Subject: Re: [PATCH] staging: bcm2835-audio: Fix memory corruption
To: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Eric Anholt <eric@anholt.net>, Stefan Wahren <stefan.wahren@i2se.com>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Florian Fainelli <f.fainelli@gmail.com>,
        Aishwarya Pant <aishpant@gmail.com>,
        linux-rpi-kernel@lists.infradead.org, devel@driverdev.osuosl.org,
        linux-kernel@vger.kernel.org
References: <1502193902-87160-1-git-send-email-phil@raspberrypi.org>
 <20170810102155.6j373u3yktnlwsqw@mwanda>
 <b88ad53c-0bd3-804e-d1db-3561fd05a850@raspberrypi.org>
 <20170810112451.nkvj7p6r7qgqqnh3@mwanda>
From: Phil Elwell <phil@raspberrypi.org>
Message-ID: <1d59de68-22c0-6f28-ef9b-3393e457e23f@raspberrypi.org>
Date: Thu, 10 Aug 2017 14:25:57 +0100
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:52.0) Gecko/20100101
 Thunderbird/52.2.1
MIME-Version: 1.0
In-Reply-To: <20170810112451.nkvj7p6r7qgqqnh3@mwanda>
Content-Type: text/plain; charset=utf-8
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3539
Lines: 85


On 10/08/2017 12:24, Dan Carpenter wrote:
> On Thu, Aug 10, 2017 at 11:52:42AM +0100, Phil Elwell wrote:
>> On 10/08/2017 11:21, Dan Carpenter wrote:
>>> The original patch did not go through the normal review process...
>>>
>>> On Tue, Aug 08, 2017 at 01:05:02PM +0100, Phil Elwell wrote:
>>>> I'm all for fixing memory leaks, but freeing a block while it is still
>>>> being used is a recipe for hard-to-debug kernel exceptions.
>>>>
>>>
>>> This bug completely breaks the driver doesn't it?  It's not very subtle
>>> so it should be easy to diagnose with git bisect.
>>
>> It's more subtle than that - the failure isn't consistent, sometimes crashing
>> and sometimes not depending on how and when memory is reused.
>>
>>>> 1) There is already a vchi method for freeing the instance, so use it.
>>>> 2) Only call it on error, and then only before initted is false.
>>>>
>>>> Signed-off-by: Phil Elwell <phil@raspberrypi.org>
>>>> Fixes: 0adbfd4694c2 ("staging: bcm2835-audio: fix memory leak in bcm2835_audio_open_connection()")
>>>> ---
>>>>  drivers/staging/vc04_services/bcm2835-audio/bcm2835-vchiq.c | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/staging/vc04_services/bcm2835-audio/bcm2835-vchiq.c b/drivers/staging/vc04_services/bcm2835-audio/bcm2835-vchiq.c
>>>> index 5f3d8f2..89f96f3 100644
>>>> --- a/drivers/staging/vc04_services/bcm2835-audio/bcm2835-vchiq.c
>>>> +++ b/drivers/staging/vc04_services/bcm2835-audio/bcm2835-vchiq.c
>>>> @@ -409,6 +409,7 @@ static int bcm2835_audio_open_connection(struct bcm2835_alsa_stream *alsa_stream
>>>>  			LOG_ERR("%s: failed to connect VCHI instance (ret=%d)\n",
>>>>  				__func__, ret);
>>>>  
>>>> +			vchi_disconnect(vchi_instance);
>>>
>>> This is ugly because why are we calling disconnect() if connect() fails?
>>> These functions should be symetric so disconnect only disconnects and
>>> we call a different function to undo vchi_initialise().
>>
>> Agreed - I'm not going to change the API, but I can add a comment.
>>
> 
> Nah...  Please don't do that.  Just create a function:
> 
> void vchiq_free(VCHIQ_INSTANCE_T *vchi_instance)
> {
> 	kfree(vchi_instance);
> }
> 
> Really vchi_initialise() is badly named and it's just a wrapper around
> vchiq_initialise().  It should be deleted and vchiq_initialise() should
> be renamed to allocate instead of initialize...  But that's for a
> separate patch.
> 
> And also we should move the kfree() out of disconnect() like I said
> and instead call vchiq_free().  But again, that's for a separate patch.
> 
> Change the goto to "return -EIO."  Leave the last error path as-is.
> Eventually we will want to break this into two functions, one that
> allocates the first vchi_instance and one that calls vc_vchi_audio_init().
> But again, there are many patches needed to fix this code.

[ static removed ]

Shouldn't that be:

void vchi_uninitialise(VCHI_INSTANCE_T instance_handle)
{
	vchiq_shutdown((VCHIQ_INSTANCE_T)instance_handle);
}
EXPORT_SYMBOL(vchi_uninitialise);

Yes, in this case the instance is just a block of memory with no users,
but if this is a general API call then you want it to work in all cases,
and calling vchiq_shutdown is the right way to free an instance. Calling
a VCHIQ method when all the other calls are to the VCHI API is grubby,
so make it a VCHI method instead.

If instead this is just another hack - a kfree with a comment in code form -
then lets keep it static within the audio driver (which is what I thought you
were initially suggesting).

Phil