Add support for multiple output formats
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 3db06125092b5fee1cbbeaa1749d839a426cab2d..ea5b43a38ba1cd65cf9547c6d4bfe7922ac30a6a 100644 (file)
--- a/stat.c
+++ b/stat.c
@@ -13,6 +13,9 @@
 #include "json.h"
 #include "lib/getrusage.h"
 #include "idletime.h"
+#include "lib/pow2.h"
+
+struct fio_mutex *stat_mutex;
 
 void update_rusage_stat(struct thread_data *td)
 {
@@ -241,15 +244,13 @@ out:
 int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
             double *mean, double *dev)
 {
-       double n = is->samples;
+       double n = (double) is->samples;
 
-       if (is->samples == 0)
+       if (n == 0)
                return 0;
 
        *min = is->min_val;
        *max = is->max_val;
-
-       n = (double) is->samples;
        *mean = is->mean.u.f;
 
        if (n > 1.0)
@@ -263,7 +264,7 @@ int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
 void show_group_stats(struct group_run_stats *rs)
 {
        char *p1, *p2, *p3, *p4;
-       const char *ddir_str[] = { "   READ", "  WRITE" , "   TRIM"};
+       const char *str[] = { "   READ", "  WRITE" , "   TRIM"};
        int i;
 
        log_info("\nRun status group %d (all jobs):\n", rs->groupid);
@@ -281,7 +282,7 @@ void show_group_stats(struct group_run_stats *rs)
 
                log_info("%s: io=%s, aggrb=%s/s, minb=%s/s, maxb=%s/s,"
                         " mint=%llumsec, maxt=%llumsec\n",
-                               rs->unified_rw_rep ? "  MIXED" : ddir_str[i],
+                               rs->unified_rw_rep ? "  MIXED" : str[i],
                                p1, p2, p3, p4,
                                (unsigned long long) rs->min_run[i],
                                (unsigned long long) rs->max_run[i]);
@@ -363,7 +364,7 @@ static void display_lat(const char *name, unsigned long min, unsigned long max,
 static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
                             int ddir)
 {
-       const char *ddir_str[] = { "read ", "write", "trim" };
+       const char *str[] = { "read ", "write", "trim" };
        unsigned long min, max, runt;
        unsigned long long bw, iops;
        double mean, dev;
@@ -386,7 +387,7 @@ static void show_ddir_status(struct group_run_stats *rs, struct thread_stat *ts,
        iops_p = num2str(iops, 6, 1, 0, 0);
 
        log_info("  %s: io=%s, bw=%s/s, iops=%s, runt=%6llumsec\n",
-                               rs->unified_rw_rep ? "mixed" : ddir_str[ddir],
+                               rs->unified_rw_rep ? "mixed" : str[ddir],
                                io_p, bw_p, iops_p,
                                (unsigned long long) ts->runtime[ddir]);
 
@@ -497,17 +498,159 @@ static void show_latencies(struct thread_stat *ts)
        show_lat_m(io_u_lat_m);
 }
 
