Skip to contentSkip to conversation

drm/amdgpu: fix xgmi perfmon a-b-a problem

PATCH In review Jonathan Kim 2 participants 7 messages Sep 15, 2020 – Sep 17, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/3 drm/amdgpu: fix xgmi perfmon a-b-a problem

Mapping hw counters per event config will cause ABA problems so map per event instead.

v2: Discontinue starting perf counters if add fails. Make it clear what's happening with pmc_start.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_df.h  |   6 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c |  42 ++++++----
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    | 105 +++++++++++-------------
 3 files changed, 78 insertions(+), 75 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_df.h +3 -3
index 373cdebe0e2f..52488bb45112 100644
@@ -44,11 +44,11 @@ struct amdgpu_df_funcs {
 	void (*enable_ecc_force_par_wr_rmw)(struct amdgpu_device *adev,
 					    bool enable);
 	int (*pmc_start)(struct amdgpu_device *adev, uint64_t config,
-					 int is_add);
+					 int counter_idx, int is_add);
 	int (*pmc_stop)(struct amdgpu_device *adev, uint64_t config,
-					 int is_remove);
+					 int counter_idx, int is_remove);
 	void (*pmc_get_count)(struct amdgpu_device *adev, uint64_t config,
-					 uint64_t *count);
+					 int counter_idx, uint64_t *count);
 	uint64_t (*get_fica)(struct amdgpu_device *adev, uint32_t ficaa_val);
 	void (*set_fica)(struct amdgpu_device *adev, uint32_t ficaa_val,
 			 uint32_t ficadl_val, uint32_t ficadh_val);
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +28 -14

100 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/df_v3_6.c +47 -58

252 lines not on this page · raw message

Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#1 lore raw

PATCH 2/3 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.

v2: add comments on sysfs structure and formatting.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h     |  13 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 342 +++++++++++++++++++++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h |   6 +-
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    |  72 +----
 drivers/gpu/drm/amd/amdgpu/df_v3_6.h    |   9 -
 5 files changed, 314 insertions(+), 128 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -13
index 13f92dea182a..f43dfdd2716a 100644
@@ -1279,19 +1279,6 @@ 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.c +300 -42

432 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h +5 -1
index 7dddb7160a11..0d214abe720e 100644
@@ -27,10 +27,14 @@
 #define _AMDGPU_PMU_H_
 
 enum amdgpu_pmu_perf_type {
-	PERF_TYPE_AMDGPU_DF = 0,
+	PERF_TYPE_AMDGPU_DF_LEGACY = 0,
+	PERF_TYPE_AMDGPU_XGMI,
 	PERF_TYPE_AMDGPU_MAX
 };
 
+#define AMDGPU_PERF_TYPE_SHIFT	56
+#define AMDGPU_PERF_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 +9 -63

80 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/df_v3_6.h +0 -9
index 76998541bc30..2505c7ef258a 100644
@@ -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;
 
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#4 lore raw

PATCH 3/3 drm/amdgpu: add xgmi perfmons for arcturus

In reply to Jonathan Kim

Add xgmi perfmons for Arcturus.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 55 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    |  3 ++
 2 files changed, 58 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +55 -0
index f3d2ac0e88a7..ec521c72e631 100644
@@ -34,6 +34,8 @@
 #define NUM_EVENTS_DF_LEGACY		8
 #define NUM_EVENTS_VEGA20_XGMI		2
 #define NUM_EVENTS_VEGA20_MAX		2
+#define NUM_EVENTS_ARCTURUS_XGMI	6
+#define NUM_EVENTS_ARCTURUS_MAX		6
 
 /* record to keep track of pmu entry per pmu type per device */
 struct amdgpu_pmu_entry {
@@ -110,6 +112,27 @@ const struct attribute_group *vega20_attr_groups[] = {
 	NULL
 };
 
+/* Arcturus events */
+static const char *arcturus_events[NUM_EVENTS_ARCTURUS_MAX][2] = {
+	{ "xgmi_link0_data_outbound", "event=0x7,instance=0x4b,umask=0x2" },
+	{ "xgmi_link1_data_outbound", "event=0x7,instance=0x4c,umask=0x2" },
+	{ "xgmi_link2_data_outbound", "event=0x7,instance=0x4d,umask=0x2" },
+	{ "xgmi_link3_data_outbound", "event=0x7,instance=0x4e,umask=0x2" },
+	{ "xgmi_link4_data_outbound", "event=0x7,instance=0x4f,umask=0x2" },
+	{ "xgmi_link5_data_outbound", "event=0x7,instance=0x50,umask=0x2" }
+};
+
+static struct attribute_group arcturus_event_attr_group = {
+	.name = "events",
+	.attrs = NULL
+};
+
+const struct attribute_group *arcturus_attr_groups[] = {
+	&amdgpu_pmu_format_attr_group,
+	&arcturus_event_attr_group,
+	NULL
+};
+
 /* All df_vega20_* items are DEPRECATED. Use vega20_ items above instead. */
 static const char *df_vega20_formats[NUM_FORMATS_DF_LEGACY][2] = {
 	{ "event", "config:0-7" },
@@ -400,6 +423,16 @@ static int init_pmu_by_type(struct amdgpu_device *adev,
 
 		pmu_entry->pmu.attr_groups = vega20_attr_groups;
 		break;
+	case CHIP_ARCTURUS:
+		amdgpu_pmu_create_attributes(evt_attr_group, evt_attr,
+				arcturus_events, 0, NUM_EVENTS_ARCTURUS_XGMI,
+				PERF_TYPE_AMDGPU_XGMI);
+		num_events += NUM_EVENTS_ARCTURUS_XGMI;
+
+		/* other events can be added here */
+
+		pmu_entry->pmu.attr_groups = arcturus_attr_groups;
+		break;
 	default:
 		return -ENODEV;
 	};
@@ -530,6 +563,28 @@ int amdgpu_pmu_init(struct amdgpu_device *adev)
 			goto err_pmu;
 		}
 
+		break;
+	case CHIP_ARCTURUS:
+		ret = amdgpu_pmu_alloc_pmu_attrs(&amdgpu_pmu_format_attr_group,
+						&fmt_attr,
+						NUM_FORMATS_AMDGPU_PMU,
+						&arcturus_event_attr_group,
+						&evt_attr,
+						NUM_EVENTS_ARCTURUS_MAX);
+
+		if (ret)
+			goto err_alloc;
+
+		ret = init_pmu_by_type(adev,
+				&amdgpu_pmu_format_attr_group, fmt_attr,
+				&arcturus_event_attr_group, evt_attr,
+				"Event", "amdgpu", PERF_TYPE_AMDGPU_MAX);
+
+		if (ret) {
+			kfree(arcturus_event_attr_group.attrs);
+			goto err_pmu;
+		}
+
 		break;
 	default:
 		return 0;
drivers/gpu/drm/amd/amdgpu/df_v3_6.c +3 -0
index 6e57ae95f997..6b4b30a8dce5 100644
@@ -513,6 +513,7 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		if (is_add)
 			return df_v3_6_pmc_add_cntr(adev, config);
 
@@ -554,6 +555,7 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		ret = df_v3_6_pmc_get_ctrl_settings(adev,
 			config,
 			counter_idx,
@@ -590,6 +592,7 @@ static void df_v3_6_pmc_get_count(struct amdgpu_device *adev,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		df_v3_6_pmc_get_read_settings(adev, config, counter_idx,
 						&lo_base_addr, &hi_base_addr);
 
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#2 lore raw

RE: [PATCH 2/3] drm/amdgpu: add per device user friendly xgmi events for vega20

In reply to Jonathan Kim

[AMD Official Use Only - Internal Distribution Only]

Some comments and a question

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Tuesday, September 15, 2020 6:00 PM
To: [email protected]
Cc: Kasiviswanathan, Harish <[email protected]>; Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 2/3] 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.

v2: add comments on sysfs structure and formatting.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h     |  13 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 342 +++++++++++++++++++++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h |   6 +-
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    |  72 +----
 drivers/gpu/drm/amd/amdgpu/df_v3_6.h    |   9 -
 5 files changed, 314 insertions(+), 128 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -13
index 13f92dea182a..f43dfdd2716a 100644
@@ -1279,19 +1279,6 @@ 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.c +213 -21

309 lines not on this page · raw message

where do all these allocations get freed?

+ *fmt_attr = kcalloc(fmt_num_attrs, sizeof(**fmt_attr), GFP_KERNEL);
+
+ if (!fmt_attr)

I think you want to check *fmt_attr

+ return -ENOMEM;
+
+ fmt_attr_group->attrs = kcalloc(fmt_num_attrs + 1,
+ sizeof(*fmt_attr_group->attrs), GFP_KERNEL);
+
+ if (!fmt_attr_group->attrs)
+ goto err_fmt_attr_grp;
+
+ *evt_attr = kcalloc(evt_num_attrs, sizeof(**evt_attr), GFP_KERNEL);
+
+ if (!evt_attr)

I think you want to check *evt_attr

+ goto err_evt_attr;
+
+ evt_attr_group->attrs = kcalloc(evt_num_attrs + 1,
+ sizeof(*evt_attr_group->attrs), GFP_KERNEL);
+
+ if (!evt_attr_group->attrs)
+ goto err_evt_attr_grp;
+
+ return 0;
+err_evt_attr_grp:
+ kfree(*evt_attr);
+err_evt_attr:
+ kfree(fmt_attr_group->attrs);
+err_fmt_attr_grp:
+ kfree(*fmt_attr);
+ return -ENOMEM;
}

/* init amdgpu_pmu */
int amdgpu_pmu_init(struct amdgpu_device *adev)
{
+ struct amdgpu_pmu_event_attribute *fmt_attr = NULL, *evt_attr = NULL,
+ *old_fmt_attr = NULL, *old_evt_attr = NULL;
int ret = 0;

+ if (adev->asic_type != CHIP_VEGA20)
+ goto init_events;
+
+ ret = amdgpu_pmu_alloc_pmu_attrs(&df_vega20_format_attr_group,
+ &old_fmt_attr,
+ NUM_FORMATS_DF_LEGACY,
+ &df_vega20_event_attr_group,
+ &old_evt_attr,
+ NUM_EVENTS_DF_LEGACY);
+
+ if (ret)
+ return ret;
+
+ ret = init_pmu_by_type(adev,
+ &df_vega20_format_attr_group, old_fmt_attr,
+ &df_vega20_event_attr_group, old_evt_attr,
+ "DF", "amdgpu_df", PERF_TYPE_AMDGPU_DF_LEGACY);
+ if (ret)
+ goto err_old_pmu;
+
+init_events:
switch (adev->asic_type) {
case CHIP_VEGA20:
- /* init df */
- ret = init_pmu_by_type(adev, df_v3_6_attr_groups,
- "DF", "amdgpu_df", PERF_TYPE_AMDGPU_DF,
- DF_V3_6_MAX_COUNTERS);
+ ret = amdgpu_pmu_alloc_pmu_attrs(&amdgpu_pmu_format_attr_group,
+ &fmt_attr,
+ NUM_FORMATS_AMDGPU_PMU,
+ &vega20_event_attr_group,
+ &evt_attr,
+ NUM_EVENTS_VEGA20_MAX);
+
+ if (ret)
+ goto err_alloc;
+
+ ret = init_pmu_by_type(adev,
+ &amdgpu_pmu_format_attr_group, fmt_attr,
+ &vega20_event_attr_group, evt_attr,
+ "Event", "amdgpu", PERF_TYPE_AMDGPU_MAX);
+
+ if (ret) {
+ kfree(vega20_event_attr_group.attrs);
+ goto err_pmu;
+ }

