Re: [PATCH] dma-buf/sw_sync: Decrement refcount on error in sw_sync_ioctl_get_deadline()

[Christian König]

Am 04.04.25 um 10:27 schrieb Dan Carpenter:
> On Mon, Mar 31, 2025 at 02:02:44PM +0200, Christian König wrote:
>> Am 31.03.25 um 11:45 schrieb Dan Carpenter:
>>> Call dma_fence_put(fence) before returning an error on this error path.
>>>
>>> Fixes: 70e67aaec2f4 ("dma-buf/sw_sync: Add fence deadline support")
>>> Signed-off-by: Dan Carpenter <dan.carpenter@xxxxxxxxxx>
>>> ---
>>>  drivers/dma-buf/sw_sync.c | 4 +++-
>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/dma-buf/sw_sync.c b/drivers/dma-buf/sw_sync.c
>>> index f5905d67dedb..b7615c5c6cac 100644
>>> --- a/drivers/dma-buf/sw_sync.c
>>> +++ b/drivers/dma-buf/sw_sync.c
>>> @@ -438,8 +438,10 @@ static int sw_sync_ioctl_get_deadline(struct sync_timeline *obj, unsigned long a
>>>  		return -EINVAL;
>>>  
>>>  	pt = dma_fence_to_sync_pt(fence);
>>> -	if (!pt)
>>> +	if (!pt) {
>>> +		dma_fence_put(fence);
>>>  		return -EINVAL;
>>> +	}
>> Good catch.
>>
>> I think it would be cleaner if we add an error label and then use "ret = -EINVAL; goto error;" here as well as a few lines below when ret is set to -ENOENT.
>>
>> This way we can also avoid the ret = 0 in the declaration and let the compiler actually check the lifetime of the assignment.
>>
> I had some issues with my email and it silently ate a bunch of outgoing
> email without saving a single trace of anything I had sent.  I see
> this was one that was eaten.

Yeah, AMD had similar problems with receiving mails at the beginning of the year.

>
> Unwind ladders don't work really well for things where you just take it
> for a little while and then drop it a few lines later.  Such as here you
> take reference and then drop it or you take a lock and then drop it.
> Normally, you can add things to anywere in the unwind ladder but if you
> add an unlock to the ladder than you to add a weird bunny hop if the goto
> isn't holding the lock.  It ends up getting confusing.  With that kind of
> thing, I prefer to do the unlock before the goto.

Yeah, completely agree. This is usually also a good indicator that something should be in a separate function.

But this case doesn't apply here, doesn't it?

I mean the solution you created below has a few more lines of code, but if you ask me that is way more readable.

The -EFAULT doesn't need any cleanup and can perfectly stay separate as far as I can see.

Regards,
Christian.

>
> free_c:
> 	free(c);
> 	goto free_b;  <-- bunny hop;
> unlock:
> 	unlock();
> free_b:
> 	free(b);
> free_a:
> 	free(a);
>
> 	return ret;
>
> regards,
> dan carpenter
>
> diff --git a/drivers/dma-buf/sw_sync.c b/drivers/dma-buf/sw_sync.c
> index f5905d67dedb..22a808995f10 100644
> --- a/drivers/dma-buf/sw_sync.c
> +++ b/drivers/dma-buf/sw_sync.c
> @@ -438,15 +438,17 @@ static int sw_sync_ioctl_get_deadline(struct sync_timeline *obj, unsigned long a
>  		return -EINVAL;
>  
>  	pt = dma_fence_to_sync_pt(fence);
> -	if (!pt)
> -		return -EINVAL;
> +	if (!pt) {
> +		ret = -EINVAL;
> +		goto put_fence;
> +	}
>  
>  	spin_lock_irqsave(fence->lock, flags);
> -	if (test_bit(SW_SYNC_HAS_DEADLINE_BIT, &fence->flags)) {
> -		data.deadline_ns = ktime_to_ns(pt->deadline);
> -	} else {
> +	if (!test_bit(SW_SYNC_HAS_DEADLINE_BIT, &fence->flags)) {
>  		ret = -ENOENT;
> +		goto unlock;
>  	}
> +	data.deadline_ns = ktime_to_ns(pt->deadline);
>  	spin_unlock_irqrestore(fence->lock, flags);
>  
>  	dma_fence_put(fence);
> @@ -458,6 +460,13 @@ static int sw_sync_ioctl_get_deadline(struct sync_timeline *obj, unsigned long a
>  		return -EFAULT;
>  
>  	return 0;
> +
> +unlock:
> +	spin_unlock_irqrestore(fence->lock, flags);
> +put_fence:
> +	dma_fence_put(fence);
> +
> +	return ret;
>  }
>  
>  static long sw_sync_ioctl(struct file *file, unsigned int cmd,
>
>