drm/amdgpu: add per device user friendly xgmi events for vega20

Non-outbound data metrics are non useful so mark them as legacy.
Bucket new perf counters into device and not device ip.
Bind events to chip instead of IP.
Report available event counters and not number of hw counter banks.
Move DF public macros to private since not needed outside of IP version.

v5: cleanup by moving per chip configs into structs

v4: After more discussion, replace *_LEGACY references with IP references
to indicate concept of pmu-typed versus event-config-typed event
registration.

v3: attr groups const array is global but attr groups are allocated per
device which doesn't work and causes problems on memory allocation and
de-allocation for pmu unregister. Switch to building const attr groups
per pmu instead to simplify solution.

v2: add comments on sysfs structure and formatting.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Harish Kasiviswanathan <harish.kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -1294,19 +1294,6 @@ bool amdgpu_device_load_pci_state(struct pci_dev *pdev);
 
 #include "amdgpu_object.h"
 
-/* used by df_v3_6.c and amdgpu_pmu.c */
-#define AMDGPU_PMU_ATTR(_name, _object)					\
-static ssize_t								\
-_name##_show(struct device *dev,					\
-			       struct device_attribute *attr,		\
-			       char *page)				\
-{									\
-	BUILD_BUG_ON(sizeof(_object) >= PAGE_SIZE - 1);			\
-	return sprintf(page, _object "\n");				\
-}									\
-									\
-static struct device_attribute pmu_attr_##_name = __ATTR_RO(_name)
-
 static inline bool amdgpu_is_tmz(struct amdgpu_device *adev)
 {
        return adev->gmc.tmz_enabled;

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

@@ -19,18 +19,38 @@
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  *
- * Author: Jonathan Kim <jonathan.kim@amd.com>
- *
  */
 
 #ifndef _AMDGPU_PMU_H_
 #define _AMDGPU_PMU_H_
 
+/* PMU types. */
 enum amdgpu_pmu_perf_type {
-	PERF_TYPE_AMDGPU_DF = 0,
-	PERF_TYPE_AMDGPU_MAX
+	AMDGPU_PMU_PERF_TYPE_NONE = 0,
+	AMDGPU_PMU_PERF_TYPE_DF,
+	AMDGPU_PMU_PERF_TYPE_ALL
 };
 
+/*
+ * PMU type AMDGPU_PMU_PERF_TYPE_ALL can hold events of different "type"
+ * configurations.  Event config types are parsed from the 64-bit raw
+ * config (See EVENT_CONFIG_TYPE_SHIFT and EVENT_CONFIG_TYPE_MASK) and
+ * are registered into the HW perf events config_base.
+ *
+ * PMU types with only a single event configuration type
+ * (non-AMDGPU_PMU_PERF_TYPE_ALL) have their event config type auto generated
+ * when the performance counter is added.
+ */
+enum amdgpu_pmu_event_config_type {
+	AMDGPU_PMU_EVENT_CONFIG_TYPE_NONE = 0,
+	AMDGPU_PMU_EVENT_CONFIG_TYPE_DF,
+	AMDGPU_PMU_EVENT_CONFIG_TYPE_XGMI,
+	AMDGPU_PMU_EVENT_CONFIG_TYPE_MAX
+};
+
+#define AMDGPU_PMU_EVENT_CONFIG_TYPE_SHIFT	56
+#define AMDGPU_PMU_EVENT_CONFIG_TYPE_MASK	0xff
+
 int amdgpu_pmu_init(struct amdgpu_device *adev);
 void amdgpu_pmu_fini(struct amdgpu_device *adev);
 

drivers/gpu/drm/amd/amdgpu/df_v3_6.c

@@ -30,71 +30,17 @@
 #define DF_3_6_SMN_REG_INST_DIST        0x8
 #define DF_3_6_INST_CNT                 8
 
-static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0,
-				       16, 32, 0, 0, 0, 2, 4, 8};
-
-/* init df format attrs */
-AMDGPU_PMU_ATTR(event,		"config:0-7");
-AMDGPU_PMU_ATTR(instance,	"config:8-15");
-AMDGPU_PMU_ATTR(umask,		"config:16-23");
-
-/* df format attributes  */
-static struct attribute *df_v3_6_format_attrs[] = {
-	&pmu_attr_event.attr,
-	&pmu_attr_instance.attr,
-	&pmu_attr_umask.attr,
-	NULL
-};
-
-/* df format attribute group */
-static struct attribute_group df_v3_6_format_attr_group = {
-	.name = "format",
-	.attrs = df_v3_6_format_attrs,
-};
+/* Defined in global_features.h as FTI_PERFMON_VISIBLE */
+#define DF_V3_6_MAX_COUNTERS		4
 
-/* df event attrs */
-AMDGPU_PMU_ATTR(cake0_pcsout_txdata,
-		      "event=0x7,instance=0x46,umask=0x2");
-AMDGPU_PMU_ATTR(cake1_pcsout_txdata,
-		      "event=0x7,instance=0x47,umask=0x2");
-AMDGPU_PMU_ATTR(cake0_pcsout_txmeta,
-		      "event=0x7,instance=0x46,umask=0x4");
-AMDGPU_PMU_ATTR(cake1_pcsout_txmeta,
-		      "event=0x7,instance=0x47,umask=0x4");
-AMDGPU_PMU_ATTR(cake0_ftiinstat_reqalloc,
-		      "event=0xb,instance=0x46,umask=0x4");
-AMDGPU_PMU_ATTR(cake1_ftiinstat_reqalloc,
-		      "event=0xb,instance=0x47,umask=0x4");
-AMDGPU_PMU_ATTR(cake0_ftiinstat_rspalloc,
-		      "event=0xb,instance=0x46,umask=0x8");
-AMDGPU_PMU_ATTR(cake1_ftiinstat_rspalloc,
-		      "event=0xb,instance=0x47,umask=0x8");
-
-/* df event attributes  */
-static struct attribute *df_v3_6_event_attrs[] = {
-	&pmu_attr_cake0_pcsout_txdata.attr,
-	&pmu_attr_cake1_pcsout_txdata.attr,
-	&pmu_attr_cake0_pcsout_txmeta.attr,
-	&pmu_attr_cake1_pcsout_txmeta.attr,
-	&pmu_attr_cake0_ftiinstat_reqalloc.attr,
-	&pmu_attr_cake1_ftiinstat_reqalloc.attr,
-	&pmu_attr_cake0_ftiinstat_rspalloc.attr,
-	&pmu_attr_cake1_ftiinstat_rspalloc.attr,
-	NULL
-};
-
-/* df event attribute group */
-static struct attribute_group df_v3_6_event_attr_group = {
-	.name = "events",
-	.attrs = df_v3_6_event_attrs
-};
+/* get flags from df perfmon config */
+#define DF_V3_6_GET_EVENT(x)		(x & 0xFFUL)
+#define DF_V3_6_GET_INSTANCE(x)		((x >> 8) & 0xFFUL)
+#define DF_V3_6_GET_UNITMASK(x)		((x >> 16) & 0xFFUL)
+#define DF_V3_6_PERFMON_OVERFLOW	0xFFFFFFFFFFFFULL
 
-/* df event attr groups  */
-const struct attribute_group *df_v3_6_attr_groups[] = {
-		&df_v3_6_format_attr_group,
-		&df_v3_6_event_attr_group,
-		NULL
-};
+static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0,
+				       16, 32, 0, 0, 0, 2, 4, 8};
 
 static uint64_t df_v3_6_get_fica(struct amdgpu_device *adev,
 				 uint32_t ficaa_val)

drivers/gpu/drm/amd/amdgpu/df_v3_6.h

@@ -35,15 +35,6 @@ enum DF_V3_6_MGCG {
 	DF_V3_6_MGCG_ENABLE_63_CYCLE_DELAY = 15
 };
 
-/* Defined in global_features.h as FTI_PERFMON_VISIBLE */
-#define DF_V3_6_MAX_COUNTERS		4
-
-/* get flags from df perfmon config */
-#define DF_V3_6_GET_EVENT(x)		(x & 0xFFUL)
-#define DF_V3_6_GET_INSTANCE(x)		((x >> 8) & 0xFFUL)
-#define DF_V3_6_GET_UNITMASK(x)		((x >> 16) & 0xFFUL)
-#define DF_V3_6_PERFMON_OVERFLOW	0xFFFFFFFFFFFFULL
-
 extern const struct attribute_group *df_v3_6_attr_groups[];
 extern const struct amdgpu_df_funcs df_v3_6_funcs;