[PATCH 1/2] of: overlay: improve error handling in of_overlay_apply_tree

[Michael Riesch]

Hi Michael,

On 9/21/22 09:57, Michael Tretter wrote:
> On Wed, 21 Sep 2022 08:55:12 +0200, Michael Riesch wrote:
>> On 9/5/22 12:07, Michael Riesch wrote:
>>> Propagate any error from of_overlay_apply_symbols and let the user
>>> know if the provided overlay is not applicable.
>>>
>>> Signed-off-by: Michael Riesch <michael.riesch at wolfvision.net>
>>> ---
>>>  drivers/of/overlay.c | 14 +++++++++-----
>>>  1 file changed, 9 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/drivers/of/overlay.c b/drivers/of/overlay.c
>>> index 20a43f5170..20686db511 100644
>>> --- a/drivers/of/overlay.c
>>> +++ b/drivers/of/overlay.c
>>> @@ -115,8 +115,8 @@ static char *of_overlay_fix_path(struct device_node *root,
>>>  	return basprintf("%s%s", target->full_name, path_tail);
>>>  }
>>>  
>>> -static void of_overlay_apply_symbols(struct device_node *root,
>>> -				     struct device_node *overlay)
>>> +static int of_overlay_apply_symbols(struct device_node *root,
>>> +				    struct device_node *overlay)
>>>  {
>>>  	const char *old_path;
>>>  	char *new_path;
>>> @@ -129,12 +129,12 @@ static void of_overlay_apply_symbols(struct device_node *root,
>>>  
>>>  	if (!overlay_symbols) {
>>>  		pr_debug("overlay doesn't have a __symbols__ node\n");
>>> -		return;
>>> +		return -EINVAL;
>>
>> Come to think of it, do all overlays need to provide a __symbols__ node?
>> If not, this check is overly strict.
> 
> Overlays don't need a __symbols__ node. It would be only required, if overlays
> are stacked and the second overlay refers to nodes of the first overlay by
> labels. Having no __symbols__ in the overlay is a success path and the message
> is just a debug message.

Thanks for the clarification. We need to fix this one, then.

Seeing that the patch is in next: Am I supposed to send an incremental
"fixup! ..." patch which can be squashed? Or should I send a proper
patch with a Fixes: tag?

>>>  	}
>>>  
>>>  	if (!root_symbols) {
>>>  		pr_info("root doesn't have a __symbols__ node\n");
>>> -		return;
>>> +		return -EINVAL;
>>
>> Ditto for the root.
> 
> I'm not sure what should happen, if the root does not have __symbols__.
> Barebox wouldn't be able to copy the __symbols__ of the overlay, but this
> still wouldn't be a problem unless overlays are stacked. In the stacking case,
> only applying the second overlay should fail.

I can reestablish the behavior before this patch, i.e., __symbols__ is
optional in both root and overlay.

> Maybe, we should add a new __symbols__ node, if the root doesn't have a
> __symbols__ node?

If this is desired, I can implement the change -- but is it desired?

>>>  	}
>>>  
>>>  	list_for_each_entry(prop, &overlay_symbols->properties, list) {
>>> @@ -148,6 +148,8 @@ static void of_overlay_apply_symbols(struct device_node *root,
>>>  			 prop->name, new_path);
>>>  		of_property_write_string(root_symbols, prop->name, new_path);
>>>  	}
>>> +
>>> +	return 0;
>>>  }
>>>  
>>>  static int of_overlay_apply_fragment(struct device_node *root,
>>> @@ -190,7 +192,9 @@ int of_overlay_apply_tree(struct device_node *root,
>>>  		goto out_err;
>>>  
>>>  	/* Copy symbols from resolved overlay to base device tree */
>>> -	of_overlay_apply_symbols(root, resolved);
>>> +	err = of_overlay_apply_symbols(root, resolved);
>>> +	if (err)
>>> +		goto out_err;
>>
>> If both checks need to be relaxed, the complete patch should be reverted
>> I guess :-/
> 
> What did you do to run into this error? What was your expectation?

Well I tried to apply an overlay without __symbols__ :-) (which did work
before).

Best regards,
Michael