Re: [PATCH v4] bluetooth: hci_h5: fix memory leak in h5_close

[Hans de Goede]

Hi,

On 10/16/20 1:55 PM, Anant Thazhemadam wrote:
> 
> Hi,
> 
> On 16/10/20 4:58 pm, Hans de Goede wrote:
>> Hi,
>>
>> On 10/7/20 5:48 AM, Anant Thazhemadam wrote:
>>> If h5_close is called when !hu->serdev, h5 is directly freed.
>>> However, h5->rx_skb is not freed, which causes a memory leak.
>>>
>>> Freeing h5->rx_skb fixes this memory leak.
>>>
>>> In case hu->serdev exists, h5->rx_skb is then set to NULL,
>>> since we do not want to risk a potential NULL pointer 
>>> dereference.
>>>
>>> Fixes: ce945552fde4 ("Bluetooth: hci_h5: Add support for serdev enumerated devices")
>>> Reported-by: syzbot+6ce141c55b2f7aafd1c4@xxxxxxxxxxxxxxxxxxxxxxxxx
>>> Tested-by: syzbot+6ce141c55b2f7aafd1c4@xxxxxxxxxxxxxxxxxxxxxxxxx
>>> Signed-off-by: Anant Thazhemadam <anant.thazhemadam@xxxxxxxxx>h5_close v4
>>> ---
>>> Changes in v4:
>>> 	* Free h5->rx_skb even when hu->serdev
>>> 	(Suggested by Hans de Goede <hdegoede@xxxxxxxxxx>)
>>> 	* If hu->serdev, then assign h5->rx_skb = NULL
>>>
>>> Changes in v3:
>>> 	* Free h5->rx_skb when !hu->serdev, and fix the memory leak
>>> 	* Do not incorrectly and unnecessarily call serdev_device_close()
>>>
>>> Changes in v2:
>>> 	* Fixed the Fixes tag
>>>
>>>  drivers/bluetooth/hci_h5.c | 4 ++++
>>>  1 file changed, 4 insertions(+)
>>>
>>> diff --git a/drivers/bluetooth/hci_h5.c b/drivers/bluetooth/hci_h5.c
>>> index e41854e0d79a..39f9553caa5c 100644
>>> --- a/drivers/bluetooth/hci_h5.c
>>> +++ b/drivers/bluetooth/hci_h5.c
>>> @@ -245,11 +245,15 @@ static int h5_close(struct hci_uart *hu)
>>>  	skb_queue_purge(&h5->rel);
>>>  	skb_queue_purge(&h5->unrel);
>>>  
>>> +	kfree_skb(h5->rx_skb);
>>> +
>>>  	if (h5->vnd && h5->vnd->close)
>>>  		h5->vnd->close(h5);
>>>  
>>>  	if (!hu->serdev)
>>>  		kfree(h5);
>>> +	else
>>> +		h5->rx_skb = NULL;
>> Please just do this unconditionally directly after
>> the kfree_skb()
> 
> Could you also please tell me why this might be necessary?
> The pointer value stored at h5->rx_skb would be freed anyways when we free h5 (since rx_skb is
> essentially a member of the structure that h5 points to).

It is necessary in the path where the struct h5 points to is not
free-ed and it is cleaner to just always do it then, as you
indicate yourself 

> Also since we're performing the *if* check, the *else* condition wouldn't exactly be taxing either,
> right?

For the computer it is not taxing, but for a human reading the code
and trying to understand the flow it makes things extra complicated
unnecessarily.

> Is there some performance metric that I'm missing where unconditionally setting it to NULL
> in this manner would be better? (I couldn't find any resources that had any similar analysis
> performed :/ )
> Or is this in interest of code readability?

Yes, it is in interest of code readability?

> Also, how about we introduce a h5 = NULL, after freeing h5 when !hu->serdev?

That is not necessary, there is no reason to have that in either code path.

Regards,

Hans