Date:   Thu, 9 Jun 2022 11:12:03 -0700
From:   Rob Clark <robdclark@...il.com>
To:     Akhil P Oommen <quic_akhilpo@...cinc.com>
Cc:     Douglas Anderson <dianders@...omium.org>,
        Jordan Crouse <jordan@...micpenguin.net>,
        Abhinav Kumar <quic_abhinavk@...cinc.com>,
        Bjorn Andersson <bjorn.andersson@...aro.org>,
        Chia-I Wu <olvaffe@...il.com>,
        Dan Carpenter <dan.carpenter@...cle.com>,
        Daniel Vetter <daniel@...ll.ch>,
        David Airlie <airlied@...ux.ie>,
        Dmitry Baryshkov <dmitry.baryshkov@...aro.org>,
        Eric Anholt <eric@...olt.net>,
        Jonathan Marek <jonathan@...ek.ca>,
        Sean Paul <sean@...rly.run>, Wang Qing <wangqing@...o.com>,
        Yangtao Li <tiny.windzz@...il.com>,
        dri-devel <dri-devel@...ts.freedesktop.org>,
        freedreno <freedreno@...ts.freedesktop.org>,
        linux-arm-msm <linux-arm-msm@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2] drm/msm: Grab the GPU runtime in a6xx routines, not
 the GMU one

On Thu, Jun 9, 2022 at 11:04 AM Akhil P Oommen <quic_akhilpo@...cinc.com> wrote:
>
> On 6/9/2022 10:17 PM, Douglas Anderson wrote:
> > >From testing on sc7180-trogdor devices, reading the GMU registers
> > needs the GMU clocks to be enabled. Those clocks get turned on in
> > a6xx_gmu_resume(). Confusingly enough, that function is called as a
> > result of the runtime_pm of the GPU "struct device", not the GMU
> > "struct device".
> >
> > Let's grab a reference to the correct device. Incidentally, this makes
> > us match the a5xx routine more closely.
> >
> > This is easily shown to fix crashes that happen if we change the GPU's
> > pm_runtime usage to not use autosuspend. It's also believed to fix
> > some long tail GPU crashes even with autosuspend.
> >
> > NOTE: the crashes I've seen were fixed by _only_ fixing
> > a6xx_gpu_busy(). However, I believe that the same arguments should be
> > made to a6xx_gmu_set_freq() so I've fixed that case too. To make that
> > fix clean, we've moved the pm runtime grabbing into the GPU file.
> >
> > As a bonus fix with this change, we change the pm_runtime get
> > functions to check for <= 0 instead of ==. This handles the case where
> > pm_runtime is disabled.
> >
> > Fixes: eadf79286a4b ("drm/msm: Check for powered down HW in the devfreq callbacks")
> > Signed-off-by: Douglas Anderson <dianders@...omium.org>
> > ---
> >
> > Changes in v2:
> > - Move the set_freq runtime pm grab to the GPU file.
> > - Use <= for the pm_runtime test, not ==.
> >
> >   drivers/gpu/drm/msm/adreno/a6xx_gmu.c |  9 ---------
> >   drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 13 +++++++++++--
> >   2 files changed, 11 insertions(+), 11 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> > index 9f76f5b15759..2410815e77b4 100644
> > --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> > +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
> > @@ -125,17 +125,9 @@ void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
> >
> >       trace_msm_gmu_freq_change(gmu->freq, perf_index);
> >
> > -     /*
> > -      * This can get called from devfreq while the hardware is idle. Don't
> > -      * bring up the power if it isn't already active
> > -      */
> > -     if (pm_runtime_get_if_in_use(gmu->dev) == 0)
> > -             return;
> > -
> >       if (!gmu->legacy) {
> >               a6xx_hfi_set_freq(gmu, perf_index);
> >               dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
> > -             pm_runtime_put(gmu->dev);
> >               return;
> >       }
> >
> > @@ -159,7 +151,6 @@ void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
> >               dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
> >
> >       dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
> > -     pm_runtime_put(gmu->dev);
> >   }
> >
> >   unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
> > diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > index 42ed9a3c4905..54efd9b76ea6 100644
> > --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.c
> > @@ -1659,7 +1659,7 @@ static u64 a6xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
> >       *out_sample_rate = 19200000;
> >
> >       /* Only read the gpu busy if the hardware is already active */
> > -     if (pm_runtime_get_if_in_use(a6xx_gpu->gmu.dev) == 0)
> > +     if (pm_runtime_get_if_in_use(&gpu->pdev->dev) <= 0)
> You are changing the behavior here when CONFIG_PM is not enabled.

I'd guess the odds of anything working with PM=n are low

BR,
-R

> -Akhil.
> >               return 0;
> >
> >       busy_cycles = gmu_read64(&a6xx_gpu->gmu,
> > @@ -1667,7 +1667,7 @@ static u64 a6xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate)
> >                       REG_A6XX_GMU_CX_GMU_POWER_COUNTER_XOCLK_0_H);
> >
> >
> > -     pm_runtime_put(a6xx_gpu->gmu.dev);
> > +     pm_runtime_put(&gpu->pdev->dev);
> >
> >       return busy_cycles;
> >   }
> > @@ -1677,9 +1677,18 @@ static void a6xx_gpu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
> >       struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
> >       struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
> >
> > +     /*
> > +      * This can get called from devfreq while the hardware is idle. Don't
> > +      * bring up the power if it isn't already active
> > +      */
> > +     if (pm_runtime_get_if_in_use(&gpu->pdev->dev) <= 0)
> > +             return;
> > +
> >       mutex_lock(&a6xx_gpu->gmu.lock);
> >       a6xx_gmu_set_freq(gpu, opp);
> >       mutex_unlock(&a6xx_gpu->gmu.lock);
> > +
> > +     pm_runtime_put(&gpu->pdev->dev);
> >   }
> >
> >   static struct msm_gem_address_space *
>

Powered by blists - more mailing lists