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

Wed Dec 1 11:04:54 PST 2021

Am Mittwoch, dem 01.12.2021 um 12:52 -0600 schrieb Adam Ford:
> On Wed, Dec 1, 2021 at 12:37 PM Lucas Stach <l.stach at 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 at 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 at 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.

That's not going to work with all the MIX domains. The handshake
requires some clocks to be enabled in the blk-ctrl (the secondary clock
gates in the blk-ctrl) to work properly. This is only done by the blk-
ctrl driver _after_ the GPC bus domain is powered up, so you can not
wait for the handshake to complete inside the GPC power up routine.

Regards,
Lucas