

            /* if all processes used this module, prepare to do a shared file reduction */

            /* set the shared file list for this module */
            memset(mod_shared_recs, 0, DARSHAN_CORE_MAX_RECORDS * sizeof(darshan_record_id));
            for(j = 0; j < DARSHAN_CORE_MAX_RECORDS && shared_recs[j] != 0; j++)
            {
                HASH_FIND(hlink, final_core->rec_hash, &shared_recs[j],
                    sizeof(darshan_record_id), ref);
                assert(ref);
                if(ref->global_mod_flags & (1 << i)) /* TODO: MACRO? */
                {
                    mod_shared_recs[shared_rec_count++] = shared_recs[j];
                }
            }

            /* if there are globally shared files, do a shared file reduction */
            if(shared_rec_count)
            {
                this_mod->mod_funcs.prepare_for_reduction(mod_shared_recs, &shared_rec_count,
                    &red_send_buf, &red_recv_buf, &rec_sz);

                if(shared_rec_count)
                {
                    /* construct a datatype for a file record.  This is serving no purpose
                     * except to make sure we can do a reduction on proper boundaries
                     */
                    DARSHAN_MPI_CALL(PMPI_Type_contiguous)(rec_sz, MPI_BYTE, &red_type);
                    DARSHAN_MPI_CALL(PMPI_Type_commit)(&red_type);

                    /* register a reduction operator for this module */
                    DARSHAN_MPI_CALL(PMPI_Op_create)(this_mod->mod_funcs.reduce_record,
                        1, &red_op);

                    /* reduce shared file records for this module */
                    DARSHAN_MPI_CALL(PMPI_Reduce)(red_send_buf, red_recv_buf,
                        shared_rec_count, red_type, red_op, 0, MPI_COMM_WORLD);

                    DARSHAN_MPI_CALL(PMPI_Type_free)(&red_type);
                    DARSHAN_MPI_CALL(PMPI_Op_free)(&red_op);
                }
            }
        }

        /* if module is registered locally, get the corresponding output buffer */
        if(this_mod)
        {
            /* get output buffer from module */
            this_mod->mod_funcs.get_output_data(&mod_buf, &mod_buf_sz);
        }

        /* compress the module buffer */
        ret = darshan_compress_buffer((void**)&mod_buf, &mod_buf_sz, 1,
            final_core->comp_buf, &comp_buf_sz);
        if(ret == 0)
        {
            /* set the starting offset of this module */
            if(tmp_off == 0)
                tmp_off = log_header.rec_map.off + log_header.rec_map.len;

            log_header.mod_map[i].off = tmp_off;

            /* write (compressed) module data buffer to the darshan log file */
            ret = darshan_log_coll_write(log_fh, final_core->comp_buf, comp_buf_sz,
                &log_header.mod_map[i]);
        }
        else
        {
            /* error in compression, participate in collective write to avoid
             * deadlock, but preserve return value to terminate darshan_shutdown.
             */
            (void)darshan_log_coll_write(log_fh, NULL, 0, &log_header.mod_map[i]);
        }

        /* error out if the compression or collective write 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;
        }

        tmp_off += log_header.mod_map[i].len;

        /* shutdown module if registered locally */
        if(this_mod)
        {
            this_mod->mod_funcs.shutdown();
        }
    }

    /* rank 0 is responsible for writing the log header */
    if(my_rank == 0)
    {
        /* initialize the remaining header fields */
        strcpy(log_header.version_string, DARSHAN_LOG_VERSION);
        log_header.magic_nr = DARSHAN_MAGIC_NR;
        log_header.comp_type = DARSHAN_GZ_COMP;

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

    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.gz, which indicates that this log file is
     * complete and ready for analysis
     */
    /* TODO: support user given logfile path/name */
    if(my_rank == 0)
    {
        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.gz", (int)(end_log_time-start_log_time+1));
            rename(logfile_name, new_logfile_name);
            /* set permissions on log file */
#ifdef __CP_GROUP_READABLE_LOGS
            chmod(new_logfile_name, (S_IRUSR|S_IRGRP));
#else
            chmod(new_logfile_name, (S_IRUSR));
#endif
            free(new_logfile_name);
        }
    }

    free(logfile_name);
    darshan_core_cleanup(final_core);

    if(internal_timing_flag)
    {
        /* TODO: what do we want to time in new darshan version? */
    }
    
    return;
}

/* 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->rec.name);
        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;
        }
    }

    free(core->trailing_data);
    free(core);

    return;
}

/* construct the darshan log file name */
static void darshan_get_logfile_name(char* logfile_name, int jobid, struct tm* start_tm)
{
    char* logpath;
    char* logname_string;
    char* logpath_override = NULL;
#ifdef __CP_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;

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

    /* 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, sizeof(cuser), "%u", uid);
    }

    /* 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 __CP_LOG_ENV
    /* just silently skip if the environment variable list is too big */
    if(strlen(__CP_LOG_ENV) < 256)
    {
        /* copy env variable list to a temporary buffer */
        strcpy(env_check, __CP_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;
}

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

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

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

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

    meta_remain = DARSHAN_JOB_METADATA_LEN -
        strlen(core->log_job.metadata) - 1;
    if(meta_remain >= (strlen(PACKAGE_VERSION) + 9))
    {
        sprintf(core->log_job.metadata, "lib_ver=%s\n", PACKAGE_VERSION);
        meta_remain -= (strlen(PACKAGE_VERSION) + 9);
    }
    if(meta_remain >= (3 + strlen(header_hints)))
    {
        m = core->log_job.metadata + strlen(core->log_job.metadata);
        /* We have room to store the hints in the metadata portion of
         * the job header.  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", header_hints);
    }
    free(header_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* trailing_data, 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 for use in header of 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(trailing_data, tmp_mnt);
        (*space_left) -= strlen(tmp_mnt);
    }

    mnt_data_count++;
    return;
}

/* darshan_get_exe_and_mounts_root()
 *
 * collects command line and list of mounted file systems into a string that
 * will be stored with the job header
 */
static void darshan_get_exe_and_mounts_root(struct darshan_core_runtime *core,
    char* trailing_data, int space_left)
{
    FILE* tab;
    struct mntent *entry;
    char* exclude;
    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
    };

    /* length of exe has already been safety checked in darshan-posix.c */
    strcat(trailing_data, core->exe);
    space_left = DARSHAN_EXE_LEN - strlen(trailing_data);

    /* 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(trailing_data, &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(trailing_data, &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;
}

/* darshan_get_exe_and_mounts()
 *
 * collects command line and list of mounted file systems into a string that
 * will be stored with the job header
 */
static char* darshan_get_exe_and_mounts(struct darshan_core_runtime *core)
{
    char* trailing_data;
    int space_left;

    space_left = DARSHAN_EXE_LEN + 1;
    trailing_data = malloc(space_left);
    if(!trailing_data)
    {
        return(NULL);
    }
    memset(trailing_data, 0, space_left);

    if(my_rank == 0)
    {
        darshan_get_exe_and_mounts_root(core, trailing_data, space_left);
    }

    /* broadcast trailing data to all nodes */
    DARSHAN_MPI_CALL(PMPI_Bcast)(trailing_data, space_left, MPI_CHAR, 0,
        MPI_COMM_WORLD);
    /* broadcast mount count to all nodes */
    DARSHAN_MPI_CALL(PMPI_Bcast)(&mnt_data_count, 1, MPI_INT, 0,
        MPI_COMM_WORLD);