From e1bcd541f63f9029f6c50116831303ad06292edc Mon Sep 17 00:00:00 2001
From: Song Liu <songliubraving@fb.com>
Date: Sun, 17 May 2020 22:39:49 -0700
Subject: [PATCH] Add option latency_run to continue enable latency_target

Currently, latency_target run will exist once fio find the highest queue
depth that meets latency_target. Add option latency_run. If set, fio will
continue running and try to meet latency_target by adusting queue depth.

Signed-off-by: Song Liu <songliubraving@fb.com>
---
 HOWTO            |  7 +++++++
 cconv.c          |  2 ++
 fio.1            |  5 +++++
 fio.h            |  1 +
 io_u.c           | 18 +++++++++++++++++-
 options.c        | 10 ++++++++++
 server.h         |  2 +-
 thread_options.h |  2 ++
 8 files changed, 45 insertions(+), 2 deletions(-)

diff --git a/HOWTO b/HOWTO
index 430c7b62..f0b4ffe4 100644
--- a/HOWTO
+++ b/HOWTO
@@ -2551,6 +2551,13 @@ I/O latency
 	defaults to 100.0, meaning that all I/Os must be equal or below to the value
 	set by :option:`latency_target`.
 
+.. option:: latency_run=bool
+
+	Used with :option:`latency_target`. If false (default), fio will find
+	the highest queue depth that meets :option:`latency_target` and exit. If
+	true, fio will continue running and try to meet :option:`latency_target`
+	by adjusting queue depth.
+
 .. option:: max_latency=time
 
 	If set, fio will exit the job with an ETIMEDOUT error if it exceeds this
diff --git a/cconv.c b/cconv.c
index 48218dc4..449bcf7b 100644
--- a/cconv.c
+++ b/cconv.c
@@ -288,6 +288,7 @@ void convert_thread_options_to_cpu(struct thread_options *o,
 	o->latency_window = le64_to_cpu(top->latency_window);
 	o->max_latency = le64_to_cpu(top->max_latency);
 	o->latency_percentile.u.f = fio_uint64_to_double(le64_to_cpu(top->latency_percentile.u.i));
+	o->latency_run = le32_to_cpu(top->latency_run);
 	o->compress_percentage = le32_to_cpu(top->compress_percentage);
 	o->compress_chunk = le32_to_cpu(top->compress_chunk);
 	o->dedupe_percentage = le32_to_cpu(top->dedupe_percentage);
@@ -487,6 +488,7 @@ void convert_thread_options_to_net(struct thread_options_pack *top,
 	top->latency_window = __cpu_to_le64(o->latency_window);
 	top->max_latency = __cpu_to_le64(o->max_latency);
 	top->latency_percentile.u.i = __cpu_to_le64(fio_double_to_uint64(o->latency_percentile.u.f));
+	top->latency_run = __cpu_to_le32(o->latency_run);
 	top->compress_percentage = cpu_to_le32(o->compress_percentage);
 	top->compress_chunk = cpu_to_le32(o->compress_chunk);
 	top->dedupe_percentage = cpu_to_le32(o->dedupe_percentage);
diff --git a/fio.1 b/fio.1
index a2379f98..3a7a359b 100644
--- a/fio.1
+++ b/fio.1
@@ -2275,6 +2275,11 @@ The percentage of I/Os that must fall within the criteria specified by
 defaults to 100.0, meaning that all I/Os must be equal or below to the value
 set by \fBlatency_target\fR.
 .TP
+.BI latency_run \fR=\fPbool
+Used with \fBlatency_target\fR. If false (default), fio will find the highest
+queue depth that meets \fBlatency_target\fR and exit. If true, fio will continue
+running and try to meet \fBlatency_target\fR by adjusting queue depth.
+.TP
 .BI max_latency \fR=\fPtime
 If set, fio will exit the job with an ETIMEDOUT error if it exceeds this
 maximum latency. When the unit is omitted, the value is interpreted in
diff --git a/fio.h b/fio.h
index bbf057c1..7610026d 100644
--- a/fio.h
+++ b/fio.h
@@ -377,6 +377,7 @@ struct thread_data {
 	unsigned int latency_qd_high;
 	unsigned int latency_qd_low;
 	unsigned int latency_failed;
+	unsigned int latency_stable_count;
 	uint64_t latency_ios;
 	int latency_end_run;
 
diff --git a/io_u.c b/io_u.c
index aa8808b8..ae1438fd 100644
--- a/io_u.c
+++ b/io_u.c
@@ -1391,6 +1391,7 @@ static bool __lat_target_failed(struct thread_data *td)
 		td->latency_qd_low--;
 
 	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
+	td->latency_stable_count = 0;
 
 	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
 
@@ -1440,6 +1441,21 @@ static void lat_target_success(struct thread_data *td)
 
 	td->latency_qd_low = td->latency_qd;
 
+	if (td->latency_qd + 1 == td->latency_qd_high) {
+		/*
+		 * latency_qd will not incease on lat_target_success(), so
+		 * called stable. If we stick with this queue depth, the
+		 * final latency is likely lower than latency_target. Fix
+		 * this by increasing latency_qd_high slowly. Use a naive
+		 * heuristic here. If we get lat_target_success() 3 times
+		 * in a row, increase latency_qd_high by 1.
+		 */
+		if (++td->latency_stable_count >= 3) {
+			td->latency_qd_high++;
+			td->latency_stable_count = 0;
+		}
+	}
+
 	/*
 	 * If we haven't failed yet, we double up to a failing value instead
 	 * of bisecting from highest possible queue depth. If we have set
@@ -1459,7 +1475,7 @@ static void lat_target_success(struct thread_data *td)
 	 * Same as last one, we are done. Let it run a latency cycle, so
 	 * we get only the results from the targeted depth.
 	 */
-	if (td->latency_qd == qd) {
+	if (!o->latency_run && td->latency_qd == qd) {
 		if (td->latency_end_run) {
 			dprint(FD_RATE, "We are done\n");
 			td->done = 1;
diff --git a/options.c b/options.c
index b18cea33..da401aed 100644
--- a/options.c
+++ b/options.c
@@ -3672,6 +3672,16 @@ struct fio_option fio_options[FIO_MAX_OPTS] = {
 		.category = FIO_OPT_C_IO,
 		.group	= FIO_OPT_G_LATPROF,
 	},
+	{
+		.name	= "latency_run",
+		.lname	= "Latency Run",
+		.type	= FIO_OPT_BOOL,
+		.off1	= offsetof(struct thread_options, latency_run),
+		.help	= "Keep adjusting queue depth to match latency_target",
+		.def	= "0",
+		.category = FIO_OPT_C_IO,
+		.group	= FIO_OPT_G_LATPROF,
+	},
 	{
 		.name	= "invalidate",
 		.lname	= "Cache invalidate",
diff --git a/server.h b/server.h
index 279b6917..de01a5c8 100644
--- a/server.h
+++ b/server.h
@@ -48,7 +48,7 @@ struct fio_net_cmd_reply {
 };
 
 enum {
-	FIO_SERVER_VER			= 82,
+	FIO_SERVER_VER			= 83,
 
 	FIO_SERVER_MAX_FRAGMENT_PDU	= 1024,
 	FIO_SERVER_MAX_CMD_MB		= 2048,
diff --git a/thread_options.h b/thread_options.h
index c78ed43d..09ccd5b2 100644
--- a/thread_options.h
+++ b/thread_options.h
@@ -324,6 +324,7 @@ struct thread_options {
 	unsigned long long latency_target;
 	unsigned long long latency_window;
 	fio_fp64_t latency_percentile;
+	uint32_t latency_run;
 
 	unsigned int sig_figs;
 
@@ -612,6 +613,7 @@ struct thread_options_pack {
 	uint64_t latency_window;
 	uint64_t max_latency;
 	fio_fp64_t latency_percentile;
+	uint32_t latency_run;
 
 	uint32_t sig_figs;
 
-- 
2.17.0