-void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
+static int block_state_category(int block_state)
+{
+       switch (block_state) {
+       case BLOCK_STATE_UNINIT:
+               return 0;
+       case BLOCK_STATE_TRIMMED:
+       case BLOCK_STATE_WRITTEN:
+               return 1;
+       case BLOCK_STATE_WRITE_FAILURE:
+       case BLOCK_STATE_TRIM_FAILURE:
+               return 2;
+       default:
+               /* Silence compile warning on some BSDs and have a return */
+               assert(0);
+               return -1;
+       }
+}
+
+static int compare_block_infos(const void *bs1, const void *bs2)
+{
+       uint32_t block1 = *(uint32_t *)bs1;
+       uint32_t block2 = *(uint32_t *)bs2;
+       int state1 = BLOCK_INFO_STATE(block1);
+       int state2 = BLOCK_INFO_STATE(block2);
+       int bscat1 = block_state_category(state1);
+       int bscat2 = block_state_category(state2);
+       int cycles1 = BLOCK_INFO_TRIMS(block1);
+       int cycles2 = BLOCK_INFO_TRIMS(block2);
+
+       if (bscat1 < bscat2)
+               return -1;
+       if (bscat1 > bscat2)
+               return 1;
+
+       if (cycles1 < cycles2)
+               return -1;
+       if (cycles1 > cycles2)
+               return 1;
+
+       if (state1 < state2)
+               return -1;
+       if (state1 > state2)
+               return 1;
+
+       assert(block1 == block2);
+       return 0;
+}
+
+static int calc_block_percentiles(int nr_block_infos, uint32_t *block_infos,
+                                 fio_fp64_t *plist, unsigned int **percentiles,
+                                 unsigned int *types)
+{
+       int len = 0;
+       int i, nr_uninit;
+
+       qsort(block_infos, nr_block_infos, sizeof(uint32_t), compare_block_infos);
+
+       while (len < FIO_IO_U_LIST_MAX_LEN && plist[len].u.f != 0.0)
+               len++;
+
+       if (!len)
+               return 0;
+
+       /*
+        * Sort the percentile list. Note that it may already be sorted if
+        * we are using the default values, but since it's a short list this
+        * isn't a worry. Also note that this does not work for NaN values.
+        */
+       if (len > 1)
+               qsort((void *)plist, len, sizeof(plist[0]), double_cmp);
+
+       nr_uninit = 0;
+       /* Start only after the uninit entries end */
+       for (nr_uninit = 0;
+            nr_uninit < nr_block_infos
+               && BLOCK_INFO_STATE(block_infos[nr_uninit]) == BLOCK_STATE_UNINIT;
+            nr_uninit ++)
+               ;
+
+       if (nr_uninit == nr_block_infos)
+               return 0;
+
+       *percentiles = calloc(len, sizeof(**percentiles));
+
+       for (i = 0; i < len; i++) {
+               int idx = (plist[i].u.f * (nr_block_infos - nr_uninit) / 100)
+                               + nr_uninit;
+               (*percentiles)[i] = BLOCK_INFO_TRIMS(block_infos[idx]);
+       }
+
+       memset(types, 0, sizeof(*types) * BLOCK_STATE_COUNT);
+       for (i = 0; i < nr_block_infos; i++)
+               types[BLOCK_INFO_STATE(block_infos[i])]++;
+
+       return len;
+}
+
+static const char *block_state_names[] = {
+       [BLOCK_STATE_UNINIT] = "unwritten",
+       [BLOCK_STATE_TRIMMED] = "trimmed",
+       [BLOCK_STATE_WRITTEN] = "written",
+       [BLOCK_STATE_TRIM_FAILURE] = "trim failure",
+       [BLOCK_STATE_WRITE_FAILURE] = "write failure",
+};
+
+static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
+                            fio_fp64_t *plist)
+{
+       int len, pos, i;
+       unsigned int *percentiles = NULL;
+       unsigned int block_state_counts[BLOCK_STATE_COUNT];
+
+       len = calc_block_percentiles(nr_block_infos, block_infos, plist,
+                                    &percentiles, block_state_counts);
+
+       log_info("  block lifetime percentiles :\n   |");
+       pos = 0;
+       for (i = 0; i < len; i++) {
+               uint32_t block_info = percentiles[i];
+#define LINE_LENGTH    75
+               char str[LINE_LENGTH];
+               int strln = snprintf(str, LINE_LENGTH, " %3.2fth=%u%c",
+                                    plist[i].u.f, block_info,
+                                    i == len - 1 ? '\n' : ',');
+               assert(strln < LINE_LENGTH);
+               if (pos + strln > LINE_LENGTH) {
+                       pos = 0;
+                       log_info("\n   |");
+               }
+               log_info("%s", str);
+               pos += strln;
+#undef LINE_LENGTH
+       }
+       if (percentiles)
+               free(percentiles);
+
+       log_info("        states               :");
+       for (i = 0; i < BLOCK_STATE_COUNT; i++)
+               log_info(" %s=%u%c",
+                        block_state_names[i], block_state_counts[i],
+                        i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
+}
+
+static void show_thread_status_normal(struct thread_stat *ts,
+                                     struct group_run_stats *rs)
 {
        double usr_cpu, sys_cpu;
        unsigned long runtime;
        double io_u_dist[FIO_IO_U_MAP_NR];
        time_t time_p;
-       char time_buf[64];
+       char time_buf[32];
 
-       if (!(ts->io_bytes[DDIR_READ] + ts->io_bytes[DDIR_WRITE] +
-           ts->io_bytes[DDIR_TRIM]) && !(ts->total_io_u[DDIR_READ] +
-           ts->total_io_u[DDIR_WRITE] + ts->total_io_u[DDIR_TRIM]))
+       if (!ddir_rw_sum(ts->io_bytes) && !ddir_rw_sum(ts->total_io_u))
                return;
 
        time(&time_p);
@@ -573,19 +716,34 @@ void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs)
                                        io_u_dist[3], io_u_dist[4],
                                        io_u_dist[5], io_u_dist[6]);
        log_info("     issued    : total=r=%llu/w=%llu/d=%llu,"
-                                " short=r=%llu/w=%llu/d=%llu\n",
+                                " short=r=%llu/w=%llu/d=%llu,"
+                                " drop=r=%llu/w=%llu/d=%llu\n",
                                        (unsigned long long) ts->total_io_u[0],
                                        (unsigned long long) ts->total_io_u[1],
                                        (unsigned long long) ts->total_io_u[2],
                                        (unsigned long long) ts->short_io_u[0],
                                        (unsigned long long) ts->short_io_u[1],
-                                       (unsigned long long) ts->short_io_u[2]);
+                                       (unsigned long long) ts->short_io_u[2],
+                                       (unsigned long long) ts->drop_io_u[0],
+                                       (unsigned long long) ts->drop_io_u[1],
+                                       (unsigned long long) ts->drop_io_u[2]);
        if (ts->continue_on_error) {
                log_info("     errors    : total=%llu, first_error=%d/<%s>\n",
                                        (unsigned long long)ts->total_err_count,
                                        ts->first_error,
                                        strerror(ts->first_error));
        }
+       if (ts->latency_depth) {
+               log_info("     latency   : target=%llu, window=%llu, percentile=%.2f%%, depth=%u\n",
+                                       (unsigned long long)ts->latency_target,
+                                       (unsigned long long)ts->latency_window,
+                                       ts->latency_percentile.u.f,
+                                       ts->latency_depth);
+       }
+
+       if (ts->nr_block_infos)
+               show_block_infos(ts->nr_block_infos, ts->block_infos,
+                                 ts->percentile_list);
 }
 
