Re: [PATCH v3 2/4] soc: amlogic: meson-gx-socinfo-sm: Add Amlogic secure-monitor SoC Information driver

[Krzysztof Kozlowski]

On 14/03/2024 13:22, Viacheslav wrote:
>> +
>>> +	soc_dev_attr = devm_kzalloc(&pdev->dev, sizeof(*soc_dev_attr),
>>> +				    GFP_KERNEL);
>>> +	if (!soc_dev_attr)
>>> +		return -ENOMEM;
>>> +
>>> +	soc_dev_attr->serial_number = socinfo_get_chipid(&pdev->dev, fw, &socinfo);
>>> +
>>> +	soc_dev_attr->family = "Amlogic Meson";
>>> +	soc_dev_attr->revision = kasprintf(GFP_KERNEL, "%x:%x - %x:%x",
>>> +					   socinfo.v1.major_id,
>>> +					   socinfo.v1.chip_rev,
>>> +					   socinfo.v1.pack_id,
>>> +					   (socinfo.v1.reserved<<4) + socinfo.v1.layout_ver);
>>> +	soc_dev_attr->soc_id = kasprintf(GFP_KERNEL, "%s (%s)",
>>> +					 socinfo_v1_to_soc_id(socinfo),
>>> +					 socinfo_v1_to_package_id(socinfo));
>>> +
>>> +	soc_dev = soc_device_register(soc_dev_attr);
>>> +
>>> +
>>> +	if (IS_ERR(soc_dev)) {
>>> +		kfree(soc_dev_attr->revision);
>>> +		kfree_const(soc_dev_attr->soc_id);
>>> +		kfree(soc_dev_attr);
>>
>> That's a double free. This was not tested.
> 
> 
> Please, describe the problem.

Test your code. What's the point of arguing over it if regular test
would show this?

> I don't quite understand what the issue is:
> 
> - kfree() releases memory allocated with kmalloc()

So point me where is kmalloc(). I don't see. I see only devm.

> - kasprintf() allocates memory using kmalloc_track_caller()
> 
> Technically, I see no difficulty in freeing the newly allocated memory. 
> In case of memory allocation issues in kasprintf, we would just get 
> NULL, which kfree should also handle properly. Considering that we don't 
> need soc_dev_attr anymore, we don't need to worry about the contents of 
> .revision and .soc_id.

Please pay attention that my comment is under specific line. We do not
discuss unrelated code.

> 
> I see that kfree_const has crept in by accident, which is essentially 
> needed here only if we replace kasprintf with kasprintf_const for 
> .soc_id, but it does not introduce any erroneous behavior.

> 
>>
>>> +		return PTR_ERR(soc_dev);
>>> +	}
>>> +
>>> +	dev = soc_device_to_device(soc_dev);
>>> +	platform_set_drvdata(pdev, soc_dev);
>>> +
>>> +	dev_info(dev, "Amlogic Meson %s Revision %x:%x (%x:%x) Detected (SM)\n",
>>> +			soc_dev_attr->soc_id,
>>> +			socinfo.v1.major_id,
>>> +			socinfo.v1.chip_rev,
>>> +			socinfo.v1.pack_id,
>>> +			(socinfo.v1.reserved<<4) + socinfo.v1.layout_ver);
>>> +
>>> +	return PTR_ERR_OR_ZERO(dev);
>>> +}
>>> +
>>> +
>>> +static int meson_gx_socinfo_sm_remove(struct platform_device *pdev)
>>> +{
>>> +	struct soc_device *soc_dev = platform_get_drvdata(pdev);
>>> +
>>> +	soc_device_unregister(soc_dev);
>>
>> If you free the memory in probe() error path, why you did not decide to
>> free it here as well? It is symmetrical, so this should make you wonder
>> - error path is wrong.
> 
> This is something I can easily explain:
> 
> In the case where we have successfully registered with 
> soc_device_register, we clean up everything that was manually allocated 
> and not used.
> In the case of unloading the driver, the cleanup should be handled by 
> the soc_device_unregister command.
> 
> Technically, it's not possible to insert memory release because until 
> the command is called, the driver is active, and afterwards, there's no 
> guarantee of the pointer's validity.

Then you do not understand lifecycle of device. There is release here
via devm. Exactly at after my comment, when } finishes.

> Perhaps it would have been better if soc_device_register copied the 
> entire soc_device_attribute structure and took care of memory allocation 
> and release itself, then we could comfortably free any excess memory 
> back in _probe.
> 
> Are you currently recommending not to release memory within the if 
> (IS_ERR(soc_dev)) section?

You have double free which will be pointed out by testing. Yes, of
course I recommend not to have double free, so not to release memory
which is being released.

Best regards,
Krzysztof