Re: [RFC PATCH 1/8] vb2: push the mmap semaphore down to __buf_prepare()

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Hans,

On Friday 22 November 2013 10:02:49 Hans Verkuil wrote:
> On 11/21/2013 08:04 PM, Laurent Pinchart wrote:
> > On Thursday 21 November 2013 16:21:59 Hans Verkuil wrote:
> >> From: Hans Verkuil <hans.verkuil@cisco.com>
> >> 
> >> Rather than taking the mmap semaphore at a relatively high-level
> >> function, push it down to the place where it is really needed.
> >> 
> >> It was placed in vb2_queue_or_prepare_buf() to prevent racing with other
> >> vb2 calls, however, I see no way that any race can happen.
> > 
> > What about the following scenario ? Both QBUF calls are performed on the
> > same buffer.
> > 
> > 	CPU 0							CPU 1
> > 	-------------------------------------------------------------------------
> > 	QBUF								QBUF
> > 		locks the queue mutex				waits for the queue mutex
> > 	vb2_qbuf
> > 	vb2_queue_or_prepare_buf
> > 	__vb2_qbuf
> > 		checks vb->state, calls
> > 	__buf_prepare
> > 	call_qop(q, wait_prepare, q);
> > 		unlocks the queue mutex
> > 		
> > 										locks the queue mutex
> > 									vb2_qbuf
> > 									vb2_queue_or_prepare_buf
> > 									__vb2_qbuf
> > 										checks vb->state, calls
> > 									__buf_prepare
> > 									call_qop(q, wait_prepare, q);
> > 										unlocks the queue mutex
> > 									queue the buffer, set buffer
> > 									 state to queue
> > 	
> > 	queue the buffer, set buffer
> > 	 state to queue
> > 
> > We would thus end up queueing the buffer twice. The vb->state check needs
> > to be performed after the brief release of the queue mutex.
> 
> Good point, I hadn't thought about that scenario. However, using mmap_sem to
> introduce a large critical section just to protect against state changes is
> IMHO not the right approach. Why not introduce a VB2_BUF_STATE_PREPARING
> state?

Note that we use the queue mutex to do so, not mmap_sem. The problem is that 
we can't release the queue mutex in the middle of a critical section without 
risking being preempted by another task. Introducing a new state might be 
possible if it effectively breaks the critical section in two independent 
parts.

> That's set at the start of __buf_prepare while the queue mutex is still
> held, and which prevents other threads of queuing the same buffer again. If
> the prepare fails, then the state is reverted back to DEQUEUED.
> 
> __fill_v4l2_buffer() will handle the PREPARING state as if it was the
> DEQUEUED state.
> 
> What do you think?

I'll have to review that in details given the potential complexity of locking 
issues :-) I'm not opposed to the idea, if it works I believe we should do it.

-- 
Regards,

Laurent Pinchart