- /* other pmu types go here*/
break;
default:
return 0;
- }
+ };

return 0;
-}
-
-
-/* destroy all pmu data associated with target device */ -void amdgpu_pmu_fini(struct amdgpu_device *adev) -{
- struct amdgpu_pmu_entry *pe, *temp;
-
- list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
- if (pe->adev == adev) {
- list_del(&pe->entry);
- perf_pmu_unregister(&pe->pmu);
- kfree(pe);
- }
- }
+err_pmu:
+ kfree(fmt_attr);
+ kfree(evt_attr);
+ kfree(amdgpu_pmu_format_attr_group.attrs);
+err_alloc:
+ if (adev->asic_type == CHIP_VEGA20)
+ amdgpu_pmu_fini(adev);
+ return ret;
+err_old_pmu:
+ kfree(old_fmt_attr);
+ kfree(old_evt_attr);
+ kfree(df_vega20_format_attr_group.attrs);
+ kfree(df_vega20_event_attr_group.attrs);
+ return ret;
}

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h +5 -1
index 7dddb7160a11..0d214abe720e 100644
@@ -27,10 +27,14 @@
 #define _AMDGPU_PMU_H_
 
 enum amdgpu_pmu_perf_type {
-	PERF_TYPE_AMDGPU_DF = 0,
+	PERF_TYPE_AMDGPU_DF_LEGACY = 0,
+	PERF_TYPE_AMDGPU_XGMI,
 	PERF_TYPE_AMDGPU_MAX
 };
 
+#define AMDGPU_PERF_TYPE_SHIFT	56
+#define AMDGPU_PERF_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 +9 -63

80 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/df_v3_6.h +0 -9
index 76998541bc30..2505c7ef258a 100644
@@ -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;
 
Signature · 5 lines
2.17.1
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Reviewed-by on v1 3/3
#5 lore raw

RE: [PATCH 3/3] drm/amdgpu: add xgmi perfmons for arcturus

In reply to Jonathan Kim

[AMD Official Use Only - Internal Distribution Only]

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Tuesday, September 15, 2020 6:00 PM
To: [email protected]
Cc: Kasiviswanathan, Harish <[email protected]>; Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: add xgmi perfmons for arcturus

Add xgmi perfmons for Arcturus.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 55 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    |  3 ++
 2 files changed, 58 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
index f3d2ac0e88a7..ec521c72e631 100644
@@ -34,6 +34,8 @@
 #define NUM_EVENTS_DF_LEGACY		8
 #define NUM_EVENTS_VEGA20_XGMI		2
 #define NUM_EVENTS_VEGA20_MAX		2
+#define NUM_EVENTS_ARCTURUS_XGMI	6
+#define NUM_EVENTS_ARCTURUS_MAX		6
 
 /* record to keep track of pmu entry per pmu type per device */  struct amdgpu_pmu_entry { @@ -110,6 +112,27 @@ const struct attribute_group *vega20_attr_groups[] = {
 	NULL

};

+/* Arcturus events */
+static const char *arcturus_events[NUM_EVENTS_ARCTURUS_MAX][2] = {
+ { "xgmi_link0_data_outbound", "event=0x7,instance=0x4b,umask=0x2" },
+ { "xgmi_link1_data_outbound", "event=0x7,instance=0x4c,umask=0x2" },
+ { "xgmi_link2_data_outbound", "event=0x7,instance=0x4d,umask=0x2" },
+ { "xgmi_link3_data_outbound", "event=0x7,instance=0x4e,umask=0x2" },
+ { "xgmi_link4_data_outbound", "event=0x7,instance=0x4f,umask=0x2" },
+ { "xgmi_link5_data_outbound", "event=0x7,instance=0x50,umask=0x2" } };
+
+static struct attribute_group arcturus_event_attr_group = {
+ .name = "events",
+ .attrs = NULL
+};
+
+const struct attribute_group *arcturus_attr_groups[] = {
+ &amdgpu_pmu_format_attr_group,
+ &arcturus_event_attr_group,
+ NULL
+};
+
/* All df_vega20_* items are DEPRECATED. Use vega20_ items above instead. */ static const char *df_vega20_formats[NUM_FORMATS_DF_LEGACY][2] = {
{ "event", "config:0-7" },
@@ -400,6 +423,16 @@ static int init_pmu_by_type(struct amdgpu_device *adev,

pmu_entry->pmu.attr_groups = vega20_attr_groups; break;
+ case CHIP_ARCTURUS:
+ amdgpu_pmu_create_attributes(evt_attr_group, evt_attr,
+ arcturus_events, 0, NUM_EVENTS_ARCTURUS_XGMI,
+ PERF_TYPE_AMDGPU_XGMI);
+ num_events += NUM_EVENTS_ARCTURUS_XGMI;
+
+ /* other events can be added here */
+
+ pmu_entry->pmu.attr_groups = arcturus_attr_groups;
+ break;
default:
return -ENODEV;
};
@@ -530,6 +563,28 @@ int amdgpu_pmu_init(struct amdgpu_device *adev)
goto err_pmu;
}

+ break;
+ case CHIP_ARCTURUS:
+ ret = amdgpu_pmu_alloc_pmu_attrs(&amdgpu_pmu_format_attr_group,
+ &fmt_attr,
+ NUM_FORMATS_AMDGPU_PMU,
+ &arcturus_event_attr_group,
+ &evt_attr,
+ NUM_EVENTS_ARCTURUS_MAX);
+
+ if (ret)
+ goto err_alloc;
+
+ ret = init_pmu_by_type(adev,
+ &amdgpu_pmu_format_attr_group, fmt_attr,
+ &arcturus_event_attr_group, evt_attr,
+ "Event", "amdgpu", PERF_TYPE_AMDGPU_MAX);
+
+ if (ret) {
+ kfree(arcturus_event_attr_group.attrs);
+ goto err_pmu;
+ }
+
break;
default:
return 0;

drivers/gpu/drm/amd/amdgpu/df_v3_6.c +3 -0
index 6e57ae95f997..6b4b30a8dce5 100644
@@ -513,6 +513,7 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		if (is_add)
 			return df_v3_6_pmc_add_cntr(adev, config);
 
@@ -554,6 +555,7 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		ret = df_v3_6_pmc_get_ctrl_settings(adev,
 			config,
 			counter_idx,
@@ -590,6 +592,7 @@ static void df_v3_6_pmc_get_count(struct amdgpu_device *adev,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		df_v3_6_pmc_get_read_settings(adev, config, counter_idx,
 						&lo_base_addr, &hi_base_addr);
 

--
2.17.1
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#3 lore raw

RE: [PATCH 2/3] drm/amdgpu: add per device user friendly xgmi events for vega20

In reply to Kasiviswanathan, Harish
Quoted 425 lines

-----Original Message-----
From: Kasiviswanathan, Harish <[email protected]>
Sent: Tuesday, September 15, 2020 10:08 PM
To: Kim, Jonathan <[email protected]>; amd- [email protected]
Subject: RE: [PATCH 2/3] drm/amdgpu: add per device user friendly xgmi events for vega20

[AMD Official Use Only - Internal Distribution Only]

Some comments and a question

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Tuesday, September 15, 2020 6:00 PM
To: [email protected]
Cc: Kasiviswanathan, Harish <[email protected]>; Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 2/3] 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.

v2: add comments on sysfs structure and formatting.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h     |  13 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 342
+++++++++++++++++++++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h |   6 +-
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    |  72 +----
 drivers/gpu/drm/amd/amdgpu/df_v3_6.h    |   9 -
 5 files changed, 314 insertions(+), 128 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -13
index 13f92dea182a..f43dfdd2716a 100644
@@ -1279,19 +1279,6 @@ 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.c +5 -1
index 1b0ec715c8ba..f3d2ac0e88a7 100644
@@ -27,9 +27,13 @@
 #include <linux/init.h>
 #include "amdgpu.h"
 #include "amdgpu_pmu.h"
-#include "df_v3_6.h"
 
 #define PMU_NAME_SIZE 32
+#define NUM_FORMATS_AMDGPU_PMU		4
+#define NUM_FORMATS_DF_LEGACY		3
+#define NUM_EVENTS_DF_LEGACY		8
+#define NUM_EVENTS_VEGA20_XGMI		2
+#define NUM_EVENTS_VEGA20_MAX		2
 
 /* record to keep track of pmu entry per pmu type per device */  struct
 amdgpu_pmu_entry { @@ -39,8 +43,106 @@ struct amdgpu_pmu_entry {
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +11 -0
 	unsigned int pmu_perf_type;
 };
 
+struct amdgpu_pmu_event_attribute {
+	struct device_attribute attr;
+	const char *event_str;
+	unsigned int type;
+};
+
+static ssize_t amdgpu_pmu_event_show(struct device *dev,
+				struct device_attribute *attr, char *buf) {
+	struct amdgpu_pmu_event_attribute *amdgpu_pmu_attr;
+
+	amdgpu_pmu_attr = container_of(attr, struct

amdgpu_pmu_event_attribute,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +7 -0
+									attr);
+
+	if (!amdgpu_pmu_attr->type)
+		return sprintf(buf, "%s\n", amdgpu_pmu_attr->event_str);
+
+	return sprintf(buf, "%s,type=0x%x\n",
+			amdgpu_pmu_attr->event_str, amdgpu_pmu_attr-
Quoted 1 line

type); }

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +9 -0
+
+static struct attribute_group amdgpu_pmu_format_attr_group = {
+	.name = "format",
+	.attrs = NULL,
+};
+
+/*
+ * Event formatting is global to all amdgpu events under sysfs folder
+ * /sys/bus/event_source/devices/amdgpu_<dev_num> where dev_num is

the
+ * primary device index. Registered events can be found in subfolder "events"
+ * and formatting under subfolder "format".
+ *
+ * Formats "event", "instance", and "umask" are currently used by xGMI

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +6 -0
+but can
+ * be for generalized for other IP usage.  If format naming is
+insufficient
+ * for newly registered IP events, append to the list below and handle
+the
+ * perf events hardware configuration (see hwc->config) as required by the

IP.
+ *
+ * Format "type" indicates IP type generated on pmu registration (see
+ * init_pmu_by_type) so non-legacy events omit this in the per-chip

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +4 -0
+event
+ * list (e.g. vega20_events).
+ */
+static const char

*amdgpu_pmu_formats[NUM_FORMATS_AMDGPU_PMU][2] = {

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +9 -0
+	{ "event", "config:0-7" },
+	{ "instance", "config:8-15" },
+	{ "umask", "config:16-23"},
+	{ "type", "config:56-63"}
+};
+
 static LIST_HEAD(amdgpu_pmu_list);
 
+/* Vega20 events */
+static const char *vega20_events[NUM_EVENTS_VEGA20_MAX][2] = {
+	{ "xgmi_link0_data_outbound",

