On Mon, Oct 28, 2024 at 11:36:15AM +0100, Konrad Dybcio wrote:
> On 28.10.2024 11:27 AM, Dmitry Baryshkov wrote:
> > On Mon, 28 Oct 2024 at 12:08, Akhil P Oommen <quic_akhilpo@xxxxxxxxxxx> wrote:
> >>
> >> On 10/28/2024 1:56 PM, Dmitry Baryshkov wrote:
> >>> On Sun, Oct 27, 2024 at 11:35:47PM +0530, Akhil P Oommen wrote:
> >>>> Clang-19 and above sometimes end up with multiple copies of the large
> >>>> a6xx_hfi_msg_bw_table structure on the stack. The problem is that
> >>>> a6xx_hfi_send_bw_table() calls a number of device specific functions to
> >>>> fill the structure, but these create another copy of the structure on
> >>>> the stack which gets copied to the first.
> >>>>
> >>>> If the functions get inlined, that busts the warning limit:
> >>>>
> >>>> drivers/gpu/drm/msm/adreno/a6xx_hfi.c:631:12: error: stack frame size (1032) exceeds limit (1024) in 'a6xx_hfi_send_bw_table' [-Werror,-Wframe-larger-than]
> >>>>
> >>>> Fix this by kmalloc-ating struct a6xx_hfi_msg_bw_table instead of using
> >>>> the stack. Also, use this opportunity to skip re-initializing this table
> >>>> to optimize gpu wake up latency.
> >>>>
> >>>> Cc: Arnd Bergmann <arnd@xxxxxxxxxx>
> >>>>
> >>>> Signed-off-by: Akhil P Oommen <quic_akhilpo@xxxxxxxxxxx>
> >>>> ---
> >>>> drivers/gpu/drm/msm/adreno/a6xx_gmu.h | 1 +
> >>>> drivers/gpu/drm/msm/adreno/a6xx_hfi.c | 34 ++++++++++++++++++++++------------
> >>>> 2 files changed, 23 insertions(+), 12 deletions(-)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
> >>>> index 94b6c5cab6f4..b4a79f88ccf4 100644
> >>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
> >>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
> >>>> @@ -99,6 +99,7 @@ struct a6xx_gmu {
> >>>> struct completion pd_gate;
> >>>>
> >>>> struct qmp *qmp;
> >>>> + struct a6xx_hfi_msg_bw_table *bw_table;
> >>>> };
> >>>>
> >>>> static inline u32 gmu_read(struct a6xx_gmu *gmu, u32 offset)
> >>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
> >>>> index cdb3f6e74d3e..55e51c81be1f 100644
> >>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
> >>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.c
> >>>> @@ -630,32 +630,42 @@ static void a6xx_build_bw_table(struct a6xx_hfi_msg_bw_table *msg)
> >>>>
> >>>> static int a6xx_hfi_send_bw_table(struct a6xx_gmu *gmu)
> >>>> {
> >>>> - struct a6xx_hfi_msg_bw_table msg = { 0 };
> >>>> + struct a6xx_hfi_msg_bw_table *msg;
> >>>> struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
> >>>> struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
> >>>>
> >>>> + if (gmu->bw_table)
> >>>> + goto send;
> >>>> +
> >>>> + msg = devm_kzalloc(gmu->dev, sizeof(*msg), GFP_KERNEL);
> >>>
> >>> Is it necessary after being sent? Isn't it better to just kzalloc() it
> >>> and then kfree() it at the end of the function?
> >>
> >> Keeping it around will help to cut down unnecessary work during
> >> subsequent gpu wake ups.
> >
> > Then, I'd say, it is better to make it a part of the a6xx_gpu struct.
>
> I think a6xx_gmu makes more logical sense here.
>
> FWIW, the driver allocates both _gmu and _gpu for all GPUs regardless
Hmm, are we expected to handle / perform BW requests in case of GMU-less
devices?
--
With best wishes
Dmitry
