Am 09.04.2017 um 14:31 schrieb Xiangliang Yu:
> For SRIOV, multi vce engine will hang when encoding. Add VMHUB
> flush to workaround it, will continue to find the root cause later.
NAK, that won't work and can cause hangs as well.
The hang you experience could be a known issue with old libdrm, do you
have updated to the latest version?
Regards,
Christian.
>
> Signed-off-by: Xiangliang Yu <Xiangliang.Yu at amd.com>
> Signed-off-by: Frank Min <Frank.Min at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/vce_v4_0.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/vce_v4_0.c b/drivers/gpu/drm/amd/amdgpu/vce_v4_0.c
> index d3448d0..b4e36f3 100644
> --- a/drivers/gpu/drm/amd/amdgpu/vce_v4_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/vce_v4_0.c
> @@ -31,6 +31,7 @@
> #include "soc15d.h"
> #include "soc15_common.h"
> #include "mmsch_v1_0.h"
> +#include "gmc_v9_0.h"
>
> #include "vega10/soc15ip.h"
> #include "vega10/VCE/vce_4_0_offset.h"
> @@ -971,6 +972,7 @@ static void vce_v4_0_emit_vm_flush(struct amdgpu_ring *ring,
> uint32_t req = ring->adev->gart.gart_funcs->get_invalidate_req(vm_id);
> unsigned eng = ring->idx;
> unsigned i;
> + struct amdgpu_device *adev = ring->adev;
>
> pd_addr = pd_addr | 0x1; /* valid bit */
> /* now only use physical base address of PDE and valid */
> @@ -979,6 +981,15 @@ static void vce_v4_0_emit_vm_flush(struct amdgpu_ring *ring,
> for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
> struct amdgpu_vmhub *hub = &ring->adev->vmhub[i];
>
> + /* Workaround multi vce engine encoding hang issue */
> + if (i && amdgpu_sriov_vf(adev)) {
> + WREG32_NO_KIQ(hub->ctx0_ptb_addr_lo32 + vm_id * 2,
> + lower_32_bits(pd_addr));
> + WREG32_NO_KIQ(hub->ctx0_ptb_addr_hi32 + vm_id * 2,
> + upper_32_bits(pd_addr));
> + gmc_v9_0_flush_vmhub(adev, hub, eng, vm_id);
> + }
> +
> amdgpu_ring_write(ring, VCE_CMD_REG_WRITE);
> amdgpu_ring_write(ring,
> (hub->ctx0_ptb_addr_hi32 + vm_id * 2) << 2);
