X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=stat.c;h=75b45733cead570f402296f38a21cafec81483bf;hp=b42e886c32aeff16aca4ca3892509b192e1aa021;hb=17df7023156b50ee7288eea0f118f1ac20b21ea2;hpb=7e419452839dee5afa86b0a6c78136c421e1d706

diff --git a/stat.c b/stat.c
index b42e886c..75b45733 100644
--- 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;
 
@@ -654,6 +657,33 @@ static void show_block_infos(int nr_block_infos, uint32_t *block_infos,
 			 i == BLOCK_STATE_COUNT - 1 ? '\n' : ',');
 }
 
+static void show_ss_normal(struct thread_stat *ts, struct buf_output *out)
+{
+	char *p1, *p2;
+	unsigned long long bw_mean, iops_mean;
+	const int i2p = is_power_of_2(ts->kb_base);
+
+	if (!ts->ss_state)
+		return;
+
+	bw_mean = steadystate_bw_mean(ts);
+	iops_mean = steadystate_iops_mean(ts);
+
+	p1 = num2str(bw_mean / ts->kb_base, 6, ts->kb_base, i2p, ts->unit_base);
+	p2 = num2str(iops_mean, 6, 1, 0, 0);
+
+	log_buf(out, "  steadystate  : attained=%s, bw=%s/s, iops=%s, %s%s=%.3f%s\n",
+		ts->ss_state & __FIO_SS_ATTAINED ? "yes" : "no",
+		p1, p2,
+		ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
+		ts->ss_state & __FIO_SS_SLOPE ? " slope": " mean dev",
+		ts->ss_criterion.u.f,
+		ts->ss_state & __FIO_SS_PCT ? "%" : "");
+
+	free(p1);
+	free(p2);
+}
+
 static void show_thread_status_normal(struct thread_stat *ts,
 				      struct group_run_stats *rs,
 				      struct buf_output *out)
@@ -758,6 +788,9 @@ static void show_thread_status_normal(struct thread_stat *ts,
 	if (ts->nr_block_infos)
 		show_block_infos(ts->nr_block_infos, ts->block_infos,
 				  ts->percentile_list, out);
+
+	if (ts->ss_dur)
+		show_ss_normal(ts, out);
 }
 
 static void show_ddir_status_terse(struct thread_stat *ts,
@@ -1252,6 +1285,56 @@ static struct json_object *show_thread_status_json(struct thread_stat *ts,
 		}
 	}
 
+	if (ts->ss_dur) {
+		struct json_object *data;
+		struct json_array *iops, *bw;
+		int i, j, k;
+		char ss_buf[64];
+
+		snprintf(ss_buf, sizeof(ss_buf), "%s%s:%f%s",
+			ts->ss_state & __FIO_SS_IOPS ? "iops" : "bw",
+			ts->ss_state & __FIO_SS_SLOPE ? "_slope" : "",
+			(float) ts->ss_limit.u.f,
+			ts->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)ts->ss_dur);
+		json_object_add_value_int(tmp, "attained", (ts->ss_state & __FIO_SS_ATTAINED) > 0);
+
+		snprintf(ss_buf, sizeof(ss_buf), "%f%s", (float) ts->ss_criterion.u.f,
+			ts->ss_state & __FIO_SS_PCT ? "%" : "");
+		json_object_add_value_string(tmp, "criterion", ss_buf);
+		json_object_add_value_float(tmp, "max_deviation", ts->ss_deviation.u.f);
+		json_object_add_value_float(tmp, "slope", ts->ss_slope.u.f);
+
+		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 ((ts->ss_state & __FIO_SS_ATTAINED) || !(ts->ss_state & __FIO_SS_BUFFER_FULL))
+			j = ts->ss_head;
+		else
+			j = ts->ss_head == 0 ? ts->ss_dur - 1 : ts->ss_head - 1;
+		for (i = 0; i < ts->ss_dur; i++) {
+			k = (j + i) % ts->ss_dur;
+			json_array_add_value_int(bw, ts->ss_bw_data[k]);
+			json_array_add_value_int(iops, ts->ss_iops_data[k]);
+		}
+		json_object_add_value_int(data, "bw_mean", steadystate_bw_mean(ts));
+		json_object_add_value_int(data, "iops_mean", steadystate_iops_mean(ts));
+		json_object_add_value_array(data, "iops", iops);
+		json_object_add_value_array(data, "bw", bw);
+	}
+
 	return root;
 }
 
@@ -1575,6 +1658,20 @@ void __show_run_stats(void)
 			ts->block_infos[k] = td->ts.block_infos[k];
 
 		sum_thread_stats(ts, &td->ts, idx == 1);
+
+		if (td->o.ss_dur) {
+			ts->ss_state = td->ss.state;
+			ts->ss_dur = td->ss.dur;
+			ts->ss_head = td->ss.head;
+			ts->ss_bw_data = td->ss.bw_data;
+			ts->ss_iops_data = td->ss.iops_data;
+			ts->ss_limit.u.f = td->ss.limit;
+			ts->ss_slope.u.f = td->ss.slope;
+			ts->ss_deviation.u.f = td->ss.deviation;
+			ts->ss_criterion.u.f = td->ss.criterion;
+		}
+		else
+			ts->ss_dur = ts->ss_state = 0;
 	}
 
 	for (i = 0; i < nr_ts; i++) {
@@ -1636,16 +1733,21 @@ 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);
@@ -1870,17 +1972,9 @@ static struct io_logs *get_new_log(struct io_log *iolog)
 			new_samples = MAX_LOG_ENTRIES;
 	}
 
-	/*
-	 * If the alloc size is sufficiently large, quiesce pending IO before
-	 * attempting it. This is to avoid spending a long time in alloc with
-	 * IO pending, which will unfairly skew the completion latencies of
-	 * inflight IO.
-	 */
 	new_size = new_samples * log_entry_sz(iolog);
-	if (new_size >= LOG_QUIESCE_SZ)
-		io_u_quiesce(iolog->td);
 
-	cur_log = malloc(sizeof(*cur_log));
+	cur_log = smalloc(sizeof(*cur_log));
 	if (cur_log) {
 		INIT_FLIST_HEAD(&cur_log->list);
 		cur_log->log = malloc(new_size);
@@ -1891,15 +1985,22 @@ static struct io_logs *get_new_log(struct io_log *iolog)
 			iolog->cur_log_max = new_samples;
 			return cur_log;
 		}
-		free(cur_log);
+		sfree(cur_log);
 	}
 
 	return NULL;
 }
 
-static struct io_logs *get_cur_log(struct io_log *iolog)
+/*
+ * 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) {
@@ -1931,7 +2032,68 @@ static struct io_logs *get_cur_log(struct io_log *iolog)
 		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,
@@ -2042,14 +2204,14 @@ 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);
 
@@ -2058,7 +2220,7 @@ 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;
 	}
 
 	/*
@@ -2072,12 +2234,19 @@ 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)
@@ -2121,7 +2290,9 @@ 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);
 
@@ -2133,6 +2304,43 @@ 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);
 }
 
@@ -2199,13 +2407,13 @@ 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);
@@ -2233,7 +2441,8 @@ 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];
@@ -2244,9 +2453,10 @@ 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,
@@ -2270,13 +2480,13 @@ 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);
@@ -2304,7 +2514,8 @@ 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];
@@ -2315,9 +2526,10 @@ 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);
 }
 
 /*
@@ -2333,7 +2545,7 @@ 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;