steadystate: get rid of ->ss_pct and encode this information in ->state via __FIO_SS_PCT
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 3d7dcde10681a99dd56a6a8dd2663cf774aae4e7..a42ee9b857f1d1745b9425749c6360f60e25a1e1 100644 (file)
--- a/stat.c
+++ b/stat.c
 #include "idletime.h"
 #include "lib/pow2.h"
 #include "lib/output_buffer.h"
+#include "helper_thread.h"
+#include "smalloc.h"
+
+#define LOG_MSEC_SLACK 10
 
 struct fio_mutex *stat_mutex;
 
+void clear_rusage_stat(struct thread_data *td)
+{
+       struct thread_stat *ts = &td->ts;
+
+       fio_getrusage(&td->ru_start);
+       ts->usr_time = ts->sys_time = 0;
+       ts->ctx = 0;
+       ts->minf = ts->majf = 0;
+}
+
 void update_rusage_stat(struct thread_data *td)
 {
        struct thread_stat *ts = &td->ts;
@@ -832,7 +846,7 @@ static void add_ddir_status_json(struct thread_stat *ts,
        unsigned int len, minv, maxv;
        int i;
        const char *ddirname[] = {"read", "write", "trim"};
-       struct json_object *dir_object, *tmp_object, *percentile_object;
+       struct json_object *dir_object, *tmp_object, *percentile_object, *clat_bins_object;
        char buf[120];
        double p_of_agg = 100.0;
 
@@ -903,6 +917,18 @@ static void add_ddir_status_json(struct thread_stat *ts,
                json_object_add_value_int(percentile_object, (const char *)buf, ovals[i]);
        }
 
+       if (output_format & FIO_OUTPUT_JSON_PLUS) {
+               clat_bins_object = json_create_object();
+               json_object_add_value_object(tmp_object, "bins", clat_bins_object);
+               for(i = 0; i < FIO_IO_U_PLAT_NR; i++) {
+                       snprintf(buf, sizeof(buf), "%d", i);
+                       json_object_add_value_int(clat_bins_object, (const char *)buf, ts->io_u_plat[ddir][i]);
+               }
+               json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_BITS", FIO_IO_U_PLAT_BITS);
+               json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_VAL", FIO_IO_U_PLAT_VAL);
+               json_object_add_value_int(clat_bins_object, "FIO_IO_U_PLAT_NR", FIO_IO_U_PLAT_NR);
+       }
+
        if (!calc_lat(&ts->lat_stat[ddir], &min, &max, &mean, &dev)) {
                min = max = 0;
                mean = dev = 0.0;
@@ -1064,8 +1090,34 @@ static void show_thread_status_terse_v3_v4(struct thread_stat *ts,
        log_buf(out, "\n");
 }
 
+void json_add_job_opts(struct json_object *root, const char *name,
+                      struct flist_head *opt_list, bool num_jobs)
+{
+       struct json_object *dir_object;
+       struct flist_head *entry;
+       struct print_option *p;
+
+       if (flist_empty(opt_list))
+               return;
+
+       dir_object = json_create_object();
+       json_object_add_value_object(root, name, dir_object);
+
+       flist_for_each(entry, opt_list) {
+               const char *pos = "";
+
+               p = flist_entry(entry, struct print_option, list);
+               if (!num_jobs && !strcmp(p->name, "numjobs"))
+                       continue;
+               if (p->value)
+                       pos = p->value;
+               json_object_add_value_string(dir_object, p->name, pos);
+       }
+}
+
 static struct json_object *show_thread_status_json(struct thread_stat *ts,
-                                   struct group_run_stats *rs)
+                                                  struct group_run_stats *rs,
+                                                  struct flist_head *opt_list)
 {
        struct json_object *root, *tmp;
        struct jobs_eta *je;
@@ -1082,10 +1134,14 @@ static struct json_object *show_thread_status_json(struct thread_stat *ts,
        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);
+       je = get_jobs_eta(true, &size);
+       if (je) {
+               json_object_add_value_int(root, "eta", je->eta_sec);
+               json_object_add_value_int(root, "elapsed", je->elapsed_sec);
+       }
 
+       if (opt_list)
+               json_add_job_opts(root, "job options", opt_list, true);
 
        add_ddir_status_json(ts, rs, DDIR_READ, root);
        add_ddir_status_json(ts, rs, DDIR_WRITE, root);
@@ -1172,7 +1228,7 @@ static struct json_object *show_thread_status_json(struct thread_stat *ts,
 
                if (len) {
                        struct json_object *block, *percentile_object, *states;
-                       int state, i;
+                       int state;
                        block = json_create_object();
                        json_object_add_value_object(root, "block", block);
 
@@ -1199,6 +1255,65 @@ static struct json_object *show_thread_status_json(struct thread_stat *ts,
                }
        }
 
+       /* steady state detection; move this behind json+? */
+       if (ts->ss) {
+               struct json_object *data;
+               struct json_array *iops, *bw;
+               struct steadystate_data *ss = ts->ss;
+               unsigned long long sum_iops, sum_bw;
+               double mean_iops, mean_bw;
+               int i, j, k;
+               char ss_buf[64];
+
+               snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
+                       ss->state & __FIO_SS_IOPS ? "iops" : "bw",
+                       ss->state & __FIO_SS_SLOPE ? "_slope" : "",
+                       (float) ss->limit,
+                       ss->state & __FIO_SS_PCT ? "%" : "");
+
+               tmp = json_create_object();
+               json_object_add_value_object(root, "steadystate", tmp);
+               json_object_add_value_string(tmp, "ss", ss_buf);
+               json_object_add_value_int(tmp, "duration", (int)ss->dur);
+               json_object_add_value_int(tmp, "steadystate_ramptime", ss->ramp_time / 1000000L);
+               json_object_add_value_int(tmp, "attained", (ss->state & __FIO_SS_ATTAINED) > 0);
+
+               snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ss->criterion,
+                       ss->state & __FIO_SS_PCT ? "%" : "");
+               json_object_add_value_string(tmp, "criterion", ss_buf);
+               json_object_add_value_float(tmp, "max_deviation", ss->deviation);
+               json_object_add_value_float(tmp, "slope", ss->slope);
+
+               data = json_create_object();
+               json_object_add_value_object(tmp, "data", data);
+               bw = json_create_array();
+               iops = json_create_array();
+
+               /*
+               ** if ss was attained or the buffer is not full,
+               ** ss->head points to the first element in the list.
+               ** otherwise it actually points to the second element
+               ** in the list
+               */
+               if ((ss->state & __FIO_SS_ATTAINED) || ss->sum_y == 0)
+                       j = ss->head;
+               else
+                       j = ss->head == 0 ? ss->dur - 1 : ss->head - 1;
+               for (i = 0, sum_iops = 0, sum_bw = 0; i < ss->dur; i++) {
+                       k = (j + i) % ss->dur;
+                       sum_bw += ss->bw_data[k];
+                       sum_iops += ss->iops_data[k];
+                       json_array_add_value_int(bw, ss->bw_data[k]);
+                       json_array_add_value_int(iops, ss->iops_data[k]);
+               }
+               mean_bw = (double) sum_bw / ss->dur;
+               mean_iops = (double) sum_iops / ss->dur;
+               json_object_add_value_float(data, "bw_mean", mean_bw);
+               json_object_add_value_float(data, "iops_mean", mean_iops);
+               json_object_add_value_array(data, "iops", iops);
+               json_object_add_value_array(data, "bw", bw);
+       }
+
        return root;
 }
 
@@ -1216,6 +1331,7 @@ static void show_thread_status_terse(struct thread_stat *ts,
 
 struct json_object *show_thread_status(struct thread_stat *ts,
                                       struct group_run_stats *rs,
+                                      struct flist_head *opt_list,
                                       struct buf_output *out)
 {
        struct json_object *ret = NULL;
@@ -1223,14 +1339,14 @@ struct json_object *show_thread_status(struct thread_stat *ts,
        if (output_format & FIO_OUTPUT_TERSE)
                show_thread_status_terse(ts, rs,  out);
        if (output_format & FIO_OUTPUT_JSON)
-               ret = show_thread_status_json(ts, rs);
+               ret = show_thread_status_json(ts, rs, opt_list);
        if (output_format & FIO_OUTPUT_NORMAL)
                show_thread_status_normal(ts, rs,  out);
 
        return ret;
 }
 
-static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
+static void sum_stat(struct io_stat *dst, struct io_stat *src, bool first)
 {
        double mean, S;
 
@@ -1245,7 +1361,7 @@ static void sum_stat(struct io_stat *dst, struct io_stat *src, int nr)
         * <http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
         *  #Parallel_algorithm>
         */
-       if (nr == 1) {
+       if (first) {
                mean = src->mean.u.f;
                S = src->S.u.f;
        } else {
@@ -1289,31 +1405,38 @@ void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src)
                dst->unit_base = src->unit_base;
 }
 
-void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr)
+void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src,
+                     bool first)
 {
        int l, k;
 
        for (l = 0; l < DDIR_RWDIR_CNT; l++) {
                if (!dst->unified_rw_rep) {
-                       sum_stat(&dst->clat_stat[l], &src->clat_stat[l], nr);
-                       sum_stat(&dst->slat_stat[l], &src->slat_stat[l], nr);
-                       sum_stat(&dst->lat_stat[l], &src->lat_stat[l], nr);
-                       sum_stat(&dst->bw_stat[l], &src->bw_stat[l], nr);
+                       sum_stat(&dst->clat_stat[l], &src->clat_stat[l], first);
+                       sum_stat(&dst->slat_stat[l], &src->slat_stat[l], first);
+                       sum_stat(&dst->lat_stat[l], &src->lat_stat[l], first);
+                       sum_stat(&dst->bw_stat[l], &src->bw_stat[l], first);
 
                        dst->io_bytes[l] += src->io_bytes[l];
 
                        if (dst->runtime[l] < src->runtime[l])
                                dst->runtime[l] = src->runtime[l];
                } else {
-                       sum_stat(&dst->clat_stat[0], &src->clat_stat[l], nr);
-                       sum_stat(&dst->slat_stat[0], &src->slat_stat[l], nr);
-                       sum_stat(&dst->lat_stat[0], &src->lat_stat[l], nr);
-                       sum_stat(&dst->bw_stat[0], &src->bw_stat[l], nr);
+                       sum_stat(&dst->clat_stat[0], &src->clat_stat[l], first);
+                       sum_stat(&dst->slat_stat[0], &src->slat_stat[l], first);
+                       sum_stat(&dst->lat_stat[0], &src->lat_stat[l], first);
+                       sum_stat(&dst->bw_stat[0], &src->bw_stat[l], first);
 
                        dst->io_bytes[0] += src->io_bytes[l];
 
                        if (dst->runtime[0] < src->runtime[l])
                                dst->runtime[0] = src->runtime[l];
+
+                       /*
+                        * We're summing to the same destination, so override
+                        * 'first' after the first iteration of the loop
+                        */
+                       first = false;
                }
        }
 
@@ -1397,11 +1520,10 @@ void __show_run_stats(void)
        struct json_object *root = NULL;
        struct json_array *array = NULL;
        struct buf_output output[FIO_OUTPUT_NR];
+       struct flist_head **opt_lists;
 
        runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
 
-       memset(output, 0, sizeof(output));
-
        for (i = 0; i < groupid + 1; i++)
                init_group_run_stat(&runstats[i]);
 
@@ -1424,9 +1546,12 @@ void __show_run_stats(void)
        }
 
        threadstats = malloc(nr_ts * sizeof(struct thread_stat));
+       opt_lists = malloc(nr_ts * sizeof(struct flist_head *));
 
-       for (i = 0; i < nr_ts; i++)
+       for (i = 0; i < nr_ts; i++) {
                init_thread_stat(&threadstats[i]);
+               opt_lists[i] = NULL;
+       }
 
        j = 0;
        last_ts = -1;
@@ -1445,6 +1570,7 @@ void __show_run_stats(void)
                ts->clat_percentiles = td->o.clat_percentiles;
                ts->percentile_precision = td->o.percentile_precision;
                memcpy(ts->percentile_list, td->o.percentile_list, sizeof(td->o.percentile_list));
+               opt_lists[j] = &td->opt_list;
 
                idx++;
                ts->members++;
@@ -1510,13 +1636,20 @@ void __show_run_stats(void)
                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);
+               sum_thread_stats(ts, &td->ts, idx == 1);
+
+               if (td->o.ss_dur)
+                       ts->ss = &td->ss;
+               else
+                       ts->ss = NULL;
        }
 
        for (i = 0; i < nr_ts; i++) {
                unsigned long long bw;
 
                ts = &threadstats[i];
+               if (ts->groupid == -1)
+                       continue;
                rs = &runstats[ts->groupid];
                rs->kb_base = ts->kb_base;
                rs->unit_base = ts->unit_base;
@@ -1560,7 +1693,7 @@ void __show_run_stats(void)
        }
 
        for (i = 0; i < FIO_OUTPUT_NR; i++)
-               buf_output_init(&output[i], i);
+               buf_output_init(&output[i]);
 
