Re: [PATCH v1] platform: Fix race condition during DMA configure at IOMMU probe time

[Bjorn Helgaas <helgaas@kernel.org>]

On Tue, Apr 22, 2025 at 04:26:49PM -0700, Will McVicker wrote:
> If devices are probed asynchronously, then there is a chance that during
> the IOMMU probe the driver is bound to the device in parallel. If this
> happens after getting the platform_driver pointer while in the function
> `platform_dma_configure()`, then the invalid `drv` pointer
> (drv==0xf...ffd8) will be de-referenced since `dev->driver != NULL`.

I need a little more hand-holding to make sense out of this.

After digging out
https://lore.kernel.org/all/aAa2Zx86yUfayPSG@google.com/, I see that
drv==0xf...ffd8 must be a result of applying to_platform_driver() to a
NULL pointer.  This patch still applies to_platform_driver(NULL), but
avoids using the result by testing drv for NULL later, which seems
prone to error.

I think this would all be clearer if we tested for the NULL pointer
explicitly before applying to_platform_driver().  I don't like setting
a pointer to an invalid value.  I think it's better if the pointer is
either valid or uninitialized because the compiler can help find uses
of uninitialized pointers.

> To avoid a kernel panic and eliminate the race condition, we should
> guard the usage of `dev->driver` by only reading it once at the
> beginning of the function.
> 
> Fixes: bcb81ac6ae3c ("iommu: Get DT/ACPI parsing into the proper probe path")
> Signed-off-by: Will McVicker <willmcvicker@google.com>
> ---
>  drivers/base/platform.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/base/platform.c b/drivers/base/platform.c
> index 1813cfd0c4bd..b948c6e8e939 100644
> --- a/drivers/base/platform.c
> +++ b/drivers/base/platform.c
> @@ -1440,7 +1440,8 @@ static void platform_shutdown(struct device *_dev)
>  
>  static int platform_dma_configure(struct device *dev)
>  {
> -	struct platform_driver *drv = to_platform_driver(dev->driver);
> +	struct device_driver *drv = READ_ONCE(dev->driver);
> +	struct platform_driver *pdrv = to_platform_driver(drv);
>  	struct fwnode_handle *fwnode = dev_fwnode(dev);
>  	enum dev_dma_attr attr;
>  	int ret = 0;
> @@ -1451,8 +1452,8 @@ static int platform_dma_configure(struct device *dev)
>  		attr = acpi_get_dma_attr(to_acpi_device_node(fwnode));
>  		ret = acpi_dma_configure(dev, attr);
>  	}
> -	/* @drv may not be valid when we're called from the IOMMU layer */
> -	if (ret || !dev->driver || drv->driver_managed_dma)
> +	/* @dev->driver may not be valid when we're called from the IOMMU layer */
> +	if (ret || !drv || pdrv->driver_managed_dma)
>  		return ret;
>  
>  	ret = iommu_device_use_default_domain(dev);
> -- 
> 2.49.0.805.g082f7c87e0-goog
>