[libcamera-devel] [PATCH] libcamera: gst: Fix double-free when acquire_buffer fails

[Marian Cichy]

On 3/10/21 6:30 PM, Laurent Pinchart wrote:
> Hi Marian
>
> Thank you for the patch.
>
> On Wed, Mar 10, 2021 at 08:35:31PM +0530, Umang Jain wrote:
>> On 3/9/21 8:05 PM, Marian Cichy wrote:
>>> If gst_buffer_pool_acquire_buffer in gst_libcamera_task_run fails, the
>>> unique_ptr to the request-object gets reset and hence, its destructor
>>> is called. However, the wrap-object points to the same object and is
>>> still alive at this moment. When the task_run-function is finished, the
>>> destructor of the wrap-object is called, which in return calls the
>>> destructor of the request-object again.
>>>
>>> Also note the wrong comment, which claims that WrapRequest does not
>>> take ownership of the request, however, actually it already has
>>> ownership.
>>>
>>> Replacing request.reset() with request.release() doesn't call the
>>> destructor on the request-object and only one free happens at the end.
>>>
>>> Signed-off-by: Marian Cichy <m.cichy at pengutronix.de>
>>> ---
>>>    src/gstreamer/gstlibcamerasrc.cpp | 6 ++++--
>>>    1 file changed, 4 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/src/gstreamer/gstlibcamerasrc.cpp b/src/gstreamer/gstlibcamerasrc.cpp
>>> index a8ed7652..b0194c2f 100644
>>> --- a/src/gstreamer/gstlibcamerasrc.cpp
>>> +++ b/src/gstreamer/gstlibcamerasrc.cpp
>>> @@ -279,10 +279,12 @@ gst_libcamera_src_task_run(gpointer user_data)
>>>    						     &buffer, nullptr);
>>>    		if (ret != GST_FLOW_OK) {
>>>    			/*
>>> -			 * RequestWrap does not take ownership, and we won't be
>>> +			 * RequestWrap has ownership, and we won't be
>>>    			 * queueing this one due to lack of buffers.
>>> +			 * So the request will be freed when RequestWrap
>>> +			 * goes out of scope.
>>>    			 */
>>> -			request.reset();
>>> +			request.release();
>> This is a bit problematic as `request` is still in-used beyond this
>> line(there is `if (request)` block below) and here it will set it as
>> NULL. You will need to take care of this situation using the RequestWrap
>> probably. However, one thing I still am not sure of ...
>>
>>>    			break;
>>>    		}
>>>    
>> I am not sure if why `RequestWrap` is written intentionally in a way, to
>> transfer the ownership of the `Request`. This fact is quite evident
>> looking at the ~RequestWrap().  The larger picture here is (which is why
>> you hit this double-free issue) that, whenever RequestWrap is used -
>> always transfer ownership of Request. So, that's a 2-step manual in
>> order to use the wrap; if the latter is forgotten - we will always hit
>> double free issues.
>>
>> Can you please look how the RequestWrap is used elsewhere in the
>> gstreamer plugin and try to see why transfer of ownership is intented in
>> RequestWrap? If it's not essential, we should probably re-fine
>> RequestWrap to really be a "wrapper" and not try to steal the reference
>> - and take the additional responsibility of free-ing it.
> The request belongs to the gstreamer element, so we need to keep track
> of it and free it. Request instances are managed with std::unique_ptr<>,
> so the best option I think would be to simple move the request to the
> wrap, using a unique pointer there.
>
> How about this untested change ?
>
> diff --git a/src/gstreamer/gstlibcamerasrc.cpp b/src/gstreamer/gstlibcamerasrc.cpp
> index 636c14dff64e..e86c3d7f74c7 100644
> --- a/src/gstreamer/gstlibcamerasrc.cpp
> +++ b/src/gstreamer/gstlibcamerasrc.cpp
> @@ -52,19 +52,19 @@ GST_DEBUG_CATEGORY_STATIC(source_debug);
>   #define GST_CAT_DEFAULT source_debug
>   
>   struct RequestWrap {
> -	RequestWrap(Request *request);
> +	RequestWrap(std::unique_ptr<Request> request);
>   	~RequestWrap();
>   
>   	void attachBuffer(GstBuffer *buffer);
>   	GstBuffer *detachBuffer(Stream *stream);
>   
>   	/* For ptr comparison only. */
> -	Request *request_;
> +	std::unique_ptr<Request> request_;
>   	std::map<Stream *, GstBuffer *> buffers_;
>   };
>   
> -RequestWrap::RequestWrap(Request *request)
> -	: request_(request)
> +RequestWrap::RequestWrap(std::unique_ptr<Request> request)
> +	: request_(std::move(request))
>   {
>   }
>   
> @@ -74,8 +74,6 @@ RequestWrap::~RequestWrap()
>   		if (item.second)
>   			gst_buffer_unref(item.second);
>   	}
> -
> -	delete request_;
>   }
>   
>   void RequestWrap::attachBuffer(GstBuffer *buffer)
> @@ -164,7 +162,7 @@ GstLibcameraSrcState::requestCompleted(Request *request)
>   	std::unique_ptr<RequestWrap> wrap = std::move(requests_.front());
>   	requests_.pop();
>   
> -	g_return_if_fail(wrap->request_ == request);
> +	g_return_if_fail(wrap->request_.get() == request);
>   
>   	if ((request->status() == Request::RequestCancelled)) {
>   		GST_DEBUG_OBJECT(src_, "Request was cancelled");
> @@ -268,8 +266,7 @@ gst_libcamera_src_task_run(gpointer user_data)
>   	GstLibcameraSrc *self = GST_LIBCAMERA_SRC(user_data);
>   	GstLibcameraSrcState *state = self->state;
>   
> -	std::unique_ptr<Request> request = state->cam_->createRequest();
> -	auto wrap = std::make_unique<RequestWrap>(request.get());
> +	auto wrap = std::make_unique<RequestWrap>(state->cam_->createRequest());
>   	for (GstPad *srcpad : state->srcpads_) {
>   		GstLibcameraPool *pool = gst_libcamera_pad_get_pool(srcpad);
>   		GstBuffer *buffer;
> @@ -279,24 +276,23 @@ gst_libcamera_src_task_run(gpointer user_data)
>   						     &buffer, nullptr);
>   		if (ret != GST_FLOW_OK) {
>   			/*
> -			 * RequestWrap does not take ownership, and we won't be
> -			 * queueing this one due to lack of buffers.
> +			 * RequestWrap has ownership of the rquest, and we
> +			 * won't be queueing this one due to lack of buffers.
>   			 */
> -			request.reset();
> +			wrap.release();
>   			break;
>   		}
>   
>   		wrap->attachBuffer(buffer);
>   	}
>   
> -	if (request) {
> +	if (wrap) {
>   		GLibLocker lock(GST_OBJECT(self));
>   		GST_TRACE_OBJECT(self, "Requesting buffers");
> -		state->cam_->queueRequest(request.get());
> +		state->cam_->queueRequest(wrap->request_.get());


Did you mean

state->cam_->queueRequest(wrap->request_)

the wrapped request is not a unique_ptr here.


With this change, I tested both the success and the error path and it 
looks fine. However, I think there should be a printed message (maybe 
GST_WARNING) when we get in the error path, as there is no indicator now 
that buffer acquistion failed.


Regards,

Marian


>   		state->requests_.push(std::move(wrap));
>   
> -		/* The request will be deleted in the completion handler. */
> -		request.release();
> +		/* The RequestWrap will be deleted in the completion handler. */
>   	}
>   
>   	GstFlowReturn ret = GST_FLOW_OK;
>
>