[PATCH 2/4] drm/amdgpu/gfx11: Implement the GFX11 KCQ pipe reset
Alex Deucher
alexdeucher at gmail.com
Wed Feb 19 19:56:56 UTC 2025
On Sun, Jan 26, 2025 at 3:38 AM Prike Liang <Prike.Liang at amd.com> wrote:
>
> Implement the GFX11 compute pipe reset. As the GFX11 CPFW
> still hasn't fully supported pipe reset yet, therefore
> disable the KCQ pipe reset temporarily.
>
> Signed-off-by: Prike Liang <Prike.Liang at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c | 139 ++++++++++++++++++++++++-
> 1 file changed, 136 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> index 395872bb1401..39fd3ce67ec9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
> @@ -66,6 +66,7 @@
> #define regPC_CONFIG_CNTL_1_BASE_IDX 1
>
> static uint32_t me_fw_start_pc;
> +static uint32_t mec_fw_start_pc;
Same comment here as on patch 1.
Alex
>
> MODULE_FIRMWARE("amdgpu/gc_11_0_0_pfp.bin");
> MODULE_FIRMWARE("amdgpu/gc_11_0_0_me.bin");
> @@ -2937,6 +2938,7 @@ static void gfx_v11_0_config_gfx_rs64(struct amdgpu_device *adev)
>
> /* cache the firmware start PC */
> me_fw_start_pc = RREG32(SOC15_REG_OFFSET(GC, 0, regCP_GFX_RS64_INSTR_PNTR1));
> + mec_fw_start_pc = RREG32(SOC15_REG_OFFSET(GC, 0, regCP_MEC_RS64_INSTR_PNTR));
> }
>
> static int gfx_v11_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev)
> @@ -3812,7 +3814,7 @@ static int gfx_v11_0_cp_compute_load_microcode(struct amdgpu_device *adev)
> le32_to_cpup(fw_data + mec_hdr->jt_offset + i));
>
> WREG32_SOC15(GC, 0, regCP_MEC_ME1_UCODE_ADDR, adev->gfx.mec_fw_version);
> -
> + mec_fw_start_pc = RREG32(SOC15_REG_OFFSET(GC, 0, regCP_MEC1_INSTR_PNTR));
> return 0;
> }
>
> @@ -6764,6 +6766,135 @@ static int gfx_v11_0_reset_kgq(struct amdgpu_ring *ring, unsigned int vmid)
> return amdgpu_ring_test_ring(ring);
> }
>
> +static int gfx_v11_0_reset_compute_pipe(struct amdgpu_ring *ring)
> +{
> +
> + struct amdgpu_device *adev = ring->adev;
> + uint32_t reset_pipe = 0, clean_pipe = 0;
> + int r;
> +
> + if (!gfx_v11_pipe_reset_support(adev))
> + return -EOPNOTSUPP;
> +
> + gfx_v11_0_set_safe_mode(adev, 0);
> + mutex_lock(&adev->srbm_mutex);
> + soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
> +
> + reset_pipe = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL);
> + clean_pipe = reset_pipe;
> +
> + if (adev->gfx.rs64_enable) {
> +
> + switch (ring->pipe) {
> + case 0:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE0_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE0_RESET, 0);
> + break;
> + case 1:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE1_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE1_RESET, 0);
> + break;
> + case 2:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE2_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE2_RESET, 0);
> + break;
> + case 3:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE3_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_RS64_CNTL,
> + MEC_PIPE3_RESET, 0);
> + break;
> + default:
> + break;
> + }
> + WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, reset_pipe);
> + WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, clean_pipe);
> + r = RREG32_SOC15(GC, 0, regCP_MEC_RS64_INSTR_PNTR) - mec_fw_start_pc;
> + } else {
> + if (ring->me == 1) {
> + switch (ring->pipe) {
> + case 0:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE0_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE0_RESET, 0);
> + break;
> + case 1:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE1_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE1_RESET, 0);
> + break;
> + case 2:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE2_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE2_RESET, 0);
> + break;
> + case 3:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE3_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME1_PIPE3_RESET, 0);
> + break;
> + default:
> + break;
> + }
> + /* mec1 fw pc: CP_MEC1_INSTR_PNTR */
> + } else {
> + switch (ring->pipe) {
> + case 0:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE0_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE0_RESET, 0);
> + break;
> + case 1:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE1_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE1_RESET, 0);
> + break;
> + case 2:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE2_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE2_RESET, 0);
> + break;
> + case 3:
> + reset_pipe = REG_SET_FIELD(reset_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE3_RESET, 1);
> + clean_pipe = REG_SET_FIELD(clean_pipe, CP_MEC_CNTL,
> + MEC_ME2_PIPE3_RESET, 0);
> + break;
> + default:
> + break;
> + }
> + /* mec2 fw pc: CP:CP_MEC2_INSTR_PNTR */
> + }
> + WREG32_SOC15(GC, 0, regCP_MEC_CNTL, reset_pipe);
> + WREG32_SOC15(GC, 0, regCP_MEC_CNTL, clean_pipe);
> + r = RREG32(SOC15_REG_OFFSET(GC, 0, regCP_MEC1_INSTR_PNTR)) - mec_fw_start_pc;
> + }
> +
> + soc21_grbm_select(adev, 0, 0, 0, 0);
> + mutex_unlock(&adev->srbm_mutex);
> + gfx_v11_0_unset_safe_mode(adev, 0);
> +
> + dev_info(adev->dev,"The ring %s pipe resets to MEC FW start PC: %s\n", ring->name,
> + r == 0 ? "successfully" : "failed");
> + /*FIXME:Sometimes driver can't cache the MEC firmware start PC correctly, so the pipe
> + * reset status relies on the compute ring test result.
> + */
> + return 0;
> +}
> +
> static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring, unsigned int vmid)
> {
> struct amdgpu_device *adev = ring->adev;
> @@ -6774,8 +6905,10 @@ static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring, unsigned int vmid)
>
> r = amdgpu_mes_reset_legacy_queue(ring->adev, ring, vmid, true);
> if (r) {
> - dev_err(adev->dev, "reset via MMIO failed %d\n", r);
> - return r;
> + dev_warn(adev->dev,"fail(%d) to reset kcq and try pipe reset\n", r);
> + r = gfx_v11_0_reset_compute_pipe(ring);
> + if (r)
> + return r;
> }
>
> r = amdgpu_bo_reserve(ring->mqd_obj, false);
> --
> 2.34.1
>
More information about the amd-gfx
mailing list