AMD General
Regards,
Prike
> -----Original Message-----
> From: amd-gfx <[email protected]> On Behalf Of Jesse
> Zhang
> Sent: Monday, June 1, 2026 1:49 PM
> To: [email protected]
> Cc: Deucher, Alexander <[email protected]>; Koenig, Christian
> <[email protected]>; Zhang, Jesse(Jie) <[email protected]>;
> Rastogi, Manu <[email protected]>; Deucher, Alexander
> <[email protected]>; Zhang, Jesse(Jie) <[email protected]>
> Subject: [PATCH v2 14/42] drm/amdgpu/gfx12: Refactor compute pipe reset and
> add
> HQD cleanup
>
> Refactor gfx_v12_0_reset_compute_pipe() to accept explicit me, pipe, and queue
> parameters instead of deriving them from the ring structure. This enables the
> function to be used in generic pipe reset flows.
>
> Introduce gfx_v12_0_clear_hqds_on_mec_pipe() to properly clear
> CP_HQD_ACTIVE and CP_HQD_DEQUEUE_REQUEST for all queues on a given
> MEC pipe while the pipe reset is asserted, ensuring the HQDs are torn down
> correctly before deasserting reset.
>
> Switch the KCQ reset path to use the common MEC pipe reset helper
> amdgpu_gfx_mec_pipe_reset_run(), which coordinates the reset sequence
> including
> KFD suspend/resume to avoid conflicts with user mode queues.
>
> v2: just update the sequence (Alex)
> v3: directly clear ACTIVE and DEQUEUE_REQUEST (Shaoyun Liu)
>
> Suggested-by: Manu Rastogi <[email protected]>
> Suggested-by: Alex Deucher <[email protected]>
> Signed-off-by: Jesse Zhang <[email protected]>
> Signed-off-by: Alex Deucher <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c | 115 +++++++++++++++----------
> 1 file changed, 69 insertions(+), 46 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c
> b/drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c
> index 4be650ce1fba..f7783129ecc8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v12_0.c
> @@ -5309,10 +5309,30 @@ static int gfx_v12_0_reset_kgq(struct amdgpu_ring
> *ring,
> return amdgpu_ring_reset_helper_end(ring, timedout_fence); }
>
> -static int gfx_v12_0_reset_compute_pipe(struct amdgpu_ring *ring)
> +/*
> + * With MEC pipe reset asserted, clear CP_HQD_ACTIVE /
> +CP_HQD_DEQUEUE_REQUEST for
> + * every queue on (me, pipe). HQDs must be torn down while pipe reset
> +stays
> + * asserted; only then clear the pipe reset bit.
> + * Caller must hold adev->srbm_mutex.
> + */
> +static void gfx_v12_0_clear_hqds_on_mec_pipe(struct amdgpu_device *adev, u32
> me,
> + u32 pipe)
> {
> - struct amdgpu_device *adev = ring->adev;
> - uint32_t reset_pipe = 0, clean_pipe = 0;
> + unsigned int q;
> + int j;
> +
> + for (q = 0; q < adev->gfx.mec.num_queue_per_pipe; q++) {
> + soc24_grbm_select(adev, me, pipe, q, 0);
> + /* Start from a clean HQD dequeue state before forcing HQD
> inactive.
