* [PATCH 1/3] drm/msm: remove explicit devfreq status reset
@ 2022-04-16 0:33 ` Chia-I Wu
0 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:33 UTC (permalink / raw)
To: freedreno
Cc: dri-devel, Sean Paul, Abhinav Kumar, David Airlie, linux-arm-msm,
Daniel Vetter, Rob Clark
It is redundant since commit 7c0ffcd40b16 ("drm/msm/gpu: Respect PM QoS
constraints") because dev_pm_qos_update_request triggers get_dev_status.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Cc: Rob Clark <robdclark@chromium.org>
---
drivers/gpu/drm/msm/msm_gpu_devfreq.c | 7 -------
1 file changed, 7 deletions(-)
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index 12641616acd3..317a95d42922 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -224,7 +224,6 @@ void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
void msm_devfreq_active(struct msm_gpu *gpu)
{
struct msm_gpu_devfreq *df = &gpu->devfreq;
- struct devfreq_dev_status status;
unsigned int idle_time;
if (!has_devfreq(gpu))
@@ -248,12 +247,6 @@ void msm_devfreq_active(struct msm_gpu *gpu)
dev_pm_qos_update_request(&df->idle_freq,
PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
-
- /*
- * Reset the polling interval so we aren't inconsistent
- * about freq vs busy/total cycles
- */
- msm_devfreq_get_dev_status(&gpu->pdev->dev, &status);
}
--
2.36.0.rc0.470.gd361397f0d-goog
^ permalink raw reply related [flat|nested] 8+ messages in thread
* [PATCH 2/3] drm/msm: simplify gpu_busy callback
2022-04-16 0:33 ` Chia-I Wu
@ 2022-04-16 0:33 ` Chia-I Wu
-1 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:33 UTC (permalink / raw)
To: freedreno
Cc: Rob Clark, David Airlie, linux-arm-msm, Abhinav Kumar, dri-devel,
Sean Paul
Move tracking and busy time calculation to msm_devfreq_get_dev_status.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Cc: Rob Clark <robdclark@chromium.org>
---
drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 19 ++++++----------
drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 15 +++++--------
drivers/gpu/drm/msm/msm_gpu.h | 9 +++-----
drivers/gpu/drm/msm/msm_gpu_devfreq.c | 32 ++++++++++++++++++++++-----
4 files changed, 41 insertions(+), 34 deletions(-)
diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c
index 407f50a15faa..217615e0e850 100644
--- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c
@@ -1662,28 +1662,23 @@ static struct msm_ringbuffer *a5xx_active_ring(struct msm_gpu *gpu)
return a5xx_gpu->cur_ring;
}
-static unsigned long a5xx_gpu_busy(struct msm_gpu *gpu)
+static u64 a5xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
{
- u64 busy_cycles, busy_time;
+ u64 busy_cycles;
/* Only read the gpu busy if the hardware is already active */
- if (pm_runtime_get_if_in_use(&gpu->pdev->dev) == 0)
+ if (pm_runtime_get_if_in_use(&gpu->pdev->dev) == 0) {
+ *out_sample_rate = 1;
return 0;
+ }
busy_cycles = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_0_LO,
REG_A5XX_RBBM_PERFCTR_RBBM_0_HI);
-
- busy_time = busy_cycles - gpu->devfreq.busy_cycles;
- do_div(busy_time, clk_get_rate(gpu->core_clk) / 1000000);
-
- gpu->devfreq.busy_cycles = busy_cycles;
+ *out_sample_rate = clk_get_rate(gpu->core_clk);
pm_runtime_put(&gpu->pdev->dev);
- if (WARN_ON(busy_time > ~0LU))
- return ~0LU;
-
- return (unsigned long)busy_time;
+ return busy_cycles;
}
static uint32_t a5xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
index ccc4fcf7a630..fba6259395dd 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -1649,12 +1649,14 @@ static void a6xx_destroy(struct msm_gpu *gpu)
kfree(a6xx_gpu);
}
-static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
+static u64 a6xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
{
struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
- u64 busy_cycles, busy_time;
+ u64 busy_cycles;
+ /* 19.2MHz */
+ *out_sample_rate = 19200000;
/* Only read the gpu busy if the hardware is already active */
if (pm_runtime_get_if_in_use(a6xx_gpu->gmu.dev) == 0)
@@ -1664,17 +1666,10 @@ static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_L,
REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
- busy_time = (busy_cycles - gpu->devfreq.busy_cycles) * 10;
- do_div(busy_time, 192);
-
- gpu->devfreq.busy_cycles = busy_cycles;
pm_runtime_put(a6xx_gpu->gmu.dev);
- if (WARN_ON(busy_time > ~0LU))
- return ~0LU;
-
- return (unsigned long)busy_time;
+ return busy_cycles;
}
static void a6xx_gpu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 02419f2ca2bc..389c6dab751b 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -62,7 +62,7 @@ struct msm_gpu_funcs {
/* for generation specific debugfs: */
void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
#endif
- unsigned long (*gpu_busy)(struct msm_gpu *gpu);
+ u64 (*gpu_busy)(struct msm_gpu *gpu, unsigned long *out_sample_rate);
struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
int (*gpu_state_put)(struct msm_gpu_state *state);
unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
@@ -106,11 +106,8 @@ struct msm_gpu_devfreq {
struct dev_pm_qos_request boost_freq;
/**
- * busy_cycles:
- *
- * Used by implementation of gpu->gpu_busy() to track the last
- * busy counter value, for calculating elapsed busy cycles since
- * last sampling period.
+ * busy_cycles: Last busy counter value, for calculating elapsed busy
+ * cycles since last sampling period.
*/
u64 busy_cycles;
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index 317a95d42922..f531015107c3 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -49,18 +49,38 @@ static unsigned long get_freq(struct msm_gpu *gpu)
return clk_get_rate(gpu->core_clk);
}
-static int msm_devfreq_get_dev_status(struct device *dev,
+static void get_raw_dev_status(struct msm_gpu *gpu,
struct devfreq_dev_status *status)
{
- struct msm_gpu *gpu = dev_to_gpu(dev);
+ struct msm_gpu_devfreq *df = &gpu->devfreq;
+ u64 busy_cycles, busy_time;
+ unsigned long sample_rate;
ktime_t time;
status->current_frequency = get_freq(gpu);
- status->busy_time = gpu->funcs->gpu_busy(gpu);
-
+ busy_cycles = gpu->funcs->gpu_busy(gpu, &sample_rate);
time = ktime_get();
- status->total_time = ktime_us_delta(time, gpu->devfreq.time);
- gpu->devfreq.time = time;
+
+ busy_time = busy_cycles - df->busy_cycles;
+ status->total_time = ktime_us_delta(time, df->time);
+
+ df->busy_cycles = busy_cycles;
+ df->time = time;
+
+ busy_time *= USEC_PER_SEC;
+ do_div(busy_time, sample_rate);
+ if (WARN_ON(busy_time > ~0LU))
+ busy_time = ~0LU;
+
+ status->busy_time = busy_time;
+}
+
+static int msm_devfreq_get_dev_status(struct device *dev,
+ struct devfreq_dev_status *status)
+{
+ struct msm_gpu *gpu = dev_to_gpu(dev);
+
+ get_raw_dev_status(gpu, status);
return 0;
}
--
2.36.0.rc0.470.gd361397f0d-goog
^ permalink raw reply related [flat|nested] 8+ messages in thread
* [PATCH 2/3] drm/msm: simplify gpu_busy callback
@ 2022-04-16 0:33 ` Chia-I Wu
0 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:33 UTC (permalink / raw)
To: freedreno
Cc: dri-devel, Sean Paul, Abhinav Kumar, David Airlie, linux-arm-msm,
Daniel Vetter, Rob Clark
Move tracking and busy time calculation to msm_devfreq_get_dev_status.
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Cc: Rob Clark <robdclark@chromium.org>
---
drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 19 ++++++----------
drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 15 +++++--------
drivers/gpu/drm/msm/msm_gpu.h | 9 +++-----
drivers/gpu/drm/msm/msm_gpu_devfreq.c | 32 ++++++++++++++++++++++-----
4 files changed, 41 insertions(+), 34 deletions(-)
diff --git a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c
index 407f50a15faa..217615e0e850 100644
--- a/drivers/gpu/drm/msm/adreno/a5xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a5xx_gpu.c
@@ -1662,28 +1662,23 @@ static struct msm_ringbuffer *a5xx_active_ring(struct msm_gpu *gpu)
return a5xx_gpu->cur_ring;
}
-static unsigned long a5xx_gpu_busy(struct msm_gpu *gpu)
+static u64 a5xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
{
- u64 busy_cycles, busy_time;
+ u64 busy_cycles;
/* Only read the gpu busy if the hardware is already active */
- if (pm_runtime_get_if_in_use(&gpu->pdev->dev) == 0)
+ if (pm_runtime_get_if_in_use(&gpu->pdev->dev) == 0) {
+ *out_sample_rate = 1;
return 0;
+ }
busy_cycles = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_0_LO,
REG_A5XX_RBBM_PERFCTR_RBBM_0_HI);
-
- busy_time = busy_cycles - gpu->devfreq.busy_cycles;
- do_div(busy_time, clk_get_rate(gpu->core_clk) / 1000000);
-
- gpu->devfreq.busy_cycles = busy_cycles;
+ *out_sample_rate = clk_get_rate(gpu->core_clk);
pm_runtime_put(&gpu->pdev->dev);
- if (WARN_ON(busy_time > ~0LU))
- return ~0LU;
-
- return (unsigned long)busy_time;
+ return busy_cycles;
}
static uint32_t a5xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
index ccc4fcf7a630..fba6259395dd 100644
--- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
+++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
@@ -1649,12 +1649,14 @@ static void a6xx_destroy(struct msm_gpu *gpu)
kfree(a6xx_gpu);
}
-static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
+static u64 a6xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
{
struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
- u64 busy_cycles, busy_time;
+ u64 busy_cycles;
+ /* 19.2MHz */
+ *out_sample_rate = 19200000;
/* Only read the gpu busy if the hardware is already active */
if (pm_runtime_get_if_in_use(a6xx_gpu->gmu.dev) == 0)
@@ -1664,17 +1666,10 @@ static unsigned long a6xx_gpu_busy(struct msm_gpu *gpu)
REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_L,
REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
- busy_time = (busy_cycles - gpu->devfreq.busy_cycles) * 10;
- do_div(busy_time, 192);
-
- gpu->devfreq.busy_cycles = busy_cycles;
pm_runtime_put(a6xx_gpu->gmu.dev);
- if (WARN_ON(busy_time > ~0LU))
- return ~0LU;
-
- return (unsigned long)busy_time;
+ return busy_cycles;
}
static void a6xx_gpu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 02419f2ca2bc..389c6dab751b 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -62,7 +62,7 @@ struct msm_gpu_funcs {
/* for generation specific debugfs: */
void (*debugfs_init)(struct msm_gpu *gpu, struct drm_minor *minor);
#endif
- unsigned long (*gpu_busy)(struct msm_gpu *gpu);
+ u64 (*gpu_busy)(struct msm_gpu *gpu, unsigned long *out_sample_rate);
struct msm_gpu_state *(*gpu_state_get)(struct msm_gpu *gpu);
int (*gpu_state_put)(struct msm_gpu_state *state);
unsigned long (*gpu_get_freq)(struct msm_gpu *gpu);
@@ -106,11 +106,8 @@ struct msm_gpu_devfreq {
struct dev_pm_qos_request boost_freq;
/**
- * busy_cycles:
- *
- * Used by implementation of gpu->gpu_busy() to track the last
- * busy counter value, for calculating elapsed busy cycles since
- * last sampling period.
+ * busy_cycles: Last busy counter value, for calculating elapsed busy
+ * cycles since last sampling period.
