reset could be handled by PM functions.
Signed-off-by: Corentin Labbe <[email protected]>
---
drivers/crypto/rockchip/rk3288_crypto.c | 19 ++++---------------
1 file changed, 4 insertions(+), 15 deletions(-)
diff --git a/drivers/crypto/rockchip/rk3288_crypto.c b/drivers/crypto/rockchip/rk3288_crypto.c
index d9258b9e71b3..a11a92e1f3fd 100644
--- a/drivers/crypto/rockchip/rk3288_crypto.c
+++ b/drivers/crypto/rockchip/rk3288_crypto.c
@@ -73,6 +73,8 @@ static int rk_crypto_pm_suspend(struct device *dev)
{
struct rk_crypto_info *rkdev = dev_get_drvdata(dev);
+ reset_control_assert(rkdev->rst);
+
rk_crypto_disable_clk(rkdev);
return 0;
}
@@ -81,6 +83,8 @@ static int rk_crypto_pm_resume(struct device *dev)
{
struct rk_crypto_info *rkdev = dev_get_drvdata(dev);
+ reset_control_deassert(rkdev->rst);
+
return rk_crypto_enable_clk(rkdev);
}
@@ -222,13 +226,6 @@ static void rk_crypto_unregister(void)
}
}
-static void rk_crypto_action(void *data)
-{
- struct rk_crypto_info *crypto_info = data;
-
- reset_control_assert(crypto_info->rst);
-}
-
static const struct of_device_id crypto_of_id_table[] = {
{ .compatible = "rockchip,rk3288-crypto" },
{}
@@ -254,14 +251,6 @@ static int rk_crypto_probe(struct platform_device *pdev)
goto err_crypto;
}
- reset_control_assert(crypto_info->rst);
- usleep_range(10, 20);
- reset_control_deassert(crypto_info->rst);
-
- err = devm_add_action_or_reset(dev, rk_crypto_action, crypto_info);
- if (err)
- goto err_crypto;
-
crypto_info->reg = devm_platform_ioremap_resource(pdev, 0);
if (IS_ERR(crypto_info->reg)) {
err = PTR_ERR(crypto_info->reg);
--
2.35.1
On Sun, May 08, 2022 at 06:59:38PM +0000, Corentin Labbe wrote:
> reset could be handled by PM functions.
Is there any further rationale for this?
After this change there is no longer a guaranteed reset pulse on probe
since the reset control may already be de-asserted. This is normally
the most important case for a reset as it's the only time when the state
of the hardware is unknown.
The original use of devm_add_action_or_reset() seems a bit weird already
since there doesn't seem to be any need to assert reset when the driver
is unloaded.
> Signed-off-by: Corentin Labbe <[email protected]>
> ---
> drivers/crypto/rockchip/rk3288_crypto.c | 19 ++++---------------
> 1 file changed, 4 insertions(+), 15 deletions(-)
>
> diff --git a/drivers/crypto/rockchip/rk3288_crypto.c b/drivers/crypto/rockchip/rk3288_crypto.c
> index d9258b9e71b3..a11a92e1f3fd 100644
> --- a/drivers/crypto/rockchip/rk3288_crypto.c
> +++ b/drivers/crypto/rockchip/rk3288_crypto.c
> @@ -73,6 +73,8 @@ static int rk_crypto_pm_suspend(struct device *dev)
> {
> struct rk_crypto_info *rkdev = dev_get_drvdata(dev);
>
> + reset_control_assert(rkdev->rst);
> +
> rk_crypto_disable_clk(rkdev);
> return 0;
> }
> @@ -81,6 +83,8 @@ static int rk_crypto_pm_resume(struct device *dev)
> {
> struct rk_crypto_info *rkdev = dev_get_drvdata(dev);
>
> + reset_control_deassert(rkdev->rst);
> +
> return rk_crypto_enable_clk(rkdev);
> }
>
> @@ -222,13 +226,6 @@ static void rk_crypto_unregister(void)
> }
> }
>
> -static void rk_crypto_action(void *data)
> -{
> - struct rk_crypto_info *crypto_info = data;
> -
> - reset_control_assert(crypto_info->rst);
> -}
> -
> static const struct of_device_id crypto_of_id_table[] = {
> { .compatible = "rockchip,rk3288-crypto" },
> {}
> @@ -254,14 +251,6 @@ static int rk_crypto_probe(struct platform_device *pdev)
> goto err_crypto;
> }
>
> - reset_control_assert(crypto_info->rst);
> - usleep_range(10, 20);
> - reset_control_deassert(crypto_info->rst);
> -
> - err = devm_add_action_or_reset(dev, rk_crypto_action, crypto_info);
> - if (err)
> - goto err_crypto;
> -
> crypto_info->reg = devm_platform_ioremap_resource(pdev, 0);
> if (IS_ERR(crypto_info->reg)) {
> err = PTR_ERR(crypto_info->reg);
> --
> 2.35.1
Le Mon, Jun 20, 2022 at 12:04:24PM +0100, John Keeping a ?crit :
> On Sun, May 08, 2022 at 06:59:38PM +0000, Corentin Labbe wrote:
> > reset could be handled by PM functions.
>
> Is there any further rationale for this?
>
> After this change there is no longer a guaranteed reset pulse on probe
> since the reset control may already be de-asserted. This is normally
> the most important case for a reset as it's the only time when the state
> of the hardware is unknown.
>
> The original use of devm_add_action_or_reset() seems a bit weird already
> since there doesn't seem to be any need to assert reset when the driver
> is unloaded.
>
I am not an hw engineer, so my knowledge on reset is low.
So why not having a reset pulse on probe is a problem ?
Do you mean I must put reset asserted on probe ?
On Tue, Jun 21, 2022 at 10:05:54AM +0200, LABBE Corentin wrote:
> Le Mon, Jun 20, 2022 at 12:04:24PM +0100, John Keeping a ?crit :
> > On Sun, May 08, 2022 at 06:59:38PM +0000, Corentin Labbe wrote:
> > > reset could be handled by PM functions.
> >
> > Is there any further rationale for this?
> >
> > After this change there is no longer a guaranteed reset pulse on probe
> > since the reset control may already be de-asserted. This is normally
> > the most important case for a reset as it's the only time when the state
> > of the hardware is unknown.
> >
> > The original use of devm_add_action_or_reset() seems a bit weird already
> > since there doesn't seem to be any need to assert reset when the driver
> > is unloaded.
> >
>
> I am not an hw engineer, so my knowledge on reset is low.
> So why not having a reset pulse on probe is a problem ?
The point of the reset is to bring the hardware back to a known state.
Since we don't know what state the hardware will be in following the
bootloader or previous OS, I think the reset in probe is the only place
that it is important.
If this patch isn't fixing anything, I suggest just dropping it.
Le Tue, Jun 21, 2022 at 02:37:35PM +0100, John Keeping a ?crit :
> On Tue, Jun 21, 2022 at 10:05:54AM +0200, LABBE Corentin wrote:
> > Le Mon, Jun 20, 2022 at 12:04:24PM +0100, John Keeping a ?crit :
> > > On Sun, May 08, 2022 at 06:59:38PM +0000, Corentin Labbe wrote:
> > > > reset could be handled by PM functions.
> > >
> > > Is there any further rationale for this?
> > >
> > > After this change there is no longer a guaranteed reset pulse on probe
> > > since the reset control may already be de-asserted. This is normally
> > > the most important case for a reset as it's the only time when the state
> > > of the hardware is unknown.
> > >
> > > The original use of devm_add_action_or_reset() seems a bit weird already
> > > since there doesn't seem to be any need to assert reset when the driver
> > > is unloaded.
> > >
> >
> > I am not an hw engineer, so my knowledge on reset is low.
> > So why not having a reset pulse on probe is a problem ?
>
> The point of the reset is to bring the hardware back to a known state.
> Since we don't know what state the hardware will be in following the
> bootloader or previous OS, I think the reset in probe is the only place
> that it is important.
>
> If this patch isn't fixing anything, I suggest just dropping it.
Thanks for the explanation, I will re-add the reset at probe.