> */
> + WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 0);
> + WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 0);
> + }
> +}
> +
> +static int gfx_v12_0_reset_compute_pipe(struct amdgpu_device *adev,
> + u32 me, u32 pipe, u32 queue)
> +{
> + uint32_t reset_val, clean_val;
> int r = 0;
>
> if (!gfx_v12_pipe_reset_support(adev))
> @@ -5320,75 +5340,78 @@ static int gfx_v12_0_reset_compute_pipe(struct
> amdgpu_ring *ring)
>
> gfx_v12_0_set_safe_mode(adev, 0);
> mutex_lock(&adev->srbm_mutex);
> - soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
> -
> - reset_pipe = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL);
> - clean_pipe = reset_pipe;
> -
> + soc24_grbm_select(adev, me, pipe, queue, 0);
> if (adev->gfx.rs64_enable) {
> - switch (ring->pipe) {
> + reset_val = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL);
> + clean_val = reset_val;
> +
> + switch (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);
> + reset_val = REG_SET_FIELD(reset_val,
> CP_MEC_RS64_CNTL,
> + MEC_PIPE0_RESET, 1);
> + clean_val = REG_SET_FIELD(clean_val,
> 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);
> + reset_val = REG_SET_FIELD(reset_val,
> CP_MEC_RS64_CNTL,
> + MEC_PIPE1_RESET, 1);
> + clean_val = REG_SET_FIELD(clean_val,
> 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);
> + reset_val = REG_SET_FIELD(reset_val,
> CP_MEC_RS64_CNTL,
> + MEC_PIPE2_RESET, 1);
> + clean_val = REG_SET_FIELD(clean_val,
> 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);
> + reset_val = REG_SET_FIELD(reset_val,
> CP_MEC_RS64_CNTL,
> + MEC_PIPE3_RESET, 1);
> + clean_val = REG_SET_FIELD(clean_val,
> 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);
> + WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, reset_val);
> + gfx_v12_0_clear_hqds_on_mec_pipe(adev, me, pipe);
Same as patch#13 which requires removing the queue selecting, other than that
the patch is Reviewed-by: Prike Liang <[email protected]>
> + soc24_grbm_select(adev, me, pipe, queue, 0);
> + WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, clean_val);
> r = (RREG32_SOC15(GC, 0, regCP_MEC_RS64_INSTR_PNTR) <<
> 2) -
> RS64_FW_UC_START_ADDR_LO;
> } else {
> - switch (ring->pipe) {
> + reset_val = RREG32_SOC15(GC, 0, regCP_MEC_CNTL);
> + clean_val = reset_val;
> +
> + switch (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);
> + reset_val = REG_SET_FIELD(reset_val, CP_MEC_CNTL,
> + MEC_ME1_PIPE0_RESET, 1);
> + clean_val = REG_SET_FIELD(clean_val, 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);
> + reset_val = REG_SET_FIELD(reset_val, CP_MEC_CNTL,
> + MEC_ME1_PIPE1_RESET, 1);
> + clean_val = REG_SET_FIELD(clean_val, CP_MEC_CNTL,
> + MEC_ME1_PIPE1_RESET, 0);
> break;
> default:
> - break;
> + break;
> }
> - WREG32_SOC15(GC, 0, regCP_MEC_CNTL, reset_pipe);
> - WREG32_SOC15(GC, 0, regCP_MEC_CNTL, clean_pipe);
> - /* Doesn't find the F32 MEC instruction pointer register, and
> suppose
> - * the driver won't run into the F32 mode.
> - */
> +
> + WREG32_SOC15(GC, 0, regCP_MEC_CNTL, reset_val);
> + gfx_v12_0_clear_hqds_on_mec_pipe(adev, me, pipe);
> + soc24_grbm_select(adev, me, pipe, queue, 0);
> + WREG32_SOC15(GC, 0, regCP_MEC_CNTL, clean_val);
> }
>
> soc24_grbm_select(adev, 0, 0, 0, 0);
> mutex_unlock(&adev->srbm_mutex);
> gfx_v12_0_unset_safe_mode(adev, 0);
>
> - dev_info(adev->dev, "The ring %s pipe resets: %s\n", ring->name,
> - r == 0 ? "successfully" : "failed");
> - /* Need the ring test to verify the pipe reset result.*/
> + dev_dbg(adev->dev, "MEC pipe me%u pipe%u queue%u resets to MEC FW
> start PC: %s\n",
> + me, pipe, queue, r == 0 ? "successfully" : "failed");
> return 0;
> }
>
> @@ -5405,7 +5428,7 @@ static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring,
> r = amdgpu_mes_reset_legacy_queue(ring->adev, ring, vmid, use_mmio, 0);
> if (r) {
> dev_warn(adev->dev, "fail(%d) to reset kcq and try pipe
> reset\n", r);
> - r = gfx_v12_0_reset_compute_pipe(ring);
> + r = gfx_v12_0_reset_compute_pipe(adev, ring->me, ring->pipe,
> +ring->queue);
> if (r)
> return r;
> }
> --
> 2.49.0