        /*
         * don't overwrite last signal output
@@ -1568,40 +1701,52 @@ void __show_run_stats(void)
        if (output_format & FIO_OUTPUT_NORMAL)
                log_buf(&output[__FIO_OUTPUT_NORMAL], "\n");
        if (output_format & FIO_OUTPUT_JSON) {
+               struct thread_data *global;
                char time_buf[32];
-               time_t time_p;
+               struct timeval now;
+               unsigned long long ms_since_epoch;
+
+               gettimeofday(&now, NULL);
+               ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
+                                (unsigned long long)(now.tv_usec) / 1000;
 
-               time(&time_p);
-               os_ctime_r((const time_t *) &time_p, time_buf,
+               os_ctime_r((const time_t *) &now.tv_sec, 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_int(root, "timestamp", now.tv_sec);
+               json_object_add_value_int(root, "timestamp_ms", ms_since_epoch);
                json_object_add_value_string(root, "time", time_buf);
+               global = get_global_options();
+               json_add_job_opts(root, "global options", &global->opt_list, false);
                array = json_create_array();
                json_object_add_value_array(root, "jobs", array);
        }
 
+       if (is_backend)
+               fio_server_send_job_options(&get_global_options()->opt_list, -1U);
+
        for (i = 0; i < nr_ts; i++) {
                ts = &threadstats[i];
                rs = &runstats[ts->groupid];
 
-               if (is_backend)
+               if (is_backend) {
+                       fio_server_send_job_options(opt_lists[i], i);
                        fio_server_send_ts(ts, rs);
-               else {
+               else {
                        if (output_format & FIO_OUTPUT_TERSE)
                                show_thread_status_terse(ts, rs, &output[__FIO_OUTPUT_TERSE]);
                        if (output_format & FIO_OUTPUT_JSON) {
-                               struct json_object *tmp = show_thread_status_json(ts, rs);
+                               struct json_object *tmp = show_thread_status_json(ts, rs, opt_lists[i]);
                                json_array_add_value_object(array, tmp);
                        }
                        if (output_format & FIO_OUTPUT_NORMAL)
                                show_thread_status_normal(ts, rs, &output[__FIO_OUTPUT_NORMAL]);
                }
        }
-       if (output_format & FIO_OUTPUT_JSON) {
+       if (!is_backend && (output_format & FIO_OUTPUT_JSON)) {
                /* disk util stats, if any */
                show_disk_util(1, root, &output[__FIO_OUTPUT_JSON]);
 
@@ -1629,15 +1774,15 @@ void __show_run_stats(void)
                show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, &output[__FIO_OUTPUT_NORMAL]);
        }
 
-       for (i = 0; i < FIO_OUTPUT_NR; i++)
+       for (i = 0; i < FIO_OUTPUT_NR; i++) {
                buf_output_flush(&output[i]);
+               buf_output_free(&output[i]);
+       }
 
        log_info_flush();
        free(runstats);
        free(threadstats);
-
-       for (i = 0; i < FIO_OUTPUT_NR; i++)
-               buf_output_free(&output[i]);
+       free(opt_lists);
 }
 
 void show_run_stats(void)
@@ -1660,19 +1805,19 @@ void __show_running_run_stats(void)
        fio_gettime(&tv, NULL);
 
        for_each_td(td, i) {
-               rt[i] = mtime_since(&td->start, &tv);
-               if (td_read(td) && td->io_bytes[DDIR_READ])
-                       td->ts.runtime[DDIR_READ] += rt[i];
-               if (td_write(td) && td->io_bytes[DDIR_WRITE])
-                       td->ts.runtime[DDIR_WRITE] += rt[i];
-               if (td_trim(td) && td->io_bytes[DDIR_TRIM])
-                       td->ts.runtime[DDIR_TRIM] += rt[i];
-
                td->update_rusage = 1;
                td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
                td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
                td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
                td->ts.total_run_time = mtime_since(&td->epoch, &tv);
+
+               rt[i] = mtime_since(&td->start, &tv);
+               if (td_read(td) && td->ts.io_bytes[DDIR_READ])
+                       td->ts.runtime[DDIR_READ] += rt[i];
+               if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
+                       td->ts.runtime[DDIR_WRITE] += rt[i];
+               if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
+                       td->ts.runtime[DDIR_TRIM] += rt[i];
        }
 
        for_each_td(td, i) {
@@ -1688,11 +1833,11 @@ void __show_running_run_stats(void)
        __show_run_stats();
 
        for_each_td(td, i) {
-               if (td_read(td) && td->io_bytes[DDIR_READ])
+               if (td_read(td) && td->ts.io_bytes[DDIR_READ])
                        td->ts.runtime[DDIR_READ] -= rt[i];
-               if (td_write(td) && td->io_bytes[DDIR_WRITE])
+               if (td_write(td) && td->ts.io_bytes[DDIR_WRITE])
                        td->ts.runtime[DDIR_WRITE] -= rt[i];
-               if (td_trim(td) && td->io_bytes[DDIR_TRIM])
+               if (td_trim(td) && td->ts.io_bytes[DDIR_TRIM])
                        td->ts.runtime[DDIR_TRIM] -= rt[i];
        }
 
@@ -1776,58 +1921,184 @@ static inline void add_stat_sample(struct io_stat *is, unsigned long data)
        is->samples++;
 }
 
+/*
+ * Return a struct io_logs, which is added to the tail of the log
+ * list for 'iolog'.
+ */
+static struct io_logs *get_new_log(struct io_log *iolog)
+{
+       size_t new_size, new_samples;
+       struct io_logs *cur_log;
+
+       /*
+        * Cap the size at MAX_LOG_ENTRIES, so we don't keep doubling
+        * forever
+        */
+       if (!iolog->cur_log_max)
+               new_samples = DEF_LOG_ENTRIES;
+       else {
+               new_samples = iolog->cur_log_max * 2;
+               if (new_samples > MAX_LOG_ENTRIES)
+                       new_samples = MAX_LOG_ENTRIES;
+       }
+
+       new_size = new_samples * log_entry_sz(iolog);
+
+       cur_log = smalloc(sizeof(*cur_log));
+       if (cur_log) {
+               INIT_FLIST_HEAD(&cur_log->list);
+               cur_log->log = malloc(new_size);
+               if (cur_log->log) {
+                       cur_log->nr_samples = 0;
+                       cur_log->max_samples = new_samples;
+                       flist_add_tail(&cur_log->list, &iolog->io_logs);
+                       iolog->cur_log_max = new_samples;
+                       return cur_log;
+               }
+               sfree(cur_log);
+       }
+
+       return NULL;
+}
+
+/*
+ * Add and return a new log chunk, or return current log if big enough
+ */
+static struct io_logs *regrow_log(struct io_log *iolog)
+{
+       struct io_logs *cur_log;
+       int i;
+
+       if (!iolog || iolog->disabled)
+               goto disable;
+
+       cur_log = iolog_cur_log(iolog);
+       if (!cur_log) {
+               cur_log = get_new_log(iolog);
+               if (!cur_log)
+                       return NULL;
+       }
+
+       if (cur_log->nr_samples < cur_log->max_samples)
+               return cur_log;
+
+       /*
+        * No room for a new sample. If we're compressing on the fly, flush
+        * out the current chunk
+        */
+       if (iolog->log_gz) {
+               if (iolog_cur_flush(iolog, cur_log)) {
+                       log_err("fio: failed flushing iolog! Will stop logging.\n");
+                       return NULL;
+               }
+       }
+
+       /*
+        * Get a new log array, and add to our list
+        */
+       cur_log = get_new_log(iolog);
+       if (!cur_log) {
+               log_err("fio: failed extending iolog! Will stop logging.\n");
+               return NULL;
+       }
+
+       if (!iolog->pending || !iolog->pending->nr_samples)
+               return cur_log;
+
+       /*
+        * Flush pending items to new log
+        */
+       for (i = 0; i < iolog->pending->nr_samples; i++) {
+               struct io_sample *src, *dst;
+
+               src = get_sample(iolog, iolog->pending, i);
+               dst = get_sample(iolog, cur_log, i);
+               memcpy(dst, src, log_entry_sz(iolog));
+       }
+       cur_log->nr_samples = iolog->pending->nr_samples;
+
+       iolog->pending->nr_samples = 0;
+       return cur_log;
+disable:
+       if (iolog)
+               iolog->disabled = true;
+       return NULL;
+}
+
+void regrow_logs(struct thread_data *td)
+{
+       regrow_log(td->slat_log);
+       regrow_log(td->clat_log);
+       regrow_log(td->clat_hist_log);
+       regrow_log(td->lat_log);
+       regrow_log(td->bw_log);
+       regrow_log(td->iops_log);
+       td->flags &= ~TD_F_REGROW_LOGS;
+}
+
+static struct io_logs *get_cur_log(struct io_log *iolog)
+{
+       struct io_logs *cur_log;
+
+       cur_log = iolog_cur_log(iolog);
+       if (!cur_log) {
+               cur_log = get_new_log(iolog);
+               if (!cur_log)
+                       return NULL;
+       }
+
+       if (cur_log->nr_samples < cur_log->max_samples)
+               return cur_log;
+
+       /*
+        * Out of space. If we're in IO offload mode, or we're not doing
+        * per unit logging (hence logging happens outside of the IO thread
+        * as well), add a new log chunk inline. If we're doing inline
+        * submissions, flag 'td' as needing a log regrow and we'll take
+        * care of it on the submission side.
+        */
+       if (iolog->td->o.io_submit_mode == IO_MODE_OFFLOAD ||
+           !per_unit_log(iolog))
+               return regrow_log(iolog);
+
+       iolog->td->flags |= TD_F_REGROW_LOGS;
+       assert(iolog->pending->nr_samples < iolog->pending->max_samples);
+       return iolog->pending;
+}
+
 static void __add_log_sample(struct io_log *iolog, unsigned long val,
                             enum fio_ddir ddir, unsigned int bs,
                             unsigned long t, uint64_t offset)
 {
-       uint64_t nr_samples = iolog->nr_samples;
-       struct io_sample *s;
+       struct io_logs *cur_log;
 