"event=0x7,instance=0x46,umask=0x2" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	{ "xgmi_link1_data_outbound",

"event=0x7,instance=0x47,umask=0x2" } };

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +14 -0
+
+static struct attribute_group vega20_event_attr_group = {
+	.name = "events",
+	.attrs = NULL
+};
+
+const struct attribute_group *vega20_attr_groups[] = {
+	&amdgpu_pmu_format_attr_group,
+	&vega20_event_attr_group,
+	NULL
+};
+
+/* All df_vega20_* items are DEPRECATED. Use vega20_ items above
+instead. */ static const char

*df_vega20_formats[NUM_FORMATS_DF_LEGACY][2] = {

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +32 -0
+	{ "event", "config:0-7" },
+	{ "instance", "config:8-15" },
+	{ "umask", "config:16-23"}
+};
+
+static const char *df_vega20_events[NUM_EVENTS_DF_LEGACY][2] = {
+	{ "cake0_pcsout_txdata", "event=0x7,instance=0x46,umask=0x2" },
+	{ "cake1_pcsout_txdata", "event=0x7,instance=0x47,umask=0x2" },
+	{ "cake0_pcsout_txmeta", "event=0x7,instance=0x46,umask=0x4" },
+	{ "cake1_pcsout_txmeta", "event=0x7,instance=0x47,umask=0x4" },
+	{ "cake0_ftiinstat_reqalloc", "event=0xb,instance=0x46,umask=0x4" },
+	{ "cake1_ftiinstat_reqalloc", "event=0xb,instance=0x47,umask=0x4" },
+	{ "cake0_ftiinstat_rspalloc", "event=0xb,instance=0x46,umask=0x8" },
+	{ "cake1_ftiinstat_rspalloc", "event=0xb,instance=0x47,umask=0x8" },
+};
+
+static struct attribute_group df_vega20_format_attr_group = {
+	.name = "format",
+	.attrs = NULL,
+};
+
+static struct attribute_group df_vega20_event_attr_group = {
+	.name = "events",
+	.attrs = NULL
+};
+
+const struct attribute_group *df_vega20_attr_groups[] = {
+	&df_vega20_format_attr_group,
+	&df_vega20_event_attr_group,
+	NULL
+};
 
 /* initialize perf counter */
 static int amdgpu_perf_event_init(struct perf_event *event) @@ -73,7
+175,8 @@ static void amdgpu_perf_start(struct perf_event *event, int

flags)

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -1
 	hwc->state = 0;
 
 	switch (pe->pmu_perf_type) {
-	case PERF_TYPE_AMDGPU_DF:
+	case PERF_TYPE_AMDGPU_DF_LEGACY:
+	case PERF_TYPE_AMDGPU_XGMI:
 		if (!(flags & PERF_EF_RELOAD)) {
 			target_cntr = pe->adev->df.funcs->pmc_start(pe-
Quoted 1 line

adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -1
 						hwc->config, 0 /* unused */,
@@ -108,7 +211,8 @@ static void amdgpu_perf_read(struct perf_event
 *event)
 		prev = local64_read(&hwc->prev_count);
 
 		switch (pe->pmu_perf_type) {
-		case PERF_TYPE_AMDGPU_DF:
+		case PERF_TYPE_AMDGPU_DF_LEGACY:
+		case PERF_TYPE_AMDGPU_XGMI:
 			pe->adev->df.funcs->pmc_get_count(pe->adev,
 						hwc->config, hwc->idx,

&count);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -1
 			break;
@@ -133,7 +237,8 @@ static void amdgpu_perf_stop(struct perf_event
 *event, int flags)
 		return;
 
 	switch (pe->pmu_perf_type) {
-	case PERF_TYPE_AMDGPU_DF:
+	case PERF_TYPE_AMDGPU_DF_LEGACY:
+	case PERF_TYPE_AMDGPU_XGMI:
 		pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc-
 >idx,
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
 									0);
 		break;
@@ -160,10 +265,15 @@ static int amdgpu_perf_add(struct perf_event
 *event, int flags)
 						  struct amdgpu_pmu_entry,
 						  pmu);
 
+	if (pe->pmu_perf_type == PERF_TYPE_AMDGPU_MAX)
+		pe->pmu_perf_type = (hwc->config >>

AMDGPU_PERF_TYPE_SHIFT) &
+

	AMDGPU_PERF_TYPE_MASK;
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +5 -2
+
 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 
 	switch (pe->pmu_perf_type) {
-	case PERF_TYPE_AMDGPU_DF:
+	case PERF_TYPE_AMDGPU_DF_LEGACY:
+	case PERF_TYPE_AMDGPU_XGMI:
 		target_cntr = pe->adev->df.funcs->pmc_start(pe->adev,
 						hwc->config, 0 /* unused */,
 						1 /* add counter */);
@@ -197,7 +307,8 @@ static void amdgpu_perf_del(struct perf_event
 *event, int flags)
 	amdgpu_perf_stop(event, PERF_EF_UPDATE);
 
 	switch (pe->pmu_perf_type) {
-	case PERF_TYPE_AMDGPU_DF:
+	case PERF_TYPE_AMDGPU_DF_LEGACY:
+	case PERF_TYPE_AMDGPU_XGMI:
 		pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc-
 >idx,
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +29 -6
 									1);
 		break;
@@ -208,18 +319,42 @@ static void amdgpu_perf_del(struct perf_event
 *event, int flags)
 	perf_event_update_userpage(event);
 }
 
-/* vega20 pmus */
+static void amdgpu_pmu_create_attributes(struct attribute_group
 *attr_group,
+				struct amdgpu_pmu_event_attribute
 *pmu_attr,
+				const char *events[][2],
+				int s_offset,
+				int e_offset,
+				unsigned int type)
+{
+	int i;
+
+	pmu_attr += s_offset;
+
+	for (i = s_offset; i < e_offset; i++) {
+		attr_group->attrs[i] = &pmu_attr->attr.attr;
+		sysfs_attr_init(&pmu_attr->attr.attr);
+		pmu_attr->attr.attr.name = events[i][0];
+		pmu_attr->attr.attr.mode = 0444;
+		pmu_attr->attr.show = amdgpu_pmu_event_show;
+		pmu_attr->event_str = events[i][1];
+		pmu_attr->type = type;
+		pmu_attr++;
+	}
+}
 
 /* init pmu tracking per pmu type */
 static int init_pmu_by_type(struct amdgpu_device *adev,
-		  const struct attribute_group *attr_groups[],
-		  char *pmu_type_name, char *pmu_file_prefix,
-		  unsigned int pmu_perf_type,
-		  unsigned int num_counters)
+			struct attribute_group *fmt_attr_group,
+			struct amdgpu_pmu_event_attribute *fmt_attr,
+			struct attribute_group *evt_attr_group,
+			struct amdgpu_pmu_event_attribute *evt_attr,
+			char *pmu_type_name, char *pmu_file_prefix,
+			unsigned int pmu_perf_type)
 {
 	char pmu_name[PMU_NAME_SIZE];
 	struct amdgpu_pmu_entry *pmu_entry;
-	int ret = 0;
+	bool is_legacy = pmu_perf_type ==

PERF_TYPE_AMDGPU_DF_LEGACY;

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	int ret = 0, num_events = 0;
 
 	pmu_entry = kzalloc(sizeof(struct amdgpu_pmu_entry),

GFP_KERNEL);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +65 -9
@@ -237,59 +372,182 @@ static int init_pmu_by_type(struct
 amdgpu_device *adev,
 		.task_ctx_nr = perf_invalid_context,
 	};
 
-	pmu_entry->pmu.attr_groups = attr_groups;
+	amdgpu_pmu_create_attributes(fmt_attr_group, fmt_attr,
+			is_legacy ? df_vega20_formats :
 amdgpu_pmu_formats, 0,
+			is_legacy ? NUM_FORMATS_DF_LEGACY :
+
 	NUM_FORMATS_AMDGPU_PMU,
+			0);
+
+	if (is_legacy) {
+		amdgpu_pmu_create_attributes(evt_attr_group, evt_attr,
+					df_vega20_events, 0,
+					NUM_EVENTS_DF_LEGACY,
+					PERF_TYPE_AMDGPU_DF_LEGACY);
+		num_events += NUM_EVENTS_DF_LEGACY;
+
+		pmu_entry->pmu.attr_groups = df_vega20_attr_groups;
+		goto legacy_register;
+	}
+
+	switch (adev->asic_type) {
+	case CHIP_VEGA20:
+		amdgpu_pmu_create_attributes(evt_attr_group, evt_attr,
+				vega20_events, 0,
 NUM_EVENTS_VEGA20_XGMI,
+				PERF_TYPE_AMDGPU_XGMI);
+		num_events += NUM_EVENTS_VEGA20_XGMI;
+
+		/* other events can be added here */
+
+		pmu_entry->pmu.attr_groups = vega20_attr_groups;
+		break;
+	default:
+		return -ENODEV;
+	};
+
+legacy_register:
 	pmu_entry->pmu_perf_type = pmu_perf_type;
-	snprintf(pmu_name, PMU_NAME_SIZE, "%s_%d",
-				pmu_file_prefix, adev_to_drm(adev)-
 >primary->index);
+	snprintf(pmu_name, PMU_NAME_SIZE, "%s_%d", pmu_file_prefix,
+					adev_to_drm(adev)->primary-
 >index);
 
 	ret = perf_pmu_register(&pmu_entry->pmu, pmu_name, -1);
 
-	if (ret) {
-		kfree(pmu_entry);
-		pr_warn("Error initializing AMDGPU %s PMUs.\n",
 pmu_type_name);
-		return ret;
-	}
+	if (ret)
+		goto err;
 
 	pr_info("Detected AMDGPU %s Counters. # of Counters = %d.\n",
-			pmu_type_name, num_counters);
+			pmu_type_name, num_events);
 
 	list_add_tail(&pmu_entry->entry, &amdgpu_pmu_list);
 
 	return 0;
+err:
+	kfree(pmu_entry);
+	pr_warn("Error initializing AMDGPU %s PMUs.\n", pmu_type_name);
+	return ret;
+}
+
+/* destroy all pmu data associated with target device */ void
+amdgpu_pmu_fini(struct amdgpu_device *adev) {
+	struct amdgpu_pmu_entry *pe, *temp;
+
+	list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
+		if (pe->adev == adev) {
+			list_del(&pe->entry);
+			perf_pmu_unregister(&pe->pmu);
+			kfree(pe);
+		}
+	}
+}
+
+static int amdgpu_pmu_alloc_pmu_attrs(
+				struct attribute_group *fmt_attr_group,
+				struct amdgpu_pmu_event_attribute
 **fmt_attr,
+				int fmt_num_attrs,
+				struct attribute_group *evt_attr_group,
+				struct amdgpu_pmu_event_attribute
 **evt_attr,
+				int evt_num_attrs)
+{
 
Quoted 1 line

where do all these allocations get freed?

Thanks for the mem leak catch. I'll free these up in amdgpu_pmu_fini after pmu unregister and fix the bad ENOMEM condition checks below.