*/
u64 busy_cycles;
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index 317a95d42922..f531015107c3 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -49,18 +49,38 @@ static unsigned long get_freq(struct msm_gpu *gpu)
return clk_get_rate(gpu->core_clk);
}
-static int msm_devfreq_get_dev_status(struct device *dev,
+static void get_raw_dev_status(struct msm_gpu *gpu,
struct devfreq_dev_status *status)
{
- struct msm_gpu *gpu = dev_to_gpu(dev);
+ struct msm_gpu_devfreq *df = &gpu->devfreq;
+ u64 busy_cycles, busy_time;
+ unsigned long sample_rate;
ktime_t time;
status->current_frequency = get_freq(gpu);
- status->busy_time = gpu->funcs->gpu_busy(gpu);
-
+ busy_cycles = gpu->funcs->gpu_busy(gpu, &sample_rate);
time = ktime_get();
- status->total_time = ktime_us_delta(time, gpu->devfreq.time);
- gpu->devfreq.time = time;
+
+ busy_time = busy_cycles - df->busy_cycles;
+ status->total_time = ktime_us_delta(time, df->time);
+
+ df->busy_cycles = busy_cycles;
+ df->time = time;
+
+ busy_time *= USEC_PER_SEC;
+ do_div(busy_time, sample_rate);
+ if (WARN_ON(busy_time > ~0LU))
+ busy_time = ~0LU;
+
+ status->busy_time = busy_time;
+}
+
+static int msm_devfreq_get_dev_status(struct device *dev,
+ struct devfreq_dev_status *status)
+{
+ struct msm_gpu *gpu = dev_to_gpu(dev);
+
+ get_raw_dev_status(gpu, status);
return 0;
}
--
2.36.0.rc0.470.gd361397f0d-goog
^ permalink raw reply related [flat|nested] 8+ messages in thread
* [PATCH 3/3] drm/msm: return the average load over the polling period
2022-04-16 0:33 ` Chia-I Wu
@ 2022-04-16 0:33 ` Chia-I Wu
-1 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:33 UTC (permalink / raw)
To: freedreno
Cc: Rob Clark, David Airlie, linux-arm-msm, Abhinav Kumar, dri-devel,
Sean Paul
simple_ondemand interacts poorly with clamp_to_idle. It only looks at
the load since the last get_dev_status call, while it should really look
at the load over polling_ms. When clamp_to_idle true, it almost always
picks the lowest frequency on active because the gpu is idle between
msm_devfreq_idle/msm_devfreq_active.
This logic could potentially be moved into devfreq core.
Fixes: 7c0ffcd40b16 ("drm/msm/gpu: Respect PM QoS constraints")
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Cc: Rob Clark <robdclark@chromium.org>
---
drivers/gpu/drm/msm/msm_gpu.h | 3 ++
drivers/gpu/drm/msm/msm_gpu_devfreq.c | 60 ++++++++++++++++++++++++++-
2 files changed, 62 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 389c6dab751b..143c56f5185b 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -9,6 +9,7 @@
#include <linux/adreno-smmu-priv.h>
#include <linux/clk.h>
+#include <linux/devfreq.h>
#include <linux/interconnect.h>
#include <linux/pm_opp.h>
#include <linux/regulator/consumer.h>
@@ -117,6 +118,8 @@ struct msm_gpu_devfreq {
/** idle_time: Time of last transition to idle: */
ktime_t idle_time;
+ struct devfreq_dev_status average_status;
+
/**
* idle_work:
*
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index f531015107c3..d2539ca78c29 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -9,6 +9,7 @@
#include <linux/devfreq.h>
#include <linux/devfreq_cooling.h>
+#include <linux/math64.h>
#include <linux/units.h>
/*
@@ -75,12 +76,69 @@ static void get_raw_dev_status(struct msm_gpu *gpu,
status->busy_time = busy_time;
}
+static void update_average_dev_status(struct msm_gpu *gpu,
+ const struct devfreq_dev_status *raw)
+{
+ struct msm_gpu_devfreq *df = &gpu->devfreq;
+ const u32 polling_ms = df->devfreq->profile->polling_ms;
+ const u32 max_history_ms = polling_ms * 11 / 10;
+ struct devfreq_dev_status *avg = &df->average_status;
+ u64 avg_freq;
+
+ /* simple_ondemand governor interacts poorly with gpu->clamp_to_idle.
+ * When we enforce the constraint on idle, it calls get_dev_status
+ * which would normally reset the stats. When we remove the
+ * constraint on active, it calls get_dev_status again where busy_time
+ * would be 0.
+ *
+ * To remedy this, we always return the average load over the past
+ * polling_ms.
+ */
+
+ /* raw is longer than polling_ms or avg has no history */
+ if (div_u64(raw->total_time, USEC_PER_MSEC) >= polling_ms ||
+ !avg->total_time) {
+ *avg = *raw;
+ return;
+ }
+
+ /* Truncate the oldest history first.
+ *
+ * Because we keep the history with a single devfreq_dev_status,
+ * rather than a list of devfreq_dev_status, we have to assume freq
+ * and load are the same over avg->total_time. We can scale down
+ * avg->busy_time and avg->total_time by the same factor to drop
+ * history.
+ */
+ if (div_u64(avg->total_time + raw->total_time, USEC_PER_MSEC) >=
+ max_history_ms) {
+ const u32 new_total_time = polling_ms * USEC_PER_MSEC -
+ raw->total_time;
+ avg->busy_time = div_u64(
+ mul_u32_u32(avg->busy_time, new_total_time),
+ avg->total_time);
+ avg->total_time = new_total_time;
+ }
+
+ /* compute the average freq over avg->total_time + raw->total_time */
+ avg_freq = mul_u32_u32(avg->current_frequency, avg->total_time);
+ avg_freq += mul_u32_u32(raw->current_frequency, raw->total_time);
+ do_div(avg_freq, avg->total_time + raw->total_time);
+
+ avg->current_frequency = avg_freq;
+ avg->busy_time += raw->busy_time;
+ avg->total_time += raw->total_time;
+}
+
static int msm_devfreq_get_dev_status(struct device *dev,
struct devfreq_dev_status *status)
{
struct msm_gpu *gpu = dev_to_gpu(dev);
+ struct devfreq_dev_status raw;
- get_raw_dev_status(gpu, status);
+ get_raw_dev_status(gpu, &raw);
+ update_average_dev_status(gpu, &raw);
+ *status = gpu->devfreq.average_status;
return 0;
}
--
2.36.0.rc0.470.gd361397f0d-goog
^ permalink raw reply related [flat|nested] 8+ messages in thread
* [PATCH 3/3] drm/msm: return the average load over the polling period
@ 2022-04-16 0:33 ` Chia-I Wu
0 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:33 UTC (permalink / raw)
To: freedreno
Cc: dri-devel, Sean Paul, Abhinav Kumar, David Airlie, linux-arm-msm,
Daniel Vetter, Rob Clark
simple_ondemand interacts poorly with clamp_to_idle. It only looks at
the load since the last get_dev_status call, while it should really look
at the load over polling_ms. When clamp_to_idle true, it almost always
picks the lowest frequency on active because the gpu is idle between
msm_devfreq_idle/msm_devfreq_active.
This logic could potentially be moved into devfreq core.
Fixes: 7c0ffcd40b16 ("drm/msm/gpu: Respect PM QoS constraints")
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Cc: Rob Clark <robdclark@chromium.org>
---
drivers/gpu/drm/msm/msm_gpu.h | 3 ++
drivers/gpu/drm/msm/msm_gpu_devfreq.c | 60 ++++++++++++++++++++++++++-
2 files changed, 62 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 389c6dab751b..143c56f5185b 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -9,6 +9,7 @@
#include <linux/adreno-smmu-priv.h>
#include <linux/clk.h>
+#include <linux/devfreq.h>
#include <linux/interconnect.h>
#include <linux/pm_opp.h>
#include <linux/regulator/consumer.h>
@@ -117,6 +118,8 @@ struct msm_gpu_devfreq {
/** idle_time: Time of last transition to idle: */
ktime_t idle_time;
+ struct devfreq_dev_status average_status;
+
/**
* idle_work:
*
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index f531015107c3..d2539ca78c29 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -9,6 +9,7 @@
#include <linux/devfreq.h>
#include <linux/devfreq_cooling.h>
+#include <linux/math64.h>
#include <linux/units.h>
/*
@@ -75,12 +76,69 @@ static void get_raw_dev_status(struct msm_gpu *gpu,
status->busy_time = busy_time;
}
+static void update_average_dev_status(struct msm_gpu *gpu,
+ const struct devfreq_dev_status *raw)
+{
+ struct msm_gpu_devfreq *df = &gpu->devfreq;
+ const u32 polling_ms = df->devfreq->profile->polling_ms;
+ const u32 max_history_ms = polling_ms * 11 / 10;
+ struct devfreq_dev_status *avg = &df->average_status;
+ u64 avg_freq;
+
+ /* simple_ondemand governor interacts poorly with gpu->clamp_to_idle.
+ * When we enforce the constraint on idle, it calls get_dev_status
+ * which would normally reset the stats. When we remove the
+ * constraint on active, it calls get_dev_status again where busy_time
+ * would be 0.
+ *
+ * To remedy this, we always return the average load over the past
+ * polling_ms.
+ */
+
+ /* raw is longer than polling_ms or avg has no history */
+ if (div_u64(raw->total_time, USEC_PER_MSEC) >= polling_ms ||
+ !avg->total_time) {
+ *avg = *raw;
+ return;
+ }
+
+ /* Truncate the oldest history first.
+ *
+ * Because we keep the history with a single devfreq_dev_status,
+ * rather than a list of devfreq_dev_status, we have to assume freq
+ * and load are the same over avg->total_time. We can scale down
+ * avg->busy_time and avg->total_time by the same factor to drop
+ * history.
+ */
+ if (div_u64(avg->total_time + raw->total_time, USEC_PER_MSEC) >=
+ max_history_ms) {
+ const u32 new_total_time = polling_ms * USEC_PER_MSEC -
+ raw->total_time;
+ avg->busy_time = div_u64(
+ mul_u32_u32(avg->busy_time, new_total_time),
+ avg->total_time);
+ avg->total_time = new_total_time;
+ }
+
+ /* compute the average freq over avg->total_time + raw->total_time */
+ avg_freq = mul_u32_u32(avg->current_frequency, avg->total_time);
+ avg_freq += mul_u32_u32(raw->current_frequency, raw->total_time);
+ do_div(avg_freq, avg->total_time + raw->total_time);
+
+ avg->current_frequency = avg_freq;
+ avg->busy_time += raw->busy_time;
+ avg->total_time += raw->total_time;
+}
+
static int msm_devfreq_get_dev_status(struct device *dev,
struct devfreq_dev_status *status)
{
struct msm_gpu *gpu = dev_to_gpu(dev);
+ struct devfreq_dev_status raw;
- get_raw_dev_status(gpu, status);
+ get_raw_dev_status(gpu, &raw);
+ update_average_dev_status(gpu, &raw);
+ *status = gpu->devfreq.average_status;
return 0;
}
--
2.36.0.rc0.470.gd361397f0d-goog
^ permalink raw reply related [flat|nested] 8+ messages in thread
* Re: [PATCH 3/3] drm/msm: return the average load over the polling period
2022-04-16 0:33 ` Chia-I Wu
@ 2022-04-16 0:50 ` Chia-I Wu
-1 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:50 UTC (permalink / raw)
To: freedreno
Cc: Rob Clark, David Airlie, linux-arm-msm, Abhinav Kumar,
ML dri-devel, Chanwoo Choi, Sean Paul
On Fri, Apr 15, 2022 at 5:33 PM Chia-I Wu <olvaffe@gmail.com> wrote:
>
> simple_ondemand interacts poorly with clamp_to_idle. It only looks at
> the load since the last get_dev_status call, while it should really look
> at the load over polling_ms. When clamp_to_idle true, it almost always
> picks the lowest frequency on active because the gpu is idle between
> msm_devfreq_idle/msm_devfreq_active.
>
> This logic could potentially be moved into devfreq core.
The idea is to extend devfreq_simple_ondemand_data to specify whether
devfreq_simple_ondemand_func should use "last status" or "average
status" to determine the target frequency. devfreq core will need to
store "struct devfreq_dev_status average_status", which will be
updated when a device uses simple_ondemand and asks for average_status
instead of last_status.
>
> Fixes: 7c0ffcd40b16 ("drm/msm/gpu: Respect PM QoS constraints")
> Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> Cc: Rob Clark <robdclark@chromium.org>
> ---
> drivers/gpu/drm/msm/msm_gpu.h | 3 ++
> drivers/gpu/drm/msm/msm_gpu_devfreq.c | 60 ++++++++++++++++++++++++++-
> 2 files changed, 62 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
> index 389c6dab751b..143c56f5185b 100644
> --- a/drivers/gpu/drm/msm/msm_gpu.h
> +++ b/drivers/gpu/drm/msm/msm_gpu.h
> @@ -9,6 +9,7 @@
>
> #include <linux/adreno-smmu-priv.h>
> #include <linux/clk.h>
> +#include <linux/devfreq.h>
> #include <linux/interconnect.h>
> #include <linux/pm_opp.h>
> #include <linux/regulator/consumer.h>
> @@ -117,6 +118,8 @@ struct msm_gpu_devfreq {
> /** idle_time: Time of last transition to idle: */
> ktime_t idle_time;
>
> + struct devfreq_dev_status average_status;
> +
> /**
> * idle_work:
> *
> diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
> index f531015107c3..d2539ca78c29 100644
> --- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
> +++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
> @@ -9,6 +9,7 @@
>
> #include <linux/devfreq.h>
> #include <linux/devfreq_cooling.h>
> +#include <linux/math64.h>
> #include <linux/units.h>
>
> /*
> @@ -75,12 +76,69 @@ static void get_raw_dev_status(struct msm_gpu *gpu,
> status->busy_time = busy_time;
> }
>
> +static void update_average_dev_status(struct msm_gpu *gpu,
> + const struct devfreq_dev_status *raw)
> +{
> + struct msm_gpu_devfreq *df = &gpu->devfreq;
> + const u32 polling_ms = df->devfreq->profile->polling_ms;
> + const u32 max_history_ms = polling_ms * 11 / 10;
> + struct devfreq_dev_status *avg = &df->average_status;
> + u64 avg_freq;
> +
> + /* simple_ondemand governor interacts poorly with gpu->clamp_to_idle.
