Re: [PATCH 2/2] drm/amdgpu: enable Navi retry fault wptr overflow

[Felix Kuehling <felix.kuehling@amd.com>]

Am 2021-11-23 um 11:03 a.m. schrieb Philip Yang:
> If xnack is on, VM retry fault interrupt send to IH ring1, and ring1
> will be full quickly. IH cannot receive other interrupts, this causes
> deadlock if migrating buffer using sdma and waiting for sdma done
> while handling retry fault.
>
> Remove VMC from IH storm client, enable ring1 write pointer
> overflow, then IH will drop retry fault interrupts and be able to receive
> other interrupts while driver is handling retry fault.
>
> IH ring1 write pointer doesn't writeback to memory by IH, and ring1
> write pointer recorded by self-irq is not updated, so always read
> the latest ring1 write pointer from register.
>
> Signed-off-by: Philip Yang <Philip.Yang@amd.com>

Looks like the same thing you did in this commit for Vega:

    b672cb1eee59 drm/amdgpu: enable retry fault wptr overflow

This series is

Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>


> ---
>  drivers/gpu/drm/amd/amdgpu/navi10_ih.c | 33 ++++++++++----------------
>  1 file changed, 12 insertions(+), 21 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/navi10_ih.c b/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
> index dafad6030947..38241cf0e1f1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
> +++ b/drivers/gpu/drm/amd/amdgpu/navi10_ih.c
> @@ -276,10 +276,8 @@ static int navi10_ih_enable_ring(struct amdgpu_device *adev,
>  	tmp = navi10_ih_rb_cntl(ih, tmp);
>  	if (ih == &adev->irq.ih)
>  		tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, RPTR_REARM, !!adev->irq.msi_enabled);
> -	if (ih == &adev->irq.ih1) {
> -		tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, WPTR_OVERFLOW_ENABLE, 0);
> +	if (ih == &adev->irq.ih1)
>  		tmp = REG_SET_FIELD(tmp, IH_RB_CNTL, RB_FULL_DRAIN_ENABLE, 1);
> -	}
>  
>  	if (amdgpu_sriov_vf(adev) && amdgpu_sriov_reg_indirect_ih(adev)) {
>  		if (psp_reg_program(&adev->psp, ih_regs->psp_reg_id, tmp)) {
> @@ -320,7 +318,6 @@ static int navi10_ih_irq_init(struct amdgpu_device *adev)
>  {
>  	struct amdgpu_ih_ring *ih[] = {&adev->irq.ih, &adev->irq.ih1, &adev->irq.ih2};
>  	u32 ih_chicken;
> -	u32 tmp;
>  	int ret;
>  	int i;
>  
> @@ -364,15 +361,6 @@ static int navi10_ih_irq_init(struct amdgpu_device *adev)
>  	adev->nbio.funcs->ih_doorbell_range(adev, ih[0]->use_doorbell,
>  					    ih[0]->doorbell_index);
>  
> -	tmp = RREG32_SOC15(OSSSYS, 0, mmIH_STORM_CLIENT_LIST_CNTL);
> -	tmp = REG_SET_FIELD(tmp, IH_STORM_CLIENT_LIST_CNTL,
> -			    CLIENT18_IS_STORM_CLIENT, 1);
> -	WREG32_SOC15(OSSSYS, 0, mmIH_STORM_CLIENT_LIST_CNTL, tmp);
> -
> -	tmp = RREG32_SOC15(OSSSYS, 0, mmIH_INT_FLOOD_CNTL);
> -	tmp = REG_SET_FIELD(tmp, IH_INT_FLOOD_CNTL, FLOOD_CNTL_ENABLE, 1);
> -	WREG32_SOC15(OSSSYS, 0, mmIH_INT_FLOOD_CNTL, tmp);
> -
>  	pci_set_master(adev->pdev);
>  
>  	/* enable interrupts */
> @@ -421,12 +409,19 @@ static u32 navi10_ih_get_wptr(struct amdgpu_device *adev,
>  	u32 wptr, tmp;
>  	struct amdgpu_ih_regs *ih_regs;
>  
> -	wptr = le32_to_cpu(*ih->wptr_cpu);
> -	ih_regs = &ih->ih_regs;
> +	if (ih == &adev->irq.ih) {
> +		/* Only ring0 supports writeback. On other rings fall back
> +		 * to register-based code with overflow checking below.
> +		 */
> +		wptr = le32_to_cpu(*ih->wptr_cpu);
>  
> -	if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
> -		goto out;
> +		if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
> +			goto out;
> +	}
>  
> +	ih_regs = &ih->ih_regs;
> +
> +	/* Double check that the overflow wasn't already cleared. */
>  	wptr = RREG32_NO_KIQ(ih_regs->ih_rb_wptr);
>  	if (!REG_GET_FIELD(wptr, IH_RB_WPTR, RB_OVERFLOW))
>  		goto out;
> @@ -514,15 +509,11 @@ static int navi10_ih_self_irq(struct amdgpu_device *adev,
>  			      struct amdgpu_irq_src *source,
>  			      struct amdgpu_iv_entry *entry)
>  {
> -	uint32_t wptr = cpu_to_le32(entry->src_data[0]);
> -
>  	switch (entry->ring_id) {
>  	case 1:
> -		*adev->irq.ih1.wptr_cpu = wptr;
>  		schedule_work(&adev->irq.ih1_work);
>  		break;
>  	case 2:
> -		*adev->irq.ih2.wptr_cpu = wptr;
>  		schedule_work(&adev->irq.ih2_work);
>  		break;
>  	default: break;