[PATCH 3/3] media: staging: rkisp1: params: in 'stop_streaming' don't release the lock while returning buffers

[Laurent Pinchart]

Hi Tomasz,

On Thu, Aug 13, 2020 at 02:50:17PM +0200, Tomasz Figa wrote:
> On Thu, Aug 13, 2020 at 12:53 PM Laurent Pinchart wrote:
> > On Thu, Aug 13, 2020 at 12:44:35PM +0200, Dafna Hirschfeld wrote:
> > > Am 26.06.20 um 18:58 schrieb Robin Murphy:
> > > > On 2020-06-26 16:59, Tomasz Figa wrote:
> > > >> On Fri, Jun 26, 2020 at 5:48 PM Dafna Hirschfeld wrote:
> > > >>> On 26.06.20 16:03, Tomasz Figa wrote:
> > > >>>> On Fri, Jun 26, 2020 at 3:32 PM Robin Murphy wrote:
> > > >>>>> On 2020-06-25 18:42, Dafna Hirschfeld wrote:
> > > >>>>>> In the stop_streaming callback 'rkisp1_params_vb2_stop_streaming'
> > > >>>>>> there is no need to release the lock 'config_lock' and then acquire
> > > >>>>>> it again at each iteration when returning all buffers.
> > > >>>>>> This is because the stream is about to end and there is no need
> > > >>>>>> to let the isr access a buffer.
> > > >>>>>>
> > > >>>>>> Signed-off-by: Dafna Hirschfeld <dafna.hirschfeld at collabora.com>
> > > >>>>>> ---
> > > >>>>>>     drivers/staging/media/rkisp1/rkisp1-params.c | 7 +------
> > > >>>>>>     1 file changed, 1 insertion(+), 6 deletions(-)
> > > >>>>>>
> > > >>>>>> diff --git a/drivers/staging/media/rkisp1/rkisp1-params.c b/drivers/staging/media/rkisp1/rkisp1-params.c
> > > >>>>>> index bf006dbeee2d..5169b02731f1 100644
> > > >>>>>> --- a/drivers/staging/media/rkisp1/rkisp1-params.c
> > > >>>>>> +++ b/drivers/staging/media/rkisp1/rkisp1-params.c
> > > >>>>>> @@ -1488,19 +1488,13 @@ static void rkisp1_params_vb2_stop_streaming(struct vb2_queue *vq)
> > > >>>>>>         /* stop params input firstly */
> > > >>>>>>         spin_lock_irqsave(&params->config_lock, flags);
> > > >>>>>>         params->is_streaming = false;
> > > >>>>>> -     spin_unlock_irqrestore(&params->config_lock, flags);
> > > >>>>>>
> > > >>>>>>         for (i = 0; i < RKISP1_ISP_PARAMS_REQ_BUFS_MAX; i++) {
> > > >>>>>> -             spin_lock_irqsave(&params->config_lock, flags);
> > > >>>>>>                 if (!list_empty(&params->params)) {
> > > >>>>>>                         buf = list_first_entry(&params->params,
> > > >>>>>>                                                struct rkisp1_buffer, queue);
> > > >>>>>>                         list_del(&buf->queue);
> > > >>>>>> -                     spin_unlock_irqrestore(&params->config_lock,
> > > >>>>>> -                                            flags);
> > > >>>>>>                 } else {
> > > >>>>>> -                     spin_unlock_irqrestore(&params->config_lock,
> > > >>>>>> -                                            flags);
> > > >>>>>>                         break;
> > > >>>>>>                 }
> > > >>>>>
> > > >>>>> Just skimming through out of idle curiosity I was going to comment that
> > > >>>>> if you end up with this pattern:
> > > >>>>>
> > > >>>>>           if (!x) {
> > > >>>>>                   //do stuff
> > > >>>>>           } else {
> > > >>>>>                   break;
> > > >>>>>           }
> > > >>>>>
> > > >>>>> it would be better as:
> > > >>>>>
> > > >>>>>           if (x)
> > > >>>>>                   break;
> > > >>>>>           //do stuff
> > > >>>>>
> > > >>>>> However I then went and looked at the whole function and frankly it's a
> > > >>>>> bit of a WTF. As best I could decipher, this whole crazy loop appears to
> > > >>>>> be a baroque reinvention of:
> > > >>>>>
> > > >>>>>           list_for_each_entry_safe(&params->params, ..., buf) {
> > > >>>>>                   list_del(&buf->queue);
> > > >>>>>                   vb2_buffer_done(&buf->vb.vb2_buf, VB2_BUF_STATE_ERROR);
> > > >>>>>           }
> > > >>> Hi, indeed this is a much simpler implementation, greping 'return_all_buffers'
> > > >>> I see that many drivers implement it this way.
> > > >>> thanks!
> > > >>>
> > > >>>>>
> > > >>>>> (assuming from context that the list should never contain more than
> > > >>>>> RKISP1_ISP_PARAMS_REQ_BUFS_MAX entries in the first place)
> > > >>>>
> > > >>>> Or if we want to avoid disabling the interrupts for the whole
> > > >>>> iteration, we could use list_splice() to move all the entries of
> > > >>>
> > > >>> But this code runs when userspace asks to stop the streaming so I don't
> > > >>> think it is important at that stage to allow the interrupts.
> > > >>
> > > >> It's generally a good practice to reduce the time spent with
> > > >> interrupts disabled. Disabling the interrupts prevents the system from
> > > >> handling external events, including timer interrupts, and scheduling
> > > >> higher priority tasks, including real time ones. How much the system
> > > >> runs with interrupts disabled is one of the factors determining the
> > > >> general system latency.
> > > >
> > > > Right, with the way we handle interrupt affinity on Arm an IRQ can't
> > > > target multiple CPUs in hardware, so any time spent with IRQs
> > > > disabled might be preventing other devices' interrupts from being
> > > > taken even if they're not explicitly affine to the current CPU.
> > > >
> > > > Now that I've looked, it appears that vb2_buffer_done() might end up
> > > > performing a DMA sync on the buffers, which, if it has to do
> > > > order-of-megabytes worth of cache maintenance for large frames, is
> > > > the kind of relatively slow operation that really doesn't want to be
> > > > done with IRQs disabled (or under a lock at all, ideally) unless it
> > > > absolutely *has* to be. If the lock is only needed here to protect
> > > > modifications to the params list itself, then moving the whole list
> > > > at once to do the cleanup "offline" sounds like a great idea to me.
> >
> > Ouch.
> >
> > > ok, that might be a problem in v4l2 in general since vb2_buffer_done
> > > is actually often used inside an irq handler
> >
> > Correct. The DMA sync should be moved to DQBUF time, there shouldn't be
> > any reason to do it in the IRQ handler. I thought this had already been
> > fixed :-(
> 
> For reference, there was a patch [1] proposed, but it moved the
> synchronization to a wrong place in the sequence, already after the
> .buf_finish queue callback, ending up breaking the drivers which need
> to access the buffer contents there.
> 
> [1] https://patchwork.linuxtv.org/project/linux-media/patch/1494255810-12672-4-git-send-email-sakari.ailus@linux.intel.com/

I think we need to fix the drivers. We just can't do cache sync in IRQ
context by default because a few drivers need to access the buffer
contents. Those drivers should instead deffer access to a work queue,
and sync explicitly. We could possibly provide helpers for that, making
it transparent if a queue flag is set.

-- 
Regards,

Laurent Pinchart