gx12.1: Add ip dump support

Mukul Joshi Tue, 07 Jul 2026 11:09:23 -0700


On 7/7/2026 1:58 PM, Alex Deucher wrote:

On Tue, Jul 7, 2026 at 12:25 PM Mukul Joshi<[email protected]> wrote:

s/gx12.1/gfx12.1 in the patch headline.


With the fixes applied and this fixed, this patch is:

Reviewed-by: Mukul Joshi <[email protected]>


On 7/6/2026 11:01 AM, Alex Deucher wrote:

Ping?

On Mon, Jun 29, 2026 at 4:28 PM Alex Deucher<[email protected]> wrote:

Ping?

On Thu, Jun 18, 2026 at 5:04 PM Alex Deucher<[email protected]> wrote:

Add support for dumping IP register state.

Signed-off-by: Alex Deucher<[email protected]>
---
  drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c | 275 +++++++++++++++++++++++++
  1 file changed, 275 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c 
b/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
index 5bdf2512f1540..ec4346a6c2af2 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v12_1.c
@@ -69,6 +69,127 @@ MODULE_FIRMWARE("amdgpu/gc_12_1_0_rlc_1.bin");
          (SH_MEM_ALIGNMENT_MODE_UNALIGNED_GFX12_1_0 << 
SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT) | \
          (3 << SH_MEM_CONFIG__INITIAL_INST_PREFETCH__SHIFT))

+
+static const struct amdgpu_hwip_reg_entry gc_reg_list_12_1[] = {
+       SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2),
+       SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS3),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT1),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT2),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT3),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STALLED_STAT1),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STALLED_STAT1),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_BUSY_STAT),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT2),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT2),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_ERROR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HPD_STATUS0),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_BASE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_BASE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_RPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_WPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_CMD_BUFSZ),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_LO),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BUFSZ),
+       SOC15_REG_ENTRY_STR(GC, 0, regCPF_UTCL1_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regCPC_UTCL1_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regCPG_UTCL1_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS_2),
+       SOC15_REG_ENTRY_STR(GC, 0, regPA_CL_CNTL_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regRMI_UTCL1_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regSQC_CACHES),
+       SOC15_REG_ENTRY_STR(GC, 0, regSQG_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regWD_UTCL1_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_CNTL2),
+       SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_LO32),
+       SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_HI32),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_DEBUG),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_CNTL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_CNTL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_INSTR_PNTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_INSTR_PNTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_INSTR_PNTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STATUS),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR0),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR1),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_RS64_INSTR_PNTR),
+       /* cp header registers */
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
+       /* SE status registers */
+       SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE0),
+       SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE1),
+};
+
+static const struct amdgpu_hwip_reg_entry gc_cp_reg_list_12_1[] = {
+       /* compute registers */
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_VMID),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PERSISTENT_STATE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PIPE_PRIORITY),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUEUE_PRIORITY),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUANTUM),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_RPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_CONTROL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_RPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_CONTROL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_REQUEST),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_CONTROL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_RPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_EVENTS),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_LO),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_CONTROL),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_OFFSET),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_SIZE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_WG_STATE_OFFSET),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_SIZE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_GDS_RESOURCE_STATE),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_ERROR),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR_MEM),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_LO),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_HI),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_OFFSET),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_DW_CNT),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_WG_STATE_OFFSET),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_STATUS),
+       /* cp header registers */
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+       SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
+};
+
  static void gfx_v12_1_xcc_disable_gpa_mode(struct amdgpu_device *adev, int 
xcc_id);
  static void gfx_v12_1_set_ring_funcs(struct amdgpu_device *adev);
  static void gfx_v12_1_set_irq_funcs(struct amdgpu_device *adev);
@@ -1148,6 +1269,153 @@ static int 
gfx_v12_1_rlc_backdoor_autoload_enable(struct amdgpu_device *adev)
         return 0;
  }

+static void gfx_v12_1_alloc_ip_dump(struct amdgpu_device *adev)
+{
+       uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_1);
+       uint32_t *ptr, inst, num_xcc;
+
+       num_xcc = NUM_XCC(adev->gfx.xcc_mask);
+
+       ptr = kcalloc(reg_count * num_xcc, sizeof(uint32_t), GFP_KERNEL);
+       if (!ptr) {
+               DRM_ERROR("Failed to allocate memory for GFX IP Dump\n");
+               adev->gfx.ip_dump_core = NULL;
+       } else {
+               adev->gfx.ip_dump_core = ptr;
+       }
+
+       /* Allocate memory for compute queue registers for all the instances */
+       reg_count = ARRAY_SIZE(gc_cp_reg_list_12_1);
+       inst = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe_per_mec *
+               adev->gfx.mec.num_queue_per_pipe;
+
+       ptr = kcalloc(reg_count * inst * num_xcc, sizeof(uint32_t), GFP_KERNEL);
+       if (!ptr) {
+               DRM_ERROR("Failed to allocate memory for Compute Queues IP 
Dump\n");
+               adev->gfx.ip_dump_compute_queues = NULL;
+       } else {
+               adev->gfx.ip_dump_compute_queues = ptr;
+       }
+}
+
+static void gfx_v12_1_ip_print(struct amdgpu_ip_block *ip_block,
+                              struct drm_printer *p)
+{
+       struct amdgpu_device *adev = ip_block->adev;
+       uint32_t i, j, k;
+       uint32_t xcc_id, xcc_offset, inst_offset;
+       uint32_t num_xcc, reg, num_inst;
+       uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_1);
+
+       if (!adev->gfx.ip_dump_core)
+               return;
+
+       num_xcc = NUM_XCC(adev->gfx.xcc_mask);
+       drm_printf(p, "Number of Instances:%d\n", num_xcc);
+       for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
+               xcc_offset = xcc_id * reg_count;
+               drm_printf(p, "\nInstance id:%d\n", xcc_id);
+               for (i = 0; i < reg_count; i++)
+                       drm_printf(p, "%-50s \t 0x%08x\n",
+                                  gc_reg_list_12_1[i].reg_name,
+                                  adev->gfx.ip_dump_core[xcc_offset + i]);
+       }
+
+       /* print compute queue registers for all instances */
+       if (!adev->gfx.ip_dump_compute_queues)
+               return;
+
+       reg_count = ARRAY_SIZE(gc_cp_reg_list_12_1);
+       drm_printf(p, "\nnum_xcc: %d num_mec: %d num_pipe: %d num_queue: %d\n",
+                  num_xcc,
+                  adev->gfx.mec.num_mec,
+                  adev->gfx.mec.num_pipe_per_mec,
+                  adev->gfx.mec.num_queue_per_pipe);
+
+       for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
+               xcc_offset = xcc_id * reg_count * num_inst;

num_inst is never initialized here.

Fixed.

+               inst_offset = 0;
+               for (i = 0; i < adev->gfx.mec.num_mec; i++) {
+                       for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) {
+                               for (k = 0; k < 
adev->gfx.mec.num_queue_per_pipe; k++) {
+                                       drm_printf(p,
+                                                  "\nxcc:%d mec:%d, pipe:%d, 
queue:%d\n",
+                                                   xcc_id, i, j, k);
+                                       for (reg = 0; reg < reg_count; reg++) {
+                                               drm_printf(p,
+                                                          "%-50s \t 0x%08x\n",
+                                                          
gc_cp_reg_list_12_1[reg].reg_name,
+                                                          
adev->gfx.ip_dump_compute_queues
+                                                          [xcc_offset + 
inst_offset +
+                                                           reg]);
+                                       }
+                                       inst_offset += reg_count;
+                               }
+                       }
+               }
+       }
+}
+
+static void gfx_v12_1_ip_dump(struct amdgpu_ip_block *ip_block)
+{
+       struct amdgpu_device *adev = ip_block->adev;
+       uint32_t i, j, k;
+       uint32_t num_xcc, reg, num_inst;
+       uint32_t xcc_id, xcc_offset, inst_offset;
+       uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_1);
+
+       if (!adev->gfx.ip_dump_core)
+               return;
+
+       num_xcc = NUM_XCC(adev->gfx.xcc_mask);
+
+       amdgpu_gfx_off_ctrl(adev, false);
+       for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
+               xcc_offset = xcc_id * reg_count;
+               for (i = 0; i < reg_count; i++)
+                       adev->gfx.ip_dump_core[xcc_offset + i] =
+                               
RREG32(SOC15_REG_ENTRY_OFFSET_INST(gc_reg_list_12_1[i],
+                                                                  GET_INST(GC, 
xcc_id)));
+       }
+       amdgpu_gfx_off_ctrl(adev, true);
+
+       /* dump compute queue registers for all instances */
+       if (!adev->gfx.ip_dump_compute_queues)
+               return;
+
+       num_inst = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe_per_mec *
+               adev->gfx.mec.num_queue_per_pipe;
+       reg_count = ARRAY_SIZE(gc_cp_reg_list_12_1);
+       amdgpu_gfx_off_ctrl(adev, false);
+       mutex_lock(&adev->srbm_mutex);
+       for (xcc_id = 0; xcc_id < num_xcc; xcc_id++) {
+               xcc_offset = xcc_id * reg_count * num_inst;
+               inst_offset = 0;
+               for (i = 0; i < adev->gfx.mec.num_mec; i++) {
+                       for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) {
+                               for (k = 0; k < 
adev->gfx.mec.num_queue_per_pipe; k++) {
+                                       /* ME0 is for GFX so start from 1 for 
CP */
+                                       soc_v1_0_grbm_select(adev, 1 + i, j, k, 
0,
+                                                            GET_INST(GC, 
xcc_id));
+
+                                       for (reg = 0; reg < reg_count; reg++) {
+                                               adev->gfx.ip_dump_compute_queues
+                                                       [xcc_offset +
+                                                        inst_offset + reg] =
+                                                       
RREG32(SOC15_REG_ENTRY_OFFSET_INST(
+                                                                      
gc_cp_reg_list_12_1[reg],
+                                                                      
GET_INST(GC, xcc_id)));
+                                       }
+                                       inst_offset += reg_count;
+                               }
+                       }
+               }
+       }
+       soc_v1_0_grbm_select(adev, 0, 0, 0, 0, 0);
+       mutex_unlock(&adev->srbm_mutex);
+       amdgpu_gfx_off_ctrl(adev, true);

Do we support GFX OFF on GFX 12.1? .

No, it's a noop if a particular chip doesn't support it.

+}
+
  static int gfx_v12_1_sw_init(struct amdgpu_ip_block *ip_block)
  {
         uint16_t major_ver, minor_ver;
@@ -1282,6 +1550,8 @@ static int gfx_v12_1_sw_init(struct amdgpu_ip_block 
*ip_block)
         if (r)
                 return r;

+       gfx_v12_1_alloc_ip_dump(adev);
+
         r = amdgpu_gfx_sysfs_init(adev);
         if (r)

You may want to free the IP dump memory if the sysfs init fails or do the alloc 
after the sysfs init.

Fixed.

Thanks,

Alex

Regards,

Mukul

                 return r;
@@ -1325,6 +1595,9 @@ static int gfx_v12_1_sw_fini(struct amdgpu_ip_block 
*ip_block)
         gfx_v12_1_free_microcode(adev);
         amdgpu_gfx_sysfs_fini(adev);

+       kfree(adev->gfx.ip_dump_core);
+       kfree(adev->gfx.ip_dump_compute_queues);
+
         return 0;
  }

@@ -3905,6 +4178,8 @@ static const struct amd_ip_funcs gfx_v12_1_ip_funcs = {
         .set_clockgating_state = gfx_v12_1_set_clockgating_state,
         .set_powergating_state = gfx_v12_1_set_powergating_state,
         .get_clockgating_state = gfx_v12_1_get_clockgating_state,
+       .dump_ip_state = gfx_v12_1_ip_dump,
+       .print_ip_state = gfx_v12_1_ip_print,
  };

  static const struct amdgpu_ring_funcs gfx_v12_1_ring_funcs_compute = {
--
2.54.0

Re: [PATCH] drm/amdgpu/gx12.1: Add ip dump support

Reply via email to