[PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

[PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

[nikhil.gupta]

From: Nikhil Gupta <nikhil.gupta@nxp.com>

1588 driver loses about 1us in adjtime operation at PTP slave.
This is because adjtime operation uses a slow non-atomic tmr_cnt_read()
followed by tmr_cnt_write() operation.
In the above sequence, since the timer counter operation loses about 1us.
Instead, tmr_offset register should be programmed with the delta
nanoseconds This won't lead to timer counter stopping and losing time
while tmr_cnt_write() is being done.
This Patch adds api for tmr_offset_read/write to program the
delta nanoseconds in the Timer offset Register.

Signed-off-by: Nikhil Gupta <nikhil.gupta@nxp.com>
Reviewed-by: Yangbo Lu <yangbo.lu@nxp.com>
---
 drivers/ptp/ptp_qoriq.c | 36 ++++++++++++++++++++++++++++++------
 1 file changed, 30 insertions(+), 6 deletions(-)

diff --git a/drivers/ptp/ptp_qoriq.c b/drivers/ptp/ptp_qoriq.c
index 08f4cf0ad9e3..5b6ea6d590be 100644
--- a/drivers/ptp/ptp_qoriq.c
+++ b/drivers/ptp/ptp_qoriq.c
@@ -48,6 +48,29 @@ static void tmr_cnt_write(struct ptp_qoriq *ptp_qoriq, u64 ns)
 	ptp_qoriq->write(&regs->ctrl_regs->tmr_cnt_h, hi);
 }
 
+static void tmr_offset_write(struct ptp_qoriq *ptp_qoriq, u64 delta_ns)
+{
+	struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
+	u32 hi = delta_ns >> 32;
+	u32 lo = delta_ns & 0xffffffff;
+
+	ptp_qoriq->write(&regs->ctrl_regs->tmroff_l, lo);
+	ptp_qoriq->write(&regs->ctrl_regs->tmroff_h, hi);
+}
+
+static u64 tmr_offset_read(struct ptp_qoriq *ptp_qoriq)
+{
+	struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
+	u64 ns;
+	u32 lo, hi;
+
+	lo = ptp_qoriq->read(&regs->ctrl_regs->tmroff_l);
+	hi = ptp_qoriq->read(&regs->ctrl_regs->tmroff_h);
+	ns = ((u64) hi) << 32;
+	ns |= lo;
+	return ns;
+}
+
 /* Caller must hold ptp_qoriq->lock. */
 static void set_alarm(struct ptp_qoriq *ptp_qoriq)
 {
@@ -55,7 +78,9 @@ static void set_alarm(struct ptp_qoriq *ptp_qoriq)
 	u64 ns;
 	u32 lo, hi;
 
-	ns = tmr_cnt_read(ptp_qoriq) + 1500000000ULL;
+	ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq)
+				     + 1500000000ULL;
+
 	ns = div_u64(ns, 1000000000UL) * 1000000000ULL;
 	ns -= ptp_qoriq->tclk_period;
 	hi = ns >> 32;
@@ -207,15 +232,12 @@ EXPORT_SYMBOL_GPL(ptp_qoriq_adjfine);
 
 int ptp_qoriq_adjtime(struct ptp_clock_info *ptp, s64 delta)
 {
-	s64 now;
 	unsigned long flags;
 	struct ptp_qoriq *ptp_qoriq = container_of(ptp, struct ptp_qoriq, caps);
 
 	spin_lock_irqsave(&ptp_qoriq->lock, flags);
 
-	now = tmr_cnt_read(ptp_qoriq);
-	now += delta;
-	tmr_cnt_write(ptp_qoriq, now);
+	tmr_offset_write(ptp_qoriq, delta);
 	set_fipers(ptp_qoriq);
 
 	spin_unlock_irqrestore(&ptp_qoriq->lock, flags);
@@ -232,7 +254,7 @@ int ptp_qoriq_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
 
 	spin_lock_irqsave(&ptp_qoriq->lock, flags);
 
-	ns = tmr_cnt_read(ptp_qoriq);
+	ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq);
 
 	spin_unlock_irqrestore(&ptp_qoriq->lock, flags);
 
@@ -251,6 +273,8 @@ int ptp_qoriq_settime(struct ptp_clock_info *ptp,
 
 	ns = timespec64_to_ns(ts);
 
+	tmr_offset_write(ptp_qoriq, 0);
+
 	spin_lock_irqsave(&ptp_qoriq->lock, flags);
 
 	tmr_cnt_write(ptp_qoriq, ns);
-- 
2.17.1

RE: [PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

[Nikhil Gupta]

++ Richard Cochran

-----Original Message-----
From: nikhil.gupta@nxp.com <nikhil.gupta@nxp.com> 
Sent: Tuesday, January 10, 2023 5:00 PM
To: linux-arm-kernel@lists.infradead.org; Y.B. Lu <yangbo.lu@nxp.com>; netdev@vger.kernel.org; linux-kernel@vger.kernel.org
Cc: Vakul Garg <vakul.garg@nxp.com>; Rajan Gupta <rajan.gupta@nxp.com>; Nikhil Gupta <nikhil.gupta@nxp.com>
Subject: [PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

From: Nikhil Gupta <nikhil.gupta@nxp.com>

1588 driver loses about 1us in adjtime operation at PTP slave.
This is because adjtime operation uses a slow non-atomic tmr_cnt_read() followed by tmr_cnt_write() operation.
In the above sequence, since the timer counter operation loses about 1us.
Instead, tmr_offset register should be programmed with the delta nanoseconds This won't lead to timer counter stopping and losing time while tmr_cnt_write() is being done.
This Patch adds api for tmr_offset_read/write to program the delta nanoseconds in the Timer offset Register.

Signed-off-by: Nikhil Gupta <nikhil.gupta@nxp.com>
Reviewed-by: Yangbo Lu <yangbo.lu@nxp.com>
---
 drivers/ptp/ptp_qoriq.c | 36 ++++++++++++++++++++++++++++++------
 1 file changed, 30 insertions(+), 6 deletions(-)

diff --git a/drivers/ptp/ptp_qoriq.c b/drivers/ptp/ptp_qoriq.c index 08f4cf0ad9e3..5b6ea6d590be 100644
--- a/drivers/ptp/ptp_qoriq.c
+++ b/drivers/ptp/ptp_qoriq.c
@@ -48,6 +48,29 @@ static void tmr_cnt_write(struct ptp_qoriq *ptp_qoriq, u64 ns)
 	ptp_qoriq->write(&regs->ctrl_regs->tmr_cnt_h, hi);  }
 
+static void tmr_offset_write(struct ptp_qoriq *ptp_qoriq, u64 delta_ns) 
+{
+	struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
+	u32 hi = delta_ns >> 32;
+	u32 lo = delta_ns & 0xffffffff;
+
+	ptp_qoriq->write(&regs->ctrl_regs->tmroff_l, lo);
+	ptp_qoriq->write(&regs->ctrl_regs->tmroff_h, hi); }
+
+static u64 tmr_offset_read(struct ptp_qoriq *ptp_qoriq) {
+	struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
+	u64 ns;
+	u32 lo, hi;
+
+	lo = ptp_qoriq->read(&regs->ctrl_regs->tmroff_l);
+	hi = ptp_qoriq->read(&regs->ctrl_regs->tmroff_h);
+	ns = ((u64) hi) << 32;
+	ns |= lo;
+	return ns;
+}
+
 /* Caller must hold ptp_qoriq->lock. */  static void set_alarm(struct ptp_qoriq *ptp_qoriq)  { @@ -55,7 +78,9 @@ static void set_alarm(struct ptp_qoriq *ptp_qoriq)
 	u64 ns;
 	u32 lo, hi;
 
-	ns = tmr_cnt_read(ptp_qoriq) + 1500000000ULL;
+	ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq)
+				     + 1500000000ULL;
+
 	ns = div_u64(ns, 1000000000UL) * 1000000000ULL;
 	ns -= ptp_qoriq->tclk_period;
 	hi = ns >> 32;
@@ -207,15 +232,12 @@ EXPORT_SYMBOL_GPL(ptp_qoriq_adjfine);
 
 int ptp_qoriq_adjtime(struct ptp_clock_info *ptp, s64 delta)  {
-	s64 now;
 	unsigned long flags;
 	struct ptp_qoriq *ptp_qoriq = container_of(ptp, struct ptp_qoriq, caps);
 
 	spin_lock_irqsave(&ptp_qoriq->lock, flags);
 
-	now = tmr_cnt_read(ptp_qoriq);
-	now += delta;
-	tmr_cnt_write(ptp_qoriq, now);
+	tmr_offset_write(ptp_qoriq, delta);
 	set_fipers(ptp_qoriq);
 
 	spin_unlock_irqrestore(&ptp_qoriq->lock, flags); @@ -232,7 +254,7 @@ int ptp_qoriq_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
 
 	spin_lock_irqsave(&ptp_qoriq->lock, flags);
 
-	ns = tmr_cnt_read(ptp_qoriq);
+	ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq);
 
 	spin_unlock_irqrestore(&ptp_qoriq->lock, flags);
 
@@ -251,6 +273,8 @@ int ptp_qoriq_settime(struct ptp_clock_info *ptp,
 
 	ns = timespec64_to_ns(ts);
 
+	tmr_offset_write(ptp_qoriq, 0);
+
 	spin_lock_irqsave(&ptp_qoriq->lock, flags);
 
 	tmr_cnt_write(ptp_qoriq, ns);
--
2.17.1

Re: [PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

[Jakub Kicinski]

please put [PATCH net-next] in the subject.

On Tue, 10 Jan 2023 17:00:24 +0530 nikhil.gupta@nxp.com wrote:
> From: Nikhil Gupta <nikhil.gupta@nxp.com>
> 
> 1588 driver loses about 1us in adjtime operation at PTP slave.
> This is because adjtime operation uses a slow non-atomic tmr_cnt_read()
> followed by tmr_cnt_write() operation.

So far so good..

> In the above sequence, since the timer counter operation loses about 1us.

s/operation/keeps incrementing after the read/ ?

but frankly I don't think this sentence adds much

> Instead, tmr_offset register should be programmed with the delta
> nanoseconds 

missing full stop at the end.
You should describe what the tmr_offset register does.

> This won't lead to timer counter stopping and losing time
> while tmr_cnt_write() is being done.

Stopping? The timer was actually stopping?

> This Patch adds api for tmr_offset_read/write to program the

Use imperative mood.

> delta nanoseconds in the Timer offset Register.
> 
> Signed-off-by: Nikhil Gupta <nikhil.gupta@nxp.com>
> Reviewed-by: Yangbo Lu <yangbo.lu@nxp.com>
> ---
>  drivers/ptp/ptp_qoriq.c | 36 ++++++++++++++++++++++++++++++------
>  1 file changed, 30 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/ptp/ptp_qoriq.c b/drivers/ptp/ptp_qoriq.c
> index 08f4cf0ad9e3..5b6ea6d590be 100644
> --- a/drivers/ptp/ptp_qoriq.c
> +++ b/drivers/ptp/ptp_qoriq.c
> @@ -48,6 +48,29 @@ static void tmr_cnt_write(struct ptp_qoriq *ptp_qoriq, u64 ns)
>  	ptp_qoriq->write(&regs->ctrl_regs->tmr_cnt_h, hi);
>  }
>  
> +static void tmr_offset_write(struct ptp_qoriq *ptp_qoriq, u64 delta_ns)
> +{
> +	struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
> +	u32 hi = delta_ns >> 32;
> +	u32 lo = delta_ns & 0xffffffff;
> +
> +	ptp_qoriq->write(&regs->ctrl_regs->tmroff_l, lo);
> +	ptp_qoriq->write(&regs->ctrl_regs->tmroff_h, hi);
> +}
> +
> +static u64 tmr_offset_read(struct ptp_qoriq *ptp_qoriq)
> +{
> +	struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
> +	u64 ns;
> +	u32 lo, hi;

Order variable lines longest to shortest 

> +	lo = ptp_qoriq->read(&regs->ctrl_regs->tmroff_l);
> +	hi = ptp_qoriq->read(&regs->ctrl_regs->tmroff_h);
> +	ns = ((u64) hi) << 32;
> +	ns |= lo;
> +	return ns;
> +}
> +
>  /* Caller must hold ptp_qoriq->lock. */
>  static void set_alarm(struct ptp_qoriq *ptp_qoriq)
>  {
> @@ -55,7 +78,9 @@ static void set_alarm(struct ptp_qoriq *ptp_qoriq)
>  	u64 ns;
>  	u32 lo, hi;
>  
> -	ns = tmr_cnt_read(ptp_qoriq) + 1500000000ULL;
> +	ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq)
> +				     + 1500000000ULL;
> +
>  	ns = div_u64(ns, 1000000000UL) * 1000000000ULL;
>  	ns -= ptp_qoriq->tclk_period;
>  	hi = ns >> 32;
> @@ -207,15 +232,12 @@ EXPORT_SYMBOL_GPL(ptp_qoriq_adjfine);
>  
>  int ptp_qoriq_adjtime(struct ptp_clock_info *ptp, s64 delta)
>  {
> -	s64 now;
>  	unsigned long flags;
>  	struct ptp_qoriq *ptp_qoriq = container_of(ptp, struct ptp_qoriq, caps);
>  
>  	spin_lock_irqsave(&ptp_qoriq->lock, flags);
>  
> -	now = tmr_cnt_read(ptp_qoriq);
> -	now += delta;
> -	tmr_cnt_write(ptp_qoriq, now);
> +	tmr_offset_write(ptp_qoriq, delta);

Writes to the offset register result in an add operation? 
Or it's a pure write? What will the offset be after a sequence of
following adjtime() calls: 
  adjtime(+100); 
  adjtime(+100); 
  adjtime(+100);
?

>  	set_fipers(ptp_qoriq);
>  
>  	spin_unlock_irqrestore(&ptp_qoriq->lock, flags);
> @@ -232,7 +254,7 @@ int ptp_qoriq_gettime(struct ptp_clock_info *ptp, struct timespec64 *ts)
>  
>  	spin_lock_irqsave(&ptp_qoriq->lock, flags);
>  
> -	ns = tmr_cnt_read(ptp_qoriq);
> +	ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq);
>  
>  	spin_unlock_irqrestore(&ptp_qoriq->lock, flags);
>  
> @@ -251,6 +273,8 @@ int ptp_qoriq_settime(struct ptp_clock_info *ptp,
>  
>  	ns = timespec64_to_ns(ts);
>  
> +	tmr_offset_write(ptp_qoriq, 0);

Shouldn't this be under the lock?

>  	spin_lock_irqsave(&ptp_qoriq->lock, flags);
>  
>  	tmr_cnt_write(ptp_qoriq, ns);

[PATCH net-next] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

[Nikhil Gupta]

-----Original Message-----
From: Jakub Kicinski <kuba@kernel.org> 
Sent: Thursday, January 12, 2023 10:39 AM
To: Nikhil Gupta <nikhil.gupta@nxp.com>
Cc: linux-arm-kernel@lists.infradead.org; Y.B. Lu <yangbo.lu@nxp.com>; netdev@vger.kernel.org; linux-kernel@vger.kernel.org; Vakul Garg <vakul.garg@nxp.com>; Rajan Gupta <rajan.gupta@nxp.com>
Subject: [EXT] Re: [PATCH] ptp_qoriq: fix latency in ptp_qoriq_adjtime() operation.

Caution: EXT Email

please put [PATCH net-next] in the subject.

On Tue, 10 Jan 2023 17:00:24 +0530 nikhil.gupta@nxp.com wrote:
> From: Nikhil Gupta <nikhil.gupta@nxp.com>
>
> 1588 driver loses about 1us in adjtime operation at PTP slave.
> This is because adjtime operation uses a slow non-atomic 
> tmr_cnt_read() followed by tmr_cnt_write() operation.

So far so good..

> In the above sequence, since the timer counter operation loses about 1us.

s/operation/keeps incrementing after the read/ ?

but frankly I don't think this sentence adds much

> Instead, tmr_offset register should be programmed with the delta 
> nanoseconds

missing full stop at the end.
You should describe what the tmr_offset register does.
[Nikhil] : current time is calculated by adding TMROFF_H/L with the timer's counter TMR_CNT_H/L register.

> This won't lead to timer counter stopping and losing time while 
> tmr_cnt_write() is being done.

Stopping? The timer was actually stopping?

> This Patch adds api for tmr_offset_read/write to program the

Use imperative mood.

> delta nanoseconds in the Timer offset Register.
>
> Signed-off-by: Nikhil Gupta <nikhil.gupta@nxp.com>
> Reviewed-by: Yangbo Lu <yangbo.lu@nxp.com>
> ---
>  drivers/ptp/ptp_qoriq.c | 36 ++++++++++++++++++++++++++++++------
>  1 file changed, 30 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/ptp/ptp_qoriq.c b/drivers/ptp/ptp_qoriq.c index 
> 08f4cf0ad9e3..5b6ea6d590be 100644
> --- a/drivers/ptp/ptp_qoriq.c
> +++ b/drivers/ptp/ptp_qoriq.c
> @@ -48,6 +48,29 @@ static void tmr_cnt_write(struct ptp_qoriq *ptp_qoriq, u64 ns)
>       ptp_qoriq->write(&regs->ctrl_regs->tmr_cnt_h, hi);  }
>
> +static void tmr_offset_write(struct ptp_qoriq *ptp_qoriq, u64 
> +delta_ns) {
> +     struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
> +     u32 hi = delta_ns >> 32;
> +     u32 lo = delta_ns & 0xffffffff;
> +
> +     ptp_qoriq->write(&regs->ctrl_regs->tmroff_l, lo);
> +     ptp_qoriq->write(&regs->ctrl_regs->tmroff_h, hi); }
> +
> +static u64 tmr_offset_read(struct ptp_qoriq *ptp_qoriq) {
> +     struct ptp_qoriq_registers *regs = &ptp_qoriq->regs;
> +     u64 ns;
> +     u32 lo, hi;

Order variable lines longest to shortest

> +     lo = ptp_qoriq->read(&regs->ctrl_regs->tmroff_l);
> +     hi = ptp_qoriq->read(&regs->ctrl_regs->tmroff_h);
> +     ns = ((u64) hi) << 32;
> +     ns |= lo;
> +     return ns;
> +}
> +
>  /* Caller must hold ptp_qoriq->lock. */  static void set_alarm(struct 
> ptp_qoriq *ptp_qoriq)  { @@ -55,7 +78,9 @@ static void 
> set_alarm(struct ptp_qoriq *ptp_qoriq)
>       u64 ns;
>       u32 lo, hi;
>
> -     ns = tmr_cnt_read(ptp_qoriq) + 1500000000ULL;
> +     ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq)
> +                                  + 1500000000ULL;
> +
>       ns = div_u64(ns, 1000000000UL) * 1000000000ULL;
>       ns -= ptp_qoriq->tclk_period;
>       hi = ns >> 32;
> @@ -207,15 +232,12 @@ EXPORT_SYMBOL_GPL(ptp_qoriq_adjfine);
>
>  int ptp_qoriq_adjtime(struct ptp_clock_info *ptp, s64 delta)  {
> -     s64 now;
>       unsigned long flags;
>       struct ptp_qoriq *ptp_qoriq = container_of(ptp, struct 
> ptp_qoriq, caps);
>
>       spin_lock_irqsave(&ptp_qoriq->lock, flags);
>
> -     now = tmr_cnt_read(ptp_qoriq);
> -     now += delta;
> -     tmr_cnt_write(ptp_qoriq, now);
> +     tmr_offset_write(ptp_qoriq, delta);

Writes to the offset register result in an add operation?
Or it's a pure write? What will the offset be after a sequence of following adjtime() calls:
  adjtime(+100);
  adjtime(+100);
  adjtime(+100);
?
[Nikhil] : It's a pure write operation, I will be sending the updated version of the patch.
	  Wherein retaining  the earlier offset value and adding to the new.

>       set_fipers(ptp_qoriq);
>
>       spin_unlock_irqrestore(&ptp_qoriq->lock, flags); @@ -232,7 
> +254,7 @@ int ptp_qoriq_gettime(struct ptp_clock_info *ptp, struct 
> timespec64 *ts)
>
>       spin_lock_irqsave(&ptp_qoriq->lock, flags);
>
> -     ns = tmr_cnt_read(ptp_qoriq);
> +     ns = tmr_cnt_read(ptp_qoriq) + tmr_offset_read(ptp_qoriq);
>
>       spin_unlock_irqrestore(&ptp_qoriq->lock, flags);
>
> @@ -251,6 +273,8 @@ int ptp_qoriq_settime(struct ptp_clock_info *ptp,
>
>       ns = timespec64_to_ns(ts);
>
> +     tmr_offset_write(ptp_qoriq, 0);

Shouldn't this be under the lock?
[Nikhil] : will update this.
>       spin_lock_irqsave(&ptp_qoriq->lock, flags);
>
>       tmr_cnt_write(ptp_qoriq, ns);