 static void show_ddir_status_terse(struct thread_stat *ts,
@@ -664,9 +822,9 @@ static void add_ddir_status_json(struct thread_stat *ts,
                struct group_run_stats *rs, int ddir, struct json_object *parent)
 {
        unsigned long min, max;
-       unsigned long long bw, iops;
+       unsigned long long bw;
        unsigned int *ovals = NULL;
-       double mean, dev;
+       double mean, dev, iops;
        unsigned int len, minv, maxv;
        int i;
        const char *ddirname[] = {"read", "write", "trim"};
@@ -683,18 +841,22 @@ static void add_ddir_status_json(struct thread_stat *ts,
        json_object_add_value_object(parent,
                ts->unified_rw_rep ? "mixed" : ddirname[ddir], dir_object);
 
-       iops = bw = 0;
+       bw = 0;
+       iops = 0.0;
        if (ts->runtime[ddir]) {
                uint64_t runt = ts->runtime[ddir];
 
                bw = ((1000 * ts->io_bytes[ddir]) / runt) / 1024;
-               iops = (1000 * (uint64_t) ts->total_io_u[ddir]) / runt;
+               iops = (1000.0 * (uint64_t) ts->total_io_u[ddir]) / runt;
        }
 
        json_object_add_value_int(dir_object, "io_bytes", ts->io_bytes[ddir] >> 10);
        json_object_add_value_int(dir_object, "bw", bw);
-       json_object_add_value_int(dir_object, "iops", iops);
+       json_object_add_value_float(dir_object, "iops", iops);
        json_object_add_value_int(dir_object, "runtime", ts->runtime[ddir]);
+       json_object_add_value_int(dir_object, "total_ios", ts->total_io_u[ddir]);
+       json_object_add_value_int(dir_object, "short_ios", ts->short_io_u[ddir]);
+       json_object_add_value_int(dir_object, "drop_ios", ts->drop_io_u[ddir]);
 
        if (!calc_lat(&ts->slat_stat[ddir], &min, &max, &mean, &dev)) {
                min = max = 0;
@@ -762,7 +924,7 @@ static void add_ddir_status_json(struct thread_stat *ts,
        }
        json_object_add_value_int(dir_object, "bw_min", min);
        json_object_add_value_int(dir_object, "bw_max", max);
-       json_object_add_value_float(dir_object, "bw_agg", mean);
+       json_object_add_value_float(dir_object, "bw_agg", p_of_agg);
        json_object_add_value_float(dir_object, "bw_mean", mean);
        json_object_add_value_float(dir_object, "bw_dev", dev);
 }
@@ -823,7 +985,7 @@ static void show_thread_status_terse_v2(struct thread_stat *ts,
        log_info("\n");
 
        /* Additional output if description is set */
-       if (ts->description)
+       if (strlen(ts->description))
                log_info(";%s", ts->description);
 
        log_info("\n");
@@ -900,17 +1062,26 @@ static struct json_object *show_thread_status_json(struct thread_stat *ts,
                                    struct group_run_stats *rs)
 {
        struct json_object *root, *tmp;
+       struct jobs_eta *je;
        double io_u_dist[FIO_IO_U_MAP_NR];
        double io_u_lat_u[FIO_IO_U_LAT_U_NR];
        double io_u_lat_m[FIO_IO_U_LAT_M_NR];
        double usr_cpu, sys_cpu;
        int i;
+       size_t size;
+
 
        root = json_create_object();
        json_object_add_value_string(root, "jobname", ts->name);
        json_object_add_value_int(root, "groupid", ts->groupid);
        json_object_add_value_int(root, "error", ts->error);
 
+       /* ETA Info */
+       je = get_jobs_eta(1, &size);
+       json_object_add_value_int(root, "eta", je->eta_sec);
+       json_object_add_value_int(root, "elapsed", je->elapsed_sec);
+
+
        add_ddir_status_json(ts, rs, DDIR_READ, root);
        add_ddir_status_json(ts, rs, DDIR_WRITE, root);
        add_ddir_status_json(ts, rs, DDIR_TRIM, root);
@@ -970,13 +1141,59 @@ static struct json_object *show_thread_status_json(struct thread_stat *ts,
        /* Additional output if continue_on_error set - default off*/
        if (ts->continue_on_error) {
                json_object_add_value_int(root, "total_err", ts->total_err_count);
-               json_object_add_value_int(root, "total_err", ts->first_error);
+               json_object_add_value_int(root, "first_error", ts->first_error);
+       }
+
+       if (ts->latency_depth) {
+               json_object_add_value_int(root, "latency_depth", ts->latency_depth);
+               json_object_add_value_int(root, "latency_target", ts->latency_target);
+               json_object_add_value_float(root, "latency_percentile", ts->latency_percentile.u.f);
+               json_object_add_value_int(root, "latency_window", ts->latency_window);
        }
 
        /* Additional output if description is set */
        if (strlen(ts->description))
                json_object_add_value_string(root, "desc", ts->description);
 
+       if (ts->nr_block_infos) {
+               /* Block error histogram and types */
+               int len;
+               unsigned int *percentiles = NULL;
+               unsigned int block_state_counts[BLOCK_STATE_COUNT];
+
+               len = calc_block_percentiles(ts->nr_block_infos, ts->block_infos,
+                                            ts->percentile_list,
+                                            &percentiles, block_state_counts);
+
+               if (len) {
+                       struct json_object *block, *percentile_object, *states;
+                       int state, i;
+                       block = json_create_object();
+                       json_object_add_value_object(root, "block", block);
+
+                       percentile_object = json_create_object();
+                       json_object_add_value_object(block, "percentiles",
+                                                    percentile_object);
+                       for (i = 0; i < len; i++) {
+                               char buf[20];
+                               snprintf(buf, sizeof(buf), "%f",
+                                        ts->percentile_list[i].u.f);
+                               json_object_add_value_int(percentile_object,
+                                                         (const char *)buf,
+                                                         percentiles[i]);
+                       }
+
+                       states = json_create_object();
+                       json_object_add_value_object(block, "states", states);
+                       for (state = 0; state < BLOCK_STATE_COUNT; state++) {
+                               json_object_add_value_int(states,
+                                       block_state_names[state],
+                                       block_state_counts[state]);
+                       }
+                       free(percentiles);
+               }
+       }
+
        return root;
 }
 
@@ -991,6 +1208,21 @@ static void show_thread_status_terse(struct thread_stat *ts,
                log_err("fio: bad terse version!? %d\n", terse_version);
 }
 
+struct json_object *show_thread_status(struct thread_stat *ts,
+                                      struct group_run_stats *rs)
+{
+       struct json_object *ret = NULL;
+
+       if (output_format & FIO_OUTPUT_TERSE)
+               show_thread_status_terse(ts, rs);
+       if (output_format & FIO_OUTPUT_JSON)
+               ret = show_thread_status_json(ts, rs);
+       if (output_format & FIO_OUTPUT_NORMAL)
+               show_thread_status_normal(ts, rs);
+
+       return ret;
+}
+
 static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
 {
        double mean, S;
@@ -1044,6 +1276,10 @@ void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
                dst->agg[i] += src->agg[i];
        }
 
+       if (!dst->kb_base)
+               dst->kb_base = src->kb_base;
+       if (!dst->unit_base)
+               dst->unit_base = src->unit_base;
 }
 
 void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
@@ -1095,9 +1331,11 @@ void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
                if (!dst->unified_rw_rep) {
                        dst->total_io_u[k] += src->total_io_u[k];
                        dst->short_io_u[k] += src->short_io_u[k];
+                       dst->drop_io_u[k] += src->drop_io_u[k];
                } else {
                        dst->total_io_u[0] += src->total_io_u[k];
                        dst->short_io_u[0] += src->short_io_u[k];
+                       dst->drop_io_u[0] += src->drop_io_u[k];
                }
        }
 
@@ -1141,17 +1379,16 @@ void init_thread_stat(struct thread_stat *ts)
        ts->groupid = -1;
 }
 
-void show_run_stats(void)
+void __show_run_stats(void)
 {
        struct group_run_stats *runstats, *rs;
        struct thread_data *td;
        struct thread_stat *threadstats, *ts;
-       int i, j, nr_ts, last_ts, idx;
+       int i, j, k, nr_ts, last_ts, idx;
        int kb_base_warned = 0;
        int unit_base_warned = 0;
        struct json_object *root = NULL;
        struct json_array *array = NULL;
-
        runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
 
        for (i = 0; i < groupid + 1; i++)
@@ -1205,12 +1442,12 @@ void show_run_stats(void)
                        /*
                         * These are per-group shared already
                         */
-                       strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE);
+                       strncpy(ts->name, td->o.name, FIO_JOBNAME_SIZE - 1);
                        if (td->o.description)
                                strncpy(ts->description, td->o.description,
-                                               FIO_JOBNAME_SIZE);
+                                               FIO_JOBDESC_SIZE - 1);
                        else
-                               memset(ts->description, 0, FIO_JOBNAME_SIZE);
+                               memset(ts->description, 0, FIO_JOBDESC_SIZE);
 