Jon

Quoted 270 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	*fmt_attr = kcalloc(fmt_num_attrs, sizeof(**fmt_attr),

GFP_KERNEL);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
+
+	if (!fmt_attr)

I think you want to check *fmt_attr

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +4 -0
+		return -ENOMEM;
+
+	fmt_attr_group->attrs = kcalloc(fmt_num_attrs + 1,
+				sizeof(*fmt_attr_group->attrs),

GFP_KERNEL);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +7 -0
+
+	if (!fmt_attr_group->attrs)
+		goto err_fmt_attr_grp;
+
+	*evt_attr = kcalloc(evt_num_attrs, sizeof(**evt_attr), GFP_KERNEL);
+
+	if (!evt_attr)

I think you want to check *evt_attr

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +17 -0
+		goto err_evt_attr;
+
+	evt_attr_group->attrs = kcalloc(evt_num_attrs + 1,
+				sizeof(*evt_attr_group->attrs), GFP_KERNEL);
+
+	if (!evt_attr_group->attrs)
+		goto err_evt_attr_grp;
+
+	return 0;
+err_evt_attr_grp:
+	kfree(*evt_attr);
+err_evt_attr:
+	kfree(fmt_attr_group->attrs);
+err_fmt_attr_grp:
+	kfree(*fmt_attr);
+	return -ENOMEM;
 }
 
 /* init amdgpu_pmu */
 int amdgpu_pmu_init(struct amdgpu_device *adev)  {
+	struct amdgpu_pmu_event_attribute *fmt_attr = NULL, *evt_attr =

NULL,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +5 -0
+				*old_fmt_attr = NULL, *old_evt_attr = NULL;
 	int ret = 0;
 
+	if (adev->asic_type != CHIP_VEGA20)
+		goto init_events;
+
+	ret =

amdgpu_pmu_alloc_pmu_attrs(&df_vega20_format_attr_group,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +11 -0
+					&old_fmt_attr,
+					NUM_FORMATS_DF_LEGACY,
+					&df_vega20_event_attr_group,
+					&old_evt_attr,
+					NUM_EVENTS_DF_LEGACY);
+
+	if (ret)
+		return ret;
+
+	ret = init_pmu_by_type(adev,
+				&df_vega20_format_attr_group,

old_fmt_attr,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
+				&df_vega20_event_attr_group, old_evt_attr,
+				"DF", "amdgpu_df",

PERF_TYPE_AMDGPU_DF_LEGACY);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +4 -3
+	if (ret)
+		goto err_old_pmu;
+
+init_events:
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
-		/* init df */
-		ret = init_pmu_by_type(adev, df_v3_6_attr_groups,
-				       "DF", "amdgpu_df",

PERF_TYPE_AMDGPU_DF,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -1
-				       DF_V3_6_MAX_COUNTERS);
+		ret =

amdgpu_pmu_alloc_pmu_attrs(&amdgpu_pmu_format_attr_group,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
+						&fmt_attr,
+

NUM_FORMATS_AMDGPU_PMU,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +3 -0
+						&vega20_event_attr_group,
+						&evt_attr,
+
	NUM_EVENTS_VEGA20_MAX);
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +6 -0
+
+		if (ret)
+			goto err_alloc;
+
+		ret = init_pmu_by_type(adev,
+				&amdgpu_pmu_format_attr_group,

fmt_attr,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
+				&vega20_event_attr_group, evt_attr,
+				"Event", "amdgpu",

PERF_TYPE_AMDGPU_MAX);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +6 -6
+
+		if (ret) {
+			kfree(vega20_event_attr_group.attrs);
+			goto err_pmu;
+		}
 
-		/* other pmu types go here*/
 		break;
 	default:
 		return 0;
-	}
+	};
 
 	return 0;
-}
-
-
-/* destroy all pmu data associated with target device */ -void

amdgpu_pmu_fini(struct amdgpu_device *adev) -{

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +14 -9
-	struct amdgpu_pmu_entry *pe, *temp;
-
-	list_for_each_entry_safe(pe, temp, &amdgpu_pmu_list, entry) {
-		if (pe->adev == adev) {
-			list_del(&pe->entry);
-			perf_pmu_unregister(&pe->pmu);
-			kfree(pe);
-		}
-	}
+err_pmu:
+	kfree(fmt_attr);
+	kfree(evt_attr);
+	kfree(amdgpu_pmu_format_attr_group.attrs);
+err_alloc:
+	if (adev->asic_type == CHIP_VEGA20)
+		amdgpu_pmu_fini(adev);
+	return ret;
+err_old_pmu:
+	kfree(old_fmt_attr);
+	kfree(old_evt_attr);
+	kfree(df_vega20_format_attr_group.attrs);
+	kfree(df_vega20_event_attr_group.attrs);
+	return ret;
 }
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.h +5 -1
index 7dddb7160a11..0d214abe720e 100644
@@ -27,10 +27,14 @@
 #define _AMDGPU_PMU_H_
 
 enum amdgpu_pmu_perf_type {
-	PERF_TYPE_AMDGPU_DF = 0,
+	PERF_TYPE_AMDGPU_DF_LEGACY = 0,
+	PERF_TYPE_AMDGPU_XGMI,
 	PERF_TYPE_AMDGPU_MAX
 };
 
+#define AMDGPU_PERF_TYPE_SHIFT	56
+#define AMDGPU_PERF_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 +9 -63
index 0ca6e176acb0..6e57ae95f997 100644
@@ -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 +0 -10
index 76998541bc30..2505c7ef258a 100644
@@ -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;
 
--
Quoted 1 line

2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#6 lore raw

RE: [PATCH 3/3] drm/amdgpu: add xgmi perfmons for arcturus

In reply to Kasiviswanathan, Harish

Hi Harish. Thanks for the review. As discussed offline, patch 2 has a problem where attr groups array is global but allocation is done per-device causing problems with mem free and pmu unregister. I'm sending out a second series that should fix this and simplify the solution as well as hopefully address your concerns. Sorry again for the churn.

