Skip to contentSkip to conversation

drm/amdgpu: stop resetting xgmi perfmons on disable

PATCH In review Jonathan Kim 2 participants 7 messages Sep 8, 2020 – Sep 15, 2020

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

PATCH 1/4 drm/amdgpu: stop resetting xgmi perfmons on disable

Disabling perf events does not specify reset in ABI so stop doing it in hardware.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_df.h |  4 ++--
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c   | 23 ++++++++++++++---------
 2 files changed, 16 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_df.h +2 -2
index 61a26c15c8dd..373cdebe0e2f 100644
@@ -44,9 +44,9 @@ 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_enable);
+					 int is_add);
 	int (*pmc_stop)(struct amdgpu_device *adev, uint64_t config,
-					 int is_disable);
+					 int is_remove);
 	void (*pmc_get_count)(struct amdgpu_device *adev, uint64_t config,
 					 uint64_t *count);
 	uint64_t (*get_fica)(struct amdgpu_device *adev, uint32_t ficaa_val);
drivers/gpu/drm/amd/amdgpu/df_v3_6.c +14 -9
index 2eab808fffeb..7b89fd2aa44a 100644
@@ -455,7 +455,8 @@ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
 					  uint32_t *lo_base_addr,
 					  uint32_t *hi_base_addr,
 					  uint32_t *lo_val,
-					  uint32_t *hi_val)
+					  uint32_t *hi_val,
+					  bool is_enable)
 {
 
 	uint32_t eventsel, instance, unitmask;
@@ -477,7 +478,8 @@ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
 	instance_5432 = (instance >> 2) & 0xf;
 	instance_76 = (instance >> 6) & 0x3;
 
-	*lo_val = (unitmask << 8) | (instance_10 << 6) | eventsel | (1 << 22);
+	*lo_val = (unitmask << 8) | (instance_10 << 6) | eventsel;
+	*lo_val = is_enable ? *lo_val | (1 << 22) : *lo_val & ~(1 << 22);
 	*hi_val = (instance_76 << 29) | instance_5432;
 
 	DRM_DEBUG_DRIVER("config=%llx addr=%08x:%08x val=%08x:%08x",
@@ -572,14 +574,14 @@ static void df_v3_6_reset_perfmon_cntr(struct amdgpu_device *adev,
 }
 
 static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
-			     int is_enable)
+			     int is_add)
 {
 	uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
 	int err = 0, ret = 0;
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
-		if (is_enable)
+		if (is_add)
 			return df_v3_6_pmc_add_cntr(adev, config);
 
 		df_v3_6_reset_perfmon_cntr(adev, config);
@@ -589,7 +591,8 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
 					&lo_base_addr,
 					&hi_base_addr,
 					&lo_val,
-					&hi_val);
+					&hi_val,
+					true);
 
 		if (ret)
 			return ret;
@@ -612,7 +615,7 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
 }
 
 static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
-			    int is_disable)
+			    int is_remove)
 {
 	uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
 	int ret = 0;
@@ -624,15 +627,17 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
 			&lo_base_addr,
 			&hi_base_addr,
 			&lo_val,
-			&hi_val);
+			&hi_val,
+			false);
 
 		if (ret)
 			return ret;
 
-		df_v3_6_reset_perfmon_cntr(adev, config);
 
-		if (is_disable)
+		if (is_remove) {
+			df_v3_6_reset_perfmon_cntr(adev, config);
 			df_v3_6_pmc_release_cntr(adev, config);
+		}
 
 		break;
 	default:
Signature · 6 lines
2.17.1

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

PATCH 2/4 drm/amdgpu: fix xgmi perfmon a-b-a problem

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

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_df.h  |   6 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c |  23 ++++--
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    | 104 +++++++++++-------------
 3 files changed, 65 insertions(+), 68 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 +16 -7
index 69af462db34d..915c580d30be 100644
@@ -74,9 +74,11 @@ static void amdgpu_perf_start(struct perf_event *event, int flags)
 	switch (pe->pmu_perf_type) {
 	case PERF_TYPE_AMDGPU_DF:
 		if (!(flags & PERF_EF_RELOAD))
-			pe->adev->df.funcs->pmc_start(pe->adev, hwc->config, 1);
+			pe->adev->df.funcs->pmc_start(pe->adev, hwc->config,
+								hwc->idx, 1);
 
-		pe->adev->df.funcs->pmc_start(pe->adev, hwc->config, 0);
+		pe->adev->df.funcs->pmc_start(pe->adev, hwc->config,
+								hwc->idx, 0);
 		break;
 	default:
 		break;
@@ -101,8 +103,8 @@ static void amdgpu_perf_read(struct perf_event *event)
 
 		switch (pe->pmu_perf_type) {
 		case PERF_TYPE_AMDGPU_DF:
-			pe->adev->df.funcs->pmc_get_count(pe->adev, hwc->config,
-							  &count);
+			pe->adev->df.funcs->pmc_get_count(pe->adev,
+						hwc->config, hwc->idx, &count);
 			break;
 		default:
 			count = 0;
@@ -126,7 +128,8 @@ static void amdgpu_perf_stop(struct perf_event *event, int flags)
 
 	switch (pe->pmu_perf_type) {
 	case PERF_TYPE_AMDGPU_DF:
-		pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, 0);
+		pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx,
+									0);
 		break;
 	default:
 		break;
@@ -157,7 +160,12 @@ static int amdgpu_perf_add(struct perf_event *event, int flags)
 	switch (pe->pmu_perf_type) {
 	case PERF_TYPE_AMDGPU_DF:
 		retval = pe->adev->df.funcs->pmc_start(pe->adev,
-						       hwc->config, 1);
+						hwc->config, hwc->idx, 1);
+		if (retval >= 0) {
+			hwc->idx = retval;
+			retval = 0;
+		}
+
 		break;
 	default:
 		return 0;
@@ -185,7 +193,8 @@ static void amdgpu_perf_del(struct perf_event *event, int flags)
 
 	switch (pe->pmu_perf_type) {
 	case PERF_TYPE_AMDGPU_DF:
-		pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, 1);
+		pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx,
+									1);
 		break;
 	default:
 		break;
drivers/gpu/drm/amd/amdgpu/df_v3_6.c +46 -58

250 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
#2 lore raw

RE: [PATCH 2/4] drm/amdgpu: fix xgmi perfmon a-b-a problem

[AMD Official Use Only - Internal Distribution Only]

Few minor comments.

-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Jonathan Kim
Sent: Tuesday, September 8, 2020 9:06 AM
To: [email protected]
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 2/4] drm/amdgpu: fix xgmi perfmon a-b-a problem

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

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_df.h  |   6 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c |  23 ++++--
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c    | 104 +++++++++++-------------
 3 files changed, 65 insertions(+), 68 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); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c

index 69af462db34d..915c580d30be 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +2 -1
@@ -74,9 +74,11 @@ static void amdgpu_perf_start(struct perf_event *event, int flags)
 	switch (pe->pmu_perf_type) {
 	case PERF_TYPE_AMDGPU_DF:
 		if (!(flags & PERF_EF_RELOAD))
-			pe->adev->df.funcs->pmc_start(pe->adev, hwc->config, 1);
+			pe->adev->df.funcs->pmc_start(pe->adev, hwc->config,
+								hwc->idx, 1);
 

The previous pmc_start() can fail if there is no slot available. The code will still continue.

- pe->adev->df.funcs->pmc_start(pe->adev, hwc->config, 0);
+ pe->adev->df.funcs->pmc_start(pe->adev, hwc->config,
+ hwc->idx, 0);
break;
default:
break;
@@ -101,8 +103,8 @@ static void amdgpu_perf_read(struct perf_event *event)

switch (pe->pmu_perf_type) {
case PERF_TYPE_AMDGPU_DF:
- pe->adev->df.funcs->pmc_get_count(pe->adev, hwc->config,
- &count);
+ pe->adev->df.funcs->pmc_get_count(pe->adev,
+ hwc->config, hwc->idx, &count);
break;
default:
count = 0;
@@ -126,7 +128,8 @@ static void amdgpu_perf_stop(struct perf_event *event, int flags)

switch (pe->pmu_perf_type) {
case PERF_TYPE_AMDGPU_DF:
- pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, 0);
+ pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx,
+ 0);
break;
default:
break;
@@ -157,7 +160,12 @@ static int amdgpu_perf_add(struct perf_event *event, int flags)
switch (pe->pmu_perf_type) {
case PERF_TYPE_AMDGPU_DF:
retval = pe->adev->df.funcs->pmc_start(pe->adev,
- hwc->config, 1);
+ hwc->config, hwc->idx, 1);

Passing hwc->idx to pmc_start() is confusing as that variable is not used in this case. Either add /*used*/ comment and/or pass 0 with /*used*/ comment. And may be add a small comment saying that when "is_add" == 1, the function returns a counter slot.

+ if (retval >= 0) {
+ hwc->idx = retval;
+ retval = 0;
+ }
+
break;
default:
return 0;
@@ -185,7 +193,8 @@ static void amdgpu_perf_del(struct perf_event *event, int flags)

switch (pe->pmu_perf_type) {
case PERF_TYPE_AMDGPU_DF:
- pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, 1);
+ pe->adev->df.funcs->pmc_stop(pe->adev, hwc->config, hwc->idx,
+ 1);
break;
default:
break;

drivers/gpu/drm/amd/amdgpu/df_v3_6.c +8 -12
index 7b89fd2aa44a..8dadcdcffba0 100644
@@ -391,33 +391,28 @@ static void df_v3_6_get_clockgating_state(struct amdgpu_device *adev,  }
 
 /* get assigned df perfmon ctr as int */ -static int df_v3_6_pmc_config_2_cntr(struct amdgpu_device *adev,
-				      uint64_t config)
+static bool df_v3_6_pmc_has_counter(struct amdgpu_device *adev,
+				      uint64_t config,
+				      int counter_idx)
 {
-	int i;
 
-	for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) {
-		if ((config & 0x0FFFFFFUL) ==
-					adev->df_perfmon_config_assign_mask[i])
-			return i;
-	}
+	return ((config & 0x0FFFFFFUL) ==
+			adev->df_perfmon_config_assign_mask[counter_idx]);
 
-	return -EINVAL;
 }
 
 /* get address based on counter assignment */  static void df_v3_6_pmc_get_addr(struct amdgpu_device *adev,
 				 uint64_t config,
+				 int counter_idx,
 				 int is_ctrl,
 				 uint32_t *lo_base_addr,
 				 uint32_t *hi_base_addr)
 {
-	int target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
-
-	if (target_cntr < 0)
+	if (!df_v3_6_pmc_has_counter(adev, config, counter_idx))
 		return;
 
-	switch (target_cntr) {
+	switch (counter_idx) {
 
 	case 0:
 		*lo_base_addr = is_ctrl ? smnPerfMonCtlLo4 : smnPerfMonCtrLo4; @@ -443,15 +438,18 @@ static void df_v3_6_pmc_get_addr(struct amdgpu_device *adev,
 /* get read counter address */
 static void df_v3_6_pmc_get_read_settings(struct amdgpu_device *adev,

uint64_t config,
+ int counter_idx,
uint32_t *lo_base_addr, uint32_t *hi_base_addr)
{
- df_v3_6_pmc_get_addr(adev, config, 0, lo_base_addr, hi_base_addr);
+ df_v3_6_pmc_get_addr(adev, config, counter_idx, 0, lo_base_addr,
+ hi_base_addr);
}

/* get control counter settings i.e. address and values to set */ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
uint64_t config,
+ int counter_idx,
uint32_t *lo_base_addr,
uint32_t *hi_base_addr,
uint32_t *lo_val,
@@ -462,7 +460,8 @@ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
uint32_t eventsel, instance, unitmask;
uint32_t instance_10, instance_5432, instance_76;

- df_v3_6_pmc_get_addr(adev, config, 1, lo_base_addr, hi_base_addr);
+ df_v3_6_pmc_get_addr(adev, config, counter_idx, 1, lo_base_addr,
+ hi_base_addr);

if ((*lo_base_addr == 0) || (*hi_base_addr == 0)) {
DRM_ERROR("[DF PMC] addressing not retrieved! Lo: %x, Hi: %x", @@ -492,18 +491,13 @@ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev, static int df_v3_6_pmc_add_cntr(struct amdgpu_device *adev,
uint64_t config)
{
- int i, target_cntr;
-
- target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
-
- if (target_cntr >= 0)
- return 0;
+ int i;

for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) {
if (adev->df_perfmon_config_assign_mask[i] == 0U) {
adev->df_perfmon_config_assign_mask[i] =
config & 0x0FFFFFFUL;
- return 0;
+ return i;
}
}

