[v3,5/5] reset: simple: read back to make sure changes are applied
diff mbox series

Message ID 20170816094701.30678-6-p.zabel@pengutronix.de
State New, archived
Headers show
Series
  • Unify simple reset drivers
Related show

Commit Message

Philipp Zabel Aug. 16, 2017, 9:47 a.m. UTC
Read back the register after setting or clearing a reset bit to make
sure that the changes are applied to the reset controller hardware.
Theoretically, this avoids the write to stay stuck in a store buffer
during the delay of an assert-delay-deassert sequence, and makes sure
that the reset really is asserted for the specified duration.

Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de>
---
 drivers/reset/reset-simple.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

Comments

Alexandru Gagniuc Aug. 16, 2017, 9 p.m. UTC | #1
On 08/16/2017 02:47 AM, Philipp Zabel wrote:
> Read back the register after setting or clearing a reset bit to make
> sure that the changes are applied to the reset controller hardware.

> Theoretically, this avoids the write to stay stuck in a store buffer

Is there hardware where this has been observed to happen, or is this 
purely theoretical? It would be nice to have a "this is needed on 
hardware XYZ because ABC, and doesn't affect other hardware" comment in 
the source.

> during the delay of an assert-delay-deassert sequence, and makes sure
> that the reset really is asserted for the specified duration.
>
> Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de>
> ---
>  drivers/reset/reset-simple.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/reset/reset-simple.c b/drivers/reset/reset-simple.c
> index 13e7d5559acc9..d98a7e7d802d1 100644
> --- a/drivers/reset/reset-simple.c
> +++ b/drivers/reset/reset-simple.c
> @@ -39,17 +39,20 @@ static int reset_simple_set(struct reset_controller_dev *rcdev,
>  	int reg_width = sizeof(u32);
>  	int bank = id / (reg_width * BITS_PER_BYTE);
>  	int offset = id % (reg_width * BITS_PER_BYTE);
> +	void __iomem *addr = data->membase + (bank * reg_width);
>  	unsigned long flags;
>  	u32 reg;
>
>  	spin_lock_irqsave(&data->lock, flags);
>
> -	reg = readl(data->membase + (bank * reg_width));
> +	reg = readl(addr);
>  	if (assert ^ data->active_low)
>  		reg |= BIT(offset);
>  	else
>  		reg &= ~BIT(offset);
> -	writel(reg, data->membase + (bank * reg_width));
> +	writel(reg, addr);
> +	/* Read back to make sure the write doesn't linger in a store buffer */
> +	readl(addr);

You're not using the returned value to check that the reset was actually 
set. This seems a very arbitrary readback workaround, which gives no 
indication if it actually succeeded or not.

Also the set() is now asymmetrical to clear(). In cases when releasing 
reset on a HW block that is about to have IO performed on it, one would 
want to make sure the reset is actually deasserted before doing any IO.

Alex

>
>  	spin_unlock_irqrestore(&data->lock, flags);
>
>
Philipp Zabel Aug. 21, 2017, 8:21 a.m. UTC | #2
Hi Alexandru,

On Wed, 2017-08-16 at 14:00 -0700, Alexandru Gagniuc wrote:
> 
> On 08/16/2017 02:47 AM, Philipp Zabel wrote:
> > Read back the register after setting or clearing a reset bit to make
> > sure that the changes are applied to the reset controller hardware.
> > Theoretically, this avoids the write to stay stuck in a store buffer
> 
> Is there hardware where this has been observed to happen, or is this 
> purely theoretical? It would be nice to have a "this is needed on 
> hardware XYZ because ABC, and doesn't affect other hardware" comment in 
> the source.

This is purely theoretical, at least in the context of reset
controllers. I'm happy to drop this patch for now.

> > during the delay of an assert-delay-deassert sequence, and makes sure
> > that the reset really is asserted for the specified duration.
> > 
> > > > Signed-off-by: Philipp Zabel <p.zabel@pengutronix.de>
> > ---
> >  drivers/reset/reset-simple.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/reset/reset-simple.c b/drivers/reset/reset-simple.c
> > index 13e7d5559acc9..d98a7e7d802d1 100644
> > --- a/drivers/reset/reset-simple.c
> > +++ b/drivers/reset/reset-simple.c
> > @@ -39,17 +39,20 @@ static int reset_simple_set(struct reset_controller_dev *rcdev,
> > > >  	int reg_width = sizeof(u32);
> > > >  	int bank = id / (reg_width * BITS_PER_BYTE);
> > > >  	int offset = id % (reg_width * BITS_PER_BYTE);
> > > > +	void __iomem *addr = data->membase + (bank * reg_width);
> > > >  	unsigned long flags;
> > > >  	u32 reg;
> > 
> > > >  	spin_lock_irqsave(&data->lock, flags);
> > 
> > > > -	reg = readl(data->membase + (bank * reg_width));
> > > > +	reg = readl(addr);
> > > >  	if (assert ^ data->active_low)
> > > >  		reg |= BIT(offset);
> > > >  	else
> > > >  		reg &= ~BIT(offset);
> > > > -	writel(reg, data->membase + (bank * reg_width));
> > > > +	writel(reg, addr);
> > > > +	/* Read back to make sure the write doesn't linger in a store buffer */
> > > > +	readl(addr);
> 
> You're not using the returned value to check that the reset was actually 
> set. This seems a very arbitrary readback workaround, which gives no 
> indication if it actually succeeded or not.

True. For those reset controllers that support status readback, this
would be a better check.

> Also the set() is now asymmetrical to clear(). In cases when releasing 
> reset on a HW block that is about to have IO performed on it, one would 
> want to make sure the reset is actually deasserted before doing any IO.

Thanks for this observation, the reset_simple_set function name is
ambiguous. I'll rename it to reset_simple_update.
The function is called both for assert and deassert operations, and
whether the bit is set or cleared depends on both the assert parameter
and the active_low flag.

regards
Philipp

Patch
diff mbox series

diff --git a/drivers/reset/reset-simple.c b/drivers/reset/reset-simple.c
index 13e7d5559acc9..d98a7e7d802d1 100644
--- a/drivers/reset/reset-simple.c
+++ b/drivers/reset/reset-simple.c
@@ -39,17 +39,20 @@  static int reset_simple_set(struct reset_controller_dev *rcdev,
 	int reg_width = sizeof(u32);
 	int bank = id / (reg_width * BITS_PER_BYTE);
 	int offset = id % (reg_width * BITS_PER_BYTE);
+	void __iomem *addr = data->membase + (bank * reg_width);
 	unsigned long flags;
 	u32 reg;
 
 	spin_lock_irqsave(&data->lock, flags);
 
-	reg = readl(data->membase + (bank * reg_width));
+	reg = readl(addr);
 	if (assert ^ data->active_low)
 		reg |= BIT(offset);
 	else
 		reg &= ~BIT(offset);
-	writel(reg, data->membase + (bank * reg_width));
+	writel(reg, addr);
+	/* Read back to make sure the write doesn't linger in a store buffer */
+	readl(addr);
 
 	spin_unlock_irqrestore(&data->lock, flags);