From: Paul Cercueil <paul@crapouillou.net> Date: 2021-03-05 17:02:29
The input_sync() function will already be called within
gpio_keys_gpio_report_event(), so there's no need to call it again after
the loop in gpio_keys_report_state().
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
---
drivers/input/keyboard/gpio_keys.c | 2 --
1 file changed, 2 deletions(-)
From: Paul Cercueil <paul@crapouillou.net> Date: 2021-03-05 17:02:29
We want to be able to report the input event as soon as the debounce
delay elapsed. However, the current code does not really ensure that,
as it uses the jiffies-based schedule_delayed_work() API. With a small
enough HZ value (HZ <= 100), this results in some input events being
lost, when a key is quickly pressed then released (on a human's time
scale).
Switching to hrtimers fixes this issue, and will work even on extremely
low HZ values (tested at HZ=24).
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
---
drivers/input/keyboard/gpio_keys.c | 33 +++++++++++++++---------------
1 file changed, 17 insertions(+), 16 deletions(-)
@@ -40,7 +39,7 @@ struct gpio_button_data {structhrtimerrelease_timer;unsignedintrelease_delay;/* in msecs, for IRQ-only buttons */-structdelayed_workwork;+structhrtimerdebounce_timer;unsignedintsoftware_debounce;/* in msecs, for GPIO-driven buttons */unsignedintirq;
From: Paul Cercueil <paul@crapouillou.net> Date: 2021-03-05 17:02:29
Dealing with input, timing is important; if the button should be
released in one millisecond, then it should be done in one millisecond
and not a hundred milliseconds.
Therefore, the standard timer API is not really suitable for this task.
Convert the gpio-keys driver to use a hrtimer instead of the standard
timer to address this issue.
Note that by using a hard IRQ for the hrtimer callback, we can get rid
of the spin_lock_irqsave() and spin_unlock_irqrestore().
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
---
drivers/input/keyboard/gpio_keys.c | 27 ++++++++++++++++-----------
1 file changed, 16 insertions(+), 11 deletions(-)
I am not sure how this works. As far as I know, even
HRTIMER_MODE_REL_SOFT do not allow sleeping in the timer handlers, and
gpio_keys_gpio_report_event() use sleeping variant of GPIOD API (and
that is not going to change).
It seems to me that if you want to use software debounce in gpio keys
driver you need to set up sufficiently high HZ for your system. Maybe we
could thrown a warning when we see low debounce delay and low HZ to
alert system developer.
Thanks.
--
Dmitry
On Fri, Mar 05, 2021 at 05:01:09PM +0000, Paul Cercueil wrote:
The input_sync() function will already be called within
gpio_keys_gpio_report_event(), so there's no need to call it again after
the loop in gpio_keys_report_state().
I'd probably go other way around and remove the sync from
gpio_keys_report_state() and add one to gpio_keys_gpio_work_func() so
that we do not have to send bunch of input_sync() when there are several
keys/buttons.
Thanks.
--
Dmitry
I am not sure how this works. As far as I know, even
HRTIMER_MODE_REL_SOFT do not allow sleeping in the timer handlers, and
gpio_keys_gpio_report_event() use sleeping variant of GPIOD API (and
that is not going to change).
Quoting <linux/hrtimers.h>, the "timer callback will be executed in
soft irq context", so sleeping should be possible.
But I guess in this case I can use HRTIMER_MODE_REL.
It seems to me that if you want to use software debounce in gpio keys
driver you need to set up sufficiently high HZ for your system. Maybe
we
could thrown a warning when we see low debounce delay and low HZ to
alert system developer.
This is exactly what we should not do. I certainly don't want to have
250+ timer interrupts per second just so that input events aren't lost,
to work around a sucky debounce implementation. Besides, if you
consider the hrtimers doc (Documentation/timers/hrtimers.rst), hrtimers
really are what should be used here.
-Paul
I am not sure how this works. As far as I know, even
HRTIMER_MODE_REL_SOFT do not allow sleeping in the timer handlers, and
gpio_keys_gpio_report_event() use sleeping variant of GPIOD API (and
that is not going to change).
Quoting <linux/hrtimers.h>, the "timer callback will be executed in soft irq
context", so sleeping should be possible.
I am afraid you misunderstand what soft irq context is, as softirqs and
tasklets still run in interrupt context and therefore can not sleep,
only code running in process context may sleep.
You can test it yourself by sticking "msleep(1)" in
gpio_keys_debounce_timer() and see if you will get "scheduling while
atomic" in logs.
But I guess in this case I can use HRTIMER_MODE_REL.
This changes selected clock source, but has no effect on whether timer
handler can sleep or not.
quoted
It seems to me that if you want to use software debounce in gpio keys
driver you need to set up sufficiently high HZ for your system. Maybe we
could thrown a warning when we see low debounce delay and low HZ to
alert system developer.
This is exactly what we should not do. I certainly don't want to have 250+
timer interrupts per second just so that input events aren't lost, to work
around a sucky debounce implementation. Besides, if you consider the
hrtimers doc (Documentation/timers/hrtimers.rst), hrtimers really are what
should be used here.
I explained why they can't. They could be if you restrict gpio_keys to
only be used with GPIOs that do not require sleep to read their state,
but I am not willing to accept such restriction. You either need to have
longer debounce, higher HZ, or see if you can use GPIO controller that
supports debounce handling. See also if you can enable dynamic
ticks/NO_HZ to limit number of timer interrupts on idle system.
Thanks.
--
Dmitry
From: Paul Cercueil <paul@crapouillou.net> Date: 2021-03-07 21:12:47
Hi Dmitry,
Le dim. 7 mars 2021 à 12:20, Dmitry Torokhov
[off-list ref] a écrit :
On Fri, Mar 05, 2021 at 08:00:43PM +0000, Paul Cercueil wrote:
quoted
Hi Dmitry,
Le ven. 5 mars 2021 à 10:35, Dmitry Torokhov
[off-list ref] a
écrit :
> Hi Paul,
>
> On Fri, Mar 05, 2021 at 05:01:11PM +0000, Paul Cercueil wrote:
> > -static void gpio_keys_gpio_work_func(struct work_struct *work)
> > +static enum hrtimer_restart gpio_keys_debounce_timer(struct
> > hrtimer *t)
> > {
> > - struct gpio_button_data *bdata =
> > - container_of(work, struct gpio_button_data, work.work);
> > + struct gpio_button_data *bdata = container_of(t,
> > + struct gpio_button_data,
> > + debounce_timer);
> >
> > gpio_keys_gpio_report_event(bdata);
>
> I am not sure how this works. As far as I know, even
> HRTIMER_MODE_REL_SOFT do not allow sleeping in the timer
handlers, and
> gpio_keys_gpio_report_event() use sleeping variant of GPIOD API
(and
> that is not going to change).
Quoting <linux/hrtimers.h>, the "timer callback will be executed in
soft irq
context", so sleeping should be possible.
I am afraid you misunderstand what soft irq context is, as softirqs
and
tasklets still run in interrupt context and therefore can not sleep,
only code running in process context may sleep.
I probably do. My understanding of "softirq" is that the callback runs
in a threaded interrupt handler.
You can test it yourself by sticking "msleep(1)" in
gpio_keys_debounce_timer() and see if you will get "scheduling while
atomic" in logs.
I tested it, it locks up.
quoted
But I guess in this case I can use HRTIMER_MODE_REL.
This changes selected clock source, but has no effect on whether timer
handler can sleep or not.
quoted
> It seems to me that if you want to use software debounce in gpio
keys
> driver you need to set up sufficiently high HZ for your system.
Maybe we
> could thrown a warning when we see low debounce delay and low HZ
to
> alert system developer.
This is exactly what we should not do. I certainly don't want to
have 250+
timer interrupts per second just so that input events aren't lost,
to work
around a sucky debounce implementation. Besides, if you consider the
hrtimers doc (Documentation/timers/hrtimers.rst), hrtimers really
are what
should be used here.
I explained why they can't. They could be if you restrict gpio_keys to
only be used with GPIOs that do not require sleep to read their state,
but I am not willing to accept such restriction. You either need to
have
longer debounce, higher HZ, or see if you can use GPIO controller that
supports debounce handling. See also if you can enable dynamic
ticks/NO_HZ to limit number of timer interrupts on idle system.
We can also use the hrtimer approach if the GPIO doesn't require sleep,
and fall back to the standard timer if it does. It's possible to detect
that with gpiod_cansleep(). The diff would be pretty slim. Would you
accept something like that?
Switching from HZ=250 to HZ=24 leads to a 3% overall performance
increase across all apps on our system, so a pretty big optimization,
and this is the only blocker.
Cheers,
-Paul
On Sun, Mar 07, 2021 at 09:11:18PM +0000, Paul Cercueil wrote:
Hi Dmitry,
Le dim. 7 mars 2021 à 12:20, Dmitry Torokhov [off-list ref] a
écrit :
quoted
On Fri, Mar 05, 2021 at 08:00:43PM +0000, Paul Cercueil wrote:
quoted
Hi Dmitry,
Le ven. 5 mars 2021 à 10:35, Dmitry Torokhov
[off-list ref] a
écrit :
> Hi Paul,
>
> On Fri, Mar 05, 2021 at 05:01:11PM +0000, Paul Cercueil wrote:
> > -static void gpio_keys_gpio_work_func(struct work_struct *work)
> > +static enum hrtimer_restart gpio_keys_debounce_timer(struct
> > hrtimer *t)
> > {
> > - struct gpio_button_data *bdata =
> > - container_of(work, struct gpio_button_data, work.work);
> > + struct gpio_button_data *bdata = container_of(t,
> > + struct gpio_button_data,
> > + debounce_timer);
> >
> > gpio_keys_gpio_report_event(bdata);
>
> I am not sure how this works. As far as I know, even
> HRTIMER_MODE_REL_SOFT do not allow sleeping in the timer
handlers, and
> gpio_keys_gpio_report_event() use sleeping variant of GPIOD API
(and
> that is not going to change).
Quoting <linux/hrtimers.h>, the "timer callback will be executed in
soft irq
context", so sleeping should be possible.
I am afraid you misunderstand what soft irq context is, as softirqs and
tasklets still run in interrupt context and therefore can not sleep,
only code running in process context may sleep.
I probably do. My understanding of "softirq" is that the callback runs in a
threaded interrupt handler.
No, you are thinking about threaded interrupts, which are separate
beasts. Softirqs are traditional bottom halfs that run after exit of
hard interrupt, but still are non-preemptible so sleeping is not
allowed.
quoted
You can test it yourself by sticking "msleep(1)" in
gpio_keys_debounce_timer() and see if you will get "scheduling while
atomic" in logs.
I tested it, it locks up.
quoted
quoted
But I guess in this case I can use HRTIMER_MODE_REL.
This changes selected clock source, but has no effect on whether timer
handler can sleep or not.
quoted
> It seems to me that if you want to use software debounce in gpio
keys
> driver you need to set up sufficiently high HZ for your system.
Maybe we
> could thrown a warning when we see low debounce delay and low HZ
to
> alert system developer.
This is exactly what we should not do. I certainly don't want to
have 250+
timer interrupts per second just so that input events aren't lost,
to work
around a sucky debounce implementation. Besides, if you consider the
hrtimers doc (Documentation/timers/hrtimers.rst), hrtimers really
are what
should be used here.
I explained why they can't. They could be if you restrict gpio_keys to
only be used with GPIOs that do not require sleep to read their state,
but I am not willing to accept such restriction. You either need to have
longer debounce, higher HZ, or see if you can use GPIO controller that
supports debounce handling. See also if you can enable dynamic
ticks/NO_HZ to limit number of timer interrupts on idle system.
We can also use the hrtimer approach if the GPIO doesn't require sleep, and
fall back to the standard timer if it does. It's possible to detect that
with gpiod_cansleep(). The diff would be pretty slim. Would you accept
something like that?
Switching from HZ=250 to HZ=24 leads to a 3% overall performance increase
across all apps on our system, so a pretty big optimization, and this is the
only blocker.
Let me take a look at the updated patch and we will see.
Thanks.
--
Dmitry