                        /*
                         * If multiple entries in this group, this is
@@ -1244,13 +1481,24 @@ void show_run_stats(void)
                        if (!td->error && td->o.continue_on_error &&
                            td->first_error) {
                                ts->error = td->first_error;
-                               strcpy(ts->verror, td->verror);
+                               ts->verror[sizeof(ts->verror) - 1] = '\0';
+                               strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
                        } else  if (td->error) {
                                ts->error = td->error;
-                               strcpy(ts->verror, td->verror);
+                               ts->verror[sizeof(ts->verror) - 1] = '\0';
+                               strncpy(ts->verror, td->verror, sizeof(ts->verror) - 1);
                        }
                }
 
+               ts->latency_depth = td->latency_qd;
+               ts->latency_target = td->o.latency_target;
+               ts->latency_percentile = td->o.latency_percentile;
+               ts->latency_window = td->o.latency_window;
+
+               ts->nr_block_infos = td->ts.nr_block_infos;
+               for (k = 0; k < ts->nr_block_infos; k++)
+                       ts->block_infos[k] = td->ts.block_infos[k];
+
                sum_thread_stats(ts, &td->ts, idx);
        }
 
@@ -1303,11 +1551,21 @@ void show_run_stats(void)
        /*
         * don't overwrite last signal output
         */
-       if (output_format == FIO_OUTPUT_NORMAL)
+       if (output_format & FIO_OUTPUT_NORMAL)
                log_info("\n");
-       else if (output_format == FIO_OUTPUT_JSON) {
+       if (output_format & FIO_OUTPUT_JSON) {
+               char time_buf[32];
+               time_t time_p;
+
+               time(&time_p);
+               os_ctime_r((const time_t *) &time_p, time_buf,
+                               sizeof(time_buf));
+               time_buf[strlen(time_buf) - 1] = '\0';
+
                root = json_create_object();
                json_object_add_value_string(root, "fio version", fio_version_string);
+               json_object_add_value_int(root, "timestamp", time_p);
+               json_object_add_value_string(root, "time", time_buf);
                array = json_create_array();
                json_object_add_value_array(root, "jobs", array);
        }
@@ -1318,15 +1576,18 @@ void show_run_stats(void)
 
                if (is_backend)
                        fio_server_send_ts(ts, rs);
-               else if (output_format == FIO_OUTPUT_TERSE)
-                       show_thread_status_terse(ts, rs);
-               else if (output_format == FIO_OUTPUT_JSON) {
-                       struct json_object *tmp = show_thread_status_json(ts, rs);
-                       json_array_add_value_object(array, tmp);
-               } else
-                       show_thread_status(ts, rs);
+               else {
+                       if (output_format & FIO_OUTPUT_TERSE)
+                               show_thread_status_terse(ts, rs);
+                       if (output_format & FIO_OUTPUT_JSON) {
+                               struct json_object *tmp = show_thread_status_json(ts, rs);
+                               json_array_add_value_object(array, tmp);
+                       }
+                       if (output_format & FIO_OUTPUT_NORMAL)
+                               show_thread_status_normal(ts, rs);
+               }
        }
-       if (output_format == FIO_OUTPUT_JSON) {
+       if (output_format & FIO_OUTPUT_JSON) {
                /* disk util stats, if any */
                show_disk_util(1, root);
 
@@ -1343,28 +1604,48 @@ void show_run_stats(void)
                rs->groupid = i;
                if (is_backend)
                        fio_server_send_gs(rs);
-               else if (output_format == FIO_OUTPUT_NORMAL)
+               else if (output_format & FIO_OUTPUT_NORMAL)
                        show_group_stats(rs);
        }
 
        if (is_backend)
                fio_server_send_du();
-       else if (output_format == FIO_OUTPUT_NORMAL) {
+       else if (output_format & FIO_OUTPUT_NORMAL) {
                show_disk_util(0, NULL);
                show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
        }
 
+       if ( !(output_format & FIO_OUTPUT_TERSE) && append_terse_output) {
+               log_info("\nAdditional Terse Output:\n");
+
+               for (i = 0; i < nr_ts; i++) {
+                       ts = &threadstats[i];
+                       rs = &runstats[ts->groupid];
+                       show_thread_status_terse(ts, rs);
+               }
+       }
+
+       log_info_flush();
        free(runstats);
        free(threadstats);
 }
 
-static void *__show_running_run_stats(void fio_unused *arg)
+void show_run_stats(void)
+{
+       fio_mutex_down(stat_mutex);
+       __show_run_stats();
+       fio_mutex_up(stat_mutex);
+}
+
+void __show_running_run_stats(void)
 {
        struct thread_data *td;
        unsigned long long *rt;
        struct timeval tv;
        int i;
 
+       fio_mutex_down(stat_mutex);
+
        rt = malloc(thread_number * sizeof(unsigned long long));
        fio_gettime(&tv, NULL);
 