> + * When we enforce the constraint on idle, it calls get_dev_status
> + * which would normally reset the stats. When we remove the
> + * constraint on active, it calls get_dev_status again where busy_time
> + * would be 0.
> + *
> + * To remedy this, we always return the average load over the past
> + * polling_ms.
> + */
> +
> + /* raw is longer than polling_ms or avg has no history */
> + if (div_u64(raw->total_time, USEC_PER_MSEC) >= polling_ms ||
> + !avg->total_time) {
> + *avg = *raw;
> + return;
> + }
> +
> + /* Truncate the oldest history first.
> + *
> + * Because we keep the history with a single devfreq_dev_status,
> + * rather than a list of devfreq_dev_status, we have to assume freq
> + * and load are the same over avg->total_time. We can scale down
> + * avg->busy_time and avg->total_time by the same factor to drop
> + * history.
> + */
> + if (div_u64(avg->total_time + raw->total_time, USEC_PER_MSEC) >=
> + max_history_ms) {
> + const u32 new_total_time = polling_ms * USEC_PER_MSEC -
> + raw->total_time;
> + avg->busy_time = div_u64(
> + mul_u32_u32(avg->busy_time, new_total_time),
> + avg->total_time);
> + avg->total_time = new_total_time;
> + }
> +
> + /* compute the average freq over avg->total_time + raw->total_time */
> + avg_freq = mul_u32_u32(avg->current_frequency, avg->total_time);
> + avg_freq += mul_u32_u32(raw->current_frequency, raw->total_time);
> + do_div(avg_freq, avg->total_time + raw->total_time);
> +
> + avg->current_frequency = avg_freq;
> + avg->busy_time += raw->busy_time;
> + avg->total_time += raw->total_time;
> +}
> +
> static int msm_devfreq_get_dev_status(struct device *dev,
> struct devfreq_dev_status *status)
> {
> struct msm_gpu *gpu = dev_to_gpu(dev);
> + struct devfreq_dev_status raw;
>
> - get_raw_dev_status(gpu, status);
> + get_raw_dev_status(gpu, &raw);
> + update_average_dev_status(gpu, &raw);
> + *status = gpu->devfreq.average_status;
>
> return 0;
> }
> --
> 2.36.0.rc0.470.gd361397f0d-goog
>
^ permalink raw reply [flat|nested] 8+ messages in thread
* Re: [PATCH 3/3] drm/msm: return the average load over the polling period
@ 2022-04-16 0:50 ` Chia-I Wu
0 siblings, 0 replies; 8+ messages in thread
From: Chia-I Wu @ 2022-04-16 0:50 UTC (permalink / raw)
To: freedreno
Cc: ML dri-devel, Sean Paul, Abhinav Kumar, David Airlie,
linux-arm-msm, Daniel Vetter, Rob Clark, Chanwoo Choi
On Fri, Apr 15, 2022 at 5:33 PM Chia-I Wu <olvaffe@gmail.com> wrote:
>
> simple_ondemand interacts poorly with clamp_to_idle. It only looks at
> the load since the last get_dev_status call, while it should really look
> at the load over polling_ms. When clamp_to_idle true, it almost always
> picks the lowest frequency on active because the gpu is idle between
> msm_devfreq_idle/msm_devfreq_active.
>
> This logic could potentially be moved into devfreq core.
The idea is to extend devfreq_simple_ondemand_data to specify whether
devfreq_simple_ondemand_func should use "last status" or "average
status" to determine the target frequency. devfreq core will need to
store "struct devfreq_dev_status average_status", which will be
updated when a device uses simple_ondemand and asks for average_status
instead of last_status.