        if (iolog->disabled)
                return;
-
-       if (!iolog->nr_samples)
+       if (flist_empty(&iolog->io_logs))
                iolog->avg_last = t;
 
-       if (iolog->nr_samples == iolog->max_samples) {
-               size_t new_size;
-               void *new_log;
+       cur_log = get_cur_log(iolog);
+       if (cur_log) {
+               struct io_sample *s;
 
-               new_size = 2 * iolog->max_samples * log_entry_sz(iolog);
+               s = get_sample(iolog, cur_log, cur_log->nr_samples);
 
-               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;
 
-       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;
 
-       if (iolog->log_offset) {
-               struct io_sample_offset *so = (void *) s;
+                       so->offset = offset;
+               }
 
-               so->offset = offset;
+               cur_log->nr_samples++;
+               return;
        }
 
-       iolog->nr_samples++;
+       iolog->disabled = true;
 }
 
 static inline void reset_io_stat(struct io_stat *ios)
@@ -1872,45 +2143,45 @@ void reset_io_stats(struct thread_data *td)
        }
 }
 
-static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed)
+static void __add_stat_to_log(struct io_log *iolog, enum fio_ddir ddir,
+                             unsigned long elapsed, bool log_max)
 {
        /*
         * 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;
+       if (iolog->avg_window[ddir].samples) {
+               unsigned long val;
 
-               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;
+               if (log_max)
+                       val = iolog->avg_window[ddir].max_val;
+               else
+                       val = iolog->avg_window[ddir].mean.u.f + 0.50;
 
-               mw = iolog->avg_window[DDIR_WRITE].mean.u.f + 0.50;
-               __add_log_sample(iolog, mw, DDIR_WRITE, 0, elapsed, 0);
+               __add_log_sample(iolog, val, ddir, 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]);
+}
 
-       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_stat_to_log(struct io_log *iolog, unsigned long elapsed,
+                            bool log_max)
+{
+       int ddir;
+
+       for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
+               __add_stat_to_log(iolog, ddir, elapsed, log_max);
 }
 
-static void add_log_sample(struct thread_data *td, struct io_log *iolog,
+static long add_log_sample(struct thread_data *td, struct io_log *iolog,
                           unsigned long val, enum fio_ddir ddir,
                           unsigned int bs, uint64_t offset)
 {
        unsigned long elapsed, this_window;
 
        if (!ddir_rw(ddir))
-               return;
+               return 0;
 
        elapsed = mtime_since_now(&td->epoch);
 
@@ -1919,7 +2190,7 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
         */
        if (!iolog->avg_msec) {
                __add_log_sample(iolog, val, ddir, bs, elapsed, offset);
-               return;
+               return 0;
        }
 
        /*
@@ -1933,30 +2204,37 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
         * need to do.
         */
        this_window = elapsed - iolog->avg_last;
-       if (this_window < iolog->avg_msec)
-               return;
+       if (elapsed < iolog->avg_last)
+               return iolog->avg_last - elapsed;
+       else if (this_window < iolog->avg_msec) {
+               int diff = iolog->avg_msec - this_window;
 
-       _add_stat_to_log(iolog, elapsed);
+               if (inline_log(iolog) || diff > LOG_MSEC_SLACK)
+                       return diff;
+       }
+
+       _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
 
-       iolog->avg_last = elapsed;
+       iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
+       return iolog->avg_msec;
 }
 
-void finalize_logs(struct thread_data *td)
+void finalize_logs(struct thread_data *td, bool unit_logs)
 {
        unsigned long elapsed;
 
        elapsed = mtime_since_now(&td->epoch);
 
-       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);
+       if (td->clat_log && unit_logs)
+               _add_stat_to_log(td->clat_log, elapsed, td->o.log_max != 0);
+       if (td->slat_log && unit_logs)
+               _add_stat_to_log(td->slat_log, elapsed, td->o.log_max != 0);
+       if (td->lat_log && unit_logs)
+               _add_stat_to_log(td->lat_log, elapsed, td->o.log_max != 0);
+       if (td->bw_log && (unit_logs == per_unit_log(td->bw_log)))
+               _add_stat_to_log(td->bw_log, elapsed, td->o.log_max != 0);
+       if (td->iops_log && (unit_logs == per_unit_log(td->iops_log)))
+               _add_stat_to_log(td->iops_log, elapsed, td->o.log_max != 0);
 }
 
 void add_agg_sample(unsigned long val, enum fio_ddir ddir, unsigned int bs)
@@ -1982,10 +2260,9 @@ 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, uint64_t offset)
 {
+       unsigned long elapsed, this_window;
        struct thread_stat *ts = &td->ts;
-
-       if (!ddir_rw(ddir))
-               return;
+       struct io_log *iolog = td->clat_hist_log;
 
        td_io_u_lock(td);
 
@@ -1997,6 +2274,43 @@ void add_clat_sample(struct thread_data *td, enum fio_ddir ddir,
        if (ts->clat_percentiles)
                add_clat_percentile_sample(ts, usec, ddir);
 
+       if (iolog && iolog->hist_msec) {
+               struct io_hist *hw = &iolog->hist_window[ddir];
+
+               hw->samples++;
+               elapsed = mtime_since_now(&td->epoch);
+               if (!hw->hist_last)
+                       hw->hist_last = elapsed;
+               this_window = elapsed - hw->hist_last;
+               
+               if (this_window >= iolog->hist_msec) {
+                       unsigned int *io_u_plat;
+                       unsigned int *dst;
+
+                       /*
+                        * Make a byte-for-byte copy of the latency histogram
+                        * stored in td->ts.io_u_plat[ddir], recording it in a
+                        * log sample. Note that the matching call to free() is
+                        * located in iolog.c after printing this sample to the
+                        * log file.
+                        */
+                       io_u_plat = (unsigned int *) td->ts.io_u_plat[ddir];
+                       dst = malloc(FIO_IO_U_PLAT_NR * sizeof(unsigned int));
+                       memcpy(dst, io_u_plat,
+                               FIO_IO_U_PLAT_NR * sizeof(unsigned int));
+                       __add_log_sample(iolog, (unsigned long )dst, ddir, bs,
+                                               elapsed, offset);
+
+                       /*
+                        * Update the last time we recorded as being now, minus
+                        * any drift in time we encountered before actually
+                        * making the record.
+                        */
+                       hw->hist_last = elapsed - (this_window - iolog->hist_msec);
+                       hw->samples = 0;
+               }
+       }
+
        td_io_u_unlock(td);
 }
 
@@ -2036,18 +2350,41 @@ void add_lat_sample(struct thread_data *td, enum fio_ddir ddir,
        td_io_u_unlock(td);
 }
 
-void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
-                  struct timeval *t)
+void add_bw_sample(struct thread_data *td, struct io_u *io_u,
+                  unsigned int bytes, unsigned long spent)
+{
+       struct thread_stat *ts = &td->ts;
+       unsigned long rate;
+
+       if (spent)
+               rate = bytes * 1000 / spent;
+       else
+               rate = 0;
+
+       td_io_u_lock(td);
+
+       add_stat_sample(&ts->bw_stat[io_u->ddir], rate);
+
+       if (td->bw_log)
+               add_log_sample(td, td->bw_log, rate, io_u->ddir, bytes, io_u->offset);
+
+       td->stat_io_bytes[io_u->ddir] = td->this_io_bytes[io_u->ddir];
+       td_io_u_unlock(td);
+}
+
+static int add_bw_samples(struct thread_data *td, struct timeval *t)
 {
        struct thread_stat *ts = &td->ts;
        unsigned long spent, rate;
+       enum fio_ddir ddir;
+       unsigned int next, next_log;
 
-       if (!ddir_rw(ddir))
-               return;
+       next_log = td->o.bw_avg_time;
 
        spent = mtime_since(&td->bw_sample_time, t);
-       if (spent < td->o.bw_avg_time)
-               return;
+       if (spent < td->o.bw_avg_time &&
+           td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
+               return td->o.bw_avg_time - spent;
 
        td_io_u_lock(td);
 
@@ -2068,28 +2405,59 @@ void add_bw_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
 
                add_stat_sample(&ts->bw_stat[ddir], rate);
 
-               if (td->bw_log)
-                       add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
+               if (td->bw_log) {
+                       unsigned int bs = 0;
+
+                       if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
+                               bs = td->o.min_bs[ddir];
+
+                       next = add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
+                       next_log = min(next_log, next);
+               }
 
                td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
        }
 
-       fio_gettime(&td->bw_sample_time, NULL);
+       timeval_add_msec(&td->bw_sample_time, td->o.bw_avg_time);
+
+       td_io_u_unlock(td);
+
+       if (spent <= td->o.bw_avg_time)
+               return min(next_log, td->o.bw_avg_time);
+
+       next = td->o.bw_avg_time - (1 + spent - td->o.bw_avg_time);
+       return min(next, next_log);
+}
+
+void add_iops_sample(struct thread_data *td, struct io_u *io_u,
+                    unsigned int bytes)
+{
+       struct thread_stat *ts = &td->ts;
+
+       td_io_u_lock(td);
+
+       add_stat_sample(&ts->iops_stat[io_u->ddir], 1);
+
+       if (td->iops_log)
+               add_log_sample(td, td->iops_log, 1, io_u->ddir, bytes, io_u->offset);
+
+       td->stat_io_blocks[io_u->ddir] = td->this_io_blocks[io_u->ddir];
        td_io_u_unlock(td);
 }
 
-void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs,
-                    struct timeval *t)
+static int add_iops_samples(struct thread_data *td, struct timeval *t)
 {
        struct thread_stat *ts = &td->ts;
        unsigned long spent, iops;
+       enum fio_ddir ddir;
+       unsigned int next, next_log;
 
-       if (!ddir_rw(ddir))
-               return;
+       next_log = td->o.iops_avg_time;
 
        spent = mtime_since(&td->iops_sample_time, t);
-       if (spent < td->o.iops_avg_time)
-               return;
+       if (spent < td->o.iops_avg_time &&
+           td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
+               return td->o.iops_avg_time - spent;
 
        td_io_u_lock(td);
 
@@ -2110,14 +2478,61 @@ void add_iops_sample(struct thread_data *td, enum fio_ddir ddir, unsigned int bs
 
                add_stat_sample(&ts->iops_stat[ddir], iops);
 
-               if (td->iops_log)
-                       add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
+               if (td->iops_log) {
+                       unsigned int bs = 0;
+
+                       if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
+                               bs = td->o.min_bs[ddir];
+
+                       next = add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
+                       next_log = min(next_log, next);
+               }
 
                td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
        }
 
-       fio_gettime(&td->iops_sample_time, NULL);
+       timeval_add_msec(&td->iops_sample_time, td->o.iops_avg_time);
+
        td_io_u_unlock(td);
+
+       if (spent <= td->o.iops_avg_time)
+               return min(next_log, td->o.iops_avg_time);
+
+       next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
+       return min(next, next_log);
+}
+
+/*
+ * Returns msecs to next event
+ */
+int calc_log_samples(void)
+{
+       struct thread_data *td;
+       unsigned int next = ~0U, tmp;
+       struct timeval now;
+       int i;
+
+       fio_gettime(&now, NULL);
+
+       for_each_td(td, i) {
+               if (in_ramp_time(td) ||
+                   !(td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING)) {
+                       next = min(td->o.iops_avg_time, td->o.bw_avg_time);
+                       continue;
+               }
+               if (!per_unit_log(td->bw_log)) {
+                       tmp = add_bw_samples(td, &now);
+                       if (tmp < next)
+                               next = tmp;
+               }
+               if (!per_unit_log(td->iops_log)) {
+                       tmp = add_iops_samples(td, &now);
+                       if (tmp < next)
+                               next = tmp;
+               }
+       }
+
+       return next == ~0U ? 0 : next;
 }
 
 void stat_init(void)
@@ -2140,8 +2555,7 @@ void stat_exit(void)
  */
 void show_running_run_stats(void)
 {
-       helper_do_stat = 1;
-       pthread_cond_signal(&helper_cond);
+       helper_do_stat();
 }
 
 uint32_t *io_u_block_info(struct thread_data *td, struct io_u *io_u)