From: Roman Beranek <hidden> Date: 2021-05-31 04:48:15
As Emil Lenngren has previously shown [1], actually only 1-2 cycles of
the prescaler-divided clock are necessary to pass before the PWM turns
off, not a full period.
To avoid having the PWM re-enabled from another thread while asleep,
ctrl_lock spinlock was converted to a mutex so that it can be released
only after the clock gate has finally been turned on.
[1] https://linux-sunxi.org/PWM_Controller_Register_Guide
Roman Beranek (6):
pwm: sun4i: enable clk prior to getting its rate
pwm: sun4i: disable EN bit prior to the delay
pwm: sun4i: replace spinlock with a mutex
pwm: sun4i: simplify calculation of the delay time
pwm: sun4i: shorten the delay to 2 cycles
pwm: sun4i: don't delay if the PWM is already off
drivers/pwm/pwm-sun4i.c | 56 +++++++++++++++++++----------------------
1 file changed, 26 insertions(+), 30 deletions(-)
--
2.31.1
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Roman Beranek <hidden> Date: 2021-05-31 04:47:48
Ensure the PWM clock is enabled prior to retrieving its rate, as is
already being done in sun4i_pwm_apply.
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 7 +++++++
1 file changed, 7 insertions(+)
From: Roman Beranek <hidden> Date: 2021-05-31 04:47:52
The reason why we wait before gating the clock is to allow for the PWM
to finish its cycle and stop. But it won't stop unless the EN bit is
disabled.
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
From: Roman Beranek <hidden> Date: 2021-05-31 04:48:27
There's no reason to expect a single jiffy has passed since writing
the CTRL register except if a preemption has occured in the meantime.
Avoid introducing unnecessary complexity and simply wait for the whole
period.
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 17 ++---------------
1 file changed, 2 insertions(+), 15 deletions(-)
@@ -314,15 +309,7 @@ static int sun4i_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm,}/* We need a full period to elapse before disabling the channel. */-now=jiffies;-if(time_before(now,sun4i_pwm->next_period[pwm->hwpwm])){-delay_us=jiffies_to_usecs(sun4i_pwm->next_period[pwm->hwpwm]--now);-if((delay_us/500)>MAX_UDELAY_MS)-msleep(delay_us/1000+1);-else-usleep_range(delay_us,delay_us*2);-}+fsleep(cstate.period/NSEC_PER_USEC+1);ctrl&=~BIT_CH(PWM_CLK_GATING,pwm->hwpwm);sun4i_pwm_writel(sun4i_pwm,ctrl,PWM_CTRL_REG);
--
2.31.1
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Roman Beranek <hidden> Date: 2021-05-31 04:48:49
Releasing ctrl_lock for the duration of the delay is not desirable as it
allows re-enabling the PWM before the delay is over. Instead, substitute
the spinlock with a mutex so that we can sleep while holding it.
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 20 +++++++++-----------
1 file changed, 9 insertions(+), 11 deletions(-)
@@ -273,7 +273,7 @@ static int sun4i_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm,ctrl|=BIT_CH(PWM_BYPASS,pwm->hwpwm);/* We can skip other parameter */sun4i_pwm_writel(sun4i_pwm,ctrl,PWM_CTRL_REG);-spin_unlock(&sun4i_pwm->ctrl_lock);+mutex_unlock(&sun4i_pwm->ctrl_lock);return0;}
@@ -308,10 +308,10 @@ static int sun4i_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm,sun4i_pwm_writel(sun4i_pwm,ctrl,PWM_CTRL_REG);-spin_unlock(&sun4i_pwm->ctrl_lock);--if(state->enabled)+if(state->enabled){+mutex_unlock(&sun4i_pwm->ctrl_lock);return0;+}/* We need a full period to elapse before disabling the channel. */now=jiffies;
From: Roman Beranek <hidden> Date: 2021-05-31 04:49:01
As Emil Lenngren has previously shown, actually only 1-2 cycles of
the prescaler-divided clock are necessary to pass before the PWM turns
off (instead of a full period). I was able to reproduce his observation
on a A64 using a logic analyzer.
Suggested-by: Emil Lenngren <redacted>
Suggested-by: Pascal Roeleven <redacted>
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 13 ++++++++-----
1 file changed, 8 insertions(+), 5 deletions(-)
@@ -277,7 +277,8 @@ static int sun4i_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm,ctrl&=~BIT_CH(PWM_BYPASS,pwm->hwpwm);}-if(PWM_REG_PRESCAL(ctrl,pwm->hwpwm)!=prescaler){+current_prescaler=PWM_REG_PRESCAL(ctrl,pwm->hwpwm);+if(current_prescaler!=prescaler){/* Prescaler changed, the clock has to be gated */ctrl&=~BIT_CH(PWM_CLK_GATING,pwm->hwpwm);sun4i_pwm_writel(sun4i_pwm,ctrl,PWM_CTRL_REG);
@@ -308,8 +309,10 @@ static int sun4i_pwm_apply(struct pwm_chip *chip, struct pwm_device *pwm,return0;}-/* We need a full period to elapse before disabling the channel. */-fsleep(cstate.period/NSEC_PER_USEC+1);+/* We need to wait 1-2 cycles before disabling the channel. */+cycle_ns=DIV_ROUND_UP(NSEC_PER_SEC,clk_get_rate(sun4i_pwm->clk))+*prescaler_table[current_prescaler];+fsleep(DIV_ROUND_UP(cycle_ns*2,NSEC_PER_USEC));ctrl&=~BIT_CH(PWM_CLK_GATING,pwm->hwpwm);sun4i_pwm_writel(sun4i_pwm,ctrl,PWM_CTRL_REG);
--
2.31.1
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
As Emil Lenngren has previously shown [1], actually only 1-2 cycles of
the prescaler-divided clock are necessary to pass before the PWM turns
off, not a full period.
To avoid having the PWM re-enabled from another thread while asleep,
ctrl_lock spinlock was converted to a mutex so that it can be released
only after the clock gate has finally been turned on.
[1] https://linux-sunxi.org/PWM_Controller_Register_Guide
Roman Beranek (6):
pwm: sun4i: enable clk prior to getting its rate
pwm: sun4i: disable EN bit prior to the delay
pwm: sun4i: replace spinlock with a mutex
pwm: sun4i: simplify calculation of the delay time
pwm: sun4i: shorten the delay to 2 cycles
pwm: sun4i: don't delay if the PWM is already off
drivers/pwm/pwm-sun4i.c | 56 +++++++++++++++++++----------------------
1 file changed, 26 insertions(+), 30 deletions(-)
Hi Roman,
Thanks for your attempt to fix this.
Unfortunately on my A10 device (Topwise A721), the controller still gets
stuck in an unrecoverable state after disabling and re-enabling the PWM
when it was already on (set in U-Boot), or when enabling it when it was
off. In this state, any changes to the period register (using devmem)
don't seem to have any effect. It seems to be stuck in the state it was
before disabling. The only thing which still works is enabling and
disabling.
I can't reproduce this behavior manually so I'm not sure what is causing
this.
Regarding the amount of cycles of sleep; Using a prescaler of 72000 the
PWM clock is 3 ms. Although timing tests using devmem seem unreliable
(too much overhead?), in U-Boot I need to 'sleep' for at least 7 ms
between the commands to make sure the output doesn't sometimes get stuck
in the enabled state. So in my case it seems to be at least 3 cycles. I
am not sure how reliable this method is. However even if I can get it
stuck in the enabled state using a shorter time, it doesn't cause the
behavior I mentioned before. I was always able to recover it manually.
Increasing the number of cycles to sleep therefore also doesn't solve my
problem. Until we can solve that I cannot confirm nor deny if 2 cycles
is enough.
Regards,
Pascal
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
From: Emil Lenngren <hidden> Date: 2021-05-31 20:03:40
Den mån 31 maj 2021 kl 21:07 skrev Pascal Roeleven [off-list ref]:
On 2021-05-31 06:46, Roman Beranek wrote:
quoted
As Emil Lenngren has previously shown [1], actually only 1-2 cycles of
the prescaler-divided clock are necessary to pass before the PWM turns
off, not a full period.
To avoid having the PWM re-enabled from another thread while asleep,
ctrl_lock spinlock was converted to a mutex so that it can be released
only after the clock gate has finally been turned on.
[1] https://linux-sunxi.org/PWM_Controller_Register_Guide
Roman Beranek (6):
pwm: sun4i: enable clk prior to getting its rate
pwm: sun4i: disable EN bit prior to the delay
pwm: sun4i: replace spinlock with a mutex
pwm: sun4i: simplify calculation of the delay time
pwm: sun4i: shorten the delay to 2 cycles
pwm: sun4i: don't delay if the PWM is already off
drivers/pwm/pwm-sun4i.c | 56 +++++++++++++++++++----------------------
1 file changed, 26 insertions(+), 30 deletions(-)
Hi Roman,
Thanks for your attempt to fix this.
Unfortunately on my A10 device (Topwise A721), the controller still gets
stuck in an unrecoverable state after disabling and re-enabling the PWM
when it was already on (set in U-Boot), or when enabling it when it was
off. In this state, any changes to the period register (using devmem)
don't seem to have any effect. It seems to be stuck in the state it was
before disabling. The only thing which still works is enabling and
disabling.
I can't reproduce this behavior manually so I'm not sure what is causing
this.
Regarding the amount of cycles of sleep; Using a prescaler of 72000 the
PWM clock is 3 ms. Although timing tests using devmem seem unreliable
(too much overhead?), in U-Boot I need to 'sleep' for at least 7 ms
between the commands to make sure the output doesn't sometimes get stuck
in the enabled state. So in my case it seems to be at least 3 cycles. I
am not sure how reliable this method is. However even if I can get it
stuck in the enabled state using a shorter time, it doesn't cause the
behavior I mentioned before. I was always able to recover it manually.
Increasing the number of cycles to sleep therefore also doesn't solve my
problem. Until we can solve that I cannot confirm nor deny if 2 cycles
is enough.
Regards,
Pascal
You could look at the devmem source code, and in C write a script that
writes to pwm register to disable the pwm, insert a usleep, then
disable the gating. This can be done for various sleep values, then
retrying with same sleep value multiple times. Assuming the overhead
is low (you can check the overhead by checking the current timestamp
at the beginning and at the end of the program, take the diff and then
subtract the sleep time), you will get one range where it never works,
one range where it works sometimes, and one range where it always
works. The uncertain range's condition for succeeding will depend on
when in the cycle you run the code.
Assuming we believe 3 cycles are enough on A10 and prescaler is 72000,
the thresholds for these ranges are 0-6 ms, 6-9 ms and 9+ ms.
About "being stuck", I'm not sure exactly what you mean but it's
expected that writes to the period register won't be visible (if you
read it after a write) when the clock gating is disabled. Three full
cycles (with the gating is on) must take place before the change is
visible (i.e. need to wait four cycles to be sure). At least on >=A13.
I documented that here:
https://linux-sunxi.org/PWM_Controller_Register_Guide.
/Emil
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
You could look at the devmem source code, and in C write a script that
writes to pwm register to disable the pwm, insert a usleep, then
disable the gating. This can be done for various sleep values, then
retrying with same sleep value multiple times. Assuming the overhead
is low (you can check the overhead by checking the current timestamp
at the beginning and at the end of the program, take the diff and then
subtract the sleep time), you will get one range where it never works,
one range where it works sometimes, and one range where it always
works. The uncertain range's condition for succeeding will depend on
when in the cycle you run the code.
Assuming we believe 3 cycles are enough on A10 and prescaler is 72000,
the thresholds for these ranges are 0-6 ms, 6-9 ms and 9+ ms.
Thank you I will give this a shot if there is still an uncertainty about
the cycles in the end. I performed my tests with a Busybox rootfs, so I
assumed the overhead was low as well.
About "being stuck", I'm not sure exactly what you mean but it's
expected that writes to the period register won't be visible (if you
read it after a write) when the clock gating is disabled. Three full
cycles (with the gating is on) must take place before the change is
visible (i.e. need to wait four cycles to be sure). At least on >=A13.
I documented that here:
https://linux-sunxi.org/PWM_Controller_Register_Guide.
By being stuck, I mean being in an state from which it can't recover.
The controller will keep outputting seemingly the same signal regardless
what you write to the period register. You can read the values back, but
they aren't effecting the output anymore. No matter in what order or
with what delay I try to re-enable and disable the gate or enable bit,
it'll keep outputting the same signal until you reset the device.
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
On Mon, May 31, 2021 at 06:46:03AM +0200, Roman Beranek wrote:
quoted hunk
Ensure the PWM clock is enabled prior to retrieving its rate, as is
already being done in sun4i_pwm_apply.
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 7 +++++++
1 file changed, 7 insertions(+)
If the clock is off, does the PWM actually run? Assuming it doesn't the
right thing to do is to ensure the clock is enabled in .probe iff the
PWM is enabled.
Best regards
Uwe
--
Pengutronix e.K. | Uwe Kleine-König |
Industrial Linux Solutions | https://www.pengutronix.de/ |
On Mon, May 31, 2021 at 06:46:04AM +0200, Roman Beranek wrote:
quoted hunk
The reason why we wait before gating the clock is to allow for the PWM
to finish its cycle and stop. But it won't stop unless the EN bit is
disabled.
Signed-off-by: Roman Beranek <redacted>
---
drivers/pwm/pwm-sun4i.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
Catching the case !state->enabled even earlier would make sense.
Otherwise you might see a needless glitch after
pwm_apply_state(mypwm, { .period = A, .duty_cycle = B, .enabled = true });
pwm_apply_state(mypwm, { .period = C, .duty_cycle = D, .enabled = false });
which might make C+D visible on the PWM before disabling.
So the comment
/* We need a full period to elapse before disabling the channel. */
is wrong?
Best regards
Uwe
--
Pengutronix e.K. | Uwe Kleine-König |
Industrial Linux Solutions | https://www.pengutronix.de/ |
As Emil Lenngren has previously shown [1], actually only 1-2 cycles of
the prescaler-divided clock are necessary to pass before the PWM turns
off, not a full period.
To avoid having the PWM re-enabled from another thread while asleep,
ctrl_lock spinlock was converted to a mutex so that it can be released
only after the clock gate has finally been turned on.
[1] https://linux-sunxi.org/PWM_Controller_Register_Guide
Roman Beranek (6):
pwm: sun4i: enable clk prior to getting its rate
pwm: sun4i: disable EN bit prior to the delay
pwm: sun4i: replace spinlock with a mutex
pwm: sun4i: simplify calculation of the delay time
pwm: sun4i: shorten the delay to 2 cycles
pwm: sun4i: don't delay if the PWM is already off
drivers/pwm/pwm-sun4i.c | 56 +++++++++++++++++++----------------------
1 file changed, 26 insertions(+), 30 deletions(-)
Hi Roman,
Thanks for your attempt to fix this.
Unfortunately on my A10 device (Topwise A721), the controller still gets
stuck in an unrecoverable state after disabling and re-enabling the PWM
when it was already on (set in U-Boot), or when enabling it when it was
off. In this state, any changes to the period register (using devmem)
don't seem to have any effect. It seems to be stuck in the state it was
before disabling. The only thing which still works is enabling and
disabling.
I can't reproduce this behavior manually so I'm not sure what is causing
this.
Regarding the amount of cycles of sleep; Using a prescaler of 72000 the
PWM clock is 3 ms. Although timing tests using devmem seem unreliable
(too much overhead?), in U-Boot I need to 'sleep' for at least 7 ms
between the commands to make sure the output doesn't sometimes get stuck
in the enabled state. So in my case it seems to be at least 3 cycles. I
am not sure how reliable this method is. However even if I can get it
stuck in the enabled state using a shorter time, it doesn't cause the
behavior I mentioned before. I was always able to recover it manually.
Increasing the number of cycles to sleep therefore also doesn't solve my
problem. Until we can solve that I cannot confirm nor deny if 2 cycles
is enough.
Regards,
Pascal
Turns out, what I'm referring to here is actually a different issue not
related to this patch series. A different series might be sent later to
address that.
So no objections from my side for this one.
Regards,
Pascal
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
@@ -304,7 +304,7 @@ static int sun4i_pwm_apply(struct pwm_chip *chip,
struct pwm_device *pwm,
sun4i_pwm_writel(sun4i_pwm, ctrl, PWM_CTRL_REG);
- if (state->enabled) {
+ if (state->enabled || !cstate.enabled) {
mutex_unlock(&sun4i_pwm->ctrl_lock);
return 0;
}
Btw, this now leaves the gate open if the controller is currently
disabled and we are only changing the period register and staying
disabled. This becomes an issue because we always expect the gate to be
disabled when the controller is disabled.
Regards,
Pascal
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel