On Sun, Jul 2, 2023 at 7:34 AM Dmitry Baryshkov
<dmitry.baryshkov@xxxxxxxxxx> wrote:
>
> On 02/07/2023 17:31, Rob Clark wrote:
> > On Sat, Jul 1, 2023 at 5:24 PM Dmitry Baryshkov
> > <dmitry.baryshkov@xxxxxxxxxx> wrote:
> >>
> >> On Sat, 1 Jul 2023 at 18:50, Rob Clark <robdclark@xxxxxxxxx> wrote:
> >>>
> >>> On Fri, Jun 30, 2023 at 4:12 PM Konrad Dybcio <konrad.dybcio@xxxxxxxxxx> wrote:
> >>>>
> >>>> Recently, a WARN_ON() was introduced to ensure that revn is filled before
> >>>> adreno_is_aXYZ is called. This however doesn't work very well when revn is
> >>>> 0 by design (such as for A635). Fill it in as a stopgap solution for
> >>>> -fixes.
> >>>>
> >>>> Fixes: cc943f43ece7 ("drm/msm/adreno: warn if chip revn is verified before being set")
> >>>> Signed-off-by: Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> >>>> ---
> >>>> Changes in v2:
> >>>> - add fixes
> >>>> - Link to v1: https://lore.kernel.org/r/20230628-topic-a635-v1-1-5056e09c08fb@xxxxxxxxxx
> >>>> ---
> >>>> drivers/gpu/drm/msm/adreno/adreno_device.c | 1 +
> >>>> 1 file changed, 1 insertion(+)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/msm/adreno/adreno_device.c b/drivers/gpu/drm/msm/adreno/adreno_device.c
> >>>> index cb94cfd137a8..8ea7eae9fc52 100644
> >>>> --- a/drivers/gpu/drm/msm/adreno/adreno_device.c
> >>>> +++ b/drivers/gpu/drm/msm/adreno/adreno_device.c
> >>>> @@ -345,6 +345,7 @@ static const struct adreno_info gpulist[] = {
> >>>> .address_space_size = SZ_16G,
> >>>> }, {
> >>>> .rev = ADRENO_REV(6, 3, 5, ANY_ID),
> >>>> + .revn = 635,
> >>>> .fw = {
> >>>> [ADRENO_FW_SQE] = "a660_sqe.fw",
> >>>> [ADRENO_FW_GMU] = "a660_gmu.bin",
> >>>>
> >>>
> >>> hmm, I realized a problem with this, it would change what
> >>> MSM_PARAM_GPU_ID and more importantly MSM_PARAM_CHIP_ID return.. The
> >>> former should be "harmless", although it isn't a good idea for uabi
> >>> changes to be a side effect of a fix. The latter is more problematic.
> >>
> >> I'd say MSM_PARAM_GPU_ID is broken for 635 anyway (won't it return 0
> >> in this case)?
> >> So the new value should be correct.
> >
> > no, it is very much intentional that GPU_ID returns 0 for newer GPUs,
> > userspace should be matching on CHIP_ID. (Also, we should be moving
> > away from trying to infer generation/etc from CHIP_ID.. userspace is
> > farther ahead of the kernel on this.)
>
> Thanks for the explanation. So in theory we can change this to always
> return 0? Or must we keep it to keep UABI / compatibility?
>
> I'm trying to understand if we can drop revn at all.
I'm not sure how likely it is that someone would try a very old mesa
with a new kernel, but that combo would break if we dropped revn
completely and returned 0 for GPU_ID. But I would like to move away
from using revn internally. (And probably move away from the patch-id
wildcard matching, and instead just explicitly list all known patch-id
values so we can start treating CHIP_ID as just an opaque identifier.)
BR,
-R
>
> >
> >> But more importantly, why are we exporting speedbin in
> >> MSM_PARAM_CHIP_ID only if there is no revn? And why are we exporting
> >> the speedbin at all as a part of CHIP_ID?
> >
> > Basically just being paranoid about not changing uabi. It probably
> > would be ok to export the speedbin for all, but I'd have to double
> > check mesa version history.
>
> Thanks!
>
> >
> > BR,
> > -R
> >
> >>>
> >>> I think I'm leaning more towards reverting commit cc943f43ece7
> >>> ("drm/msm/adreno: warn if chip revn is verified before being set") for
> >>> -fixes. I'm still thinking about options for a longer term fix.
> >>>
> >>> BR,
> >>> -R
> >>>
> >>>
> >>>> ---
> >>>> base-commit: 5c875096d59010cee4e00da1f9c7bdb07a025dc2
> >>>> change-id: 20230628-topic-a635-1b3c2c987417
> >>>>
> >>>> Best regards,
> >>>> --
> >>>> Konrad Dybcio <konrad.dybcio@xxxxxxxxxx>
> >>>>
> >>
> >>
> >>
> >> --
> >> With best wishes
> >> Dmitry
>
> --
> With best wishes
> Dmitry
>
