Re: [PATCH v3 1/2] driver core: detach device's pm_domain after devres_release_all

[Rafael J. Wysocki]

On Wednesday, August 30, 2017 3:34:26 AM CEST Shawn Lin wrote:
> The intention of this patch is to move dev_pm_domain_detach after
> devres_release_all to avoid possible accessing device's registers
> with genpd been powered off.
> 
> Many common IP drivers use devm_request_irq to request irq for either
> shared irq or non-shared irq. So we rely on devres_release_all to
> free irq automatically. However we could see a situation that if the
> driver use devm_request_irq to register a shared irq and unbind the
> driver later, the irq could be triggerd cocurrently just between
> finishing dev_pm_domain_detach and calling devm_irq_release, so that
> driver's ISR should be called and try to access device's register, which
> may hang up the system. The reason is that some SoCs, including Rockchips'
> SoCs, couldn't support accessing controllers' registers w/o clk and power
> domain enabled.
> 
> Also we have CONFIG_DEBUG_SHIRQ to theoretically expose this kind of
> problem as devm_free_irq -> __free_irq says: "It's a shared IRQ -- the
> driver ought to be prepared for an IRQ event to happen even now it's being
> freed". That will simulate the aforementioned situation as it fires
> extra irq action to make sure driver/system is robust enough to deal
> with this kind of problem.
> 
> So now we face a two possible choices to fix this by
> (1) either using request_irq and free_irq directly
> (2) or moving dev_pm_domain_detach after devres_release_all which
> makes sure we free the irq before powering off power domain.
> 
> However, choice (1) implies that devm_request_irq shouldn't exist
> or at least shouldn't be used for shared irq case. Meanwhile we don't
> know how many drivers have this kind of issue and need to fix. So
> choice (2) makes more sense to me, and that is the reason for why we
> need to fix it like what this patch does.
> 
> Signed-off-by: Shawn Lin <shawn.lin@xxxxxxxxxxxxxx>
> 
> ---
> 
> Changes in v3:
> - fix the code path for consolidating the attach for both of driver
>   and bus driver, and then move detach to the error path
> - rework the changelog
> 
>  drivers/base/dd.c       | 16 ++++++++++++++++
>  drivers/base/platform.c | 18 ++----------------
>  2 files changed, 18 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> index ad44b40..aea0bb1 100644
> --- a/drivers/base/dd.c
> +++ b/drivers/base/dd.c
> @@ -25,7 +25,9 @@
>  #include <linux/kthread.h>
>  #include <linux/wait.h>
>  #include <linux/async.h>
> +#include <linux/platform_device.h>
>  #include <linux/pm_runtime.h>
> +#include <linux/pm_domain.h>
>  #include <linux/pinctrl/devinfo.h>
>  
>  #include "base.h"
> @@ -356,6 +358,7 @@ static int really_probe(struct device *dev, struct device_driver *drv)
>  	int local_trigger_count = atomic_read(&deferred_trigger_count);
>  	bool test_remove = IS_ENABLED(CONFIG_DEBUG_TEST_DRIVER_REMOVE) &&
>  			   !drv->suppress_bind_attrs;
> +	struct platform_driver *pdrv;

Oh.

>  
>  	if (defer_all_probes) {
>  		/*
> @@ -409,6 +412,16 @@ static int really_probe(struct device *dev, struct device_driver *drv)
>  	 */
>  	devices_kset_move_last(dev);
>  
> +	ret = dev_pm_domain_attach(dev, true);

Are you sure this can go here?

> +	pdrv = to_platform_driver(dev->driver);

Well, no.  There has to be a better way.

> +	/* don't fail if just dev_pm_domain_attach failed */
> +	if (pdrv && pdrv->prevent_deferred_probe &&
> +	    ret == -EPROBE_DEFER) {
> +		dev_warn(dev, "probe deferral not supported\n");
> +		ret = -ENXIO;
> +		goto probe_failed;
> +	}
> +
>  	if (dev->bus->probe) {
>  		ret = dev->bus->probe(dev);
>  		if (ret)
> @@ -428,6 +441,7 @@ static int really_probe(struct device *dev, struct device_driver *drv)
>  			drv->remove(dev);
>  
>  		devres_release_all(dev);
> +		dev_pm_domain_detach(dev, true);
>  		driver_sysfs_remove(dev);
>  		dev->driver = NULL;
>  		dev_set_drvdata(dev, NULL);
> @@ -458,6 +472,7 @@ static int really_probe(struct device *dev, struct device_driver *drv)
>  pinctrl_bind_failed:
>  	device_links_no_driver(dev);
>  	devres_release_all(dev);
> +	dev_pm_domain_detach(dev, true);
>  	driver_sysfs_remove(dev);
>  	dev->driver = NULL;
>  	dev_set_drvdata(dev, NULL);
> @@ -864,6 +879,7 @@ static void __device_release_driver(struct device *dev, struct device *parent)
>  		dma_deconfigure(dev);
>  
>  		devres_release_all(dev);
> +		dev_pm_domain_detach(dev, true);
>  		dev->driver = NULL;
>  		dev_set_drvdata(dev, NULL);
>  		if (dev->pm_domain && dev->pm_domain->dismiss)
> diff --git a/drivers/base/platform.c b/drivers/base/platform.c
> index d1bd992..8fa688d 100644
> --- a/drivers/base/platform.c
> +++ b/drivers/base/platform.c
> @@ -572,22 +572,8 @@ static int platform_drv_probe(struct device *_dev)
>  	if (ret < 0)
>  		return ret;
>  
> -	ret = dev_pm_domain_attach(_dev, true);

There are other bus types using dev_pm_domain_attach() IIRC, so why is
platform special?

> -	if (ret != -EPROBE_DEFER) {
> -		if (drv->probe) {
> -			ret = drv->probe(dev);
> -			if (ret)
> -				dev_pm_domain_detach(_dev, true);
> -		} else {
> -			/* don't fail if just dev_pm_domain_attach failed */
> -			ret = 0;
> -		}
> -	}
> -
> -	if (drv->prevent_deferred_probe && ret == -EPROBE_DEFER) {
> -		dev_warn(_dev, "probe deferral not supported\n");
> -		ret = -ENXIO;
> -	}
> +	if (drv->probe)
> +		ret = drv->probe(dev);
>  
>  	return ret;
>  }
> 

The last part is not related to PM domains.  Why does it need to be changed?

Thanks,
Rafael