@@ -1385,6 +1666,8 @@ static void *__show_running_run_stats(void fio_unused *arg)
        }
 
        for_each_td(td, i) {
+               if (td->runstate >= TD_EXITED)
+                       continue;
                if (td->rusage_sem) {
                        td->update_rusage = 1;
                        fio_mutex_down(td->rusage_sem);
@@ -1392,7 +1675,7 @@ static void *__show_running_run_stats(void fio_unused *arg)
                td->update_rusage = 0;
        }
 
-       show_run_stats();
+       __show_run_stats();
 
        for_each_td(td, i) {
                if (td_read(td) && td->io_bytes[DDIR_READ])
@@ -1404,20 +1687,64 @@ static void *__show_running_run_stats(void fio_unused *arg)
        }
 
        free(rt);
-       return NULL;
+       fio_mutex_up(stat_mutex);
 }
 
-/*
- * Called from signal handler. It _should_ be safe to just run this inline
- * in the sig handler, but we should be disturbing the system less by just
- * creating a thread to do it.
- */
-void show_running_run_stats(void)
+static int status_interval_init;
+static struct timeval status_time;
+static int status_file_disabled;
+
+#define FIO_STATUS_FILE                "fio-dump-status"
+
+static int check_status_file(void)
 {
-       pthread_t thread;
+       struct stat sb;
+       const char *temp_dir;
+       char fio_status_file_path[PATH_MAX];
+
+       if (status_file_disabled)
+               return 0;
+
+       temp_dir = getenv("TMPDIR");
+       if (temp_dir == NULL) {
+               temp_dir = getenv("TEMP");
+               if (temp_dir && strlen(temp_dir) >= PATH_MAX)
+                       temp_dir = NULL;
+       }
+       if (temp_dir == NULL)
+               temp_dir = "/tmp";
+
+       snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
 
-       pthread_create(&thread, NULL, __show_running_run_stats, NULL);
-       pthread_detach(thread);
+       if (stat(fio_status_file_path, &sb))
+               return 0;
+
+       if (unlink(fio_status_file_path) < 0) {
+               log_err("fio: failed to unlink %s: %s\n", fio_status_file_path,
+                                                       strerror(errno));
+               log_err("fio: disabling status file updates\n");
+               status_file_disabled = 1;
+       }
+
+       return 1;
+}
+
+void check_for_running_stats(void)
+{
+       if (status_interval) {
+               if (!status_interval_init) {
+                       fio_gettime(&status_time, NULL);
+                       status_interval_init = 1;
+               } else if (mtime_since_now(&status_time) >= status_interval) {
+                       show_running_run_stats();
+                       fio_gettime(&status_time, NULL);
+                       return;
+               }
+       }
+       if (check_status_file()) {
+               show_running_run_stats();
+               return;
+       }
 }
 
 static inline void add_stat_sample(struct io_stat *is, unsigned long data)
@@ -1441,24 +1768,55 @@ static inline void add_stat_sample(struct io_stat *is, unsigned long data)
 
 static void __add_log_sample(struct io_log *iolog, unsigned long val,
                             enum fio_ddir ddir, unsigned int bs,
-                            unsigned long t)
+                            unsigned long t, uint64_t offset)
 {
-       const int nr_samples = iolog->nr_samples;
+       uint64_t nr_samples = iolog->nr_samples;
+       struct io_sample *s;
+
+       if (iolog->disabled)
+               return;
 
        if (!iolog->nr_samples)
                iolog->avg_last = t;
 
        if (iolog->nr_samples == iolog->max_samples) {
-               int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
+               size_t new_size;
+               void *new_log;
+
+               new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
+
+               if (iolog->log_gz && (new_size > iolog->log_gz)) {
+                       if (iolog_flush(iolog, 0)) {
+                               log_err("fio: failed flushing iolog! Will stop logging.\n");
+                               iolog->disabled = 1;
+                               return;
+                       }
+                       nr_samples = iolog->nr_samples;
+               } else {
+                       new_log = realloc(iolog->log, new_size);
+                       if (!new_log) {
+                               log_err("fio: failed extending iolog! Will stop logging.\n");
+                               iolog->disabled = 1;
+                               return;
+                       }
+                       iolog->log = new_log;
+                       iolog->max_samples <<= 1;
+               }
+       }
+
+       s = get_sample(iolog, nr_samples);
+
+       s->val = val;
+       s->time = t;
+       io_sample_set_ddir(iolog, s, ddir);
+       s->bs = bs;
+
+       if (iolog->log_offset) {
+               struct io_sample_offset *so = (void *) s;
 
-               iolog->log = realloc(iolog->log, new_size);
-               iolog->max_samples <<= 1;
+               so->offset = offset;
        }
 
-       iolog->log[nr_samples].val = val;
-       iolog->log[nr_samples].time = t;
-       iolog->log[nr_samples].ddir = ddir;
-       iolog->log[nr_samples].bs = bs;
        iolog->nr_samples++;
 }
 
@@ -1468,9 +1826,76 @@ static inline void reset_io_stat(struct io_stat *ios)
        ios->mean.u.f = ios->S.u.f = 0;
 }
 
+void reset_io_stats(struct thread_data *td)
+{
+       struct thread_stat *ts = &td->ts;
+       int i, j;
+
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+               reset_io_stat(&ts->clat_stat[i]);
+               reset_io_stat(&ts->slat_stat[i]);
+               reset_io_stat(&ts->lat_stat[i]);
+               reset_io_stat(&ts->bw_stat[i]);
+               reset_io_stat(&ts->iops_stat[i]);
+
+               ts->io_bytes[i] = 0;
+               ts->runtime[i] = 0;
+
+               for (j = 0; j < FIO_IO_U_PLAT_NR; j++)
+                       ts->io_u_plat[i][j] = 0;
+       }
+
+       for (i = 0; i < FIO_IO_U_MAP_NR; i++) {
+               ts->io_u_map[i] = 0;
+               ts->io_u_submit[i] = 0;
+               ts->io_u_complete[i] = 0;
+               ts->io_u_lat_u[i] = 0;
+               ts->io_u_lat_m[i] = 0;
+               ts->total_submit = 0;
+               ts->total_complete = 0;
+       }
+
+       for (i = 0; i < 3; i++) {
+               ts->total_io_u[i] = 0;
+               ts->short_io_u[i] = 0;
+               ts->drop_io_u[i] = 0;
+       }
+}
+
+static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
+{
+       /*
+        * Note an entry in the log. Use the mean from the logged samples,
+        * making sure to properly round up. Only write a log entry if we
+        * had actual samples done.
+        */
+       if (iolog->avg_window[DDIR_READ].samples) {
+               unsigned long mr;
+
+               mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
+               __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed, 0);
+       }
+       if (iolog->avg_window[DDIR_WRITE].samples) {
+               unsigned long mw;
+
+               mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
+               __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
+       }
+       if (iolog->avg_window[DDIR_TRIM].samples) {
+               unsigned long mw;
+
+               mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
+               __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed, 0);
+       }
+
+       reset_io_stat(&iolog->avg_window[DDIR_READ]);
+       reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
+       reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
+}
+
 static void add_log_sample(struct thread_data *td, struct io_log *iolog,
                           unsigned long val, enum fio_ddir ddir,
-                          unsigned int bs)
+                          unsigned int bs, uint64_t offset)
 {
        unsigned long elapsed, this_window;
 
@@ -1483,7 +1908,7 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
         * If no time averaging, just add the log sample.
         */
        if (!iolog->avg_msec) {
-               __add_log_sample(iolog, val, ddir, bs, elapsed);
+               __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
                return;
        }
 
@@ -1501,35 +1926,27 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
        if (this_window < iolog->avg_msec)
                return;
 
-       /*
-        * Note an entry in the log. Use the mean from the logged samples,
-        * making sure to properly round up. Only write a log entry if we
-        * had actual samples done.
-        */
-       if (iolog->avg_window[DDIR_READ].samples) {
-               unsigned long mr;
-
-               mr = iolog->avg_window[DDIR_READ].mean.u.f + 0.50;
-               __add_log_sample(iolog, mr, DDIR_READ, 0, elapsed);
-       }
-       if (iolog->avg_window[DDIR_WRITE].samples) {
-               unsigned long mw;
+       _add_stat_to_log(iolog, elapsed);
 
-               mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
-               __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed);
-       }
-       if (iolog->avg_window[DDIR_TRIM].samples) {
-               unsigned long mw;
+       iolog->avg_last = elapsed;
+}
 
-               mw = iolog->avg_window[DDIR_TRIM].mean.u.f + 0.50;
-               __add_log_sample(iolog, mw, DDIR_TRIM, 0, elapsed);
-       }
+void finalize_logs(struct thread_data *td)
+{
+       unsigned long elapsed;
 
+       elapsed = mtime_since_now(&td->epoch);
 
-       reset_io_stat(&iolog->avg_window[DDIR_READ]);
-       reset_io_stat(&iolog->avg_window[DDIR_WRITE]);
-       reset_io_stat(&iolog->avg_window[DDIR_TRIM]);
-       iolog->avg_last = elapsed;
+       if (td->clat_log)
+               _add_stat_to_log(td->clat_log, elapsed);
+       if (td->slat_log)
+               _add_stat_to_log(td->slat_log, elapsed);
+       if (td->lat_log)
+               _add_stat_to_log(td->lat_log, elapsed);
+       if (td->bw_log)
+               _add_stat_to_log(td->bw_log, elapsed);
+       if (td->iops_log)
+               _add_stat_to_log(td->iops_log, elapsed);
 }
 
 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
@@ -1540,7 +1957,7 @@ void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
                return;
 
        iolog = agg_io_log[ddir];
-       __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis());
+       __add_log_sample(iolog, val, ddir, bs, mtime_since_genesis(), 0);
 }
 
 static void add_clat_percentile_sample(struct thread_stat *ts,
@@ -1553,48 +1970,60 @@ static void add_clat_percentile_sample(struct thread_stat *ts,
 }
 
 void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
-                    unsigned long usec, unsigned int bs)
+                    unsigned long usec, unsigned int bs, uint64_t offset)
 {
        struct thread_stat *ts = &td->ts;
 
        if (!ddir_rw(ddir))
                return;
 
+       td_io_u_lock(td);
+
        add_stat_sample(&ts->clat_stat[ddir], usec);
 
        if (td->clat_log)
-               add_log_sample(td, td->clat_log, usec, ddir, bs);
+               add_log_sample(td, td->clat_log, usec, ddir, bs, offset);
 
        if (ts->clat_percentiles)
                add_clat_percentile_sample(ts, usec, ddir);
+
+       td_io_u_unlock(td);
 }
 
 void add_slat_sample(struct thread_data *td, enum fio_ddir ddir,
-                    unsigned long usec, unsigned int bs)
+                    unsigned long usec, unsigned int bs, uint64_t offset)
 {
        struct thread_stat *ts = &td->ts;
 
        if (!ddir_rw(ddir))
                return;
 
+       td_io_u_lock(td);
+
        add_stat_sample(&ts->slat_stat[ddir], usec);
 
        if (td->slat_log)
-               add_log_sample(td, td->slat_log, usec, ddir, bs);
+               add_log_sample(td, td->slat_log, usec, ddir, bs, offset);
+
+       td_io_u_unlock(td);
 }
 
 void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
-                   unsigned long usec, unsigned int bs)
+                   unsigned long usec, unsigned int bs, uint64_t offset)
 {
        struct thread_stat *ts = &td->ts;
 
        if (!ddir_rw(ddir))
                return;
 
+       td_io_u_lock(td);
+
        add_stat_sample(&ts->lat_stat[ddir], usec);
 
        if (td->lat_log)
-               add_log_sample(td, td->lat_log, usec, ddir, bs);
+               add_log_sample(td, td->lat_log, usec, ddir, bs, offset);
+
+       td_io_u_unlock(td);
 }
 
 void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
@@ -1610,6 +2039,8 @@ void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
        if (spent < td->o.bw_avg_time)
                return;
 
+       td_io_u_lock(td);
+
        /*
         * Compute both read and write rates for the interval.
         */
@@ -1620,19 +2051,24 @@ void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
                if (!delta)
                        continue; /* No entries for interval */
 
-               rate = delta * 1000 / spent / 1024;
+               if (spent)
+                       rate = delta * 1000 / spent / 1024;
+               else
+                       rate = 0;
+
                add_stat_sample(&ts->bw_stat[ddir], rate);
 
                if (td->bw_log)
-                       add_log_sample(td, td->bw_log, rate, ddir, bs);
+                       add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
 
                td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
        }
 
        fio_gettime(&td->bw_sample_time, NULL);
+       td_io_u_unlock(td);
 }
 
-void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
+void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
                     struct timeval *t)
 {
        struct thread_stat *ts = &td->ts;
@@ -1645,6 +2081,8 @@ void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
        if (spent < td->o.iops_avg_time)
                return;
 
+       td_io_u_lock(td);
+
        /*
         * Compute both read and write rates for the interval.
         */
@@ -1655,14 +2093,54 @@ void add_iops_sample(struct thread_data *td, enum fio_ddir ddir,
                if (!delta)
                        continue; /* No entries for interval */
 
-               iops = (delta * 1000) / spent;
+               if (spent)
+                       iops = (delta * 1000) / spent;
+               else
+                       iops = 0;
+
                add_stat_sample(&ts->iops_stat[ddir], iops);
 
                if (td->iops_log)
-                       add_log_sample(td, td->iops_log, iops, ddir, 0);
+                       add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
 
                td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
        }
 
        fio_gettime(&td->iops_sample_time, NULL);
+       td_io_u_unlock(td);
+}
+
+void stat_init(void)
+{
+       stat_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
+}
+
+void stat_exit(void)
+{
+       /*
+        * When we have the mutex, we know out-of-band access to it
+        * have ended.
+        */
+       fio_mutex_down(stat_mutex);
+       fio_mutex_remove(stat_mutex);
+}
+
+/*
+ * Called from signal handler. Wake up status thread.
+ */
+void show_running_run_stats(void)
+{
+       helper_do_stat = 1;
+       pthread_cond_signal(&helper_cond);
+}
+
+uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)
+{
+       /* Ignore io_u's which span multiple blocks--they will just get
+        * inaccurate counts. */
+       int idx = (io_u->offset - io_u->file->file_offset)
+                       / td->o.bs[DDIR_TRIM];
+       uint32_t *info = &td->ts.block_infos[idx];
+       assert(idx < td->ts.nr_block_infos);
+       return info;
 }