Re: [PATCH] kernel/resource: Invalid memory access in __release_resource

[Bjorn Helgaas <bhelgaas@google.com>]

[+cc Grant (author of ac80a51e2ce5)]

Hi Ricardo,

On Mon, Apr 20, 2015 at 06:22:52PM +0200, Ricardo Ribalda Delgado wrote:
> When a resource is initialized via of_platform_populate.
> resource->parent is initialized to NULL via kzalloc.
> (of_platform_populate->of_device_alloc->of_address_to_resource)
> 
> If of_platform_depopulate is called later, resource->parent is
> accessed (Offset 0x30 of address 0), causing a kernel error.

Interesting; how'd you find this?  It looks like the
of_platform_depopulate() code has been this way for a long time, so we
must be doing something new that makes us trip over this now.  More
analysis below...

> This patch evaluates resouce->parent before accessing it. If it
> is not initialized, -EACCESS is returned.
> 
> Fixes:
> BUG: unable to handle kernel NULL pointer deference at 0000000000000030
> IP: release_resource+0x26/0x90
> Signed-off-by: Ricardo Ribalda Delgado <ricardo.ribalda@gmail.com>
> ---
>  kernel/resource.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/kernel/resource.c b/kernel/resource.c
> index 90552aa..35dc716 100644
> --- a/kernel/resource.c
> +++ b/kernel/resource.c
> @@ -237,6 +237,9 @@ static int __release_resource(struct resource *old)
>  {
>  	struct resource *tmp, **p;
>  
> +	if (!old->parent)
> +		return -EINVAL;

This path has been fine for a long time without testing for a NULL
pointer, so I suspect this change papers over an issue that would be
better fixed elsewhere.

>From reading drivers/base/platform.c, it looks like the intent is
that platform device users would use these interfaces:

  - platform_device_alloc()

  - platform_device_add_resources(platform_device *pdev)
      pdev->num_resources = num

  - platform_device_add(platform_device *pdev)
      for (i = 0; i < pdev->num_resources; i++)
	insert_resource()
      device_add(&pdev->dev)

  - platform_device_unregister(platform_device *pdev)
      platform_device_del(platform_device *pdev)
	for (i = 0; i < pdev->num_resources; i++)
	  release_resource()

Resources are added by platform_device_add_resources() and inserted
into the resource tree by platform_device_add().  In this usage,
resources are removed from the resource tree by
platform_device_unregister(), and there's no issue with
resource->parent being NULL.

OF uses platform_device_alloc() and platform_device_unregister(), but 
not platform_device_add().  It doesn't call insert_resource(), and that
breaks the platform_device_unregister() assumption that the resources
are in the resource tree:

  - of_platform_populate()
      ...
	of_device_alloc()
	  pdev = platform_device_alloc()
	  # set pdev->resource, similar to platform_device_add_resources()
	  of_device_add(platform_device *pdev)
	    # similar to platform_device_add(), but note there's no
	    # insert_resource() in this path
	    device_add(&pdev->dev)

  - of_platform_depopulate()
      of_platform_device_destroy()
        platform_device_unregister(platform_device *pdev)
	  platform_device_del(platform_device *pdev)
	    for (i = 0; i < pdev->num_resources; i++)
	      release_resource()

I cc'd Grant because ac80a51e2ce5 ("of/device: populate
platform_device (of_device) resource table on allocation") added the
pdev->resource management in of_device_alloc(), so maybe he has more
insight into this.

>  	p = &old->parent->child;
>  	for (;;) {
>  		tmp = *p;
> -- 
> 2.1.4
>