Re: [RFC PATCH v3 4/6] media: tegra: Add Tegra210 Video input driver

[Sowjanya Komatineni]

On 2/16/20 11:54 AM, Sowjanya Komatineni wrote:
> On 2/16/20 3:03 AM, Hans Verkuil wrote:
> > External email: Use caution opening links or attachments
> >
> >
> > On 2/14/20 7:23 PM, Sowjanya Komatineni wrote:
> > > Tegra210 contains a powerful Video Input (VI) hardware controller
> > > which can support up to 6 MIPI CSI camera sensors.
> > >
> > > Each Tegra CSI port can be one-to-one mapped to VI channel and can
> > > capture from an external camera sensor connected to CSI or from
> > > built-in test pattern generator.
> > >
> > > Tegra210 supports built-in test pattern generator from CSI to VI.
> > >
> > > This patch adds a V4L2 media controller and capture driver support
> > > for Tegra210 built-in CSI to VI test pattern generator.
> > >
> > > Signed-off-by: Sowjanya Komatineni <skomatineni@xxxxxxxxxx>
> > > ---
> > > Â drivers/staging/media/KconfigÂÂÂÂÂÂÂÂÂÂÂÂÂ |ÂÂÂ 2 +
> > > Â drivers/staging/media/MakefileÂÂÂÂÂÂÂÂÂÂÂÂ |ÂÂÂ 1 +
> > > Â drivers/staging/media/tegra/KconfigÂÂÂÂÂÂÂ |ÂÂ 10 +
> > > Â drivers/staging/media/tegra/MakefileÂÂÂÂÂÂ |ÂÂÂ 8 +
> > > Â drivers/staging/media/tegra/TODOÂÂÂÂÂÂÂÂÂÂ |ÂÂ 10 +
> > > Â drivers/staging/media/tegra/tegra-common.h |Â 239 +++++++
> > > Â drivers/staging/media/tegra/tegra-csi.cÂÂÂ |Â 374 ++++++++++
> > > Â drivers/staging/media/tegra/tegra-csi.hÂÂÂ |Â 115 ++++
> > > Â drivers/staging/media/tegra/tegra-vi.cÂÂÂÂ | 1019 
> > > ++++++++++++++++++++++++++++
> > > Â drivers/staging/media/tegra/tegra-vi.hÂÂÂÂ |ÂÂ 79 +++
> > >  drivers/staging/media/tegra/tegra-video.c | 118 ++++
> > >  drivers/staging/media/tegra/tegra-video.h | 32 +
> > > Â drivers/staging/media/tegra/tegra210.cÂÂÂÂ |Â 767 
> > > +++++++++++++++++++++
> > > Â drivers/staging/media/tegra/tegra210.hÂÂÂÂ |Â 190 ++++++
> > > Â 14 files changed, 2964 insertions(+)
> > > Â create mode 100644 drivers/staging/media/tegra/Kconfig
> > > Â create mode 100644 drivers/staging/media/tegra/Makefile
> > > Â create mode 100644 drivers/staging/media/tegra/TODO
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-common.h
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-csi.c
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-csi.h
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-vi.c
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-vi.h
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-video.c
> > > Â create mode 100644 drivers/staging/media/tegra/tegra-video.h
> > > Â create mode 100644 drivers/staging/media/tegra/tegra210.c
> > > Â create mode 100644 drivers/staging/media/tegra/tegra210.h
> > <snip>
> >
> > > +/*
> > > + * videobuf2 queue operations
> > > + */
> > > +static int tegra_channel_queue_setup(struct vb2_queue *vq,
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ unsigned int *nbuffers,
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ unsigned int *nplanes,
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ unsigned int sizes[],
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ struct device *alloc_devs[])
> > > +{
> > > +ÂÂÂÂ struct tegra_vi_channel *chan = vb2_get_drv_priv(vq);
> > > +
> > > +ÂÂÂÂ if (*nplanes)
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂ return sizes[0] < chan->format.sizeimage ? -EINVAL : 0;
> > > +
> > > +ÂÂÂÂ *nplanes = 1;
> > > +ÂÂÂÂ sizes[0] = chan->format.sizeimage;
> > > +ÂÂÂÂ alloc_devs[0] = chan->vi->dev;
> > > +
> > > +ÂÂÂÂ /*
> > > +ÂÂÂÂÂ * allocate min 3 buffers in queue to avoid race between DMA
> > > +ÂÂÂÂÂ * writes and userspace reads.
> > > +ÂÂÂÂÂ */
> > > +ÂÂÂÂ if (*nbuffers < 3)
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂ *nbuffers = 3;
> > First of all, don't check this here, instead set the struct vb2_queue 
> > field
> > 'min_buffers_needed' to 3 instead.
> >
> > But the reason given for this check is peculiar: there should not be any
> > race at all. Usually the reason for requiring a specific minimum 
> > number of
> > buffers is that the DMA engine needs at least 2 buffers before it can 
> > start
> > streaming: it can't give back a buffer to userspace (vb2_buffer_done())
> > unless there is a second buffer it can start to capture to next. So 
> > for many
> > DMA implementations you need a minimum of 2 buffers: two buffers for the
> > DMA engine, one buffer being processed by userspace.
> >
> > If the driver is starved of buffers it will typically keep capturing to
> > the last buffer until a new buffer is queued.
> >
> > In any case, once the driver releases a buffer via vb2_buffer_done() the
> > buffer memory is no longer owned by the driver.
> >
> > To be precise, buffer ownership is as follows:
> >
> > userspace -> VIDIOC_QBUF -> vb2 -> buf_queue -> driver -> 
> > vb2_buffer_done() -> vb2 -> VIDIOC_DQBUF -> userspace
> >
> > (vb2 == videobuf2 framework)
> >
> > Note that vb2 never touches the buffer memory.
> >
> > So if you get a race condition in this driver, then there is something
> > strange going on. It looks like vb2_buffer_done() is called while DMA is
> > still ongoing, or because the driver really needs to keep one buffer
> > available at all times.
> >
> > Regards,
> >
> > ÂÂÂÂÂÂÂÂ Hans
>
> Thanks Hans.
>
> On running v4l2-compliance streaming tests for longer run, I noticed 
> kernel reporting unable to write to read-only memory and with debugs I 
> observed when this error was reported, I see 2 buffers queued and both 
> using same address.
>
> for first buffer capture start thread initiates capture and wakes done 
> thread to wait for memory write ack and once its done buffer is 
> released to user space but I see upon buffer released to user space 
> immediate next buffer capture single shot gets issued (as soon as 
> single shot is issued frame capture data is written to memory by DMA) 
> and I see this kernel error of unable to write to read-only memory.
>
> This error happens rare and happens on long run and all the times of 
> repro's, I see when other thread releases buffer immediate I see 
> single shot gets issued as 2 buffers are queued up at the same time 
> with same DMA address.
Just to be clear, I meant all the times when kernel reports error unable 
to write to read-only memory, I see 2 buffers gets queued and as the 
capture start thread and done thread are parallel and when capture 
thread wakes done thread on receiving FS event, done thread for waiting 
for memory write happens parallel to next frame capture and I see while 
vb2_buffer_done happens in done thread next frame single shot has been 
issues by capture start thread in parallel when it hits this error.

> With using minimum 3 buffers, this issue doesnt happen at all from 
> almost 72 hours of testing.
>
>
> Will try with setting vb2 queue field min_buffers_needed as 3 instead 
> of adding check in queue setup.

> > > +
> > > +ÂÂÂÂ return 0;
> > > +}