"media: vb2: add (un)prepare_streaming queue ops" causes vb2_queue->streaming to be set earlier ?
Hi Hans,
I just fixed a crash in the out of tree driver which is caused by:
a10b21532574 ("media: vb2: add (un)prepare_streaming queue ops")
I know that breaking out of tree code is fine / not our problem, but I wonder if this maybe affects some in tree code too ?
The problem with the out of tree driver is that it used vb2_queue->streaming inside the buf_queue() callback to determine if the buffer was being pre-queued (from vb2_start_streaming() calling __enqueue_in_driver()) before the start_streaming queue-op is called.
Or if it was being queued after the start_streaming queue-op has been called.
With vb2_queue->streaming now being set before the
__enqueue_in_driver() calls in vb2_start_streaming() this
broke and it went down the route to see if there was space available in the hw-queue which NULL pointer deref-ed because the hw-queue had not been setup yet (more or less).
This is fixed now as far as the ipu6 driver is concerned, but I wonder if we may not have in tree drivers making similar assumptions about vb2_queue->streaming in their buf_queue() callbacks ?
Regards,
Hans
Hi Hans,
Quoted 24 lines · Hans de Goede
Hi Hans,
I just fixed a crash in the out of tree driver which is caused by:
a10b21532574("media: vb2: add (un)prepare_streaming queue ops")I know that breaking out of tree code is fine / not our problem, but I wonder if this maybe affects some in tree code too ?
The problem with the out of tree driver is that it used
vb2_queue->streaminginside thebuf_queue()callback to determine if the buffer was being pre-queued (fromvb2_start_streaming()calling__enqueue_in_driver()) before thestart_streamingqueue-op is called.Or if it was being queued after the
start_streamingqueue-op has been called.With
vb2_queue->streamingnow being set before the
__enqueue_in_driver()calls invb2_start_streaming()this
broke and it went down the route to see if there was space available in the hw-queue which NULL pointer deref-ed because the hw-queue had not been setup yet (more or less).
Quoted 4 lines
This is fixed now as far as the ipu6 driver is concerned, but I wonder if we may not have in tree drivers making similar assumptions about
vb2_queue->streamingin theirbuf_queue()callbacks ?
I checked, and there is nobody else that's doing that.
It's definitely wrong to use q->streaming like that, it was never meant as a check for whether or not the buffer was pre-queued.
Thanks for the heads-up, though! It is a real change in vb2 behavior that I hadn't realized could cause problems.
It's a good change, since having buf_queue called when q->streaming is false makes no sense. However, you can use q->start_streaming_called instead if for some reason a driver needs to know.
Regards,
Hans
Quoted 27 lines · Hans Verkuil
Hi Hans,
Quoted 24 lines · Hans de Goede
Hi Hans,
I just fixed a crash in the out of tree driver which is caused by:
a10b21532574("media: vb2: add (un)prepare_streaming queue ops")I know that breaking out of tree code is fine / not our problem, but I wonder if this maybe affects some in tree code too ?
The problem with the out of tree driver is that it used
vb2_queue->streaminginside thebuf_queue()callback to determine if the buffer was being pre-queued (fromvb2_start_streaming()calling__enqueue_in_driver()) before thestart_streamingqueue-op is called.Or if it was being queued after the
start_streamingqueue-op has been called.With
vb2_queue->streamingnow being set before the
__enqueue_in_driver()calls invb2_start_streaming()this
broke and it went down the route to see if there was space available in the hw-queue which NULL pointer deref-ed because the hw-queue had not been setup yet (more or less).
Quoted 6 lines
Quoted 4 lines
This is fixed now as far as the ipu6 driver is concerned, but I wonder if we may not have in tree drivers making similar assumptions about
vb2_queue->streamingin theirbuf_queue()callbacks ?I checked, and there is nobody else that's doing that.
Unfortunately, commit a10b21532574 ("media: vb2: add (un)prepare_streaming queue ops") broke the vsp1 driver, which oopses when starting streaming due to the issue described above.
Quoted 2 lines
It's definitely wrong to use
q->streaminglike that, it was never meant as a check for whether or not the buffer was pre-queued.
It's done though :-) The vsp1 .buf_queue() handler reads as
static void vsp1_video_buffer_queue(struct vb2_buffer *vb)
{
struct vb2_v4l2_buffer *vbuf = to_vb2_v4l2_buffer(vb);
struct vsp1_video *video = vb2_get_drv_priv(vb->vb2_queue);
struct vsp1_pipeline *pipe = video->rwpf->entity.pipe;
struct vsp1_vb2_buffer *buf = to_vsp1_vb2_buffer(vbuf);
unsigned long flags;
bool empty;
spin_lock_irqsave(&video->irqlock, flags);
empty = list_empty(&video->irqqueue);
list_add_tail(&buf->queue, &video->irqqueue);
spin_unlock_irqrestore(&video->irqlock, flags);
if (!empty)
return;
spin_lock_irqsave(&pipe->irqlock, flags);
video->rwpf->mem = buf->mem;
pipe->buffers_ready |= 1 << video->pipe_index;
if (vb2_is_streaming(&video->queue) &&
vsp1_pipeline_ready(pipe))
vsp1_video_pipeline_run(pipe);
spin_unlock_irqrestore(&pipe->irqlock, flags);
} Quoted 6 lines
Thanks for the heads-up, though! It is a real change in vb2 behavior that I hadn't realized could cause problems.
It's a good change, since having
buf_queuecalled whenq->streamingis false makes no sense. However, you can useq->start_streaming_calledinstead if for some reason a driver needs to know.
This fixes the issue in the vsp1 driver. I'll submit a patch, but I'm worried about other breakages thought, *lots* of drivers call vb2_is_streaming(), including in the .buf_queue() handler. Even the m2m core does so. I'm not sure we'll be able to fix all that in time for v6.2.
Signature · 3 lines
Regards, Laurent Pinchart