Re: [PATCH v2] i2c: fix reference leak in MP2 PCI device

[Ma Ke <make24@iscas.ac.cn>]

On Thu, 2 Oct 2025 at 07:56, Andi Shyti <andi.shyti@kernel.org> wrote:
> Hi,
> 
> > diff --git a/drivers/i2c/busses/i2c-amd-mp2-pci.c b/drivers/i2c/busses/i2c-amd-mp2-pci.c
> > index ef7370d3dbea..60edbabc2986 100644
> > --- a/drivers/i2c/busses/i2c-amd-mp2-pci.c
> > +++ b/drivers/i2c/busses/i2c-amd-mp2-pci.c
> > @@ -458,13 +458,16 @@ struct amd_mp2_dev *amd_mp2_find_device(void)
> >  {
> >  	struct device *dev;
> >  	struct pci_dev *pci_dev;
> > +	struct amd_mp2_dev *mp2_dev;
> >  
> >  	dev = driver_find_next_device(&amd_mp2_pci_driver.driver, NULL);
> >  	if (!dev)
> >  		return NULL;
> >  
> >  	pci_dev = to_pci_dev(dev);
> > -	return (struct amd_mp2_dev *)pci_get_drvdata(pci_dev);
> > +	mp2_dev = (struct amd_mp2_dev *)pci_get_drvdata(pci_dev);
> > +	put_device(dev);
> > +	return mp2_dev;
> 
> the patch is good, but I don't think you need to declare mp2_dev
> because to_pci_dev(dev) should work even without hodling the
> reference of dev.
Thank you for your feedback. The declaration of the temporary variable
mp2_dev in the patch may be necessary because we need to save the 
result of pci_get_drvdata(pci_dev) before calling put_device(dev). If 
we do not do this and instead call put_device(dev) first before 
returning pci_get_drvdata(pci_dev), it may lead to the deallocation of
dev, thereby invalidating pci_dev. Accessing its driver data at this 
point could potentially trigger a use-after-free error. Therefore, the
temporary variable ensures that the driver data remains safely 
accessible even after the reference is released. So perhaps we need to
retain the declaration of this temporary variable?

Best regards,
Ma Ke
> 
> I also have to agree with Markus that something like:
> 
> struct device *dev __free(put_device) = ...; /* it can also be NULL */
> 
> would work nicer.
> 
> Thanks,
> Andi
> 
> >  }
> >  EXPORT_SYMBOL_GPL(amd_mp2_find_device);
> >  
> > -- 
> > 2.17.1
> >