When all references are dropped, callback function plx_dma_release()
for put_device() will call kfree(plxdev) to release memory.
Fix the error path to fix the possible double free.

Fixes: 07503e6aefe4 ("dmaengine: plx_dma: add a missing put_device() on error path")
Signed-off-by: Miaoqian Lin <linmq006@gmail.com>
---
Please correct me if I make mistakes, thanks for your review.
---
 drivers/dma/plx_dma.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/dma/plx_dma.c b/drivers/dma/plx_dma.c
index 12725fa1655f..bce724ff4e16 100644
--- a/drivers/dma/plx_dma.c
+++ b/drivers/dma/plx_dma.c
@@ -546,8 +546,9 @@ static int plx_dma_create(struct pci_dev *pdev)
 	return 0;
 
 put_device:
-	put_device(&pdev->dev);
 	free_irq(pci_irq_vector(pdev, 0),  plxdev);
+	put_device(&pdev->dev);
+	return rc;
 free_plx:
 	kfree(plxdev);
 
-- 
2.25.1

On 2022-12-19 23:17, Miaoqian Lin wrote:
> When all references are dropped, callback function plx_dma_release()
> for put_device() will call kfree(plxdev) to release memory.
> Fix the error path to fix the possible double free.
>
> Fixes: 07503e6aefe4 ("dmaengine: plx_dma: add a missing put_device() on error path")
> Signed-off-by: Miaoqian Lin <linmq006@gmail.com>
> ---
> Please correct me if I make mistakes, thanks for your review.
> ---
>  drivers/dma/plx_dma.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/dma/plx_dma.c b/drivers/dma/plx_dma.c
> index 12725fa1655f..bce724ff4e16 100644
> --- a/drivers/dma/plx_dma.c
> +++ b/drivers/dma/plx_dma.c
> @@ -546,8 +546,9 @@ static int plx_dma_create(struct pci_dev *pdev)
>  	return 0;
>  
>  put_device:
> -	put_device(&pdev->dev);
>  	free_irq(pci_irq_vector(pdev, 0),  plxdev);
> +	put_device(&pdev->dev);
> +	return rc;
>  free_plx:
>  	kfree(plxdev);
>  


Sorry, after reviewing, I don't think this patch is correct.

plx_dma_release() is called from dma_async_device_unregister() which
won't be called if dma_async_device_register() fails. It does not get
freed when the pci device is put. So I don't think this is a possible
double free.

Logan

On Tue, Dec 20, 2022 at 09:35:38AM -0700, Logan Gunthorpe wrote:
> 
> 
> On 2022-12-19 23:17, Miaoqian Lin wrote:
> > When all references are dropped, callback function plx_dma_release()
> > for put_device() will call kfree(plxdev) to release memory.
> > Fix the error path to fix the possible double free.
> >
> > Fixes: 07503e6aefe4 ("dmaengine: plx_dma: add a missing put_device() on error path")
> > Signed-off-by: Miaoqian Lin <linmq006@gmail.com>
> > ---
> > Please correct me if I make mistakes, thanks for your review.
> > ---
> >  drivers/dma/plx_dma.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/dma/plx_dma.c b/drivers/dma/plx_dma.c
> > index 12725fa1655f..bce724ff4e16 100644
> > --- a/drivers/dma/plx_dma.c
> > +++ b/drivers/dma/plx_dma.c
> > @@ -546,8 +546,9 @@ static int plx_dma_create(struct pci_dev *pdev)
> >  	return 0;
> >  
> >  put_device:
> > -	put_device(&pdev->dev);
> >  	free_irq(pci_irq_vector(pdev, 0),  plxdev);
> > +	put_device(&pdev->dev);
> > +	return rc;
> >  free_plx:
> >  	kfree(plxdev);
> >  
> 
> 
> Sorry, after reviewing, I don't think this patch is correct.
> 
> plx_dma_release() is called from dma_async_device_unregister() which
> won't be called if dma_async_device_register() fails. It does not get
> freed when the pci device is put. So I don't think this is a possible
> double free.

I think instead of "double free" it Miaoqian meant "use after free".  It
does look like a use after free to me.  Certainly there is no harm from
applying the patch and it makes the code more obviously correct.

regards,
dan carpenter

On 2022-12-23 03:42, Dan Carpenter wrote:
> On Tue, Dec 20, 2022 at 09:35:38AM -0700, Logan Gunthorpe wrote:
>>
>>
>> On 2022-12-19 23:17, Miaoqian Lin wrote:
>>> When all references are dropped, callback function plx_dma_release()
>>> for put_device() will call kfree(plxdev) to release memory.
>>> Fix the error path to fix the possible double free.
>>>
>>> Fixes: 07503e6aefe4 ("dmaengine: plx_dma: add a missing put_device() on error path")
>>> Signed-off-by: Miaoqian Lin <linmq006@gmail.com>
>>> ---
>>> Please correct me if I make mistakes, thanks for your review.
>>> ---
>>>  drivers/dma/plx_dma.c | 3 ++-
>>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/dma/plx_dma.c b/drivers/dma/plx_dma.c
>>> index 12725fa1655f..bce724ff4e16 100644
>>> --- a/drivers/dma/plx_dma.c
>>> +++ b/drivers/dma/plx_dma.c
>>> @@ -546,8 +546,9 @@ static int plx_dma_create(struct pci_dev *pdev)
>>>  	return 0;
>>>  
>>>  put_device:
>>> -	put_device(&pdev->dev);
>>>  	free_irq(pci_irq_vector(pdev, 0),  plxdev);
>>> +	put_device(&pdev->dev);
>>> +	return rc;
>>>  free_plx:
>>>  	kfree(plxdev);
>>>  
>>
>>
>> Sorry, after reviewing, I don't think this patch is correct.
>>
>> plx_dma_release() is called from dma_async_device_unregister() which
>> won't be called if dma_async_device_register() fails. It does not get
>> freed when the pci device is put. So I don't think this is a possible
>> double free.
> 
> I think instead of "double free" it Miaoqian meant "use after free".  It
> does look like a use after free to me.  Certainly there is no harm from
> applying the patch and it makes the code more obviously correct.

Yeah, moving the put_device() up would certainly look more correct.

But I think it's actually not a bug either because plx_dma_create() is
called from the pci probe function, so the pci device cannot go away.

Logan