@@ -512,59 +506,50 @@ static int df_v3_6_pmc_add_cntr(struct amdgpu_device *adev,

#define DEFERRED_ARM_MASK (1 << 31)
static int df_v3_6_pmc_set_deferred(struct amdgpu_device *adev,

-				    uint64_t config, bool is_deferred)
+				    int counter_idx, uint64_t config,
+				    bool is_deferred)

{
- int target_cntr;
-
- target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);

- if (target_cntr < 0)
+ if (!df_v3_6_pmc_has_counter(adev, config, counter_idx))
return -EINVAL;

if (is_deferred)
- adev->df_perfmon_config_assign_mask[target_cntr] |=
+ adev->df_perfmon_config_assign_mask[counter_idx] |=
DEFERRED_ARM_MASK;
else
- adev->df_perfmon_config_assign_mask[target_cntr] &=
+ adev->df_perfmon_config_assign_mask[counter_idx] &=
~DEFERRED_ARM_MASK;

return 0;
}

static bool df_v3_6_pmc_is_deferred(struct amdgpu_device *adev,
+ int counter_idx,
uint64_t config)
{
- int target_cntr;
-
- target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
-
- /*
- * we never get target_cntr < 0 since this funciton is only called in
- * pmc_count for now but we should check anyways.
- */
- return (target_cntr >= 0 &&
- (adev->df_perfmon_config_assign_mask[target_cntr]
- & DEFERRED_ARM_MASK));
+ return (df_v3_6_pmc_has_counter(adev, config, counter_idx) &&
+ (adev->df_perfmon_config_assign_mask[counter_idx]
+ & DEFERRED_ARM_MASK));

}

/* release performance counter */
static void df_v3_6_pmc_release_cntr(struct amdgpu_device *adev,

-				     uint64_t config)
+				     uint64_t config,
+				     int counter_idx)

{
- int target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
-
- if (target_cntr >= 0)
- adev->df_perfmon_config_assign_mask[target_cntr] = 0ULL;
+ if (df_v3_6_pmc_has_counter(adev, config, counter_idx))
+ adev->df_perfmon_config_assign_mask[counter_idx] = 0ULL;
}

static void df_v3_6_reset_perfmon_cntr(struct amdgpu_device *adev,
- uint64_t config)
+ uint64_t config,
+ int counter_idx)
{
uint32_t lo_base_addr = 0, hi_base_addr = 0;

- df_v3_6_pmc_get_read_settings(adev, config, &lo_base_addr,
+ df_v3_6_pmc_get_read_settings(adev, config, counter_idx,
+&lo_base_addr,
&hi_base_addr);

if ((lo_base_addr == 0) || (hi_base_addr == 0)) @@ -574,7 +559,7 @@ static void df_v3_6_reset_perfmon_cntr(struct amdgpu_device *adev, }

static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
- int is_add)
+ int counter_idx, int is_add)
{
uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
int err = 0, ret = 0;
@@ -584,10 +569,9 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
if (is_add)
return df_v3_6_pmc_add_cntr(adev, config);

- df_v3_6_reset_perfmon_cntr(adev, config);
-
ret = df_v3_6_pmc_get_ctrl_settings(adev,
config,
+ counter_idx,
&lo_base_addr,
&hi_base_addr,
&lo_val,
@@ -604,7 +588,8 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
hi_val);

if (err)
- ret = df_v3_6_pmc_set_deferred(adev, config, true);
+ ret = df_v3_6_pmc_set_deferred(adev, config,
+ counter_idx, true);

break;
default:
@@ -615,7 +600,7 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config, }

static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
- int is_remove)
+ int counter_idx, int is_remove)
{
uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
int ret = 0;
@@ -624,6 +609,7 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
case CHIP_VEGA20:
ret = df_v3_6_pmc_get_ctrl_settings(adev,
config,
+ counter_idx,
&lo_base_addr,
&hi_base_addr,
&lo_val,
@@ -635,8 +621,8 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,

if (is_remove) {
- df_v3_6_reset_perfmon_cntr(adev, config);
- df_v3_6_pmc_release_cntr(adev, config);
+ df_v3_6_reset_perfmon_cntr(adev, config, counter_idx);
+ df_v3_6_pmc_release_cntr(adev, config, counter_idx);
}

break;
@@ -649,6 +635,7 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,

static void df_v3_6_pmc_get_count(struct amdgpu_device *adev,
uint64_t config,
+ int counter_idx,
uint64_t *count)
{
uint32_t lo_base_addr = 0, hi_base_addr = 0, lo_val = 0, hi_val = 0; @@ -656,14 +643,14 @@ static void df_v3_6_pmc_get_count(struct amdgpu_device *adev,

switch (adev->asic_type) {
case CHIP_VEGA20:
- df_v3_6_pmc_get_read_settings(adev, config, &lo_base_addr,
- &hi_base_addr);
+ df_v3_6_pmc_get_read_settings(adev, config, counter_idx,
+ &lo_base_addr, &hi_base_addr);

if ((lo_base_addr == 0) || (hi_base_addr == 0))
return;

/* rearm the counter or throw away count value on failure */
- if (df_v3_6_pmc_is_deferred(adev, config)) {
+ if (df_v3_6_pmc_is_deferred(adev, config, counter_idx)) {
int rearm_err = df_v3_6_perfmon_arm_with_status(adev,
lo_base_addr, lo_val, hi_base_addr, hi_val);
@@ -671,7 +658,8 @@ static void df_v3_6_pmc_get_count(struct amdgpu_device *adev,
if (rearm_err)
return;

- df_v3_6_pmc_set_deferred(adev, config, false);
+ df_v3_6_pmc_set_deferred(adev, config, counter_idx,
+ false);
}

df_v3_6_perfmon_rreg(adev, lo_base_addr, &lo_val,
--
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Charish.kasiviswanathan%40amd.com%7Cdd87ca2efb7547914bba08d853f80f67%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637351672175311161&amp;sdata=5LsHUQogyAkmCKwgLsmuyNGsK%2BB9ng5T%2F9d45CPc%2BcA%3D&amp;reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#3 lore raw

PATCH 3/4 drm/amdgpu: add per device user friendly xgmi events for vega20

In reply to Jonathan Kim

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.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h     |  13 -
 drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c | 327 +++++++++++++++++++++---
 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, 299 insertions(+), 128 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -13
index 49ea9fa780c4..0cb2f12a1fbd 100644
@@ -1265,19 +1265,6 @@ void amdgpu_unregister_gpu_instance(struct amdgpu_device *adev);
 
 #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 +285 -42

417 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 4/4 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 82f57bd38716..4adf9c6e3944 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 {
@@ -95,6 +97,27 @@ 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" },
@@ -380,6 +403,16 @@ 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;
 	};
@@ -510,6 +543,28 @@ 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 569c40be6e75..23af431de997 100644
@@ -512,6 +512,7 @@ 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);
 
@@ -553,6 +554,7 @@ 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,
@@ -589,6 +591,7 @@ 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
#5 lore raw

RE: [PATCH 4/4] drm/amdgpu: add xgmi perfmons for arcturus

[AMD Official Use Only - Internal Distribution Only]

Ping.

Thanks,

Jon

Quoted 151 lines

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Tuesday, September 8, 2020 9:07 AM
To: [email protected]
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 4/4] 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 82f57bd38716..4adf9c6e3944 100644
@@ -34,6 +34,8 @@
 #define NUM_EVENTS_DF_LEGACY8
 #define NUM_EVENTS_VEGA20_XGMI2
 #define NUM_EVENTS_VEGA20_MAX2
+#define NUM_EVENTS_ARCTURUS_XGMI6
+#define NUM_EVENTS_ARCTURUS_MAX6
 
 /* record to keep track of pmu entry per pmu type per device */  struct

amdgpu_pmu_entry { @@ -95,6 +97,27 @@ const struct attribute_group *vega20_attr_groups[] = {
NULL
};

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +3 -0
+/* 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" } };

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] = {
{ "event", "config:0-7" },

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +10 -0
@@ -380,6 +403,16 @@ 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;
};

drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c +14 -0
@@ -510,6 +543,28 @@ 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);

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 569c40be6e75..23af431de997 100644
@@ -512,6 +512,7 @@ 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);
@@ -553,6 +554,7 @@ 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,

drivers/gpu/drm/amd/amdgpu/df_v3_6.c +1 -0
@@ -589,6 +591,7 @@ 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

Reviewed-by on v1 1/4
#6 lore raw
In reply to Jonathan Kim

[AMD Official Use Only - Internal Distribution Only]

-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Jonathan Kim
Sent: Tuesday, September 8, 2020 9:06 AM
To: [email protected]
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 1/4] drm/amdgpu: stop resetting xgmi perfmons on disable

Disabling perf events does not specify reset in ABI so stop doing it in hardware.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_df.h |  4 ++--
 drivers/gpu/drm/amd/amdgpu/df_v3_6.c   | 23 ++++++++++++++---------
 2 files changed, 16 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_df.h +2 -2
index 61a26c15c8dd..373cdebe0e2f 100644
@@ -44,9 +44,9 @@ 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_enable);
+					 int is_add);
 	int (*pmc_stop)(struct amdgpu_device *adev, uint64_t config,
-					 int is_disable);
+					 int is_remove);
 	void (*pmc_get_count)(struct amdgpu_device *adev, uint64_t config,
 					 uint64_t *count);
 	uint64_t (*get_fica)(struct amdgpu_device *adev, uint32_t ficaa_val);
drivers/gpu/drm/amd/amdgpu/df_v3_6.c +14 -9
index 2eab808fffeb..7b89fd2aa44a 100644
@@ -455,7 +455,8 @@ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
 					  uint32_t *lo_base_addr,
 					  uint32_t *hi_base_addr,
 					  uint32_t *lo_val,
-					  uint32_t *hi_val)
+					  uint32_t *hi_val,
+					  bool is_enable)
 {
 
 	uint32_t eventsel, instance, unitmask;
@@ -477,7 +478,8 @@ static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
 	instance_5432 = (instance >> 2) & 0xf;
 	instance_76 = (instance >> 6) & 0x3;
 
-	*lo_val = (unitmask << 8) | (instance_10 << 6) | eventsel | (1 << 22);
+	*lo_val = (unitmask << 8) | (instance_10 << 6) | eventsel;
+	*lo_val = is_enable ? *lo_val | (1 << 22) : *lo_val & ~(1 << 22);
 	*hi_val = (instance_76 << 29) | instance_5432;
 
 	DRM_DEBUG_DRIVER("config=%llx addr=%08x:%08x val=%08x:%08x",
@@ -572,14 +574,14 @@ static void df_v3_6_reset_perfmon_cntr(struct amdgpu_device *adev,
 }
 
 static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
-			     int is_enable)
+			     int is_add)
 {
 	uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
 	int err = 0, ret = 0;
 
 	switch (adev->asic_type) {
 	case CHIP_VEGA20:
-		if (is_enable)
+		if (is_add)
 			return df_v3_6_pmc_add_cntr(adev, config);
 
 		df_v3_6_reset_perfmon_cntr(adev, config);
@@ -589,7 +591,8 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
 					&lo_base_addr,
 					&hi_base_addr,
 					&lo_val,
-					&hi_val);
+					&hi_val,
+					true);
 
 		if (ret)
 			return ret;
@@ -612,7 +615,7 @@ static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
 }
 
 static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
-			    int is_disable)
+			    int is_remove)
 {
 	uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
 	int ret = 0;
@@ -624,15 +627,17 @@ static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
 			&lo_base_addr,
 			&hi_base_addr,
 			&lo_val,
-			&hi_val);
+			&hi_val,
+			false);
 
 		if (ret)
 			return ret;
 
-		df_v3_6_reset_perfmon_cntr(adev, config);
 
-		if (is_disable)
+		if (is_remove) {
+			df_v3_6_reset_perfmon_cntr(adev, config);
 			df_v3_6_pmc_release_cntr(adev, config);
+		}
 
 		break;
 	default:
Signature · 10 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Charish.kasiviswanathan%40amd.com%7Cbaec81e867af49b73e8d08d853f80fa2%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637351672173732796&amp;sdata=cd5%2BhKe%2Fz9QWs0QIH5zaTG3aWyLOUsOuViawyz9awNs%3D&amp;reserved=0
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx