2019-04-06 12:06:40

by Nicholas Mc Guire

[permalink] [raw]
Subject: [PATCH RFC] staging: wilc1000: give usleep_range a range

usleep_range() is called in non-atomic context so there is little point
in setting min==max the jitter of hrtimer is determined by interruptions
anyway. usleep_range can only perform the intended coalescence if some
room for placing the hrtimer is provided. Given the range of milliseconds
the delay will be anything from 2 to a few anyway - so make it 2-5 ms.

Signed-off-by: Nicholas Mc Guire <[email protected]>
---

Problem located with an experimental coccinelle script
./drivers/staging/wilc1000/wilc_wlan.c:411:4-16: WARNING: inefficient usleep_range with range 0 (min==max)
./drivers/staging/wilc1000/wilc_wlan.c:426:4-16: WARNING: inefficient usleep_range with range 0 (min==max)

Someone that knows the motivation for setting the time to 2 millisecond
might need to check if the 2 milliseconds where seen as tollerable max or
min - I'm assuming it was the min so extending.

Patch was compile tested with: x86_64_defconfig + Staging=y,
WILC1000_SDIO=m, WILC1000_SPI=m, WILC1000=m

Patch is against 5.1-rc3 (localversion-next is -next-20190405)

drivers/staging/wilc1000/wilc_wlan.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/staging/wilc1000/wilc_wlan.c b/drivers/staging/wilc1000/wilc_wlan.c
index c238969..42da533 100644
--- a/drivers/staging/wilc1000/wilc_wlan.c
+++ b/drivers/staging/wilc1000/wilc_wlan.c
@@ -408,7 +408,7 @@ void chip_wakeup(struct wilc *wilc)
wilc->hif_func->hif_write_reg(wilc, 1, reg & ~BIT(1));

do {
- usleep_range(2 * 1000, 2 * 1000);
+ usleep_range(2 * 1000, 5 * 1000);
wilc_get_chipid(wilc, true);
} while (wilc_get_chipid(wilc, true) == 0);
} while (wilc_get_chipid(wilc, true) == 0);
@@ -423,7 +423,7 @@ void chip_wakeup(struct wilc *wilc)
&clk_status_reg);

while ((clk_status_reg & 0x1) == 0) {
- usleep_range(2 * 1000, 2 * 1000);
+ usleep_range(2 * 1000, 5 * 1000);

wilc->hif_func->hif_read_reg(wilc, 0xf1,
&clk_status_reg);
--
2.1.4



2019-04-08 21:01:29

by Adham Abozaeid

[permalink] [raw]
Subject: Re: [PATCH RFC] staging: wilc1000: give usleep_range a range

Hi Nicholas

On 4/6/19 5:01 AM, Nicholas Mc Guire wrote:
> External E-Mail
>
>
> usleep_range() is called in non-atomic context so there is little point
> in setting min==max the jitter of hrtimer is determined by interruptions
> anyway. usleep_range can only perform the intended coalescence if some
> room for placing the hrtimer is provided. Given the range of milliseconds
> the delay will be anything from 2 to a few anyway - so make it 2-5 ms.
>
> Signed-off-by: Nicholas Mc Guire <[email protected]>
> ---
>
> Problem located with an experimental coccinelle script
> ./drivers/staging/wilc1000/wilc_wlan.c:411:4-16: WARNING: inefficient usleep_range with range 0 (min==max)
> ./drivers/staging/wilc1000/wilc_wlan.c:426:4-16: WARNING: inefficient usleep_range with range 0 (min==max)
>
> Someone that knows the motivation for setting the time to 2 millisecond
> might need to check if the 2 milliseconds where seen as tollerable max or
> min - I'm assuming it was the min so extending.
>
> Patch was compile tested with: x86_64_defconfig + Staging=y,
> WILC1000_SDIO=m, WILC1000_SPI=m, WILC1000=m
>
> Patch is against 5.1-rc3 (localversion-next is -next-20190405)
>
> drivers/staging/wilc1000/wilc_wlan.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)


This

> diff --git a/drivers/staging/wilc1000/wilc_wlan.c b/drivers/staging/wilc1000/wilc_wlan.c
> index c238969..42da533 100644
> --- a/drivers/staging/wilc1000/wilc_wlan.c
> +++ b/drivers/staging/wilc1000/wilc_wlan.c
> @@ -408,7 +408,7 @@ void chip_wakeup(struct wilc *wilc)
> wilc->hif_func->hif_write_reg(wilc, 1, reg & ~BIT(1));
>
> do {
> - usleep_range(2 * 1000, 2 * 1000);
> + usleep_range(2 * 1000, 5 * 1000);
> wilc_get_chipid(wilc, true);
> } while (wilc_get_chipid(wilc, true) == 0);
> } while (wilc_get_chipid(wilc, true) == 0);
> @@ -423,7 +423,7 @@ void chip_wakeup(struct wilc *wilc)
> &clk_status_reg);
>
> while ((clk_status_reg & 0x1) == 0) {
> - usleep_range(2 * 1000, 2 * 1000);
> + usleep_range(2 * 1000, 5 * 1000);
>
> wilc->hif_func->hif_read_reg(wilc, 0xf1,
> &clk_status_reg);

2019-04-08 21:10:10

by Adham Abozaeid

[permalink] [raw]
Subject: Re: [PATCH RFC] staging: wilc1000: give usleep_range a range

Hi Nicholas

On 4/6/19 5:01 AM, Nicholas Mc Guire wrote:
> External E-Mail
>
>
> Someone that knows the motivation for setting the time to 2 millisecond
> might need to check if the 2 milliseconds where seen as tollerable max or
> min - I'm assuming it was the min so extending.

2 msec is the time the chip takes to wake up from sleep.

Increasing the maximum to 5 msec will impact the throughput since this call is on the transmit path.

> diff --git a/drivers/staging/wilc1000/wilc_wlan.c b/drivers/staging/wilc1000/wilc_wlan.c
> index c238969..42da533 100644
> --- a/drivers/staging/wilc1000/wilc_wlan.c
> +++ b/drivers/staging/wilc1000/wilc_wlan.c
> @@ -408,7 +408,7 @@ void chip_wakeup(struct wilc *wilc)
> wilc->hif_func->hif_write_reg(wilc, 1, reg & ~BIT(1));
>
> do {
> - usleep_range(2 * 1000, 2 * 1000);
> + usleep_range(2 * 1000, 5 * 1000);
> wilc_get_chipid(wilc, true);
> } while (wilc_get_chipid(wilc, true) == 0);
> } while (wilc_get_chipid(wilc, true) == 0);
> @@ -423,7 +423,7 @@ void chip_wakeup(struct wilc *wilc)
> &clk_status_reg);
>
> while ((clk_status_reg & 0x1) == 0) {
> - usleep_range(2 * 1000, 2 * 1000);
> + usleep_range(2 * 1000, 5 * 1000);
>
> wilc->hif_func->hif_read_reg(wilc, 0xf1,
> &clk_status_reg);


Thanks,

Adham

2019-04-09 01:37:49

by Nicholas Mc Guire

[permalink] [raw]
Subject: Re: [PATCH RFC] staging: wilc1000: give usleep_range a range

On Mon, Apr 08, 2019 at 09:10:00PM +0000, [email protected] wrote:
> Hi Nicholas
>
> On 4/6/19 5:01 AM, Nicholas Mc Guire wrote:
> > External E-Mail
> >
> >
> > Someone that knows the motivation for setting the time to 2 millisecond
> > might need to check if the 2 milliseconds where seen as tollerable max or
> > min - I'm assuming it was the min so extending.
>
> 2 msec is the time the chip takes to wake up from sleep.
>
> Increasing the maximum to 5 msec will impact the throughput since this call is on the transmit path.
>

ok - would it be tollerable to make it 2 - 2.5 ms ?
even that would allow for the hrtimer subsystem to optimize
a lot. In any case the min==max case gives you very little
if you run a test-case with usleep_range(1000,1000) and
a loop with usleep_range(1000,2000) and look at the distribution
you will have a hard time seeing any difference.

I doubt you would readily see the change from usleep_range(2000,2000)
to usleep_range(2000,3000) in benchmarks - maybe (2000,5000) would
be visible.

My assumption (I have not analyzed it in detail) is that if
you have a high re-use of existing timers that the setup of the timer
is faster and thats why increasing the range > 0 can actually result
in better jitter distribution.

thx!
hofrat

> > diff --git a/drivers/staging/wilc1000/wilc_wlan.c b/drivers/staging/wilc1000/wilc_wlan.c
> > index c238969..42da533 100644
> > --- a/drivers/staging/wilc1000/wilc_wlan.c
> > +++ b/drivers/staging/wilc1000/wilc_wlan.c
> > @@ -408,7 +408,7 @@ void chip_wakeup(struct wilc *wilc)
> > wilc->hif_func->hif_write_reg(wilc, 1, reg & ~BIT(1));
> >
> > do {
> > - usleep_range(2 * 1000, 2 * 1000);
> > + usleep_range(2 * 1000, 5 * 1000);
> > wilc_get_chipid(wilc, true);
> > } while (wilc_get_chipid(wilc, true) == 0);
> > } while (wilc_get_chipid(wilc, true) == 0);
> > @@ -423,7 +423,7 @@ void chip_wakeup(struct wilc *wilc)
> > &clk_status_reg);
> >
> > while ((clk_status_reg & 0x1) == 0) {
> > - usleep_range(2 * 1000, 2 * 1000);
> > + usleep_range(2 * 1000, 5 * 1000);
> >
> > wilc->hif_func->hif_read_reg(wilc, 0xf1,
> > &clk_status_reg);
>
>
> Thanks,
>
> Adham
>

2019-04-10 18:31:30

by Adham Abozaeid

[permalink] [raw]
Subject: Re: [PATCH RFC] staging: wilc1000: give usleep_range a range

Hi Nicolas

On 4/8/19 6:36 PM, Nicholas Mc Guire wrote:
> On Mon, Apr 08, 2019 at 09:10:00PM +0000, [email protected] wrote:
>> Hi Nicholas
>>
>> On 4/6/19 5:01 AM, Nicholas Mc Guire wrote:
>>> External E-Mail
>>>
>>>
>>> Someone that knows the motivation for setting the time to 2 millisecond
>>> might need to check if the 2 milliseconds where seen as tollerable max or
>>> min - I'm assuming it was the min so extending.
>> 2 msec is the time the chip takes to wake up from sleep.
>>
>> Increasing the maximum to 5 msec will impact the throughput since this call is on the transmit path.
>>
> ok - would it be tollerable to make it 2 - 2.5 ms ?
> even that would allow for the hrtimer subsystem to optimize
> a lot. In any case the min==max case gives you very little
> if you run a test-case with usleep_range(1000,1000) and
> a loop with usleep_range(1000,2000) and look at the distribution
> you will have a hard time seeing any difference.

yes, I believe 2.5 shouldn't be a problem.

Thanks,

Adham


2019-04-11 02:45:38

by Nicholas Mc Guire

[permalink] [raw]
Subject: Re: [PATCH RFC] staging: wilc1000: give usleep_range a range

On Wed, Apr 10, 2019 at 06:31:21PM +0000, [email protected] wrote:
> Hi Nicolas
>
> On 4/8/19 6:36 PM, Nicholas Mc Guire wrote:
> > On Mon, Apr 08, 2019 at 09:10:00PM +0000, [email protected] wrote:
> >> Hi Nicholas
> >>
> >> On 4/6/19 5:01 AM, Nicholas Mc Guire wrote:
> >>> External E-Mail
> >>>
> >>>
> >>> Someone that knows the motivation for setting the time to 2 millisecond
> >>> might need to check if the 2 milliseconds where seen as tollerable max or
> >>> min - I'm assuming it was the min so extending.
> >> 2 msec is the time the chip takes to wake up from sleep.
> >>
> >> Increasing the maximum to 5 msec will impact the throughput since this call is on the transmit path.
> >>
> > ok - would it be tollerable to make it 2 - 2.5 ms ?
> > even that would allow for the hrtimer subsystem to optimize
> > a lot. In any case the min==max case gives you very little
> > if you run a test-case with usleep_range(1000,1000) and
> > a loop with usleep_range(1000,2000) and look at the distribution
> > you will have a hard time seeing any difference.
>
> yes, I believe 2.5 shouldn't be a problem.
>
thanks - will send out a V2 then shortly.

thx!
hofrat