drifting in output of interval-averaged values was eventually causing IOP samples...
[fio.git] / stat.c
diff --git a/stat.c b/stat.c
index 7db41376913170fec8faeefd7bc457f2bad8228a..08a402a74c6d12e6c7bb339acce638dec5234514 100644 (file)
--- a/stat.c
+++ b/stat.c
@@ -15,6 +15,8 @@
 #include "idletime.h"
 #include "lib/pow2.h"
 #include "lib/output_buffer.h"
+#include "helper_thread.h"
+#include "smalloc.h"
 
 struct fio_mutex *stat_mutex;
 
@@ -1635,16 +1637,21 @@ 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;
 
-               time(&time_p);
-               os_ctime_r((const time_t *) &time_p, time_buf,
+               gettimeofday(&now, NULL);
+               ms_since_epoch = (unsigned long long)(now.tv_sec) * 1000 +
+                                (unsigned long long)(now.tv_usec) / 1000;
+
+               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);
@@ -1848,66 +1855,183 @@ 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->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)
@@ -2018,7 +2142,7 @@ static void add_log_sample(struct thread_data *td, struct io_log *iolog,
 
        _add_stat_to_log(iolog, elapsed, td->o.log_max != 0);
 
-       iolog->avg_last = elapsed;
+       iolog->avg_last = elapsed - (this_window - iolog->avg_msec);
 }
 
 void finalize_logs(struct thread_data *td, bool unit_logs)
@@ -2141,9 +2265,6 @@ static int add_bw_samples(struct thread_data *td, struct timeval *t)
        unsigned long spent, rate;
        enum fio_ddir ddir;
 
-       if (per_unit_log(td->bw_log))
-               return 0;
-
        spent = mtime_since(&td->bw_sample_time, t);
        if (spent < td->o.bw_avg_time &&
            td->o.bw_avg_time - spent >= 10)
@@ -2168,8 +2289,14 @@ static int add_bw_samples(struct thread_data *td, struct timeval *t)
 
                add_stat_sample(&ts->bw_stat[ddir], rate);
 
-               if (td->bw_log)
-                       add_log_sample(td, td->bw_log, rate, ddir, 0, 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];
+
+                       add_log_sample(td, td->bw_log, rate, ddir, bs, 0);
+               }
 
                td->stat_io_bytes[ddir] = td->this_io_bytes[ddir];
        }
@@ -2206,9 +2333,6 @@ static int add_iops_samples(struct thread_data *td, struct timeval *t)
        unsigned long spent, iops;
        enum fio_ddir ddir;
 
-       if (per_unit_log(td->iops_log))
-               return 0;
-
        spent = mtime_since(&td->iops_sample_time, t);
        if (spent < td->o.iops_avg_time &&
            td->o.iops_avg_time - spent >= 10)
@@ -2233,8 +2357,14 @@ static int add_iops_samples(struct thread_data *td, struct timeval *t)
 
                add_stat_sample(&ts->iops_stat[ddir], iops);
 
-               if (td->iops_log)
-                       add_log_sample(td, td->iops_log, iops, ddir, 0, 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];
+
+                       add_log_sample(td, td->iops_log, iops, ddir, bs, 0);
+               }
 
                td->stat_io_blocks[ddir] = td->this_io_blocks[ddir];
        }
@@ -2262,7 +2392,7 @@ 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;