drm/amdgpu: Support umc node harvest config on umc v8_10

Don't need to query error count and error address on harvest umc nodes.
v2: Fix code bug, use active_mask instead of harvsest_config
    and remove unnecessary argument in LOOP macro.
v3: Leave adev->gmc.num_umc unchanged.

Signed-off-by: Candice Li <[email protected]>
Reviewed-by: Tao Zhou <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

Author: candicelicy

drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c

@@ -543,6 +543,7 @@ static void amdgpu_discovery_read_from_harvest_table(struct amdgpu_device *adev,
 	struct harvest_table *harvest_info;
 	u16 offset;
 	int i;
+	uint32_t umc_harvest_config = 0;
 
 	bhdr = (struct binary_header *)adev->mman.discovery_bin;
 	offset = le16_to_cpu(bhdr->table_list[HARVEST_INFO].offset);
@@ -570,12 +571,17 @@ static void amdgpu_discovery_read_from_harvest_table(struct amdgpu_device *adev,
 			adev->harvest_ip_mask |= AMD_HARVEST_IP_DMU_MASK;
 			break;
 		case UMC_HWID:
+			umc_harvest_config |=
+				1 << (le16_to_cpu(harvest_info->list[i].number_instance));
 			(*umc_harvest_count)++;
 			break;
 		default:
 			break;
 		}
 	}
+
+	adev->umc.active_mask = ((1 << adev->umc.node_inst_num) - 1) &
+				~umc_harvest_config;
 }
 
 /* ================================================== */
@@ -1156,8 +1162,10 @@ static int amdgpu_discovery_reg_base_init(struct amdgpu_device *adev)
 						AMDGPU_MAX_SDMA_INSTANCES);
 			}
 
-			if (le16_to_cpu(ip->hw_id) == UMC_HWID)
+			if (le16_to_cpu(ip->hw_id) == UMC_HWID) {
 				adev->gmc.num_umc++;
+				adev->umc.node_inst_num++;
+			}
 
 			for (k = 0; k < num_base_address; k++) {
 				/*

drivers/gpu/drm/amd/amdgpu/amdgpu_umc.h

@@ -42,7 +42,7 @@
 #define LOOP_UMC_INST_AND_CH(umc_inst, ch_inst) LOOP_UMC_INST((umc_inst)) LOOP_UMC_CH_INST((ch_inst))
 
 #define LOOP_UMC_NODE_INST(node_inst) \
-		for ((node_inst) = 0; (node_inst) < adev->umc.node_inst_num; (node_inst)++)
+		for_each_set_bit((node_inst), &(adev->umc.active_mask), adev->umc.node_inst_num)
 
 #define LOOP_UMC_EACH_NODE_INST_AND_CH(node_inst, umc_inst, ch_inst) \
 		LOOP_UMC_NODE_INST((node_inst)) LOOP_UMC_INST_AND_CH((umc_inst), (ch_inst))
@@ -69,7 +69,7 @@ struct amdgpu_umc {
 	/* number of umc instance with memory map register access */
 	uint32_t umc_inst_num;
 
-	/*number of umc node instance with memory map register access*/
+	/* Total number of umc node instance including harvest one */
 	uint32_t node_inst_num;
 
 	/* UMC regiser per channel offset */
@@ -82,6 +82,9 @@ struct amdgpu_umc {
 
 	const struct amdgpu_umc_funcs *funcs;
 	struct amdgpu_umc_ras *ras;
+
+	/* active mask for umc node instance */
+	unsigned long active_mask;
 };
 
 int amdgpu_umc_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block);

drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c

@@ -567,7 +567,6 @@ static void gmc_v11_0_set_umc_funcs(struct amdgpu_device *adev)
 	case IP_VERSION(8, 10, 0):
 		adev->umc.channel_inst_num = UMC_V8_10_CHANNEL_INSTANCE_NUM;
 		adev->umc.umc_inst_num = UMC_V8_10_UMC_INSTANCE_NUM;
-		adev->umc.node_inst_num = adev->gmc.num_umc;
 		adev->umc.max_ras_err_cnt_per_query = UMC_V8_10_TOTAL_CHANNEL_NUM(adev);
 		adev->umc.channel_offs = UMC_V8_10_PER_CHANNEL_OFFSET;
 		adev->umc.retire_unit = UMC_V8_10_NA_COL_2BITS_POWER_OF_2_NUM;

drivers/gpu/drm/amd/amdgpu/umc_v8_10.h

@@ -31,9 +31,9 @@
 /* number of umc instance with memory map register access */
 #define UMC_V8_10_UMC_INSTANCE_NUM		2
 
-/* Total channel instances for all umc nodes */
+/* Total channel instances for all available umc nodes */
 #define UMC_V8_10_TOTAL_CHANNEL_NUM(adev) \
-	(UMC_V8_10_CHANNEL_INSTANCE_NUM * UMC_V8_10_UMC_INSTANCE_NUM * (adev)->umc.node_inst_num)
+	(UMC_V8_10_CHANNEL_INSTANCE_NUM * UMC_V8_10_UMC_INSTANCE_NUM * (adev)->gmc.num_umc)
 
 /* UMC regiser per channel offset */
 #define UMC_V8_10_PER_CHANNEL_OFFSET	0x400