[PATCH v3 0/2] clocksource/drivers/dw_apb_timer: improve performance a bit

[PATCH v3 0/2] clocksource/drivers/dw_apb_timer: improve performance a bit

[Jisheng Zhang]

These two patches try to improve the dw_apb_timer clocksource/clockevent
performance.

since v2:
 - only use relaxed version in critical path
 - inline apbt_readl/apbt_writel
 - add some performance numbers

since v1:
 - correct the wrong sentence in commit msg about writel performance on
   CA9 with outer L2 cache. Thank Arnd for pointing out this in another
   thread.

Jisheng Zhang (2):
  clocksource/drivers/dw_apb_timer: Inline apbt_readl and apbt_writel
  clocksource/drivers/dw_apb_timer: Use {readl|writel}_relaxed in
    critical path

 drivers/clocksource/dw_apb_timer.c | 30 +++++++++++++++++++++---------
 1 file changed, 21 insertions(+), 9 deletions(-)

-- 
2.6.2

[PATCH v3 1/2] clocksource/drivers/dw_apb_timer: Inline apbt_readl and apbt_writel

[Jisheng Zhang]

It seems gcc can automatically inline apbt_writel() for us, but
apbt_real isn't inlined. This patch makes them inline to get a trivial
performance improvement: the time __apbt_read_clocksource() spends on
Marvell BG4CT platform:

 before the patch 1275240ns on average
 after the patch 1263240ns on average

so we get 1% performance improvement.

Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
---
 drivers/clocksource/dw_apb_timer.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/clocksource/dw_apb_timer.c b/drivers/clocksource/dw_apb_timer.c
index 3a6d9db..114696b 100644
--- a/drivers/clocksource/dw_apb_timer.c
+++ b/drivers/clocksource/dw_apb_timer.c
@@ -49,13 +49,13 @@ clocksource_to_dw_apb_clocksource(struct clocksource *cs)
 	return container_of(cs, struct dw_apb_clocksource, cs);
 }
 
-static u32 apbt_readl(struct dw_apb_timer *timer, unsigned long offs)
+static inline u32 apbt_readl(struct dw_apb_timer *timer, unsigned long offs)
 {
 	return readl(timer->base + offs);
 }
 
-static void apbt_writel(struct dw_apb_timer *timer, u32 val,
-		 unsigned long offs)
+static inline void apbt_writel(struct dw_apb_timer *timer, u32 val,
+			unsigned long offs)
 {
 	writel(val, timer->base + offs);
 }
-- 
2.6.2

[PATCH v3 2/2] clocksource/drivers/dw_apb_timer: Use {readl|writel}_relaxed in critical path

[Jisheng Zhang]

It's safe to use the relaxed version. From another side, the relaxed io
accessor macros are available on all architectures now, so we can use
the relaxed versions to get a trivial system performance improvement,
we measured time the following functions spent on Marvell BG4CT:

__apbt_read_clocksource():

before the patch: 1263240ns on average
after the patch: 1250080ns on average
improved by 1%

apbt_eoi():

before the patch: 1290960ns on average
after the patch: 1248240ns on average

apbt_next_event():

before the patch: 3333660ns on average
after the patch: 1322040ns on average

improved by 60%!

Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
---
 drivers/clocksource/dw_apb_timer.c | 24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)

diff --git a/drivers/clocksource/dw_apb_timer.c b/drivers/clocksource/dw_apb_timer.c
index 114696b..6334526 100644
--- a/drivers/clocksource/dw_apb_timer.c
+++ b/drivers/clocksource/dw_apb_timer.c
@@ -60,6 +60,17 @@ static inline void apbt_writel(struct dw_apb_timer *timer, u32 val,
 	writel(val, timer->base + offs);
 }
 
+static inline u32 apbt_readl_relaxed(struct dw_apb_timer *timer, unsigned long offs)
+{
+	return readl_relaxed(timer->base + offs);
+}
+
+static inline void apbt_writel_relaxed(struct dw_apb_timer *timer, u32 val,
+			unsigned long offs)
+{
+	writel_relaxed(val, timer->base + offs);
+}
+
 static void apbt_disable_int(struct dw_apb_timer *timer)
 {
 	u32 ctrl = apbt_readl(timer, APBTMR_N_CONTROL);
@@ -81,7 +92,7 @@ void dw_apb_clockevent_pause(struct dw_apb_clock_event_device *dw_ced)
 
 static void apbt_eoi(struct dw_apb_timer *timer)
 {
-	apbt_readl(timer, APBTMR_N_EOI);
+	apbt_readl_relaxed(timer, APBTMR_N_EOI);
 }
 
 static irqreturn_t dw_apb_clockevent_irq(int irq, void *data)
@@ -200,13 +211,13 @@ static int apbt_next_event(unsigned long delta,
 	struct dw_apb_clock_event_device *dw_ced = ced_to_dw_apb_ced(evt);
 
 	/* Disable timer */
-	ctrl = apbt_readl(&dw_ced->timer, APBTMR_N_CONTROL);
+	ctrl = apbt_readl_relaxed(&dw_ced->timer, APBTMR_N_CONTROL);
 	ctrl &= ~APBTMR_CONTROL_ENABLE;
-	apbt_writel(&dw_ced->timer, ctrl, APBTMR_N_CONTROL);
+	apbt_writel_relaxed(&dw_ced->timer, ctrl, APBTMR_N_CONTROL);
 	/* write new count */
-	apbt_writel(&dw_ced->timer, delta, APBTMR_N_LOAD_COUNT);
+	apbt_writel_relaxed(&dw_ced->timer, delta, APBTMR_N_LOAD_COUNT);
 	ctrl |= APBTMR_CONTROL_ENABLE;
-	apbt_writel(&dw_ced->timer, ctrl, APBTMR_N_CONTROL);
+	apbt_writel_relaxed(&dw_ced->timer, ctrl, APBTMR_N_CONTROL);
 
 	return 0;
 }
@@ -342,7 +353,8 @@ static cycle_t __apbt_read_clocksource(struct clocksource *cs)
 	struct dw_apb_clocksource *dw_cs =
 		clocksource_to_dw_apb_clocksource(cs);
 
-	current_count = apbt_readl(&dw_cs->timer, APBTMR_N_CURRENT_VALUE);
+	current_count = apbt_readl_relaxed(&dw_cs->timer,
+					APBTMR_N_CURRENT_VALUE);
 
 	return (cycle_t)~current_count;
 }
-- 
2.6.2

[PATCH v3 0/2] clocksource/drivers/dw_apb_timer: improve performance a bit

[Jisheng Zhang]

On Wed, 25 Nov 2015 23:48:32 +0800
Jisheng Zhang wrote:

> These two patches try to improve the dw_apb_timer clocksource/clockevent
> performance.

These patches depends on the apbt_readl return value fix patch:

http://lists.infradead.org/pipermail/linux-arm-kernel/2015-November/388250.html

thanks,
Jisheng

> 
> since v2:
>  - only use relaxed version in critical path
>  - inline apbt_readl/apbt_writel
>  - add some performance numbers
> 
> since v1:
>  - correct the wrong sentence in commit msg about writel performance on
>    CA9 with outer L2 cache. Thank Arnd for pointing out this in another
>    thread.
> 
> Jisheng Zhang (2):
>   clocksource/drivers/dw_apb_timer: Inline apbt_readl and apbt_writel
>   clocksource/drivers/dw_apb_timer: Use {readl|writel}_relaxed in
>     critical path
> 
>  drivers/clocksource/dw_apb_timer.c | 30 +++++++++++++++++++++---------
>  1 file changed, 21 insertions(+), 9 deletions(-)
>