>
> Fixes: 7c0ffcd40b16 ("drm/msm/gpu: Respect PM QoS constraints")
> Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
> Cc: Rob Clark <robdclark@chromium.org>
> ---
> drivers/gpu/drm/msm/msm_gpu.h | 3 ++
> drivers/gpu/drm/msm/msm_gpu_devfreq.c | 60 ++++++++++++++++++++++++++-
> 2 files changed, 62 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
> index 389c6dab751b..143c56f5185b 100644
> --- a/drivers/gpu/drm/msm/msm_gpu.h
> +++ b/drivers/gpu/drm/msm/msm_gpu.h
> @@ -9,6 +9,7 @@
>
> #include <linux/adreno-smmu-priv.h>
> #include <linux/clk.h>
> +#include <linux/devfreq.h>
> #include <linux/interconnect.h>
> #include <linux/pm_opp.h>
> #include <linux/regulator/consumer.h>
> @@ -117,6 +118,8 @@ struct msm_gpu_devfreq {
> /** idle_time: Time of last transition to idle: */
> ktime_t idle_time;
>
> + struct devfreq_dev_status average_status;
> +
> /**
> * idle_work:
> *
> diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
> index f531015107c3..d2539ca78c29 100644
> --- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
> +++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
> @@ -9,6 +9,7 @@
>
> #include <linux/devfreq.h>
> #include <linux/devfreq_cooling.h>
> +#include <linux/math64.h>
> #include <linux/units.h>
>
> /*
> @@ -75,12 +76,69 @@ static void get_raw_dev_status(struct msm_gpu *gpu,
> status->busy_time = busy_time;
> }
>
> +static void update_average_dev_status(struct msm_gpu *gpu,
> + const struct devfreq_dev_status *raw)
> +{
> + struct msm_gpu_devfreq *df = &gpu->devfreq;
> + const u32 polling_ms = df->devfreq->profile->polling_ms;
> + const u32 max_history_ms = polling_ms * 11 / 10;
> + struct devfreq_dev_status *avg = &df->average_status;
> + u64 avg_freq;
> +
> + /* simple_ondemand governor interacts poorly with gpu->clamp_to_idle.
> + * When we enforce the constraint on idle, it calls get_dev_status
> + * which would normally reset the stats. When we remove the
> + * constraint on active, it calls get_dev_status again where busy_time
> + * would be 0.
> + *
> + * To remedy this, we always return the average load over the past
> + * polling_ms.
> + */
> +
> + /* raw is longer than polling_ms or avg has no history */
> + if (div_u64(raw->total_time, USEC_PER_MSEC) >= polling_ms ||
> + !avg->total_time) {
> + *avg = *raw;
> + return;
> + }
> +
> + /* Truncate the oldest history first.
> + *
> + * Because we keep the history with a single devfreq_dev_status,
> + * rather than a list of devfreq_dev_status, we have to assume freq
> + * and load are the same over avg->total_time. We can scale down
> + * avg->busy_time and avg->total_time by the same factor to drop
> + * history.
> + */
> + if (div_u64(avg->total_time + raw->total_time, USEC_PER_MSEC) >=
> + max_history_ms) {
> + const u32 new_total_time = polling_ms * USEC_PER_MSEC -
> + raw->total_time;
> + avg->busy_time = div_u64(
> + mul_u32_u32(avg->busy_time, new_total_time),
> + avg->total_time);
> + avg->total_time = new_total_time;
> + }
> +
> + /* compute the average freq over avg->total_time + raw->total_time */
> + avg_freq = mul_u32_u32(avg->current_frequency, avg->total_time);
> + avg_freq += mul_u32_u32(raw->current_frequency, raw->total_time);
> + do_div(avg_freq, avg->total_time + raw->total_time);
> +
> + avg->current_frequency = avg_freq;
> + avg->busy_time += raw->busy_time;
> + avg->total_time += raw->total_time;
> +}
> +
> static int msm_devfreq_get_dev_status(struct device *dev,
> struct devfreq_dev_status *status)
> {
> struct msm_gpu *gpu = dev_to_gpu(dev);
> + struct devfreq_dev_status raw;
>
> - get_raw_dev_status(gpu, status);
> + get_raw_dev_status(gpu, &raw);
> + update_average_dev_status(gpu, &raw);
> + *status = gpu->devfreq.average_status;
>
> return 0;
> }
> --
> 2.36.0.rc0.470.gd361397f0d-goog
>
^ permalink raw reply [flat|nested] 8+ messages in thread