Re: [PATCH v2] drm/msm/dpu: make "vblank timeout" more useful

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, 21 Feb 2024 at 01:04, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
>
>
>
> On 2/20/2024 2:42 PM, Dmitry Baryshkov wrote:
> > On Wed, 21 Feb 2024 at 00:40, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
> >>
> >>
> >>
> >> On 2/19/2024 3:52 AM, Dmitry Baryshkov wrote:
> >>> On Wed, 14 Feb 2024 at 22:36, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 2/14/2024 11:20 AM, Dmitry Baryshkov wrote:
> >>>>> On Wed, 14 Feb 2024 at 20:02, Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:
> >>>>>>
> >>>>>>
> >>>>>>
> >>>>>> On 2/8/2024 6:50 AM, Dmitry Baryshkov wrote:
> >>>>>>> We have several reports of vblank timeout messages. However after some
> >>>>>>> debugging it was found that there might be different causes to that.
> >>>>>>> To allow us to identify the DPU block that gets stuck, include the
> >>>>>>> actual CTL_FLUSH value into the timeout message and trigger the devcore
> >>>>>>> snapshot capture.
> >>>>>>>
> >>>>>>> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx>
> >>>>>>> ---
> >>>>>>> Changes in v2:
> >>>>>>> - Added a call to msm_disp_snapshot_state() to trigger devcore dump
> >>>>>>>       (Abhinav)
> >>>>>>> - Link to v1: https://lore.kernel.org/r/20240106-fd-dpu-debug-timeout-v1-1-6d9762884641@xxxxxxxxxx
> >>>>>>> ---
> >>>>>>>      drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c | 3 ++-
> >>>>>>>      1 file changed, 2 insertions(+), 1 deletion(-)
> >>>>>>>
> >>>>>>> diff --git a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
> >>>>>>> index d0f56c5c4cce..a8d6165b3c0a 100644
> >>>>>>> --- a/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
> >>>>>>> +++ b/drivers/gpu/drm/msm/disp/dpu1/dpu_encoder_phys_vid.c
> >>>>>>> @@ -489,7 +489,8 @@ static int dpu_encoder_phys_vid_wait_for_commit_done(
> >>>>>>>                  (hw_ctl->ops.get_flush_register(hw_ctl) == 0),
> >>>>>>>                  msecs_to_jiffies(50));
> >>>>>>>          if (ret <= 0) {
> >>>>>>> -             DPU_ERROR("vblank timeout\n");
> >>>>>>> +             DPU_ERROR("vblank timeout: %x\n", hw_ctl->ops.get_flush_register(hw_ctl));
> >>>>>>> +             msm_disp_snapshot_state(phys_enc->parent->dev);
> >>>>>>
> >>>>>>
> >>>>>> There is no rate limiting in this piece of code unfortunately. So this
> >>>>>> will flood the number of snapshots.
> >>>>>
> >>>>> Well... Yes and no. The devcoredump will destroy other snapshots if
> >>>>> there is a pending one. So only the console will be flooded and only
> >>>>> in case when MSM_DISP_SNAPSHOT_DUMP_IN_CONSOLE is enabled.
> >>>>>
> >>>>
> >>>> Yes, true but at the same time this makes it hard to capture a good dump
> >>>> as potentially every vblank you could timeout so this destroy/create
> >>>> cycle wont end.
> >>>
> >>> Excuse me, maybe I miss something. On the first timeout the snapshot
> >>> is created. It is held by the kernel until it is fully read out from
> >>> the userspace. Other snapshots will not interfere with this snapshot.
> >>>
> >>
> >> For every new snapshot a new devcoredump device will be created which
> >> should remain till it has been read. But now this will be created every
> >> blank. IMO, this is really too much data for no reason.
> >
> > No-no-no. If there is a devcoredump for a device, the next one will
> > not be created. See dev_coredumpm().
> > So all the snapshots will be created and then destroyed immediately.
> >
>
> hmm ... I have certainly seen devcd_count go higher than one (but not
> more than 2). I am wondering whether this was because of some race
> condition of the previous destroy / new create.
>
> But anyway, this part is clear now. thanks.
>
> >>
> >> Subsequent vblank timeouts are not going to give any new information
> >> compared to the existing snapshot of the first vblank timeout thats why
> >> we should just create the snapshot when the first error happens and stop.
> >>
> >> For other frame done timeouts, infact subsequent timeouts without any
> >> sort of recovery in between are quite misleading because hardware was
> >> already not able to fetch the previous frame so it will most likely not
> >> fetch the next one either till it has recovered. Typically thats why
> >> these vblank timeouts happen in a flurry as the hardware never really
> >> recovered from the first timeout.
> >>
> >>> Or are you worried that snapshotting takes time, so taking a snapshot
> >>> will also interfere with the vblank timings for the next vblank?
> >>>
> >>
> >> Yes this is another point.
> >
>
> snapshots will still be captured every vblank timeout and reading
> through the entire DPU reg space every vblank timeout is certainly
> something we can avoid.

Ack.

-- 
With best wishes
Dmitry



[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux