[PATCH v5] can: Convert to runtime_pm

[Sören Brinkmann]

On Tue, 2015-01-13 at 06:44PM +0100, Marc Kleine-Budde wrote:
> On 01/13/2015 06:24 PM, Sören Brinkmann wrote:
> > On Tue, 2015-01-13 at 06:17PM +0100, Marc Kleine-Budde wrote:
> >> On 01/13/2015 06:08 PM, Sören Brinkmann wrote:
> >>> On Tue, 2015-01-13 at 12:08PM +0100, Marc Kleine-Budde wrote:
> >>>> On 01/12/2015 07:45 PM, Sören Brinkmann wrote:
> >>>>> On Mon, 2015-01-12 at 08:34PM +0530, Kedareswara rao Appana wrote:
> >>>>>> Instead of enabling/disabling clocks at several locations in the driver,
> >>>>>> Use the runtime_pm framework. This consolidates the actions for runtime PM
> >>>>>> In the appropriate callbacks and makes the driver more readable and mantainable.
> >>>>>>
> >>>>>> Signed-off-by: Soren Brinkmann <soren.brinkmann at xilinx.com>
> >>>>>> Signed-off-by: Kedareswara rao Appana <appanad at xilinx.com>
> >>>>>> ---
> >>>>>> Changes for v5:
> >>>>>>  - Updated with the review comments.
> >>>>>>    Updated the remove fuction to use runtime_pm.
> >>>>>> Chnages for v4:
> >>>>>>  - Updated with the review comments.
> >>>>>> Changes for v3:
> >>>>>>   - Converted the driver to use runtime_pm.
> >>>>>> Changes for v2:
> >>>>>>   - Removed the struct platform_device* from suspend/resume
> >>>>>>     as suggest by Lothar.
> >>>>>>
> >>>>>>  drivers/net/can/xilinx_can.c |  157 ++++++++++++++++++++++++++++-------------
> >>>>>>  1 files changed, 107 insertions(+), 50 deletions(-)
> >>>>> [..]
> >>>>>> +static int __maybe_unused xcan_runtime_resume(struct device *dev)
> >>>>>>  {
> >>>>>> -	struct platform_device *pdev = dev_get_drvdata(dev);
> >>>>>> -	struct net_device *ndev = platform_get_drvdata(pdev);
> >>>>>> +	struct net_device *ndev = dev_get_drvdata(dev);
> >>>>>>  	struct xcan_priv *priv = netdev_priv(ndev);
> >>>>>>  	int ret;
> >>>>>> +	u32 isr, status;
> >>>>>>  
> >>>>>>  	ret = clk_enable(priv->bus_clk);
> >>>>>>  	if (ret) {
> >>>>>> @@ -1014,15 +1030,28 @@ static int __maybe_unused xcan_resume(struct device *dev)
> >>>>>>  	ret = clk_enable(priv->can_clk);
> >>>>>>  	if (ret) {
> >>>>>>  		dev_err(dev, "Cannot enable clock.\n");
> >>>>>> -		clk_disable_unprepare(priv->bus_clk);
> >>>>>> +		clk_disable(priv->bus_clk);
> >>>>> [...]
> >>>>>> @@ -1173,12 +1219,23 @@ static int xcan_remove(struct platform_device *pdev)
> >>>>>>  {
> >>>>>>  	struct net_device *ndev = platform_get_drvdata(pdev);
> >>>>>>  	struct xcan_priv *priv = netdev_priv(ndev);
> >>>>>> +	int ret;
> >>>>>> +
> >>>>>> +	ret = pm_runtime_get_sync(&pdev->dev);
> >>>>>> +	if (ret < 0) {
> >>>>>> +		netdev_err(ndev, "%s: pm_runtime_get failed(%d)\n",
> >>>>>> +				__func__, ret);
> >>>>>> +		return ret;
> >>>>>> +	}
> >>>>>>  
> >>>>>>  	if (set_reset_mode(ndev) < 0)
> >>>>>>  		netdev_err(ndev, "mode resetting failed!\n");
> >>>>>>  
> >>>>>>  	unregister_candev(ndev);
> >>>>>> +	pm_runtime_disable(&pdev->dev);
> >>>>>>  	netif_napi_del(&priv->napi);
> >>>>>> +	clk_disable_unprepare(priv->bus_clk);
> >>>>>> +	clk_disable_unprepare(priv->can_clk);
> >>>>>
> >>>>> Shouldn't pretty much all these occurrences of clk_disable/enable
> >>>>> disappear? This should all be handled by the runtime_pm framework now.
> >>>>
> >>>> We have:
> >>>> - clk_prepare_enable() in probe
> >>>
> >>> This should become something like pm_runtime_get_sync(), shouldn't it?
> >>>
> >>>> - clk_disable_unprepare() in remove
> >>>
> >>> pm_runtime_put()
> >>>
> >>>> - clk_enable() in runtime_resume
> >>>> - clk_disable() in runtime_suspend
> >>>
> >>> These are the ones needed.
> >>>
> >>> The above makes me suspect that the clocks are always on, regardless of
> >>
> >> Define "on" :)
> >> The clocks are prepared after probe() exists, but not enabled. The first
> >> pm_runtime_get_sync() will enable the clocks.
> >>
> >>> the runtime suspend state since they are enabled in probe and disabled
> >>> in remove, is that right? Ideally, the usage in probe and remove should
> >>> be migrated to runtime_pm and clocks should really only be running when
> >>> needed and not throughout the whole lifetime of the driver.
> >>
> >> The clocks are not en/disabled via pm_runtime, because
> >> pm_runtime_get_sync() is called from atomic contect. We can have another
> >> look into the driver and try to change this.
> 
> > Wasn't that why the call to pm_runtime_irq_safe() was added?
> 
> Good question. That should be investigated.
> 
> > Also, clk_enable/disable should be okay to be run from atomic context.
> > And if the clock are already prepared after the exit of probe that
> > should be enough. Then remove() should just have to do the unprepare.
> > But I don't see why runtime_pm shouldn't be able to do the
> > enable/disable.
> 
> runtime_pm does call the clk_{enable,disable} function. But you mean
> clk_prepare() + pm_runtime_get_sync() should be used in probe() instead
> of calling clk_prepare_enable(). Good idea! I think the
> "pm_runtime_set_active(&pdev->dev);" has to be removed from the patch.

Right, that's what I was thinking. The proposed changes make sense, IMHO.

> 
> Coming back whether blocking calls are allowed or not.
> If you make a call to pm_runtime_irq_safe(), you state that it's okay to
> call pm_runtime_get_sync() from atomic context. But it's only called in
> open, probe, remove and in xcan_get_berr_counter, which is not called
> from atomic either. So let's try to remove the pm_runtime_irq_safe() and
> use clk_prepare_enable() clk_disable_unprepare() in the runtime_resume()
> runtime_suspend() functions.

IIRC, xcan_get_berr_counter() is called from atomic context. I think
that was how this got started.

	Sören