Re: [PATCH 2/6] drm/etnaviv: determine product, customer and eco id

[Lucas Stach]

On Mo, 2020-01-06 at 12:40 +0100, Christian Gmeiner wrote:
> Hi Lucas
> 
> Am Mo., 6. Jan. 2020 um 12:22 Uhr schrieb Lucas Stach <l.stach@xxxxxxxxxxxxxx>:
> > On Mo, 2020-01-06 at 11:57 +0100, Christian Gmeiner wrote:
> > > Hi Lucas
> > > 
> > > Am Mo., 6. Jan. 2020 um 11:03 Uhr schrieb Lucas Stach <l.stach@xxxxxxxxxxxxxx>:
> > > > On Do, 2020-01-02 at 11:02 +0100, Christian Gmeiner wrote:
> > > > > They will be used for extended HWDB support. The eco id logic was taken
> > > > > from galcore kernel driver sources.
> > > > > 
> > > > > Signed-off-by: Christian Gmeiner <christian.gmeiner@xxxxxxxxx>
> > > > > ---
> > > > >  drivers/gpu/drm/etnaviv/etnaviv_gpu.c | 17 +++++++++++++++++
> > > > >  drivers/gpu/drm/etnaviv/etnaviv_gpu.h |  6 +++---
> > > > >  2 files changed, 20 insertions(+), 3 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> > > > > index d47d1a8e0219..253301be9e95 100644
> > > > > --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> > > > > +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> > > > > @@ -321,6 +321,18 @@ static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
> > > > >               gpu->identity.varyings_count -= 1;
> > > > >  }
> > > > > 
> > > > > +static void etnaviv_hw_eco_id(struct etnaviv_gpu *gpu)
> > > > > +{
> > > > > +     const u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
> > > > > +     gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
> > > > > +
> > > > > +     if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617))
> > > > > +             gpu->identity.eco_id = 1;
> > > > > +
> > > > > +     if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511))
> > > > > +             gpu->identity.eco_id = 1;
> > > > 
> > > > I'm not sure if those two checks warrant a separate function. Maybe
> > > > just place them besides the other ID fixups?
> > > > 
> > > 
> > > This is almost a 1:1 copy of _GetEcoID(..) but will try to move the fixups.
> > > 
> > > 
> > > > > +}
> > > > > +
> > > > >  static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
> > > > >  {
> > > > >       u32 chipIdentity;
> > > > > @@ -362,6 +374,8 @@ static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
> > > > >                       }
> > > > >               }
> > > > > 
> > > > > +             gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
> > > > > +
> > > > >               /*
> > > > >                * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
> > > > >                * reality it's just a re-branded GC3000. We can identify this
> > > > > @@ -375,6 +389,9 @@ static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
> > > > >               }
> > > > >       }
> > > > > 
> > > > > +     etnaviv_hw_eco_id(gpu);
> > > > > +     gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
> > > > 
> > > > I don't like this scattering of identity register reads. Please move
> > > > all of those reads to the else clause where we currently read
> > > > model/rev. I doubt that the customer ID register is available on the
> > > > really early cores, that only have the VIVS_HI_CHIP_IDENTITY register.
> > > > 
> > > 
> > > There is feature bit for it: chipMinorFeatures5_HAS_PRODUCTID
> > > Will change the code to make use of it. Shall I still put it in the
> > > else clause then?
> > 
> > If there's a feature bit we need to move the read toward the end of the
> > function, as we currently read the features as the last step in the
> > hw_identify.
> > 
> > But then I'm not sure if the HAS_PRODUCTID feature bit is correct. At
> > least wumpus' gpus_comparison says that none of the known <= GC3000
> > cores has it set, which seems... suspicious.
> > 
> 
> Hmm... what if I just mimic what is done here?
> https://github.com/etnaviv/vivante_kernel_drivers/blob/master/imx8_v6.2.3.129602/hal/kernel/arch/gc_hal_kernel_hardware.c#L220
> 
> Unconditionally read the product id at the end of the else clause. The
> same is done in the stm32 galcore kernel driver.

If we read it unconditionally, just move it to the start of the else
clause, next to the model/rev reads and be done with it.

Regards,
Lucas