Re: [RFC V2 0/2] arm64: imx8mm: Enable Hantro VPUs

[Adam Ford <aford173@gmail.com>]

On Wed, Dec 1, 2021 at 12:37 PM Lucas Stach <l.stach@pengutronix.de> wrote:
>
> Am Mittwoch, dem 01.12.2021 um 10:16 -0800 schrieb Tim Harvey:
> > On Wed, Dec 1, 2021 at 9:32 AM Lucas Stach <l.stach@pengutronix.de> wrote:
> > >
> > > Hi Tim,
> > >
> > > Am Mittwoch, dem 01.12.2021 um 09:23 -0800 schrieb Tim Harvey:
> > > > On Tue, Nov 30, 2021 at 5:33 PM Adam Ford <aford173@gmail.com> wrote:
> > > > >
> > > > > The i.MX8M has two Hantro video decoders, called G1 and G2 which appear
> > > > > to be related to the video decoders used on the i.MX8MQ, but because of
> > > > > how the Mini handles the power domains, the VPU driver does not need to
> > > > > handle all the functions, nor does it support the post-processor,
> > > > > so a new compatible flag is required.
> > > > >
> > > > > With the suggestion from Hans Verkuil, I was able to get the G2 splat to go away
> > > > > with changes to FORCE_MAX_ZONEORDER, but I found I could also set cma=512M, however
> > > > > it's unclear to me if that's an acceptable alternative.
> > > > >
> > > > > At the suggestion of Ezequiel Garcia and Nicolas Dufresne I have some
> > > > > results from Fluster. However, the G2 VPU appears to fail most tests.
> > > > >
> > > > > ./fluster.py run -dGStreamer-H.264-V4L2SL-Gst1.0
> > > > > Ran 90/135 tests successfully               in 76.431 secs
> > > > >
> > > > >  ./fluster.py run -d GStreamer-VP8-V4L2SL-Gst1.0
> > > > > Ran 55/61 tests successfully               in 21.454 secs
> > > > >
> > > > > ./fluster.py run -d GStreamer-VP9-V4L2SL-Gst1.0
> > > > > Ran 0/303 tests successfully               in 20.016 secs
> > > > >
> > > > > Each day seems to show more and more G2 submissions, and gstreamer seems to be
> > > > > still working on the VP9, so I am not sure if I should drop G2 as well.
> > > > >
> > > > >
> > > > > Adam Ford (2):
> > > > >   media: hantro: Add support for i.MX8M Mini
> > > > >   arm64: dts: imx8mm: Enable VPU-G1 and VPU-G2
> > > > >
> > > > >  arch/arm64/boot/dts/freescale/imx8mm.dtsi   | 41 +++++++++++++++
> > > > >  drivers/staging/media/hantro/hantro_drv.c   |  2 +
> > > > >  drivers/staging/media/hantro/hantro_hw.h    |  2 +
> > > > >  drivers/staging/media/hantro/imx8m_vpu_hw.c | 57 +++++++++++++++++++++
> > > > >  4 files changed, 102 insertions(+)
> > > > >
> > > >
> > > > Adam,
> > > >
> > > > That's for the patches!
> > > >
> > > > I tested just this series on top of v5.16-rc3 on an
> > > > imx8mm-venice-gw73xx-0x and found that if I loop fluster I can end up
> > > > getting a hang within 10 to 15 mins or so when imx8m_blk_ctrl_power_on
> > > > is called for VPUMIX pd :
> > > > while [ 1 ]; do uptime; ./fluster.py run -d GStreamer-VP8-V4L2SL-Gst1.0; done
> > > > ...
> > > > [  618.838436] imx-pgc imx-pgc-domain.6: failed to command PGC
> > > > [  618.844407] imx8m-blk-ctrl 38330000.blk-ctrl: failed to power up bus domain
> > > >
> > > > I added prints in imx_pgc_power_{up,down} and
> > > > imx8m_blk_ctrl_power_{on,off} to get some more context
> > > > ...
> > > > Ran 55/61 tests successfully               in 8.685 secs
> > > >  17:16:34 up 17 min,  0 users,  load average: 3.97, 2.11, 0.93
> > > > ********************************************************************************
> > > > ********************
> > > > Running test suite VP8-TEST-VECTORS with decoder GStreamer-VP8-V4L2SL-Gst1.0
> > > > Using 4 parallel job(s)
> > > > ********************************************************************************
> > > > ********************
> > > >
> > > > [TEST SUITE      ] (DECODER                    ) TEST VECTOR               ... R
> > > > ESULT
> > > > ----------------------------------------------------------------------
> > > > [ 1023.114806] imx8m_blk_ctrl_power_on vpublk-g1
> > > > [ 1023.119669] imx_pgc_power_up vpumix
> > > > [ 1023.124307] imx-pgc imx-pgc-domain.6: failed to command PGC
> > > > [ 1023.130006] imx8m-blk-ctrl 38330000.blk-ctrl: failed to power up bus domain
> > > >
> > > > While this wouldn't be an issue with this series it does indicate we
> > > > still have something racy in blk-ctrl. Can you reproduce this (and if
> > > > not what kernel are you based on)? Perhaps you or Lucas have some
> > > > ideas?
> > > >
> > > Did you have "[PATCH] soc: imx: gpcv2: Synchronously suspend MIX
> > > domains" applied when running those tests? It has only recently been
> > > picked up by Shawn and may have an influence on the bus domain
> > > behavior.
> > >
> >
> > Lucas,
> >
> > Good point. I did have that originally before I started pruning down
> > to the bare minimum to reproduce the issue.
> >
> > I added it back and now I have the following:
> > arm64: dts: imx8mm: Enable VPU-G1 and VPU-G2
> > media: hantro: Add support for i.MX8M Mini
> > soc: imx: gpcv2: keep i.MX8MM VPU-H1 bus clock active
> > soc: imx: gpcv2: Synchronously suspend MIX domains
> > Linux 5.16-rc3
> >
> > Here's the latest with that patch:
> > ...
> > [VP8-TEST-VECTORS] (GStreamer-VP8-V4L2SL-Gst1.0)
> > vp80-00-comprehensive-007 ... Success
> > [  316.632373] imx8m_blk_ctrl_power_off vpublk-g1
> > [  316.636908] imx_pgc_power_down vpu-g1
> > [  316.640983] imx_pgc_power_down vpumix
> > [  316.756869] imx8m_blk_ctrl_power_on vpublk-g1
> > [  316.761360] imx_pgc_power_up vpumix
> > [  316.765985] imx-pgc imx-pgc-domain.6: failed to command PGC
> > [  316.772743] imx8m-blk-ctrl 38330000.blk-ctrl: failed to power up bus domain
> > ^^^ hang
>
> Hm, I wonder if there's some broken error handling here somewhere, as a
> failure to power up a domain shouldn't lead to a hang.
>
> However, that doesn't explain why the PGC isn't completing the request.
> Can you try to extend the timeout some more. Even though I think that
> 1msec should already be generous. Can you dump the content of the
> GPC_PU_PGC_SW_PUP_REQ and GPC_A53_PU_PGC_PUP_STATUSn (all 3 of them)
> registers, when the failure condition is hit?

I submitted a patch [1]  to enable the commented-out if statement
which waits for the handshake if the gpc domain was invoked by the
blk-ctrl or we knew if the bus clock was operational.

I am not 100% certain it can work as-is with the vpumix, but based on
what I've seen from my testing, it's not hanging or causing errors.

[1] - https://lore.kernel.org/linux-arm-kernel/20211120194900.1309914-1-aford173@gmail.com/T/

I didn't have it applied to my latest RFC for the G1 and G2 because I
had not noticed a change in behavior one way or the other with that
patch.

adam
>
> Regards,
> Lucas
>