[PATCH] drm/amdgpu: Use PSP to program IH_RB_CNTL_RING1/2 on SRIOV

Christian König ckoenig.leichtzumerken at gmail.com
Mon Jun 7 17:40:05 UTC 2021 

That won't work either.

We still need to initialize the control registers and tell the hardware 
that we have properly setup the ring buffers.

Just add the error message to psp_reg_program() instead of duplicating 
that over and over again.

Christian.

Am 07.06.21 um 19:33 schrieb Khaire, Rohit:
> [AMD Public Use]
>
> OK.
>
> I will just skip the function call for SRIOV and resend.
>
> Rohit
>
> -----Original Message-----
> From: Koenig, Christian <Christian.Koenig at amd.com>
> Sent: June 7, 2021 12:42 PM
> To: Kuehling, Felix <Felix.Kuehling at amd.com>; Khaire, Rohit <Rohit.Khaire at amd.com>; amd-gfx at lists.freedesktop.org; Deucher, Alexander <Alexander.Deucher at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com>; Deng, Emily <Emily.Deng at amd.com>; Liu, Monk <Monk.Liu at amd.com>; Zhou, Peng Ju <PengJu.Zhou at amd.com>; Chen, Horace <Horace.Chen at amd.com>
> Cc: Ming, Davis <Davis.Ming at amd.com>
> Subject: Re: [PATCH] drm/amdgpu: Use PSP to program IH_RB_CNTL_RING1/2 on SRIOV
>
> Ah, good point. In this case we should probably rather save than sorry.
>
> Then I suggest to clean up this patch, repeating the psp_reg_program() and error message is pretty horrible coding style.
>
> Christian.
>
> Am 07.06.21 um 18:36 schrieb Felix Kuehling:
>> With SRIOV, the interrupt routing is setup by the hypervisor driver.
>> We need the secondary IH rings in case the hypervisor enabled
>> rerouting of page fault interrupts. I'm not sure what the hypervisor driver does today.
>>
>> Regards,
>>     Felix
>>
>>
>> Am 2021-06-07 um 12:29 p.m. schrieb Christian König:
>>> That's a workaround for bare metal and as far as I know doesn't apply
>>> to SRIOV.
>>>
>>> We only need the additional IH rings for page fault handling or log
>>> handling and as far as I know that is incompatible with SRIOV for the
>>> moment. But Felix might have some more updates on this.
>>>
>>> So as long as we don't support that under SRIOV we don't need this
>>> patch either.
>>>
>>> Christian.
>>>
>>> Am 07.06.21 um 17:59 schrieb Khaire, Rohit:
>>>> [AMD Public Use]
>>>>
>>>> The hash is 5ea6f9c
>>>>
>>>> Rohit
>>>>
>>>> -----Original Message-----
>>>> From: Koenig, Christian <Christian.Koenig at amd.com>
>>>> Sent: June 7, 2021 11:58 AM
>>>> To: Khaire, Rohit <Rohit.Khaire at amd.com>;
>>>> amd-gfx at lists.freedesktop.org; Deucher, Alexander
>>>> <Alexander.Deucher at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com>;
>>>> Deng, Emily <Emily.Deng at amd.com>; Liu, Monk <Monk.Liu at amd.com>;
>>>> Zhou, Peng Ju <PengJu.Zhou at amd.com>; Chen, Horace
>>>> <Horace.Chen at amd.com>
>>>> Cc: Ming, Davis <Davis.Ming at amd.com>
>>>> Subject: Re: [PATCH] drm/amdgpu: Use PSP to program
>>>> IH_RB_CNTL_RING1/2 on SRIOV
>>>>
>>>> Do you have the hash for this commit?
>>>>
>>>> Thanks,
>>>> Christian.
>>>>
>>>> Am 07.06.21 um 17:30 schrieb Khaire, Rohit:
>>>>> [AMD Public Use]
>>>>>
>>>>> We don't need RING1 and RING2 functionality for SRIOV afaik.
>>>>>
>>>>> But looking at the description of the original commit message it
>>>>> affects RING0 too?
>>>>>
>>>>> " drm/amdgpu: add timeout flush mechanism to update wptr for self
>>>>> interrupt (v2)
>>>>>
>>>>> outstanding log reaches threshold will trigger IH ring1/2's wptr
>>>>> reported, that will avoid generating interrupts to ring0 too frequent.
>>>>> But if ring1/2's wptr hasn't been increased for a long time, the
>>>>> outstanding log can't reach threshold so that driver can't get
>>>>> latest wptr info and miss some interrupts."
>>>>>
>>>>> Rohit
>>>>>
>>>>> -----Original Message-----
>>>>> From: Koenig, Christian <Christian.Koenig at amd.com>
>>>>> Sent: June 7, 2021 10:31 AM
>>>>> To: Khaire, Rohit <Rohit.Khaire at amd.com>;
>>>>> amd-gfx at lists.freedesktop.org; Deucher, Alexander
>>>>> <Alexander.Deucher at amd.com>; Zhang, Hawking
>>>>> <Hawking.Zhang at amd.com>; Deng, Emily <Emily.Deng at amd.com>; Liu,
>>>>> Monk <Monk.Liu at amd.com>; Zhou, Peng Ju <PengJu.Zhou at amd.com>; Chen,
>>>>> Horace <Horace.Chen at amd.com>
>>>>> Cc: Ming, Davis <Davis.Ming at amd.com>
>>>>> Subject: Re: [PATCH] drm/amdgpu: Use PSP to program
>>>>> IH_RB_CNTL_RING1/2 on SRIOV
>>>>>
>>>>> Why are the ring 1&2 enabled on SRIOV in the first place?
>>>>>
>>>>> Christian.
>>>>>
>>>>> Am 07.06.21 um 16:23 schrieb Rohit Khaire:
>>>>>> This is similar to IH_RB_CNTL programming in
>>>>>> navi10_ih_toggle_ring_interrupts
>>>>>>
>>>>>> Signed-off-by: Rohit Khaire <rohit.khaire at amd.com>
>>>>>> ---
>>>>>>       drivers/gpu/drm/amd/amdgpu/navi10_ih.c | 20
>>>>>> ++++++++++++++++++--
>>>>>>       1 file changed, 18 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
>>>>>> b/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
>>>>>> index eac564e8dd52..e41188c04846 100644
>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
>>>>>> @@ -120,11 +120,27 @@ force_update_wptr_for_self_int(struct
>>>>>> amdgpu_device *adev,
>>>>>>           ih_rb_cntl = REG_SET_FIELD(ih_rb_cntl, IH_RB_CNTL_RING1,
>>>>>>                          RB_USED_INT_THRESHOLD, threshold);
>>>>>>       -    WREG32_SOC15(OSSSYS, 0, mmIH_RB_CNTL_RING1, ih_rb_cntl);
>>>>>> +    if (amdgpu_sriov_vf(adev) &&
>>>>>> amdgpu_sriov_reg_indirect_ih(adev)) {
>>>>>> +        if (psp_reg_program(&adev->psp, PSP_REG_IH_RB_CNTL_RING1,
>>>>>> ih_rb_cntl)) {
>>>>>> +            DRM_ERROR("PSP program IH_RB_CNTL_RING1 failed!\n");
>>>>>> +            return;
>>>>>> +        }
>>>>>> +    } else {
>>>>>> +        WREG32_SOC15(OSSSYS, 0, mmIH_RB_CNTL_RING1, ih_rb_cntl);
>>>>>> +    }
>>>>>> +
>>>>>>           ih_rb_cntl = RREG32_SOC15(OSSSYS, 0, mmIH_RB_CNTL_RING2);
>>>>>>           ih_rb_cntl = REG_SET_FIELD(ih_rb_cntl, IH_RB_CNTL_RING2,
>>>>>>                          RB_USED_INT_THRESHOLD, threshold);
>>>>>> -    WREG32_SOC15(OSSSYS, 0, mmIH_RB_CNTL_RING2, ih_rb_cntl);
>>>>>> +    if (amdgpu_sriov_vf(adev) &&
>>>>>> amdgpu_sriov_reg_indirect_ih(adev)) {
>>>>>> +        if (psp_reg_program(&adev->psp, PSP_REG_IH_RB_CNTL_RING2,
>>>>>> ih_rb_cntl)) {
>>>>>> +            DRM_ERROR("PSP program IH_RB_CNTL_RING2 failed!\n");
>>>>>> +            return;
>>>>>> +        }
>>>>>> +    } else {
>>>>>> +        WREG32_SOC15(OSSSYS, 0, mmIH_RB_CNTL_RING2, ih_rb_cntl);
>>>>>> +    }
>>>>>> +
>>>>>>           WREG32_SOC15(OSSSYS, 0, mmIH_CNTL2, ih_cntl);
>>>>>>       }
>>>>>>       
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

More information about the amd-gfx mailing list