Jon

Quoted 163 lines

-----Original Message-----
From: Kasiviswanathan, Harish <[email protected]>
Sent: Tuesday, September 15, 2020 10:10 PM
To: Kim, Jonathan <[email protected]>; amd- [email protected]
Subject: RE: [PATCH 3/3] drm/amdgpu: add xgmi perfmons for arcturus

[AMD Official Use Only - Internal Distribution Only]

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Tuesday, September 15, 2020 6:00 PM
To: [email protected]
Cc: Kasiviswanathan, Harish <[email protected]>; Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 3/3] drm/amdgpu: add xgmi perfmons for arcturus

Add xgmi perfmons for Arcturus.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 55
+++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    |  3 ++
 2 files changed, 58 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
index f3d2ac0e88a7..ec521c72e631 100644
@@ -34,6 +34,8 @@
 #define NUM_EVENTS_DF_LEGACY		8
 #define NUM_EVENTS_VEGA20_XGMI		2
 #define NUM_EVENTS_VEGA20_MAX		2
+#define NUM_EVENTS_ARCTURUS_XGMI	6
+#define NUM_EVENTS_ARCTURUS_MAX		6
 
 /* record to keep track of pmu entry per pmu type per device */  struct

amdgpu_pmu_entry { @@ -110,6 +112,27 @@ const struct attribute_group *vega20_attr_groups[] = {

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +3 -0
 	NULL
 };
 
+/* Arcturus events */
+static const char *arcturus_events[NUM_EVENTS_ARCTURUS_MAX][2] = {
+	{ "xgmi_link0_data_outbound",

"event=0x7,instance=0x4b,umask=0x2" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	{ "xgmi_link1_data_outbound",

"event=0x7,instance=0x4c,umask=0x2" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	{ "xgmi_link2_data_outbound",

"event=0x7,instance=0x4d,umask=0x2" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	{ "xgmi_link3_data_outbound",

"event=0x7,instance=0x4e,umask=0x2" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	{ "xgmi_link4_data_outbound",

"event=0x7,instance=0x4f,umask=0x2" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +1 -0
+	{ "xgmi_link5_data_outbound",

"event=0x7,instance=0x50,umask=0x2" } };

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +12 -0
+
+static struct attribute_group arcturus_event_attr_group = {
+	.name = "events",
+	.attrs = NULL
+};
+
+const struct attribute_group *arcturus_attr_groups[] = {
+	&amdgpu_pmu_format_attr_group,
+	&arcturus_event_attr_group,
+	NULL
+};
+
 /* All df_vega20_* items are DEPRECATED. Use vega20_ items above

instead. */ static const char
*df_vega20_formats[NUM_FORMATS_DF_LEGACY][2] = {

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +10 -0
 	{ "event", "config:0-7" },
@@ -400,6 +423,16 @@ static int init_pmu_by_type(struct amdgpu_device
 *adev,
 
 		pmu_entry->pmu.attr_groups = vega20_attr_groups;
 		break;
+	case CHIP_ARCTURUS:
+		amdgpu_pmu_create_attributes(evt_attr_group, evt_attr,
+				arcturus_events, 0,
 NUM_EVENTS_ARCTURUS_XGMI,
+				PERF_TYPE_AMDGPU_XGMI);
+		num_events += NUM_EVENTS_ARCTURUS_XGMI;
+
+		/* other events can be added here */
+
+		pmu_entry->pmu.attr_groups = arcturus_attr_groups;
+		break;
 	default:
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +14 -0
 		return -ENODEV;
 	};
@@ -530,6 +563,28 @@ int amdgpu_pmu_init(struct amdgpu_device *adev)
 			goto err_pmu;
 		}
 
+		break;
+	case CHIP_ARCTURUS:
+		ret =
 amdgpu_pmu_alloc_pmu_attrs(&amdgpu_pmu_format_attr_group,
+						&fmt_attr,
+
 	NUM_FORMATS_AMDGPU_PMU,
+						&arcturus_event_attr_group,
+						&evt_attr,
+
 	NUM_EVENTS_ARCTURUS_MAX);
+
+		if (ret)
+			goto err_alloc;
+
+		ret = init_pmu_by_type(adev,
+				&amdgpu_pmu_format_attr_group,

fmt_attr,

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -0
+				&arcturus_event_attr_group, evt_attr,
+				"Event", "amdgpu",

PERF_TYPE_AMDGPU_MAX);

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +6 -0
+
+		if (ret) {
+			kfree(arcturus_event_attr_group.attrs);
+			goto err_pmu;
+		}
+
 		break;
 	default:
 		return 0;
drivers/gpu/drm/amd/amdgpu/df_v3_6.c +2 -0
index 6e57ae95f997..6b4b30a8dce5 100644
@@ -513,6 +513,7 @@ static int df_v3_6_pmc_start(struct amdgpu_device
 *adev, uint64_t config,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		if (is_add)
 			return df_v3_6_pmc_add_cntr(adev, config);
@@ -554,6 +555,7 @@ static int df_v3_6_pmc_stop(struct amdgpu_device
 *adev, uint64_t config,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		ret = df_v3_6_pmc_get_ctrl_settings(adev,
 			config,
drivers/gpu/drm/amd/amdgpu/df_v3_6.c +1 -0
 			counter_idx,
@@ -590,6 +592,7 @@ static void df_v3_6_pmc_get_count(struct
 amdgpu_device *adev,
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
+	case CHIP_ARCTURUS:
 		df_v3_6_pmc_get_read_settings(adev, config, counter_idx,
 						&lo_base_addr,

&hi_base_addr);

Quoted 2 lines

--
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx