Merge branch 'histograms-PR' of https://github.com/cronburg/fio
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 4d87c29a6fe19afbf9661c84c1adb555e48f2b2b..96af94ba683deb5b4de2420e7fd04cf10e7b7d88 100644 (file)
--- a/stat.c
+++ b/stat.c
@@ -16,6 +16,9 @@
 #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;
 
@@ -1636,16 +1639,21 @@ void __show_run_stats(void)
        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);
@@ -1849,66 +1857,184 @@ 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, new_samples;
-               void *new_log;
+       cur_log = get_cur_log(iolog);
+       if (cur_log) {
+               struct io_sample *s;
 
-               if (!iolog->max_samples)
-                       new_samples = DEF_LOG_ENTRIES;
-               else
-                       new_samples = iolog->max_samples * 2;
-
-               new_size = new_samples * log_entry_sz(iolog);
-
-               if (iolog->log_gz && (new_size > iolog->log_gz)) {
-                       if (!iolog->log) {
-                               iolog->log = malloc(new_size);
-                               iolog->max_samples = new_samples;
-                       } else 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 = new_samples;
-               }
-       }
+               s = get_sample(iolog, cur_log, cur_log->nr_samples);
 
-       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)
@@ -1984,14 +2110,14 @@ static void _add_stat_to_log(struct io_log *iolog, unsigned long elapsed,
                __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);
 
@@ -2000,7 +2126,7 @@ 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;
        }
 
        /*
@@ -2014,12 +2140,19 @@ 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;
+
+               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, bool unit_logs)
@@ -2063,7 +2196,9 @@ 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;
+       struct io_log *iolog = td->clat_hist_log;
 
        td_io_u_lock(td);
 
@@ -2075,6 +2210,35 @@ 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) {
+                       /*
+                        * 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.
+                        */
+                       unsigned int *io_u_plat = (unsigned int *)(td->ts.io_u_plat[ddir]);
+                       unsigned int *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, (uint64_t)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);
 }
 
@@ -2141,13 +2305,13 @@ 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 (per_unit_log(td->bw_log))
-               return 0;
+       next_log = td->o.bw_avg_time;
 
        spent = mtime_since(&td->bw_sample_time, t);
        if (spent < td->o.bw_avg_time &&
-           td->o.bw_avg_time - spent >= 10)
+           td->o.bw_avg_time - spent >= LOG_MSEC_SLACK)
                return td->o.bw_avg_time - spent;
 
        td_io_u_lock(td);
@@ -2175,7 +2339,8 @@ static int add_bw_samples(struct thread_data *td, struct timeval *t)
                        if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
                                bs = td->o.min_bs[ddir];
 
-                       add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
+                       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];
@@ -2186,9 +2351,10 @@ static int add_bw_samples(struct thread_data *td, struct timeval *t)
        td_io_u_unlock(td);
 
        if (spent <= td->o.bw_avg_time)
-               return td->o.bw_avg_time;
+               return min(next_log, td->o.bw_avg_time);
 
-       return td->o.bw_avg_time - (1 + spent - 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,
@@ -2212,13 +2378,13 @@ 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 (per_unit_log(td->iops_log))
-               return 0;
+       next_log = td->o.iops_avg_time;
 
        spent = mtime_since(&td->iops_sample_time, t);
        if (spent < td->o.iops_avg_time &&
-           td->o.iops_avg_time - spent >= 10)
+           td->o.iops_avg_time - spent >= LOG_MSEC_SLACK)
                return td->o.iops_avg_time - spent;
 
        td_io_u_lock(td);
@@ -2246,7 +2412,8 @@ static int add_iops_samples(struct thread_data *td, struct timeval *t)
                        if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
                                bs = td->o.min_bs[ddir];
 
-                       add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
+                       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];
@@ -2257,9 +2424,10 @@ static int add_iops_samples(struct thread_data *td, struct timeval *t)
        td_io_u_unlock(td);
 
        if (spent <= td->o.iops_avg_time)
-               return td->o.iops_avg_time;
+               return min(next_log, td->o.iops_avg_time);
 
-       return td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
+       next = td->o.iops_avg_time - (1 + spent - td->o.iops_avg_time);
+       return min(next, next_log);
 }
 
 /*
@@ -2275,7 +2443,7 @@ int calc_log_samples(void)
        fio_gettime(&now, NULL);
 
        for_each_td(td, i) {
-               if (!ramp_time_over(td) ||
+               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;