Skip to contentSkip to conversation

iio: adc: exynos: do not rely on 'users' counter in ISR

PATCH Superseded [email protected] 3 participants 6 messages Oct 5, 2020 – Oct 6, 2020

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 iio: adc: exynos: do not rely on 'users' counter in ISR

The order in which 'users' counter is decremented vs calling drivers' close() method is implementation specific, and we should not rely on it. Let's introduce driver private flag and use it to signal ISR to exit when device is being closed.

This has a side-effect of fixing issue of accessing inut->users outside of input->mutex protection.

---
 drivers/iio/adc/exynos_adc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/iio/adc/exynos_adc.c +7 -1
index 22131a677445..7eb2a5df6e98 100644
@@ -135,6 +135,8 @@ struct exynos_adc {
 	u32			value;
 	unsigned int            version;
 
+	bool			ts_enabled;
+
 	bool			read_ts;
 	u32			ts_x;
 	u32			ts_y;
@@ -633,7 +635,7 @@ static irqreturn_t exynos_ts_isr(int irq, void *dev_id)
 	bool pressed;
 	int ret;
 
-	while (info->input->users) {
+	while (info->ts_enabled) {
 		ret = exynos_read_s3c64xx_ts(dev, &x, &y);
 		if (ret == -ETIMEDOUT)
 			break;
@@ -712,6 +714,8 @@ static int exynos_adc_ts_open(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
+	info->ts_enabled = true;
+	mb();
 	enable_irq(info->tsirq);
 
 	return 0;
@@ -721,6 +725,8 @@ static void exynos_adc_ts_close(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
+	info->ts_enabled = false;
+	mb();
 	disable_irq(info->tsirq);
 }
 
Signature · 5 lines
2.28.0.806.g8561365e88-goog


-- 
Dmitry
maintainer Acked-by on v1 1/1
#1 lore raw
Quoted 5 lines · On Sun, Oct 04, 2020 at 10:24:20PM -0700, [email protected] wrote:

The order in which 'users' counter is decremented vs calling drivers' close() method is implementation specific, and we should not rely on it. Let's introduce driver private flag and use it to signal ISR to exit when device is being closed.

Quoted 8 lines

This has a side-effect of fixing issue of accessing inut->users outside of input->mutex protection.

---
 drivers/iio/adc/exynos_adc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

Best regards,
Krzysztof

#2 lore raw
In reply to [email protected]
Quoted 46 lines · On Sun, Oct 04, 2020 at 10:24:20PM -0700, [email protected] wrote:

The order in which 'users' counter is decremented vs calling drivers' close() method is implementation specific, and we should not rely on it. Let's introduce driver private flag and use it to signal ISR to exit when device is being closed.

This has a side-effect of fixing issue of accessing inut->users outside of input->mutex protection.

---
 drivers/iio/adc/exynos_adc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/iio/adc/exynos_adc.c +5 -1
index 22131a677445..7eb2a5df6e98 100644
@@ -135,6 +135,8 @@ struct exynos_adc {
 	u32			value;
 	unsigned int            version;
 
+	bool			ts_enabled;
+
 	bool			read_ts;
 	u32			ts_x;
 	u32			ts_y;
@@ -633,7 +635,7 @@ static irqreturn_t exynos_ts_isr(int irq, void *dev_id)
 	bool pressed;
 	int ret;
 
-	while (info->input->users) {
+	while (info->ts_enabled) {
 		ret = exynos_read_s3c64xx_ts(dev, &x, &y);
 		if (ret == -ETIMEDOUT)
 			break;
@@ -712,6 +714,8 @@ static int exynos_adc_ts_open(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
+	info->ts_enabled = true;
+	mb();
 	enable_irq(info->tsirq);
 
 	return 0;

drivers/iio/adc/exynos_adc.c static void exynos_adc_ts_close(struct input_dev *dev) · line 730

Quoted 6 lines
drivers/iio/adc/exynos_adc.c +2 -0
@@ -721,6 +725,8 @@ static void exynos_adc_ts_close(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
+	info->ts_enabled = false;
+	mb();
 	disable_irq(info->tsirq);

This should be WRITE_ONCE paired with READ_ONCE in the ISR.

But is the check really needed? I see that this is to break waiting for a touch release event, so I would assume this shouldn't wait forever (unless the hardware is buggy) and breaking the loop will desync touch state (I would guess this would be noticable by next user).

Best Regards,
Michał Mirosław

#3 lore raw

Hi Michał,

Quoted 50 lines · Michał Mirosław
Quoted 49 lines · On Sun, Oct 04, 2020 at 10:24:20PM -0700, [email protected] wrote:

The order in which 'users' counter is decremented vs calling drivers' close() method is implementation specific, and we should not rely on it. Let's introduce driver private flag and use it to signal ISR to exit when device is being closed.

This has a side-effect of fixing issue of accessing inut->users outside of input->mutex protection.

---
 drivers/iio/adc/exynos_adc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/iio/adc/exynos_adc.c +5 -1
index 22131a677445..7eb2a5df6e98 100644
@@ -135,6 +135,8 @@ struct exynos_adc {
 	u32			value;
 	unsigned int            version;
 
+	bool			ts_enabled;
+
 	bool			read_ts;
 	u32			ts_x;
 	u32			ts_y;
@@ -633,7 +635,7 @@ static irqreturn_t exynos_ts_isr(int irq, void *dev_id)
 	bool pressed;
 	int ret;
 
-	while (info->input->users) {
+	while (info->ts_enabled) {
 		ret = exynos_read_s3c64xx_ts(dev, &x, &y);
 		if (ret == -ETIMEDOUT)
 			break;
@@ -712,6 +714,8 @@ static int exynos_adc_ts_open(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
+	info->ts_enabled = true;
+	mb();
 	enable_irq(info->tsirq);
 
 	return 0;
@@ -721,6 +725,8 @@ static void exynos_adc_ts_close(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
Quoted 5 lines

drivers/iio/adc/exynos_adc.c static void exynos_adc_ts_close(struct input_dev *dev) · line 730

Quoted 3 lines
drivers/iio/adc/exynos_adc.c +2 -0
+	info->ts_enabled = false;
+	mb();
 	disable_irq(info->tsirq);

This should be WRITE_ONCE paired with READ_ONCE in the ISR.

Why? I can see that maybe full memory barrier is too heavy when we set the flag to true, but the only requirement is for the flag to be set before we disable IRQ, so any additional guarantees provided by WRITE_ONCE are not needed. On the read side we want the flag to be noticed eventually, and there is no additional dependencies on the data, so it is unclear what READ_ONCE() will give us here.

Quoted 4 lines

But is the check really needed? I see that this is to break waiting for a touch release event, so I would assume this shouldn't wait forever (unless the hardware is buggy)

It is not hardware, it is user. Do you want to delay indefinitely close() just because user has a finger on the touchscreen?

Quoted 2 lines

and breaking the loop will desync touch
state (I would guess this would be noticable by next user).

Upon next open driver will service the interrupt and provide new set of touch coordinates. Userspace is supposed to query current state of device when opening it before starting processing events. Or you are concerned about some other state?

In any case, this is current driver behavior and if it needs to be changed it is a topic for a separate patch.

Thanks.

Signature · 1 line
Dmitry
#4 lore raw
Quoted 59 lines · On Mon, Oct 05, 2020 at 10:36:36AM -0700, [email protected] wrote:

Hi Michał,

Quoted 50 lines · Michał Mirosław
Quoted 49 lines · On Sun, Oct 04, 2020 at 10:24:20PM -0700, [email protected] wrote:

The order in which 'users' counter is decremented vs calling drivers' close() method is implementation specific, and we should not rely on it. Let's introduce driver private flag and use it to signal ISR to exit when device is being closed.

This has a side-effect of fixing issue of accessing inut->users outside of input->mutex protection.

---
 drivers/iio/adc/exynos_adc.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/iio/adc/exynos_adc.c +5 -1
index 22131a677445..7eb2a5df6e98 100644
@@ -135,6 +135,8 @@ struct exynos_adc {
 	u32			value;
 	unsigned int            version;
 
+	bool			ts_enabled;
+
 	bool			read_ts;
 	u32			ts_x;
 	u32			ts_y;
@@ -633,7 +635,7 @@ static irqreturn_t exynos_ts_isr(int irq, void *dev_id)
 	bool pressed;
 	int ret;
 
-	while (info->input->users) {
+	while (info->ts_enabled) {
 		ret = exynos_read_s3c64xx_ts(dev, &x, &y);
 		if (ret == -ETIMEDOUT)
 			break;
@@ -712,6 +714,8 @@ static int exynos_adc_ts_open(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
+	info->ts_enabled = true;
+	mb();
 	enable_irq(info->tsirq);
 
 	return 0;
@@ -721,6 +725,8 @@ static void exynos_adc_ts_close(struct input_dev *dev)
 {
 	struct exynos_adc *info = input_get_drvdata(dev);
 
Quoted 5 lines

drivers/iio/adc/exynos_adc.c static void exynos_adc_ts_close(struct input_dev *dev) · line 730

Quoted 3 lines
drivers/iio/adc/exynos_adc.c +2 -0
+	info->ts_enabled = false;
+	mb();
 	disable_irq(info->tsirq);

This should be WRITE_ONCE paired with READ_ONCE in the ISR.

Quoted 6 lines

Why? I can see that maybe full memory barrier is too heavy when we set the flag to true, but the only requirement is for the flag to be set before we disable IRQ, so any additional guarantees provided by WRITE_ONCE are not needed. On the read side we want the flag to be noticed eventually, and there is no additional dependencies on the data, so it is unclear what READ_ONCE() will give us here.

Without READ_ONCE you have no guarantee that the compiler won't optimize 'while (flag) ...' to 'if (flag) for(;;) ...'.

Maybe the platform has stronger coherency guarantees than Linux memory model assumes, but if not, a CPU running the ISR (without paired memory barrier) might not ever see the store from another CPU (both in current and proposed code).

Quoted 6 lines
Quoted 3 lines

But is the check really needed? I see that this is to break waiting for a touch release event, so I would assume this shouldn't wait forever (unless the hardware is buggy)

It is not hardware, it is user. Do you want to delay indefinitely close() just because user has a finger on the touchscreen?

Ack. This covers the why of loop breaking.

Quoted 6 lines
Quoted 2 lines

and breaking the loop will desync touch
state (I would guess this would be noticable by next user).

Upon next open driver will service the interrupt and provide new set of touch coordinates. Userspace is supposed to query current state of device when opening it before starting processing events. Or you are concerned about some other state?

From the code I would expect that there is a slight window, wher when the user releases the touch between close() and open(), the client that open()s will see a 'pressed' state until the ISR runs again (probably immediately because of pending interrupt). OTOH, maybe the app should be prepared for that anyway?

Quoted 2 lines

In any case, this is current driver behavior and if it needs to be changed it is a topic for a separate patch.

Agreed.

Quoted 5 lines

Thanks.

Signature · 1 line
Dmitry
#5 lore raw
Quoted 11 lines · Michał Mirosław
Quoted 9 lines · On Mon, Oct 05, 2020 at 10:36:36AM -0700, [email protected] wrote:

Hi Michał,

Quoted 2 lines · Michał Mirosław

and breaking the loop will desync touch
state (I would guess this would be noticable by next user).

Upon next open driver will service the interrupt and provide new set of touch coordinates. Userspace is supposed to query current state of device when opening it before starting processing events. Or you are concerned about some other state?

Quoted 5 lines

From the code I would expect that there is a slight window, wher when the user releases the touch between close() and open(), the client that open()s will see a 'pressed' state until the ISR runs again (probably immediately because of pending interrupt). OTOH, maybe the app should be prepared for that anyway?

I think users should be prepared for it. There could be many users, so anyone opening interface device (evdev or similar) can not expect that the hardware device is in a quiesce state; it could just happen that finger was on the surface and will be released as someone is opening the device anyway.

Thanks.

Signature · 1 line
Dmitry