[PATCH V2] accel/amdxdna: Support sensors for column utilization

[PATCH V2] accel/amdxdna: Support sensors for column utilization

[Lizhi Hou]

From: "Mario Limonciello (AMD)" <superm1@kernel.org>

The AMD PMF driver provides realtime column utilization (npu_busy)
metrics for the NPU. Extend the DRM_IOCTL_AMDXDNA_GET_INFO sensor
query to expose these metrics to userspace.

Add AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION to the sensor type enum
and update aie2_get_sensors() to return both the total power and up
to 8 column utilization sensors if the user buffer permits.

Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
Reviewed-by: Lizhi Hou <lizhi.hou@amd.com>
[lizhi: support legacy tool which uses small buffer. checkpatch cleanup]
Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
---
 drivers/accel/amdxdna/aie2_pci.c | 34 +++++++++++++++++++++++++++-----
 drivers/accel/amdxdna/aie2_pci.h |  8 ++++++++
 include/uapi/drm/amdxdna_accel.h |  3 ++-
 3 files changed, 39 insertions(+), 6 deletions(-)

diff --git a/drivers/accel/amdxdna/aie2_pci.c b/drivers/accel/amdxdna/aie2_pci.c
index a2e586512e26..c57c785a2d15 100644
--- a/drivers/accel/amdxdna/aie2_pci.c
+++ b/drivers/accel/amdxdna/aie2_pci.c
@@ -787,16 +787,18 @@ static int aie2_get_clock_metadata(struct amdxdna_client *client,
 static int aie2_get_sensors(struct amdxdna_client *client,
 			    struct amdxdna_drm_get_info *args)
 {
+	struct amdxdna_dev_hdl *ndev = client->xdna->dev_handle;
 	struct amdxdna_drm_query_sensor sensor = {};
+	struct amd_pmf_npu_metrics npu_metrics;
+	u32 sensors_count = 0, i;
 	int ret;
 
-	if (args->buffer_size < sizeof(sensor))
-		return -EINVAL;
-
-	ret = AIE2_GET_PMF_NPU_DATA(npu_power, sensor.input);
+	ret = AIE2_GET_PMF_NPU_METRICS(&npu_metrics);
 	if (ret)
 		return ret;
+
 	sensor.type = AMDXDNA_SENSOR_TYPE_POWER;
+	sensor.input = npu_metrics.npu_power;
 	sensor.unitm = -3;
 	scnprintf(sensor.label, sizeof(sensor.label), "Total Power");
 	scnprintf(sensor.units, sizeof(sensor.units), "mW");
@@ -804,7 +806,29 @@ static int aie2_get_sensors(struct amdxdna_client *client,
 	if (copy_to_user(u64_to_user_ptr(args->buffer), &sensor, sizeof(sensor)))
 		return -EFAULT;
 
-	args->buffer_size = sizeof(sensor);
+	sensors_count++;
+	if (args->buffer_size <= sensors_count * sizeof(sensor))
+		goto out;
+
+	for (i = 0; i < min_t(u32, ndev->total_col, 8); i++) {
+		memset(&sensor, 0, sizeof(sensor));
+		sensor.input = npu_metrics.npu_busy[i];
+		sensor.type = AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION;
+		sensor.unitm = 0;
+		scnprintf(sensor.label, sizeof(sensor.label), "Column %d Utilization", i);
+		scnprintf(sensor.units, sizeof(sensor.units), "%%");
+
+		if (copy_to_user(u64_to_user_ptr(args->buffer) + sensors_count * sizeof(sensor),
+				 &sensor, sizeof(sensor)))
+			return -EFAULT;
+
+		sensors_count++;
+		if (args->buffer_size <= sensors_count * sizeof(sensor))
+			goto out;
+	}
+
+out:
+	args->buffer_size = sensors_count * sizeof(sensor);
 
 	return 0;
 }
diff --git a/drivers/accel/amdxdna/aie2_pci.h b/drivers/accel/amdxdna/aie2_pci.h
index 1bb88711bedb..0ae174862592 100644
--- a/drivers/accel/amdxdna/aie2_pci.h
+++ b/drivers/accel/amdxdna/aie2_pci.h
@@ -48,6 +48,7 @@
 })
 
 #if IS_ENABLED(CONFIG_AMD_PMF)
+#define AIE2_GET_PMF_NPU_METRICS(metrics) amd_pmf_get_npu_data(metrics)
 #define AIE2_GET_PMF_NPU_DATA(field, val)				\
 ({									\
 	struct amd_pmf_npu_metrics _npu_metrics;			\
@@ -58,6 +59,13 @@
 	(_ret);								\
 })
 #else
+#define AIE2_GET_PMF_NPU_METRICS(metrics)				\
+({									\
+	typeof(metrics) _m = metrics;					\
+	memset(_m, 0xff, sizeof(*_m));					\
+	(-EOPNOTSUPP);							\
+})
+
 #define SENSOR_DEFAULT_npu_power	U32_MAX
 #define AIE2_GET_PMF_NPU_DATA(field, val)				\
 ({									\
diff --git a/include/uapi/drm/amdxdna_accel.h b/include/uapi/drm/amdxdna_accel.h
index 9c44db2b3dcd..5bd13f4435f5 100644
--- a/include/uapi/drm/amdxdna_accel.h
+++ b/include/uapi/drm/amdxdna_accel.h
@@ -353,7 +353,8 @@ struct amdxdna_drm_query_clock_metadata {
 };
 
 enum amdxdna_sensor_type {
-	AMDXDNA_SENSOR_TYPE_POWER
+	AMDXDNA_SENSOR_TYPE_POWER,
+	AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION
 };
 
 /**
-- 
2.34.1

Re: [PATCH V2] accel/amdxdna: Support sensors for column utilization

[Mario Limonciello]

On 3/11/26 12:18, Lizhi Hou wrote:
> From: "Mario Limonciello (AMD)" <superm1@kernel.org>
> 
> The AMD PMF driver provides realtime column utilization (npu_busy)
> metrics for the NPU. Extend the DRM_IOCTL_AMDXDNA_GET_INFO sensor
> query to expose these metrics to userspace.
> 
> Add AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION to the sensor type enum
> and update aie2_get_sensors() to return both the total power and up
> to 8 column utilization sensors if the user buffer permits.
> 
> Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
> Reviewed-by: Lizhi Hou <lizhi.hou@amd.com>
> [lizhi: support legacy tool which uses small buffer. checkpatch cleanup]
> Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>

Thanks, no concerns on the cleanups.

> ---
>   drivers/accel/amdxdna/aie2_pci.c | 34 +++++++++++++++++++++++++++-----
>   drivers/accel/amdxdna/aie2_pci.h |  8 ++++++++
>   include/uapi/drm/amdxdna_accel.h |  3 ++-
>   3 files changed, 39 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/accel/amdxdna/aie2_pci.c b/drivers/accel/amdxdna/aie2_pci.c
> index a2e586512e26..c57c785a2d15 100644
> --- a/drivers/accel/amdxdna/aie2_pci.c
> +++ b/drivers/accel/amdxdna/aie2_pci.c
> @@ -787,16 +787,18 @@ static int aie2_get_clock_metadata(struct amdxdna_client *client,
>   static int aie2_get_sensors(struct amdxdna_client *client,
>   			    struct amdxdna_drm_get_info *args)
>   {
> +	struct amdxdna_dev_hdl *ndev = client->xdna->dev_handle;
>   	struct amdxdna_drm_query_sensor sensor = {};
> +	struct amd_pmf_npu_metrics npu_metrics;
> +	u32 sensors_count = 0, i;
>   	int ret;
>   
> -	if (args->buffer_size < sizeof(sensor))
> -		return -EINVAL;
> -
> -	ret = AIE2_GET_PMF_NPU_DATA(npu_power, sensor.input);
> +	ret = AIE2_GET_PMF_NPU_METRICS(&npu_metrics);
>   	if (ret)
>   		return ret;
> +
>   	sensor.type = AMDXDNA_SENSOR_TYPE_POWER;
> +	sensor.input = npu_metrics.npu_power;
>   	sensor.unitm = -3;
>   	scnprintf(sensor.label, sizeof(sensor.label), "Total Power");
>   	scnprintf(sensor.units, sizeof(sensor.units), "mW");
> @@ -804,7 +806,29 @@ static int aie2_get_sensors(struct amdxdna_client *client,
>   	if (copy_to_user(u64_to_user_ptr(args->buffer), &sensor, sizeof(sensor)))
>   		return -EFAULT;
>   
> -	args->buffer_size = sizeof(sensor);
> +	sensors_count++;
> +	if (args->buffer_size <= sensors_count * sizeof(sensor))
> +		goto out;
> +
> +	for (i = 0; i < min_t(u32, ndev->total_col, 8); i++) {
> +		memset(&sensor, 0, sizeof(sensor));
> +		sensor.input = npu_metrics.npu_busy[i];
> +		sensor.type = AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION;
> +		sensor.unitm = 0;
> +		scnprintf(sensor.label, sizeof(sensor.label), "Column %d Utilization", i);
> +		scnprintf(sensor.units, sizeof(sensor.units), "%%");
> +
> +		if (copy_to_user(u64_to_user_ptr(args->buffer) + sensors_count * sizeof(sensor),
> +				 &sensor, sizeof(sensor)))
> +			return -EFAULT;
> +
> +		sensors_count++;
> +		if (args->buffer_size <= sensors_count * sizeof(sensor))
> +			goto out;
> +	}
> +
> +out:
> +	args->buffer_size = sensors_count * sizeof(sensor);
>   
>   	return 0;
>   }
> diff --git a/drivers/accel/amdxdna/aie2_pci.h b/drivers/accel/amdxdna/aie2_pci.h
> index 1bb88711bedb..0ae174862592 100644
> --- a/drivers/accel/amdxdna/aie2_pci.h
> +++ b/drivers/accel/amdxdna/aie2_pci.h
> @@ -48,6 +48,7 @@
>   })
>   
>   #if IS_ENABLED(CONFIG_AMD_PMF)
> +#define AIE2_GET_PMF_NPU_METRICS(metrics) amd_pmf_get_npu_data(metrics)
>   #define AIE2_GET_PMF_NPU_DATA(field, val)				\
>   ({									\
>   	struct amd_pmf_npu_metrics _npu_metrics;			\
> @@ -58,6 +59,13 @@
>   	(_ret);								\
>   })
>   #else
> +#define AIE2_GET_PMF_NPU_METRICS(metrics)				\
> +({									\
> +	typeof(metrics) _m = metrics;					\
> +	memset(_m, 0xff, sizeof(*_m));					\
> +	(-EOPNOTSUPP);							\
> +})
> +
>   #define SENSOR_DEFAULT_npu_power	U32_MAX
>   #define AIE2_GET_PMF_NPU_DATA(field, val)				\
>   ({									\
> diff --git a/include/uapi/drm/amdxdna_accel.h b/include/uapi/drm/amdxdna_accel.h
> index 9c44db2b3dcd..5bd13f4435f5 100644
> --- a/include/uapi/drm/amdxdna_accel.h
> +++ b/include/uapi/drm/amdxdna_accel.h
> @@ -353,7 +353,8 @@ struct amdxdna_drm_query_clock_metadata {
>   };
>   
>   enum amdxdna_sensor_type {
> -	AMDXDNA_SENSOR_TYPE_POWER
> +	AMDXDNA_SENSOR_TYPE_POWER,
> +	AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION
>   };
>   
>   /**

Re: [PATCH V2] accel/amdxdna: Support sensors for column utilization

[Lizhi Hou]

Applied to drm-misc-next

On 3/11/26 10:21, Mario Limonciello wrote:
>
>
> On 3/11/26 12:18, Lizhi Hou wrote:
>> From: "Mario Limonciello (AMD)" <superm1@kernel.org>
>>
>> The AMD PMF driver provides realtime column utilization (npu_busy)
>> metrics for the NPU. Extend the DRM_IOCTL_AMDXDNA_GET_INFO sensor
>> query to expose these metrics to userspace.
>>
>> Add AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION to the sensor type enum
>> and update aie2_get_sensors() to return both the total power and up
>> to 8 column utilization sensors if the user buffer permits.
>>
>> Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
>> Reviewed-by: Lizhi Hou <lizhi.hou@amd.com>
>> [lizhi: support legacy tool which uses small buffer. checkpatch cleanup]
>> Signed-off-by: Lizhi Hou <lizhi.hou@amd.com>
>
> Thanks, no concerns on the cleanups.
>
>> ---
>>   drivers/accel/amdxdna/aie2_pci.c | 34 +++++++++++++++++++++++++++-----
>>   drivers/accel/amdxdna/aie2_pci.h |  8 ++++++++
>>   include/uapi/drm/amdxdna_accel.h |  3 ++-
>>   3 files changed, 39 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/accel/amdxdna/aie2_pci.c 
>> b/drivers/accel/amdxdna/aie2_pci.c
>> index a2e586512e26..c57c785a2d15 100644
>> --- a/drivers/accel/amdxdna/aie2_pci.c
>> +++ b/drivers/accel/amdxdna/aie2_pci.c
>> @@ -787,16 +787,18 @@ static int aie2_get_clock_metadata(struct 
>> amdxdna_client *client,
>>   static int aie2_get_sensors(struct amdxdna_client *client,
>>                   struct amdxdna_drm_get_info *args)
>>   {
>> +    struct amdxdna_dev_hdl *ndev = client->xdna->dev_handle;
>>       struct amdxdna_drm_query_sensor sensor = {};
>> +    struct amd_pmf_npu_metrics npu_metrics;
>> +    u32 sensors_count = 0, i;
>>       int ret;
>>   -    if (args->buffer_size < sizeof(sensor))
>> -        return -EINVAL;
>> -
>> -    ret = AIE2_GET_PMF_NPU_DATA(npu_power, sensor.input);
>> +    ret = AIE2_GET_PMF_NPU_METRICS(&npu_metrics);
>>       if (ret)
>>           return ret;
>> +
>>       sensor.type = AMDXDNA_SENSOR_TYPE_POWER;
>> +    sensor.input = npu_metrics.npu_power;
>>       sensor.unitm = -3;
>>       scnprintf(sensor.label, sizeof(sensor.label), "Total Power");
>>       scnprintf(sensor.units, sizeof(sensor.units), "mW");
>> @@ -804,7 +806,29 @@ static int aie2_get_sensors(struct 
>> amdxdna_client *client,
>>       if (copy_to_user(u64_to_user_ptr(args->buffer), &sensor, 
>> sizeof(sensor)))
>>           return -EFAULT;
>>   -    args->buffer_size = sizeof(sensor);
>> +    sensors_count++;
>> +    if (args->buffer_size <= sensors_count * sizeof(sensor))
>> +        goto out;
>> +
>> +    for (i = 0; i < min_t(u32, ndev->total_col, 8); i++) {
>> +        memset(&sensor, 0, sizeof(sensor));
>> +        sensor.input = npu_metrics.npu_busy[i];
>> +        sensor.type = AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION;
>> +        sensor.unitm = 0;
>> +        scnprintf(sensor.label, sizeof(sensor.label), "Column %d 
>> Utilization", i);
>> +        scnprintf(sensor.units, sizeof(sensor.units), "%%");
>> +
>> +        if (copy_to_user(u64_to_user_ptr(args->buffer) + 
>> sensors_count * sizeof(sensor),
>> +                 &sensor, sizeof(sensor)))
>> +            return -EFAULT;
>> +
>> +        sensors_count++;
>> +        if (args->buffer_size <= sensors_count * sizeof(sensor))
>> +            goto out;
>> +    }
>> +
>> +out:
>> +    args->buffer_size = sensors_count * sizeof(sensor);
>>         return 0;
>>   }
>> diff --git a/drivers/accel/amdxdna/aie2_pci.h 
>> b/drivers/accel/amdxdna/aie2_pci.h
>> index 1bb88711bedb..0ae174862592 100644
>> --- a/drivers/accel/amdxdna/aie2_pci.h
>> +++ b/drivers/accel/amdxdna/aie2_pci.h
>> @@ -48,6 +48,7 @@
>>   })
>>     #if IS_ENABLED(CONFIG_AMD_PMF)
>> +#define AIE2_GET_PMF_NPU_METRICS(metrics) amd_pmf_get_npu_data(metrics)
>>   #define AIE2_GET_PMF_NPU_DATA(field, val)                \
>>   ({                                    \
>>       struct amd_pmf_npu_metrics _npu_metrics;            \
>> @@ -58,6 +59,13 @@
>>       (_ret);                                \
>>   })
>>   #else
>> +#define AIE2_GET_PMF_NPU_METRICS(metrics)                \
>> +({                                    \
>> +    typeof(metrics) _m = metrics;                    \
>> +    memset(_m, 0xff, sizeof(*_m));                    \
>> +    (-EOPNOTSUPP);                            \
>> +})
>> +
>>   #define SENSOR_DEFAULT_npu_power    U32_MAX
>>   #define AIE2_GET_PMF_NPU_DATA(field, val)                \
>>   ({                                    \
>> diff --git a/include/uapi/drm/amdxdna_accel.h 
>> b/include/uapi/drm/amdxdna_accel.h
>> index 9c44db2b3dcd..5bd13f4435f5 100644
>> --- a/include/uapi/drm/amdxdna_accel.h
>> +++ b/include/uapi/drm/amdxdna_accel.h
>> @@ -353,7 +353,8 @@ struct amdxdna_drm_query_clock_metadata {
>>   };
>>     enum amdxdna_sensor_type {
>> -    AMDXDNA_SENSOR_TYPE_POWER
>> +    AMDXDNA_SENSOR_TYPE_POWER,
>> +    AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION
>>   };
>>     /**
>

Claude review: accel/amdxdna: Support sensors for column utilization

[Claude Code Review Bot]

Overall Series Review

Subject: accel/amdxdna: Support sensors for column utilization
Author: Lizhi Hou <lizhi.hou@amd.com>
Patches: 3
Reviewed: 2026-03-12T06:48:17.277058

---

This is a single-patch series (V2) adding column utilization sensor support to the `accel/amdxdna` driver. It extends the existing `aie2_get_sensors()` function to return up to 8 column utilization sensors alongside the existing power sensor, using NPU busy metrics from the AMD PMF driver. The approach is reasonable and the patch is relatively small, but there are a few issues worth addressing around buffer validation, the `!CONFIG_AMD_PMF` stub, and the hard-coded array size.

---
Generated by Claude Code Patch Reviewer

Claude review: accel/amdxdna: Support sensors for column utilization

[Claude Code Review Bot]

Patch Review

**1. Removed buffer size validation (medium severity)**

The original code had:
```c
if (args->buffer_size < sizeof(sensor))
    return -EINVAL;
```

This is removed by the patch. The new code proceeds to `copy_to_user()` of the first (power) sensor without checking whether the user buffer is large enough for even one sensor. If userspace passes `buffer_size = 0`, the copy proceeds unconditionally and the `<=` check at line 170 (`args->buffer_size <= sensors_count * sizeof(sensor)`) only fires *after* the copy. The minimum buffer size check should be retained before the first `copy_to_user`.

**2. Off-by-one / boundary check logic**

```c
sensors_count++;
if (args->buffer_size <= sensors_count * sizeof(sensor))
    goto out;
```

This check is after copying, so it guards the *next* sensor, not the current one. For the first sensor (power), there's no pre-check at all (see point 1). The logic works for subsequent column sensors since the check prevents writing beyond the buffer, but it's fragile and inconsistent — the first sensor has no guard while subsequent ones do.

**3. Hard-coded `8` instead of using `ARRAY_SIZE`**

```c
for (i = 0; i < min_t(u32, ndev->total_col, 8); i++) {
```

The `8` corresponds to the size of `npu_metrics.npu_busy[8]` in `struct amd_pmf_npu_metrics`. This should use `ARRAY_SIZE(npu_metrics.npu_busy)` to avoid a silent bug if the struct is ever expanded.

**4. `!CONFIG_AMD_PMF` stub uses wrong error return**

```c
#define AIE2_GET_PMF_NPU_METRICS(metrics)				\
({									\
	typeof(metrics) _m = metrics;					\
	memset(_m, 0xff, sizeof(*_m));					\
	(-EOPNOTSUPP);							\
})
```

The `memset` to `0xff` is presumably for debug purposes, but since the caller checks `if (ret) return ret;` this data is never used. More importantly, the existing `AIE2_GET_PMF_NPU_DATA` stub uses `-EOPNOTSUPP` too but also fills in a `U32_MAX` default value and returns `0` (success) — the two stubs have inconsistent error semantics. The new `AIE2_GET_PMF_NPU_METRICS` returns an error, meaning sensors will be completely unavailable when `CONFIG_AMD_PMF` is disabled, whereas the old `AIE2_GET_PMF_NPU_DATA` returned success with a sentinel value. If the original behavior was to still report sensors (with a sentinel) when PMF is unavailable, this changes that contract.

**5. `sensor.input` type mismatch**

Looking at the structs:
- `struct amd_pmf_npu_metrics` has `u16 npu_busy[8]`
- `struct amdxdna_drm_query_sensor` has `__u32 input`

```c
sensor.input = npu_metrics.npu_busy[i];
```

This is fine (implicit widening from `u16` to `u32`), but worth noting that the `npu_power` field is also `u16` in the metrics struct. Both assignments are safe.

**6. `ndev->total_col` could exceed 8**

```c
for (i = 0; i < min_t(u32, ndev->total_col, 8); i++) {
```

The `min_t` clamp is correct since `npu_busy` has only 8 entries. But if hardware with >8 columns exists, userspace will silently get incomplete data with no indication. A comment would be helpful here.

**7. The old `AIE2_GET_PMF_NPU_DATA` macro is now dead code**

The patch adds `AIE2_GET_PMF_NPU_METRICS` and converts the sole caller, but doesn't remove the now-unused `AIE2_GET_PMF_NPU_DATA` macro (both the `CONFIG_AMD_PMF` and `!CONFIG_AMD_PMF` variants) or the related `SENSOR_DEFAULT_npu_power` define. These should be cleaned up.

**8. Minor: goto label placement**

```c
out:
	args->buffer_size = sensors_count * sizeof(sensor);

	return 0;
```

The `out:` label with its blank line above the assignment is fine stylistically, but placing it directly above the assignment makes the flow clearer that `buffer_size` is always updated regardless of how many sensors were written.

---
Generated by Claude Code Patch Reviewer

[PATCH 0/2] Add some sensor data to amdxdna

[Mario Limonciello (AMD)]

Previously a series [1] was sent to add power sensor data for amdxdna
through the PMF mailbox.  The PMF patch was merged but the XDNA
patches were planned for later.

Re-send the XDNA patch, and then add in another patch for getting
column utilization.

Link: https://lore.kernel.org/platform-driver-x86/20260105172956.3732123-3-lizhi.hou@amd.com/ [1]

Lizhi Hou (1):
  accel/amdxdna: Add IOCTL to retrieve realtime NPU power estimate

Mario Limonciello (AMD) (1):
  accel/amdxdna: Support sensors for column utilization

 drivers/accel/amdxdna/aie2_pci.c        | 50 +++++++++++++++++++++++++
 drivers/accel/amdxdna/aie2_pci.h        | 22 +++++++++++
 drivers/accel/amdxdna/amdxdna_pci_drv.c |  3 +-
 include/uapi/drm/amdxdna_accel.h        |  3 +-
 4 files changed, 76 insertions(+), 2 deletions(-)

-- 
2.43.0

[PATCH 2/2] accel/amdxdna: Support sensors for column utilization

[Mario Limonciello (AMD)]

The AMD PMF driver provides realtime column utilization (npu_busy)
metrics for the NPU. Extend the DRM_IOCTL_AMDXDNA_GET_INFO sensor
query to expose these metrics to userspace.

Add AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION to the sensor type enum
and update aie2_get_sensors() to return both the total power and up
to 8 column utilization sensors if the user buffer permits.

Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
---
 drivers/accel/amdxdna/aie2_pci.c | 27 ++++++++++++++++++++++++---
 drivers/accel/amdxdna/aie2_pci.h |  2 ++
 include/uapi/drm/amdxdna_accel.h |  3 ++-
 3 files changed, 28 insertions(+), 4 deletions(-)

diff --git a/drivers/accel/amdxdna/aie2_pci.c b/drivers/accel/amdxdna/aie2_pci.c
index a2e586512e26c..1f11822f0bfda 100644
--- a/drivers/accel/amdxdna/aie2_pci.c
+++ b/drivers/accel/amdxdna/aie2_pci.c
@@ -787,16 +787,24 @@ static int aie2_get_clock_metadata(struct amdxdna_client *client,
 static int aie2_get_sensors(struct amdxdna_client *client,
 			    struct amdxdna_drm_get_info *args)
 {
+	struct amdxdna_dev_hdl *ndev = client->xdna->dev_handle;
 	struct amdxdna_drm_query_sensor sensor = {};
+	struct amd_pmf_npu_metrics npu_metrics;
+	u32 sensors_count, util_count, i;
 	int ret;
 
-	if (args->buffer_size < sizeof(sensor))
+	util_count = min_t(u32, ndev->total_col, 8);
+	sensors_count = 1 + util_count;
+
+	if (args->buffer_size < sensors_count * sizeof(sensor))
 		return -EINVAL;
 
-	ret = AIE2_GET_PMF_NPU_DATA(npu_power, sensor.input);
+	ret = AIE2_GET_PMF_NPU_METRICS(&npu_metrics);
 	if (ret)
 		return ret;
+
 	sensor.type = AMDXDNA_SENSOR_TYPE_POWER;
+	sensor.input = npu_metrics.npu_power;
 	sensor.unitm = -3;
 	scnprintf(sensor.label, sizeof(sensor.label), "Total Power");
 	scnprintf(sensor.units, sizeof(sensor.units), "mW");
@@ -804,7 +812,20 @@ static int aie2_get_sensors(struct amdxdna_client *client,
 	if (copy_to_user(u64_to_user_ptr(args->buffer), &sensor, sizeof(sensor)))
 		return -EFAULT;
 
-	args->buffer_size = sizeof(sensor);
+	for (i = 0; i < util_count; i++) {
+		memset(&sensor, 0, sizeof(sensor));
+		sensor.input = npu_metrics.npu_busy[i];
+		sensor.type = AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION;
+		sensor.unitm = 0;
+		scnprintf(sensor.label, sizeof(sensor.label), "Column %d Utilization", i);
+		scnprintf(sensor.units, sizeof(sensor.units), "%%");
+
+		if (copy_to_user(u64_to_user_ptr(args->buffer) + (i + 1) * sizeof(sensor),
+				 &sensor, sizeof(sensor)))
+			return -EFAULT;
+	}
+
+	args->buffer_size = sensors_count * sizeof(sensor);
 
 	return 0;
 }
diff --git a/drivers/accel/amdxdna/aie2_pci.h b/drivers/accel/amdxdna/aie2_pci.h
index 1bb88711bedb3..424fd6d852191 100644
--- a/drivers/accel/amdxdna/aie2_pci.h
+++ b/drivers/accel/amdxdna/aie2_pci.h
@@ -48,6 +48,7 @@
 })
 
 #if IS_ENABLED(CONFIG_AMD_PMF)
+#define AIE2_GET_PMF_NPU_METRICS(metrics) amd_pmf_get_npu_data(metrics)
 #define AIE2_GET_PMF_NPU_DATA(field, val)				\
 ({									\
 	struct amd_pmf_npu_metrics _npu_metrics;			\
@@ -58,6 +59,7 @@
 	(_ret);								\
 })
 #else
+#define AIE2_GET_PMF_NPU_METRICS(metrics) (-EOPNOTSUPP)
 #define SENSOR_DEFAULT_npu_power	U32_MAX
 #define AIE2_GET_PMF_NPU_DATA(field, val)				\
 ({									\
diff --git a/include/uapi/drm/amdxdna_accel.h b/include/uapi/drm/amdxdna_accel.h
index 9c44db2b3dcd4..5bd13f4435f55 100644
--- a/include/uapi/drm/amdxdna_accel.h
+++ b/include/uapi/drm/amdxdna_accel.h
@@ -353,7 +353,8 @@ struct amdxdna_drm_query_clock_metadata {
 };
 
 enum amdxdna_sensor_type {
-	AMDXDNA_SENSOR_TYPE_POWER
+	AMDXDNA_SENSOR_TYPE_POWER,
+	AMDXDNA_SENSOR_TYPE_COLUMN_UTILIZATION
 };
 
 /**
-- 
2.43.0

Claude review: accel/amdxdna: Support sensors for column utilization

[Claude Code Review Bot]

Patch Review

**Rigid buffer size requirement breaks forward/backward compatibility** (`aie2_pci.c`):

```c
util_count = min_t(u32, ndev->total_col, 8);
sensors_count = 1 + util_count;

if (args->buffer_size < sensors_count * sizeof(sensor))
    return -EINVAL;
```

This is the most significant issue in the series. Patch 1 requires space for 1 sensor. Patch 2 changes this to require space for `1 + min(total_col, 8)` sensors. Problems:

1. **No discovery mechanism**: Userspace has no way to query how many sensors exist before allocating the buffer. It must know `total_col` a priori, which is a driver implementation detail.
2. **All-or-nothing**: If userspace only wants the power sensor, it cannot query just that one — it must allocate for all sensors or get `-EINVAL`.
3. **Hardware-dependent size**: Different hardware with different `total_col` values require different buffer sizes for the same IOCTL parameter.

A better design would be: fill as many sensors as fit in the provided buffer and set `args->buffer_size` to the number of bytes actually written. Userspace can then allocate a larger buffer and retry if it wants more. This is a common pattern in kernel ioctls.

**`AIE2_GET_PMF_NPU_DATA` becomes dead code** (`aie2_pci.h`):

Patch 2 adds:
```c
#define AIE2_GET_PMF_NPU_METRICS(metrics) amd_pmf_get_npu_data(metrics)
```

and changes `aie2_get_sensors()` to use `AIE2_GET_PMF_NPU_METRICS()` instead of `AIE2_GET_PMF_NPU_DATA()`. But the old macro (and its `SENSOR_DEFAULT_npu_power` helper) are left in place with no remaining callers. These should be removed.

Also, `AIE2_GET_PMF_NPU_METRICS` is a trivial one-line wrapper that adds no value — it just calls `amd_pmf_get_npu_data()`. Consider calling the function directly (guarded by `#if IS_ENABLED(CONFIG_AMD_PMF)`) instead of adding a macro that just aliases the function.

**Partial copy_to_user on failure** (`aie2_pci.c`):

```c
if (copy_to_user(u64_to_user_ptr(args->buffer), &sensor, sizeof(sensor)))
    return -EFAULT;

for (i = 0; i < util_count; i++) {
    ...
    if (copy_to_user(u64_to_user_ptr(args->buffer) + (i + 1) * sizeof(sensor),
                     &sensor, sizeof(sensor)))
        return -EFAULT;
}

args->buffer_size = sensors_count * sizeof(sensor);
```

If the power sensor is successfully copied but a subsequent column utilization copy fails, the function returns `-EFAULT` without setting `args->buffer_size`. Userspace has received partial data in its buffer but has no indication of how much was written. This is a minor concern in practice (a partial `copy_to_user` failure typically indicates a corrupted user pointer), but building into a local buffer and doing a single `copy_to_user` would be cleaner and avoid partial writes entirely.

**`npu_busy` is `u16` but `sensor.input` is `__u32`** — this is fine, the implicit widening is correct. Just noting for completeness that the PMF metrics use `u16` values while the UAPI struct uses `__u32`.

**Missing sensor index/identifier**: The column utilization sensors are only distinguished by their `label` string ("Column 0 Utilization", etc.). There is no numeric index field in the sensor struct. Userspace must parse the label string to determine which column a sensor corresponds to. Adding a field or using a sub-type would be more robust, though this would require a UAPI change.

**Version bump scope**: The version bump to 0.7 is in patch 1 with the comment `"Support getting power and utilization data"`, but utilization support is only added in patch 2. The version bump should either be in patch 2 (when the full feature is present) or the comment in patch 1 should only mention power data.

---
Generated by Claude Code Patch Reviewer