Hello Ahmad,

On Wed, Jun 07, 2023 at 10:08:18AM +0200, Ahmad Fatoum wrote:
> of_get_devicepath code flow is split into two:
> 
>   A) Either the device tree node in question has a direct udev_device
>      associated with it
> 
>   B) Or we assume it's a partition and lookup udev_device for the parent
>      first, before finding a child udev_device or setting a partition
>      offset within the parent udev_device.
> 
> Since v2017.03.0, we have had a fallthrough from case A into case B:
> If we have a udev_device, but it's neither a EEPROMs, MTDs or block
> device, we just consider it a partition. This is problematic, because
> this may result in us pointing at a very different device:
> 
>   - backend points at a SD-Card host. Host is enabled, but SD-Card
>     is not inserted, so no block device
> 
>   - case A fails, so it's assumed it's a partition and case B
>     uses parent SoC bus to lookup appropriate device
> 
>   - We fall through into the second device_find_block_device, which
>     will take the first matching block device across the SoC. So
>     we could end up with the eMMC: a completely different device
>     than what was pointed at.

So another surprise is that device_find_block_device() recurses to find
a device when starting on /soc, isn't it? Is this worth addressing?

> Fixes: 929ed64cb42f ("of_get_devicepath: make partition finding more robust")
> Signed-off-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
> ---
>  src/libdt.c | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/src/libdt.c b/src/libdt.c
> index e54d7fb5649d..7b99efe5b2de 100644
> --- a/src/libdt.c
> +++ b/src/libdt.c
> @@ -2492,9 +2492,11 @@ int of_get_devicepath(struct device_node *partition_node, char **devpath, off_t
>  		}
>  
>  		/*
> -		 * If we found a device but couldn't classify it above, we fall
> -		 * through.
> +		 * If we find a udev_device but couldn't classify it above,
> +		 * it's an error. Falling through would mean to handle it as a
> +		 * partition and could lead us to return an arbitrary sibling device
>  		 */
> +		return -ENODEV;

I don't remember the details of 929ed64cb42f any more, but probably I
didn't have a specific case that were fixed by that commit. Your
rationale makes sense.

Reviewed-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |