[PATCH] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[PATCH] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Karim Eshapa]

Convert the jiffies into usecs then use it with usleep_range
such that instead of stuck doing nothing until action happens,
sleep with range improves responsiveness and power usage
and avoid hacking jiffies. it's used for approximate time.
You can check /kernel/time/timer.c.

Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
---
 drivers/soc/fsl/qbman/qman.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
index 6f509f6..e0df4d1 100644
--- a/drivers/soc/fsl/qbman/qman.c
+++ b/drivers/soc/fsl/qbman/qman.c
@@ -1084,11 +1084,9 @@ static int drain_mr_fqrni(struct qm_portal *p)
 		 * entries well before the ring has been fully consumed, so
 		 * we're being *really* paranoid here.
 		 */
-		u64 now, then = jiffies;
+		unsigned int udel_time = jiffies_to_usecs(10000);
 
-		do {
-			now = jiffies;
-		} while ((then + 10000) > now);
+		usleep_range(udel_time/2, udel_time);
 		msg = qm_mr_current(p);
 		if (!msg)
 			return 0;
-- 
2.7.4

[PATCH] drivers:soc:fsl:qbman:qman.c: unsigned long jiffies value.

[Karim Eshapa]

unsigned long jiffies value sorry for that.

Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
---
 drivers/soc/fsl/qbman/qman.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
index e0df4d1..6e1a44a 100644
--- a/drivers/soc/fsl/qbman/qman.c
+++ b/drivers/soc/fsl/qbman/qman.c
@@ -1084,7 +1084,7 @@ static int drain_mr_fqrni(struct qm_portal *p)
 		 * entries well before the ring has been fully consumed, so
 		 * we're being *really* paranoid here.
 		 */
-		unsigned int udel_time = jiffies_to_usecs(10000);
+		unsigned long udel_time = jiffies_to_usecs(10000);
 
 		usleep_range(udel_time/2, udel_time);
 		msg = qm_mr_current(p);
-- 
2.7.4

Re: [PATCH] drivers:soc:fsl:qbman:qman.c: unsigned long jiffies value.

[Scott Wood]

On Sat, 2017-04-29 at 22:43 +0200, Karim Eshapa wrote:
> unsigned long jiffies value sorry for that.

You mean unsigned long msecs?

> 
> Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
> ---
>  drivers/soc/fsl/qbman/qman.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
> index e0df4d1..6e1a44a 100644
> --- a/drivers/soc/fsl/qbman/qman.c
> +++ b/drivers/soc/fsl/qbman/qman.c
> @@ -1084,7 +1084,7 @@ static int drain_mr_fqrni(struct qm_portal *p)
>  		 * entries well before the ring has been fully consumed, so
>  		 * we're being *really* paranoid here.
>  		 */
> -		unsigned int udel_time = jiffies_to_usecs(10000);
> +		unsigned long udel_time = jiffies_to_usecs(10000);
>  
>  		usleep_range(udel_time/2, udel_time);
>  		msg = qm_mr_current(p);

If unsigned int isn't big enough, then unsigned long won't be either on 32-
bit.  With such a long delay why not use msleep()?

As for the previous patch[1], you're halving the minimum timeout which may not
be correct.

For the NXP people: Is there *really* no better way to handle this than
waiting for so long?  Nothing that can be checked to exit the loop early (at
least, you could exit early if there is more work to do so only the final
iteration takes the full timeout)?  And why is the desired timeout specified
in jiffies, the duration of which can change based on kernel config and
doesn't reflect anything about the hardware?

-Scott

[1] When fixing a patch you've already posted that hasn't yet been applied,
send a replacement (v2) patch rather than a separate fix.

RE:drivers:soc:fsl:qbman:qman.c: unsigned long jiffies value

[Karim Eshapa]

On Sat, 29 Apr 2017 18:32:55 -0500, Scott Wood wrote:
>On Sat, 2017-04-29 at 22:43 +0200, Karim Eshapa wrote:
>
>> unsigned long jiffies value sorry for that.
>>
> You mean unsigned long msecs?
>

Yes, I mean usecs. 

>>
>> Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
>> ---
>>  drivers/soc/fsl/qbman/qman.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
>> index e0df4d1..6e1a44a 100644
>> --- a/drivers/soc/fsl/qbman/qman.c
>> +++ b/drivers/soc/fsl/qbman/qman.c
>> @@ -1084,7 +1084,7 @@ static int drain_mr_fqrni(struct qm_portal *p)
>>                * entries well before the ring has been fully consumed, so
>>                * we're being *really* paranoid here.
>>                */
>> -             unsigned int udel_time = jiffies_to_usecs(10000);
>> +             unsigned long udel_time = jiffies_to_usecs(10000);
>>  
>>               usleep_range(udel_time/2, udel_time);
>>               msg = qm_mr_current(p);

>If unsigned int isn't big enough, then unsigned long won't be either on 32-
>bit.  With such a long delay why not use msleep()?
>

I agree with you in long int.
After looking at Documentation/timers/timers-howto.txt I think
the msleep() is better as we actually have long delay.
may be we can use jiffies_to_msecs() with msleep() in case of we still have 
this large jiffies difference.

>As for the previous patch[1], you're halving the minimum timeout which may not
>be correct.
>
>

For the minimum timeout, I've read the following comments.

		 /*
                  * if MR was full and h/w had other FQRNI entries to produce, we
                  * need to allow it time to produce those entries once the
                  * existing entries are consumed. A worst-case situation
                  * (fully-loaded system) means h/w sequencers may have to do 3-4
                  * other things before servicing the portal's MR pump, each of
                  * which (if slow) may take ~50 qman cycles (which is ~200
                  * processor cycles). So rounding up and then multiplying this
                  * worst-case estimate by a factor of 10, just to be
                  * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
                  * one entry at a time, so h/w has an opportunity to produce new
                  * entries well before the ring has been fully consumed, so
                  * we're being *really* paranoid here.
		  */
		  u64 now, then = jiffies;
		
		do {
	                now = jiffies;
                 } while ((then + 10000) > now);  

He needs to guarantee certain action so, he made a very large factor of saftey
therefore I've used sleep in range with approximate delay. But we still need 
the driver owner to define appropriate value to put.

>[1] When fixing a patch you've already posted that hasn't yet been applied,
>send a replacement (v2) patch rather than a separate fix.
>

Thanks so much, I'm just newbies :)


Karim   

[PATCH v2] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Karim Eshapa]

Avoid stuck and hacking jiffies for a long time and using msleep()
for certatin numeber of cylces without the factor of safety
but using the the long delay considering the factor of safety
with the while loop such that after msleep for a short period
of time we check the  msg if it's OK, breaking the big loop delay.

Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
---
 drivers/soc/fsl/qbman/qman.c | 47 ++++++++++++++++++++++----------------------
 1 file changed, 24 insertions(+), 23 deletions(-)

diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
index 6f509f6..4f99298 100644
--- a/drivers/soc/fsl/qbman/qman.c
+++ b/drivers/soc/fsl/qbman/qman.c
@@ -1067,32 +1067,33 @@ static irqreturn_t portal_isr(int irq, void *ptr)
 static int drain_mr_fqrni(struct qm_portal *p)
 {
 	const union qm_mr_entry *msg;
+	unsigned long stop;
+	unsigned int timeout = jiffies_to_msecs(1000);
 loop:
 	msg = qm_mr_current(p);
-	if (!msg) {
-		/*
-		 * if MR was full and h/w had other FQRNI entries to produce, we
-		 * need to allow it time to produce those entries once the
-		 * existing entries are consumed. A worst-case situation
-		 * (fully-loaded system) means h/w sequencers may have to do 3-4
-		 * other things before servicing the portal's MR pump, each of
-		 * which (if slow) may take ~50 qman cycles (which is ~200
-		 * processor cycles). So rounding up and then multiplying this
-		 * worst-case estimate by a factor of 10, just to be
-		 * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
-		 * one entry at a time, so h/w has an opportunity to produce new
-		 * entries well before the ring has been fully consumed, so
-		 * we're being *really* paranoid here.
-		 */
-		u64 now, then = jiffies;
-
-		do {
-			now = jiffies;
-		} while ((then + 10000) > now);
+	stop = jiffies + 10000;
+	/*
+	 * if MR was full and h/w had other FQRNI entries to produce, we
+	 * need to allow it time to produce those entries once the
+	 * existing entries are consumed. A worst-case situation
+	 * (fully-loaded system) means h/w sequencers may have to do 3-4
+	 * other things before servicing the portal's MR pump, each of
+	 * which (if slow) may take ~50 qman cycles (which is ~200
+	 * processor cycles). So rounding up and then multiplying this
+	 * worst-case estimate by a factor of 10, just to be
+	 * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
+	 * one entry at a time, so h/w has an opportunity to produce new
+	 * entries well before the ring has been fully consumed, so
+	 * we're being *really* paranoid here.
+	 */
+	do {
+		if (msg)
+			break;
+		msleep(timeout);
 		msg = qm_mr_current(p);
-		if (!msg)
-			return 0;
-	}
+	} while (time_before(jiffies, stop));
+	if (!msg)
+		return 0;
 	if ((msg->verb & QM_MR_VERB_TYPE_MASK) != QM_MR_VERB_FQRNI) {
 		/* We aren't draining anything but FQRNIs */
 		pr_err("Found verb 0x%x in MR\n", msg->verb);
-- 
2.7.4

Re: [PATCH v2] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Scott Wood]

On Thu, 2017-05-04 at 06:58 +0200, Karim Eshapa wrote:
> +	stop = jiffies + 10000;
> +	/*
> +	 * if MR was full and h/w had other FQRNI entries to produce, we
> +	 * need to allow it time to produce those entries once the
> +	 * existing entries are consumed. A worst-case situation
> +	 * (fully-loaded system) means h/w sequencers may have to do 3-4
> +	 * other things before servicing the portal's MR pump, each of
> +	 * which (if slow) may take ~50 qman cycles (which is ~200
> +	 * processor cycles). So rounding up and then multiplying this
> +	 * worst-case estimate by a factor of 10, just to be
> +	 * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
> +	 * one entry at a time, so h/w has an opportunity to produce new
> +	 * entries well before the ring has been fully consumed, so
> +	 * we're being *really* paranoid here.
> +	 */

OK, upon reading this more closely it seems the intent was to delay for 10,000
*processor cycles* and somehow that got turned into 10,000 jiffies (which is
40 seconds at the default Hz!).  We could just replace this whole thing with
msleep(1) and still be far more paranoid than was originally intended.

Claudiu and Roy, any comments?

-Scott

Re: [PATCH v2] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Roy Pledge]

On 5/4/2017 5:07 PM, Scott Wood wrote:
> On Thu, 2017-05-04 at 06:58 +0200, Karim Eshapa wrote:
>> +	stop = jiffies + 10000;
>> +	/*
>> +	 * if MR was full and h/w had other FQRNI entries to produce, we
>> +	 * need to allow it time to produce those entries once the
>> +	 * existing entries are consumed. A worst-case situation
>> +	 * (fully-loaded system) means h/w sequencers may have to do 3-4
>> +	 * other things before servicing the portal's MR pump, each of
>> +	 * which (if slow) may take ~50 qman cycles (which is ~200
>> +	 * processor cycles). So rounding up and then multiplying this
>> +	 * worst-case estimate by a factor of 10, just to be
>> +	 * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
>> +	 * one entry at a time, so h/w has an opportunity to produce new
>> +	 * entries well before the ring has been fully consumed, so
>> +	 * we're being *really* paranoid here.
>> +	 */
> OK, upon reading this more closely it seems the intent was to delay for 10,000
> *processor cycles* and somehow that got turned into 10,000 jiffies (which is
> 40 seconds at the default Hz!).  We could just replace this whole thing with
> msleep(1) and still be far more paranoid than was originally intended.
>
> Claudiu and Roy, any comments?
Yes the timing here is certainly off, the code changed a few times since
the comment was originally written.
An msleep(1) seems reasonable here to me.

Roy
>
> -Scott
>
>

RE: [PATCH v3] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Karim Eshapa]

Use msleep() instead of stucking with
long delay will be more efficient.

Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
---
 drivers/soc/fsl/qbman/qman.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
index 3d891db..18d391e 100644
--- a/drivers/soc/fsl/qbman/qman.c
+++ b/drivers/soc/fsl/qbman/qman.c
@@ -1084,11 +1084,7 @@ static int drain_mr_fqrni(struct qm_portal *p)
 		 * entries well before the ring has been fully consumed, so
 		 * we're being *really* paranoid here.
 		 */
-		u64 now, then = jiffies;
-
-		do {
-			now = jiffies;
-		} while ((then + 10000) > now);
+		msleep(1);
 		msg = qm_mr_current(p);
 		if (!msg)
 			return 0;
-- 
2.7.4

Re: [v3] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Scott Wood]

On Fri, May 05, 2017 at 07:45:18AM +0200, Karim Eshapa wrote:
> Use msleep() instead of stucking with
> long delay will be more efficient.
> 
> Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
> ---
>  drivers/soc/fsl/qbman/qman.c | 6 +-----
>  1 file changed, 1 insertion(+), 5 deletions(-)

Acked-by: Scott Wood <oss@buserror.net>

(though the subject line should be "soc/qman: ...")

Leo, are you going to send this patch (and other qman patches) via
arm-soc?

-Scott

RE: [v3] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Leo Li]

> -----Original Message-----
> From: Scott Wood [mailto:oss@buserror.net]
> Sent: Saturday, June 24, 2017 9:47 PM
> To: Karim Eshapa <karim.eshapa@gmail.com>
> Cc: Roy Pledge <roy.pledge@nxp.com>; linux-kernel@vger.kernel.org;
> Claudiu Manoil <claudiu.manoil@nxp.com>; colin.king@canonical.com;
> linuxppc-dev@lists.ozlabs.org; linux-arm-kernel@lists.infradead.org; Leo Li
> <leoyang.li@nxp.com>
> Subject: Re: [v3] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck
> hacking jiffies.
> 
> On Fri, May 05, 2017 at 07:45:18AM +0200, Karim Eshapa wrote:
> > Use msleep() instead of stucking with
> > long delay will be more efficient.
> >
> > Signed-off-by: Karim Eshapa <karim.eshapa@gmail.com>
> > ---
> >  drivers/soc/fsl/qbman/qman.c | 6 +-----
> >  1 file changed, 1 insertion(+), 5 deletions(-)
> 
> Acked-by: Scott Wood <oss@buserror.net>
> 
> (though the subject line should be "soc/qman: ...")
> 
> Leo, are you going to send this patch (and other qman patches) via arm-soc?

Yes.  I can take it through the pull request for soc/fsl via arm-soc.  As mentioned in the feedback from David in another email, probably we should update the comment and commit message to mention how 10000 cycles becomes 1ms.

Regards,
Leo

RE: [PATCH v2] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Karim Eshapa]

>On 5/4/2017 5:07 PM, Scott Wood wrote:
>> On Thu, 2017-05-04 at 06:58 +0200, Karim Eshapa wrote:
>>> +    stop = jiffies + 10000;
>>> +    /*
>>> +     * if MR was full and h/w had other FQRNI entries to produce, we
>>> +     * need to allow it time to produce those entries once the
>>> +     * existing entries are consumed. A worst-case situation
>>> +     * (fully-loaded system) means h/w sequencers may have to do 3-4
>>> +     * other things before servicing the portal's MR pump, each of
>>> +     * which (if slow) may take ~50 qman cycles (which is ~200
>>> +     * processor cycles). So rounding up and then multiplying this
>>> +     * worst-case estimate by a factor of 10, just to be
>>> +     * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
>>> +     * one entry at a time, so h/w has an opportunity to produce new
>>> +     * entries well before the ring has been fully consumed, so
>>> +     * we're being *really* paranoid here.
>>> +     */
>> OK, upon reading this more closely it seems the intent was to delay for 10,000
>> *processor cycles* and somehow that got turned into 10,000 jiffies (which is
>> 40 seconds at the default Hz!).  We could just replace this whole thing with
>> msleep(1) and still be far more paranoid than was originally intended.
>>
>> Claudiu and Roy, any comments?
>Yes the timing here is certainly off, the code changed a few times since
>the comment was originally written.
>An msleep(1) seems reasonable here to me.

If the previous patch with msleep(1) is OK.
can I send a patch to slightly change the comments.

Thanks,
Karim

Re: [PATCH v2] drivers:soc:fsl:qbman:qman.c: Sleep instead of stuck hacking jiffies.

[Scott Wood]

On Fri, 2017-05-05 at 08:01 +0200, Karim Eshapa wrote:
> > On 5/4/2017 5:07 PM, Scott Wood wrote:
> > > On Thu, 2017-05-04 at 06:58 +0200, Karim Eshapa wrote:
> > > > +    stop = jiffies + 10000;
> > > > +    /*
> > > > +     * if MR was full and h/w had other FQRNI entries to produce, we
> > > > +     * need to allow it time to produce those entries once the
> > > > +     * existing entries are consumed. A worst-case situation
> > > > +     * (fully-loaded system) means h/w sequencers may have to do 3-4
> > > > +     * other things before servicing the portal's MR pump, each of
> > > > +     * which (if slow) may take ~50 qman cycles (which is ~200
> > > > +     * processor cycles). So rounding up and then multiplying this
> > > > +     * worst-case estimate by a factor of 10, just to be
> > > > +     * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
> > > > +     * one entry at a time, so h/w has an opportunity to produce new
> > > > +     * entries well before the ring has been fully consumed, so
> > > > +     * we're being *really* paranoid here.
> > > > +     */
> > > 
> > > OK, upon reading this more closely it seems the intent was to delay for
> > > 10,000
> > > *processor cycles* and somehow that got turned into 10,000 jiffies
> > > (which is
> > > 40 seconds at the default Hz!).  We could just replace this whole thing
> > > with
> > > msleep(1) and still be far more paranoid than was originally intended.
> > > 
> > > Claudiu and Roy, any comments?
> > 
> > Yes the timing here is certainly off, the code changed a few times since
> > the comment was originally written.
> > An msleep(1) seems reasonable here to me.
> 
> If the previous patch with msleep(1) is OK.
> can I send a patch to slightly change the comments.

Yes.

-Scott