Skip to contentSkip to conversation

Reduce the number of GPU resumes

PATCH In review Pierre-Eric Pelloux-Prayer 5 participants 16 messages Jun 18, 2024 – Jun 26, 2024

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover Reduce the number of GPU resumes

The goal I'm aiming for is to be able to open a dri node and use amdgpu_device_initialize without waking up the GPU. One visible outcome is that it would become possible to call vkEnumeratePhysicalDevices without having to resume all the GPUs in the system.

This series implements some of the changes required to achieve this goal with a focus on not waking up the GPU for ioctls that don't need the GPU to be active.
The output of AMD_DEBUG=info doesn't change with these patches applied.

The other changes required are found in following patchset:
https://lists.freedesktop.org/archives/amd-gfx/2024-June/109796.html

Pierre-Eric Pelloux-Prayer (6):
drm/amdgpu: allow ioctls to opt-out of runtime pm drm/amdgpu: skip runtime pm for selected ioctls drm/amdgpu: refactor amdgpu_info_ioctl to allow finer pm drm/amdgpu: add AMDGPU_INFO_GB_ADDR_CONFIG query drm/amdgpu: cache mclk/sclk min/max values drm/amdgpu: resume the device from amdgpu_gem_fault

 drivers/gpu/drm/amd/amdgpu/amdgpu.h           |  12 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_atombios.c  |   8 +-
 .../gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c  |   8 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |   7 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c       |  69 ++++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c       |  13 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 116 +++++++++++++-----
 drivers/gpu/drm/amd/pm/legacy-dpm/kv_dpm.c    |   8 +-
 drivers/gpu/drm/amd/pm/legacy-dpm/si_dpm.c    |  12 +-
 include/uapi/drm/amdgpu_drm.h                 |   2 +
 10 files changed, 198 insertions(+), 57 deletions(-)
Signature · 1 line
2.40.1
#1 lore raw

PATCH 1/6 drm/amdgpu: allow ioctls to opt-out of runtime pm

Waking up a device can take multiple seconds, so if it's not going to be used we might as well not resume it.

The safest default behavior for all ioctls is to resume the GPU, so this change allows specific ioctls to opt-out of generic runtime pm.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 25 ++++++++++++++++++++-----
 1 file changed, 20 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +20 -5
index 60d5758939ae..a9831b243bfc 100644
@@ -2855,18 +2855,33 @@ long amdgpu_drm_ioctl(struct file *filp,
 {
 	struct drm_file *file_priv = filp->private_data;
 	struct drm_device *dev;
+	bool needs_device;
 	long ret;
 
 	dev = file_priv->minor->dev;
-	ret = pm_runtime_get_sync(dev->dev);
-	if (ret < 0)
-		goto out;
+
+	/* Some ioctl can opt-out of powermanagement handling
+	 * if they don't require the device to be resumed.
+	 */
+	switch (cmd) {
+	default:
+		needs_device = true;
+	}
+
+	if (needs_device) {
+		ret = pm_runtime_get_sync(dev->dev);
+		if (ret < 0)
+			goto out;
+	}
 
 	ret = drm_ioctl(filp, cmd, arg);
 
-	pm_runtime_mark_last_busy(dev->dev);
 out:
-	pm_runtime_put_autosuspend(dev->dev);
+	if (needs_device) {
+		pm_runtime_mark_last_busy(dev->dev);
+		pm_runtime_put_autosuspend(dev->dev);
+	}
+
 	return ret;
 }
 
Signature · 1 line
2.40.1
maintainer
#2 lore raw

Re: [PATCH 1/6] drm/amdgpu: allow ioctls to opt-out of runtime pm

Am 18.06.24 um 17:23 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 6 lines

Waking up a device can take multiple seconds, so if it's not going to be used we might as well not resume it.

The safest default behavior for all ioctls is to resume the GPU, so this change allows specific ioctls to opt-out of generic runtime pm.

I'm really wondering if we shouldn't put that into the IOCTL description.

See amdgpu_ioctls_kms and DRM_IOCTL_DEF_DRV() for what I mean.

Regards,
Christian.

Quoted 49 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 25 ++++++++++++++++++++-----
  1 file changed, 20 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +20 -5
index 60d5758939ae..a9831b243bfc 100644
@@ -2855,18 +2855,33 @@ long amdgpu_drm_ioctl(struct file *filp,
  {
  	struct drm_file *file_priv = filp->private_data;
  	struct drm_device *dev;
+	bool needs_device;
  	long ret;
 
  	dev = file_priv->minor->dev;
-	ret = pm_runtime_get_sync(dev->dev);
-	if (ret < 0)
-		goto out;
+
+	/* Some ioctl can opt-out of powermanagement handling
+	 * if they don't require the device to be resumed.
+	 */
+	switch (cmd) {
+	default:
+		needs_device = true;
+	}
+
+	if (needs_device) {
+		ret = pm_runtime_get_sync(dev->dev);
+		if (ret < 0)
+			goto out;
+	}
 
  	ret = drm_ioctl(filp, cmd, arg);
 
-	pm_runtime_mark_last_busy(dev->dev);
  out:
-	pm_runtime_put_autosuspend(dev->dev);
+	if (needs_device) {
+		pm_runtime_mark_last_busy(dev->dev);
+		pm_runtime_put_autosuspend(dev->dev);
+	}
+
  	return ret;
  }
 
#3 lore raw

Re: [PATCH 1/6] drm/amdgpu: allow ioctls to opt-out of runtime pm

Le 19/06/2024 à 11:26, Christian König a écrit :

Quoted 11 lines

Am 18.06.24 um 17:23 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 6 lines

Waking up a device can take multiple seconds, so if it's not going to be used we might as well not resume it.

The safest default behavior for all ioctls is to resume the GPU, so this change allows specific ioctls to opt-out of generic runtime pm.

I'm really wondering if we shouldn't put that into the IOCTL description.

See amdgpu_ioctls_kms and DRM_IOCTL_DEF_DRV() for what I mean.

Are you suggesting to add a new entry in enum drm_ioctl_flags to indicate ioctls which need the device to be awake?

Something like: "DRM_NO_DEVICE = BIT(6)" and then use it for both core and amdgpu ioctls?

Pierre-Eric

Quoted 49 lines

Regards,
Christian.

Quoted 45 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 25 ++++++++++++++++++++-----

  1 file changed, 20 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +15 -4
index 60d5758939ae..a9831b243bfc 100644
@@ -2855,18 +2855,33 @@ long amdgpu_drm_ioctl(struct file *filp,
   {
       struct drm_file *file_priv = filp->private_data;
       struct drm_device *dev;
+    bool needs_device;
       long ret;
       dev = file_priv->minor->dev;
-    ret = pm_runtime_get_sync(dev->dev);
-    if (ret < 0)
-        goto out;
+
+    /* Some ioctl can opt-out of powermanagement handling
+     * if they don't require the device to be resumed.
+     */
+    switch (cmd) {
+    default:
+        needs_device = true;
+    }
+
+    if (needs_device) {
+        ret = pm_runtime_get_sync(dev->dev);
+        if (ret < 0)
+            goto out;
+    }
       ret = drm_ioctl(filp, cmd, arg);
-    pm_runtime_mark_last_busy(dev->dev);

  out:
-    pm_runtime_put_autosuspend(dev->dev);
+    if (needs_device) {
+        pm_runtime_mark_last_busy(dev->dev);
+        pm_runtime_put_autosuspend(dev->dev);
+    }
+
      return ret;
  }

maintainer
#4 lore raw

Re: [PATCH 1/6] drm/amdgpu: allow ioctls to opt-out of runtime pm

Am 20.06.24 um 15:06 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 9 lines

Le 19/06/2024 à 11:26, Christian König a écrit :

Quoted 8 lines

Am 18.06.24 um 17:23 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 6 lines

Waking up a device can take multiple seconds, so if it's not going to be used we might as well not resume it.

The safest default behavior for all ioctls is to resume the GPU, so this change allows specific ioctls to opt-out of generic runtime pm.

Quoted 8 lines
Quoted 3 lines

I'm really wondering if we shouldn't put that into the IOCTL description.

See amdgpu_ioctls_kms and DRM_IOCTL_DEF_DRV() for what I mean.

Are you suggesting to add a new entry in enum drm_ioctl_flags to indicate ioctls which need the device to be awake?

Something like: "DRM_NO_DEVICE = BIT(6)" and then use it for both core and amdgpu ioctls?

Yeah something like that. Maybe name that DRM_SW_ONLY or something like that.

Christian.

Quoted 56 lines

Pierre-Eric

Quoted 50 lines

Regards,
Christian.

Quoted 46 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 25
++++++++++++++++++++-----

  1 file changed, 20 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +15 -4
index 60d5758939ae..a9831b243bfc 100644
@@ -2855,18 +2855,33 @@ long amdgpu_drm_ioctl(struct file *filp,
   {
       struct drm_file *file_priv = filp->private_data;
       struct drm_device *dev;
+    bool needs_device;
       long ret;
       dev = file_priv->minor->dev;
-    ret = pm_runtime_get_sync(dev->dev);
-    if (ret < 0)
-        goto out;
+
+    /* Some ioctl can opt-out of powermanagement handling
+     * if they don't require the device to be resumed.
+     */
+    switch (cmd) {
+    default:
+        needs_device = true;
+    }
+
+    if (needs_device) {
+        ret = pm_runtime_get_sync(dev->dev);
+        if (ret < 0)
+            goto out;
+    }
       ret = drm_ioctl(filp, cmd, arg);
-    pm_runtime_mark_last_busy(dev->dev);

  out:
-    pm_runtime_put_autosuspend(dev->dev);
+    if (needs_device) {
+        pm_runtime_mark_last_busy(dev->dev);
+        pm_runtime_put_autosuspend(dev->dev);
+    }
+
      return ret;
  }

#5 lore raw

Re: [PATCH 1/6] drm/amdgpu: allow ioctls to opt-out of runtime pm

Le 20/06/2024 à 15:36, Christian König a écrit :

Quoted 12 lines

Am 20.06.24 um 15:06 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 11 lines

Le 19/06/2024 à 11:26, Christian König a écrit :

Quoted 10 lines

Am 18.06.24 um 17:23 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 6 lines

Waking up a device can take multiple seconds, so if it's not going to be used we might as well not resume it.

The safest default behavior for all ioctls is to resume the GPU, so this change allows specific ioctls to opt-out of generic runtime pm.

I'm really wondering if we shouldn't put that into the IOCTL description.

Quoted 8 lines
Quoted 6 lines
Quoted 1 line

See amdgpu_ioctls_kms and DRM_IOCTL_DEF_DRV() for what I mean.

Are you suggesting to add a new entry in enum drm_ioctl_flags to indicate ioctls which need the device to be awake?

Something like: "DRM_NO_DEVICE = BIT(6)" and then use it for both core and amdgpu ioctls?

Yeah something like that. Maybe name that DRM_SW_ONLY or something like that.

+ dri-devel to gauge interest in adding such a flag in shared code.

Pierre-Eric

Quoted 59 lines

Christian.

Quoted 56 lines

Pierre-Eric

Quoted 50 lines

Regards,
Christian.

Quoted 46 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 25
++++++++++++++++++++-----

  1 file changed, 20 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +15 -4
index 60d5758939ae..a9831b243bfc 100644
@@ -2855,18 +2855,33 @@ long amdgpu_drm_ioctl(struct file *filp,
   {
       struct drm_file *file_priv = filp->private_data;
       struct drm_device *dev;
+    bool needs_device;
       long ret;
       dev = file_priv->minor->dev;
-    ret = pm_runtime_get_sync(dev->dev);
-    if (ret < 0)
-        goto out;
+
+    /* Some ioctl can opt-out of powermanagement handling
+     * if they don't require the device to be resumed.
+     */
+    switch (cmd) {
+    default:
+        needs_device = true;
+    }
+
+    if (needs_device) {
+        ret = pm_runtime_get_sync(dev->dev);
+        if (ret < 0)
+            goto out;
+    }
       ret = drm_ioctl(filp, cmd, arg);
-    pm_runtime_mark_last_busy(dev->dev);

  out:
-    pm_runtime_put_autosuspend(dev->dev);
+    if (needs_device) {
+        pm_runtime_mark_last_busy(dev->dev);
+        pm_runtime_put_autosuspend(dev->dev);
+    }
+
      return ret;
  }

#6 lore raw

Re: [PATCH 1/6] drm/amdgpu: allow ioctls to opt-out of runtime pm

Quoted 20 lines · Pierre-Eric Pelloux-Prayer

Le 20/06/2024 à 15:36, Christian König a écrit :

Quoted 18 lines

Am 20.06.24 um 15:06 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 17 lines

Le 19/06/2024 à 11:26, Christian König a écrit :

Quoted 4 lines

Am 18.06.24 um 17:23 schrieb Pierre-Eric Pelloux-Prayer:

Quoted 3 lines

Waking up a device can take multiple seconds, so if it's not going to be used we might as well not resume it.

Quoted 8 lines
Quoted 3 lines

The safest default behavior for all ioctls is to resume the GPU, so this change allows specific ioctls to opt-out of generic runtime pm.

I'm really wondering if we shouldn't put that into the IOCTL description.

See amdgpu_ioctls_kms and DRM_IOCTL_DEF_DRV() for what I mean.

Are you suggesting to add a new entry in enum drm_ioctl_flags to indicate ioctls which need the device to be awake?

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

Something like: "DRM_NO_DEVICE = BIT(6)" and then use it for both core and amdgpu ioctls?

Yeah something like that. Maybe name that DRM_SW_ONLY or something like that.

+ dri-devel to gauge interest in adding such a flag in shared code.

Eh please no. That's just fundamentally the wrong way round to do runtime pm, but amdgpu goes way, way back in desing in that record to the vga switcheroo, where the simple hack was to just add runtime pm handling at the entry points and call it done.

If you look at any other drm driver, the runtime pm handling is done way down when touching the actual relevant hardware blocks. Because especially on arm it's actually a pile of runtime pm domains.

So essentially this is like pushing a big driver lock down the callchain until it's actually at the right level. First step would be to push it into each of the amdgpu ioctl callbacks, so that you can drop amdgpu_drm_ioctl and use drm_ioctl again directly. And then push it further until you can remove it from all the places where it's not needed.

DRM_SW_ONLY otoh is pure midlayer mistake design.

Cheers, Sima

Quoted 66 lines

Pierre-Eric

Quoted 61 lines

Christian.

Quoted 58 lines

Pierre-Eric

Quoted 52 lines

Regards,
Christian.

Quoted 48 lines

Signed-off-by: Pierre-Eric Pelloux-Prayer
<[email protected]>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 25
++++++++++++++++++++-----

  1 file changed, 20 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +15 -4
index 60d5758939ae..a9831b243bfc 100644
@@ -2855,18 +2855,33 @@ long amdgpu_drm_ioctl(struct file *filp,
   {
       struct drm_file *file_priv = filp->private_data;
       struct drm_device *dev;
+    bool needs_device;
       long ret;
       dev = file_priv->minor->dev;
-    ret = pm_runtime_get_sync(dev->dev);
-    if (ret < 0)
-        goto out;
+
+    /* Some ioctl can opt-out of powermanagement handling
+     * if they don't require the device to be resumed.
+     */
+    switch (cmd) {
+    default:
+        needs_device = true;
+    }
+
+    if (needs_device) {
+        ret = pm_runtime_get_sync(dev->dev);
+        if (ret < 0)
+            goto out;
+    }
       ret = drm_ioctl(filp, cmd, arg);
-    pm_runtime_mark_last_busy(dev->dev);

  out:
-    pm_runtime_put_autosuspend(dev->dev);
+    if (needs_device) {
+        pm_runtime_mark_last_busy(dev->dev);
+        pm_runtime_put_autosuspend(dev->dev);
+    }
+
      return ret;
  }

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#7 lore raw

PATCH 2/6 drm/amdgpu: skip runtime pm for selected ioctls

In reply to Pierre-Eric Pelloux-Prayer

These ioctls don't need to GPU, so don't resume it.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 45 +++++++++++++++++++++++--
 1 file changed, 42 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +42 -3
index a9831b243bfc..d21d5af7f187 100644
@@ -2855,7 +2855,9 @@ long amdgpu_drm_ioctl(struct file *filp,
 {
 	struct drm_file *file_priv = filp->private_data;
 	struct drm_device *dev;
+	bool is_driver_ioctl;
 	bool needs_device;
+	unsigned int nr;
 	long ret;
 
 	dev = file_priv->minor->dev;
@@ -2863,9 +2865,46 @@ long amdgpu_drm_ioctl(struct file *filp,
 	/* Some ioctl can opt-out of powermanagement handling
 	 * if they don't require the device to be resumed.
 	 */
-	switch (cmd) {
-	default:
-		needs_device = true;
+	nr = DRM_IOCTL_NR(cmd);
+
+	is_driver_ioctl = nr >= DRM_COMMAND_BASE && nr < DRM_COMMAND_END;
+
+	if (is_driver_ioctl) {
+		switch (nr - DRM_COMMAND_BASE) {
+		/* These 4 only operate on kernel data structure. */
+		case DRM_AMDGPU_VM:
+		case DRM_AMDGPU_SCHED:
+		case DRM_AMDGPU_BO_LIST:
+		case DRM_AMDGPU_FENCE_TO_HANDLE:
+		/* All the waits don't need to resume up the device. If there are
+		 * jobs in progress, the device can't be in suspended state. And if
+		 * there's nothing no in-flight jobs, then the waits are no-op.
+		 */
+		case DRM_AMDGPU_GEM_WAIT_IDLE:
+		case DRM_AMDGPU_WAIT_CS:
+		case DRM_AMDGPU_WAIT_FENCES:
+			needs_device = false;
+			break;
+		default:
+			needs_device = true;
+		}
+	} else {
+		/* Most drm core ioctls don't need the device, but to avoid missing one
+		 * that requires it, implement the "can skip pm" logic as an allow list.
+		 */
+		switch (nr) {
+		case DRM_IOCTL_NR(DRM_IOCTL_VERSION):
+		case DRM_IOCTL_NR(DRM_IOCTL_AUTH_MAGIC):
+		case DRM_IOCTL_NR(DRM_IOCTL_GET_CAP):
+		case DRM_IOCTL_NR(DRM_IOCTL_SYNCOBJ_CREATE):
+		/* Same logic as DRM_AMDGPU_WAIT_* */
+		case DRM_IOCTL_NR(DRM_IOCTL_SYNCOBJ_WAIT):
+		case DRM_IOCTL_NR(DRM_IOCTL_SYNCOBJ_DESTROY):
+			needs_device = false;
+			break;
+		default:
+			needs_device = true;
+		}
 	}
 
 	if (needs_device) {
Signature · 1 line
2.40.1
#8 lore raw

PATCH 3/6 drm/amdgpu: refactor amdgpu_info_ioctl to allow finer pm

In reply to Pierre-Eric Pelloux-Prayer

AMDGPU_INFO_xxx has lots of different queries, and only a small number of them actually reaches out to the GPU.

This commit extract the amdgpu_info_ioctl implementation to a helper function, and then wrap it with the runtime pm logic each query type needs.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c |  2 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 94 ++++++++++++++++++++-----
 2 files changed, 80 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +2 -0
index d21d5af7f187..f51f121d804e 100644
@@ -2871,6 +2871,8 @@ long amdgpu_drm_ioctl(struct file *filp,
 
 	if (is_driver_ioctl) {
 		switch (nr - DRM_COMMAND_BASE) {
+		/* amdgpu_info_ioctl will resume the device if it needs to. */
+		case DRM_AMDGPU_INFO:
 		/* These 4 only operate on kernel data structure. */
 		case DRM_AMDGPU_VM:
 		case DRM_AMDGPU_SCHED:
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +78 -16

106 lines not on this page · raw message

Signature · 1 line
2.40.1
#9 lore raw

PATCH 4/6 drm/amdgpu: add AMDGPU_INFO_GB_ADDR_CONFIG query

In reply to Pierre-Eric Pelloux-Prayer

libdrm_amdgpu uses AMDGPU_INFO_READ_MMR_REG to fill the dev->info.gb_addr_cfg value.
Since this value is already known by the kernel, this commit implements a new query to return it.

The libdrm MR to use this query is:
https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/368

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 5 +++++
 include/uapi/drm/amdgpu_drm.h           | 2 ++
 3 files changed, 9 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +2 -1
index f51f121d804e..403add7f05af 100644
@@ -117,9 +117,10 @@
  * - 3.56.0 - Update IB start address and size alignment for decode and encode
  * - 3.57.0 - Compute tunneling on GFX10+
  * - 3.58.0 - Add AMDGPU_IDS_FLAGS_MODE_PF, AMDGPU_IDS_FLAGS_MODE_VF & AMDGPU_IDS_FLAGS_MODE_PT
+ * - 3.59.0 - Add AMDGPU_INFO_GB_ADDR_CONFIG support
  */
 #define KMS_DRIVER_MAJOR	3
-#define KMS_DRIVER_MINOR	58
+#define KMS_DRIVER_MINOR	59
 #define KMS_DRIVER_PATCHLEVEL	0
 
 /*
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -0
index b32ff6e1baaf..dbb05d51682b 100644
@@ -1256,6 +1256,10 @@ static int amdgpu_info(struct drm_device *dev, void *data, struct drm_file *filp
 		return copy_to_user(out, &gpuvm_fault,
 				    min((size_t)size, sizeof(gpuvm_fault))) ? -EFAULT : 0;
 	}
+	case AMDGPU_INFO_GB_ADDR_CONFIG: {
+		ui32 = adev->gfx.config.gb_addr_config;
+		return copy_to_user(out, &ui32, min(size, 4u)) ? -EFAULT : 0;
+	}
 	default:
 		DRM_DEBUG_KMS("Invalid request %d\n", info->query);
 		return -EINVAL;
@@ -1310,6 +1314,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	case AMDGPU_INFO_VIDEO_CAPS:
 	case AMDGPU_INFO_MAX_IBS:
 	case AMDGPU_INFO_GPUVM_FAULT:
+	case AMDGPU_INFO_GB_ADDR_CONFIG:
 		need_runtime_pm = false;
 		break;
 
include/uapi/drm/amdgpu_drm.h +2 -0
index 3e488b0119eb..680492cd73d8 100644
@@ -933,6 +933,8 @@ struct drm_amdgpu_cs_chunk_cp_gfx_shadow {
 #define AMDGPU_INFO_MAX_IBS			0x22
 /* query last page fault info */
 #define AMDGPU_INFO_GPUVM_FAULT			0x23
+/* Query GB_ADDR_CONFIG */
+#define AMDGPU_INFO_GB_ADDR_CONFIG		0x24
 
 #define AMDGPU_INFO_MMR_SE_INDEX_SHIFT	0
 #define AMDGPU_INFO_MMR_SE_INDEX_MASK	0xff
Signature · 1 line
2.40.1
#10 lore raw

Re: [PATCH 4/6] drm/amdgpu: add AMDGPU_INFO_GB_ADDR_CONFIG query

I would put this into drm_amdgpu_info_device. That structure can grow in size.

Marek

Quoted 71 lines · Pierre-Eric Pelloux-Prayer

libdrm_amdgpu uses AMDGPU_INFO_READ_MMR_REG to fill the dev->info.gb_addr_cfg value.
Since this value is already known by the kernel, this commit implements a new query to return it.

The libdrm MR to use this query is:
https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/368

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 5 +++++
 include/uapi/drm/amdgpu_drm.h           | 2 ++
 3 files changed, 9 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +2 -1
index f51f121d804e..403add7f05af 100644
@@ -117,9 +117,10 @@
  * - 3.56.0 - Update IB start address and size alignment for decode and encode
  * - 3.57.0 - Compute tunneling on GFX10+
  * - 3.58.0 - Add AMDGPU_IDS_FLAGS_MODE_PF, AMDGPU_IDS_FLAGS_MODE_VF & AMDGPU_IDS_FLAGS_MODE_PT
+ * - 3.59.0 - Add AMDGPU_INFO_GB_ADDR_CONFIG support
  */
 #define KMS_DRIVER_MAJOR       3
-#define KMS_DRIVER_MINOR       58
+#define KMS_DRIVER_MINOR       59
 #define KMS_DRIVER_PATCHLEVEL  0
 
 /*
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -0
index b32ff6e1baaf..dbb05d51682b 100644
@@ -1256,6 +1256,10 @@ static int amdgpu_info(struct drm_device *dev, void *data, struct drm_file *filp
                return copy_to_user(out, &gpuvm_fault,
                                    min((size_t)size, sizeof(gpuvm_fault))) ? -EFAULT : 0;
        }
+       case AMDGPU_INFO_GB_ADDR_CONFIG: {
+               ui32 = adev->gfx.config.gb_addr_config;
+               return copy_to_user(out, &ui32, min(size, 4u)) ? -EFAULT : 0;
+       }
        default:
                DRM_DEBUG_KMS("Invalid request %d\n", info->query);
                return -EINVAL;
@@ -1310,6 +1314,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
        case AMDGPU_INFO_VIDEO_CAPS:
        case AMDGPU_INFO_MAX_IBS:
        case AMDGPU_INFO_GPUVM_FAULT:
+       case AMDGPU_INFO_GB_ADDR_CONFIG:
                need_runtime_pm = false;
                break;
 
include/uapi/drm/amdgpu_drm.h +2 -1
index 3e488b0119eb..680492cd73d8 100644
@@ -933,6 +933,8 @@ struct drm_amdgpu_cs_chunk_cp_gfx_shadow {
 #define AMDGPU_INFO_MAX_IBS                    0x22
 /* query last page fault info */
 #define AMDGPU_INFO_GPUVM_FAULT                        0x23
+/* Query GB_ADDR_CONFIG */
+#define AMDGPU_INFO_GB_ADDR_CONFIG             0x24
 
 #define AMDGPU_INFO_MMR_SE_INDEX_SHIFT 0
 #define AMDGPU_INFO_MMR_SE_INDEX_MASK  0xff
--

2.40.1

maintainer
#11 lore raw

Re: [PATCH 4/6] drm/amdgpu: add AMDGPU_INFO_GB_ADDR_CONFIG query

I would try to avoid that.

Putting everything into amdgpu_info_device was a mistake only done because people assumed that IOCTLs on Linux are to expensive to query all information separately.

We should rather have distinct IOCTLs for each value because that is way more flexible and we won't find later that we have to deprecate fields and work around issues because of legacy hw.

Regards,
Christian.

Am 19.06.24 um 02:34 schrieb Marek Olšák:

Quoted 76 lines

I would put this into drm_amdgpu_info_device. That structure can grow in size.

Marek

Quoted 70 lines · Pierre-Eric Pelloux-Prayer

libdrm_amdgpu uses AMDGPU_INFO_READ_MMR_REG to fill the dev->info.gb_addr_cfg value.
Since this value is already known by the kernel, this commit implements a new query to return it.

The libdrm MR to use this query is:
https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/368

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 ++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 5 +++++
  include/uapi/drm/amdgpu_drm.h           | 2 ++
  3 files changed, 9 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +2 -1
index f51f121d804e..403add7f05af 100644
@@ -117,9 +117,10 @@
   * - 3.56.0 - Update IB start address and size alignment for decode and encode
   * - 3.57.0 - Compute tunneling on GFX10+
   * - 3.58.0 - Add AMDGPU_IDS_FLAGS_MODE_PF, AMDGPU_IDS_FLAGS_MODE_VF & AMDGPU_IDS_FLAGS_MODE_PT
+ * - 3.59.0 - Add AMDGPU_INFO_GB_ADDR_CONFIG support
   */
  #define KMS_DRIVER_MAJOR       3
-#define KMS_DRIVER_MINOR       58
+#define KMS_DRIVER_MINOR       59
  #define KMS_DRIVER_PATCHLEVEL  0
 
  /*
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -0
index b32ff6e1baaf..dbb05d51682b 100644
@@ -1256,6 +1256,10 @@ static int amdgpu_info(struct drm_device *dev, void *data, struct drm_file *filp
                 return copy_to_user(out, &gpuvm_fault,
                                     min((size_t)size, sizeof(gpuvm_fault))) ? -EFAULT : 0;
         }
+       case AMDGPU_INFO_GB_ADDR_CONFIG: {
+               ui32 = adev->gfx.config.gb_addr_config;
+               return copy_to_user(out, &ui32, min(size, 4u)) ? -EFAULT : 0;
+       }
         default:
                 DRM_DEBUG_KMS("Invalid request %d\n", info->query);
                 return -EINVAL;
@@ -1310,6 +1314,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
         case AMDGPU_INFO_VIDEO_CAPS:
         case AMDGPU_INFO_MAX_IBS:
         case AMDGPU_INFO_GPUVM_FAULT:
+       case AMDGPU_INFO_GB_ADDR_CONFIG:
                 need_runtime_pm = false;
                 break;
 
include/uapi/drm/amdgpu_drm.h +2 -1
index 3e488b0119eb..680492cd73d8 100644
@@ -933,6 +933,8 @@ struct drm_amdgpu_cs_chunk_cp_gfx_shadow {
  #define AMDGPU_INFO_MAX_IBS                    0x22
  /* query last page fault info */
  #define AMDGPU_INFO_GPUVM_FAULT                        0x23
+/* Query GB_ADDR_CONFIG */
+#define AMDGPU_INFO_GB_ADDR_CONFIG             0x24
 
  #define AMDGPU_INFO_MMR_SE_INDEX_SHIFT 0
  #define AMDGPU_INFO_MMR_SE_INDEX_MASK  0xff
--

2.40.1

#12 lore raw

Re: [PATCH 4/6] drm/amdgpu: add AMDGPU_INFO_GB_ADDR_CONFIG query

The INFO ioctl was designed to allow increasing the sizes of all info structures. GB_ADDR_CONFIG isn't that special to justify a separate query.

Marek

Quoted 92 lines · Christian König

I would try to avoid that.

Putting everything into amdgpu_info_device was a mistake only done because people assumed that IOCTLs on Linux are to expensive to query all information separately.

We should rather have distinct IOCTLs for each value because that is way more flexible and we won't find later that we have to deprecate fields and work around issues because of legacy hw.

Regards,
Christian.

Am 19.06.24 um 02:34 schrieb Marek Olšák:

Quoted 76 lines

I would put this into drm_amdgpu_info_device. That structure can grow in size.

Marek

Quoted 70 lines · Pierre-Eric Pelloux-Prayer

libdrm_amdgpu uses AMDGPU_INFO_READ_MMR_REG to fill the dev->info.gb_addr_cfg value.
Since this value is already known by the kernel, this commit implements a new query to return it.

The libdrm MR to use this query is:
https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/368

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 ++-
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 5 +++++
  include/uapi/drm/amdgpu_drm.h           | 2 ++
  3 files changed, 9 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +2 -1
index f51f121d804e..403add7f05af 100644
@@ -117,9 +117,10 @@
   * - 3.56.0 - Update IB start address and size alignment for decode and encode
   * - 3.57.0 - Compute tunneling on GFX10+
   * - 3.58.0 - Add AMDGPU_IDS_FLAGS_MODE_PF, AMDGPU_IDS_FLAGS_MODE_VF & AMDGPU_IDS_FLAGS_MODE_PT
+ * - 3.59.0 - Add AMDGPU_INFO_GB_ADDR_CONFIG support
   */
  #define KMS_DRIVER_MAJOR       3
-#define KMS_DRIVER_MINOR       58
+#define KMS_DRIVER_MINOR       59
  #define KMS_DRIVER_PATCHLEVEL  0
 
  /*
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -0
index b32ff6e1baaf..dbb05d51682b 100644
@@ -1256,6 +1256,10 @@ static int amdgpu_info(struct drm_device *dev, void *data, struct drm_file *filp
                 return copy_to_user(out, &gpuvm_fault,
                                     min((size_t)size, sizeof(gpuvm_fault))) ? -EFAULT : 0;
         }
+       case AMDGPU_INFO_GB_ADDR_CONFIG: {
+               ui32 = adev->gfx.config.gb_addr_config;
+               return copy_to_user(out, &ui32, min(size, 4u)) ? -EFAULT : 0;
+       }
         default:
                 DRM_DEBUG_KMS("Invalid request %d\n", info->query);
                 return -EINVAL;
@@ -1310,6 +1314,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
         case AMDGPU_INFO_VIDEO_CAPS:
         case AMDGPU_INFO_MAX_IBS:
         case AMDGPU_INFO_GPUVM_FAULT:
+       case AMDGPU_INFO_GB_ADDR_CONFIG:
                 need_runtime_pm = false;
                 break;
 
include/uapi/drm/amdgpu_drm.h +2 -1
index 3e488b0119eb..680492cd73d8 100644
@@ -933,6 +933,8 @@ struct drm_amdgpu_cs_chunk_cp_gfx_shadow {
  #define AMDGPU_INFO_MAX_IBS                    0x22
  /* query last page fault info */
  #define AMDGPU_INFO_GPUVM_FAULT                        0x23
+/* Query GB_ADDR_CONFIG */
+#define AMDGPU_INFO_GB_ADDR_CONFIG             0x24
 
  #define AMDGPU_INFO_MMR_SE_INDEX_SHIFT 0
  #define AMDGPU_INFO_MMR_SE_INDEX_MASK  0xff
--

2.40.1

#13 lore raw

Re: [PATCH 4/6] drm/amdgpu: add AMDGPU_INFO_GB_ADDR_CONFIG query

Both versions are fine by me, so I'll update the code to match whatever you agree on.

Pierre-Eric

Le 19/06/2024 à 20:44, Marek Olšák a écrit :

Quoted 100 lines

The INFO ioctl was designed to allow increasing the sizes of all info structures. GB_ADDR_CONFIG isn't that special to justify a separate query.

Marek

Quoted 92 lines · Christian König

I would try to avoid that.

Putting everything into amdgpu_info_device was a mistake only done because people assumed that IOCTLs on Linux are to expensive to query all information separately.

We should rather have distinct IOCTLs for each value because that is way more flexible and we won't find later that we have to deprecate fields and work around issues because of legacy hw.

Regards,
Christian.

Am 19.06.24 um 02:34 schrieb Marek Olšák:

Quoted 76 lines

I would put this into drm_amdgpu_info_device. That structure can grow in size.

Marek

Quoted 70 lines · Pierre-Eric Pelloux-Prayer

libdrm_amdgpu uses AMDGPU_INFO_READ_MMR_REG to fill the dev->info.gb_addr_cfg value.
Since this value is already known by the kernel, this commit implements a new query to return it.

The libdrm MR to use this query is:
https://gitlab.freedesktop.org/mesa/drm/-/merge_requests/368

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 ++-
   drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 5 +++++
   include/uapi/drm/amdgpu_drm.h           | 2 ++
   3 files changed, 9 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +2 -1
index f51f121d804e..403add7f05af 100644
@@ -117,9 +117,10 @@
    * - 3.56.0 - Update IB start address and size alignment for decode and encode
    * - 3.57.0 - Compute tunneling on GFX10+
    * - 3.58.0 - Add AMDGPU_IDS_FLAGS_MODE_PF, AMDGPU_IDS_FLAGS_MODE_VF & AMDGPU_IDS_FLAGS_MODE_PT
+ * - 3.59.0 - Add AMDGPU_INFO_GB_ADDR_CONFIG support
    */
   #define KMS_DRIVER_MAJOR       3
-#define KMS_DRIVER_MINOR       58
+#define KMS_DRIVER_MINOR       59
   #define KMS_DRIVER_PATCHLEVEL  0
 
   /*
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -0
index b32ff6e1baaf..dbb05d51682b 100644
@@ -1256,6 +1256,10 @@ static int amdgpu_info(struct drm_device *dev, void *data, struct drm_file *filp
                  return copy_to_user(out, &gpuvm_fault,
                                      min((size_t)size, sizeof(gpuvm_fault))) ? -EFAULT : 0;
          }
+       case AMDGPU_INFO_GB_ADDR_CONFIG: {
+               ui32 = adev->gfx.config.gb_addr_config;
+               return copy_to_user(out, &ui32, min(size, 4u)) ? -EFAULT : 0;
+       }
          default:
                  DRM_DEBUG_KMS("Invalid request %d\n", info->query);
                  return -EINVAL;
@@ -1310,6 +1314,7 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
          case AMDGPU_INFO_VIDEO_CAPS:
          case AMDGPU_INFO_MAX_IBS:
          case AMDGPU_INFO_GPUVM_FAULT:
+       case AMDGPU_INFO_GB_ADDR_CONFIG:
                  need_runtime_pm = false;
                  break;
 
include/uapi/drm/amdgpu_drm.h +2 -1
index 3e488b0119eb..680492cd73d8 100644
@@ -933,6 +933,8 @@ struct drm_amdgpu_cs_chunk_cp_gfx_shadow {
   #define AMDGPU_INFO_MAX_IBS                    0x22
   /* query last page fault info */
   #define AMDGPU_INFO_GPUVM_FAULT                        0x23
+/* Query GB_ADDR_CONFIG */
+#define AMDGPU_INFO_GB_ADDR_CONFIG             0x24
 
   #define AMDGPU_INFO_MMR_SE_INDEX_SHIFT 0
   #define AMDGPU_INFO_MMR_SE_INDEX_MASK  0xff
--

2.40.1

#14 lore raw

PATCH 5/6 drm/amdgpu: cache mclk/sclk min/max values

In reply to Pierre-Eric Pelloux-Prayer

This way these values can be returned directly when using AMDGPU_INFO_DEV_INFO, without waking up the GPU.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h           | 12 ++++++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_atombios.c  |  8 ++++----
 .../gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c  |  8 ++++----
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  7 +++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 19 +++++--------------
 drivers/gpu/drm/amd/pm/legacy-dpm/kv_dpm.c    |  8 ++++----
 drivers/gpu/drm/amd/pm/legacy-dpm/si_dpm.c    | 12 ++++++------
 7 files changed, 40 insertions(+), 34 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +10 -2
index 083f353cff6e..75db8eba73d9 100644
@@ -418,8 +418,16 @@ struct amdgpu_clock {
 	struct amdgpu_pll spll;
 	struct amdgpu_pll mpll;
 	/* 10 Khz units */
-	uint32_t default_mclk;
-	uint32_t default_sclk;
+	struct {
+		uint32_t min;
+		uint32_t max;
+		uint32_t def;
+	} mclk;
+	struct {
+		uint32_t min;
+		uint32_t max;
+		uint32_t def;
+	} sclk;
 	uint32_t default_dispclk;
 	uint32_t current_dispclk;
 	uint32_t dp_extclk;
drivers/gpu/drm/amd/amdgpu/amdgpu_atombios.c +4 -4
index 7dc102f0bc1d..f2c2b05233f0 100644
@@ -658,9 +658,9 @@ int amdgpu_atombios_get_clock_info(struct amdgpu_device *adev)
 		mpll->pll_in_max =
 		    le16_to_cpu(firmware_info->info.usMaxMemoryClockPLL_Input);
 
-		adev->clock.default_sclk =
+		adev->clock.sclk.def = adev->clock.sclk.min = adev->clock.sclk.max =
 		    le32_to_cpu(firmware_info->info.ulDefaultEngineClock);
-		adev->clock.default_mclk =
+		adev->clock.mclk.def = adev->clock.mclk.min = adev->clock.mclk.max =
 		    le32_to_cpu(firmware_info->info.ulDefaultMemoryClock);
 
 		mpll->min_post_div = 1;
@@ -699,8 +699,8 @@ int amdgpu_atombios_get_clock_info(struct amdgpu_device *adev)
 		ret = 0;
 	}
 
-	adev->pm.current_sclk = adev->clock.default_sclk;
-	adev->pm.current_mclk = adev->clock.default_mclk;
+	adev->pm.current_sclk = adev->clock.sclk.def;
+	adev->pm.current_mclk = adev->clock.mclk.def;
 
 	return ret;
 }
drivers/gpu/drm/amd/amdgpu/amdgpu_atomfirmware.c +4 -4
index f932bec6e534..6eb125b1bd08 100644
@@ -719,13 +719,13 @@ int amdgpu_atomfirmware_get_clock_info(struct amdgpu_device *adev)
 			(union firmware_info *)(mode_info->atom_context->bios +
 						data_offset);
 
-		adev->clock.default_sclk =
+		adev->clock.sclk.def = adev->clock.sclk.min = adev->clock.sclk.max =
 			le32_to_cpu(firmware_info->v31.bootup_sclk_in10khz);
-		adev->clock.default_mclk =
+		adev->clock.mclk.def = adev->clock.mclk.min = adev->clock.mclk.max =
 			le32_to_cpu(firmware_info->v31.bootup_mclk_in10khz);
 
-		adev->pm.current_sclk = adev->clock.default_sclk;
-		adev->pm.current_mclk = adev->clock.default_mclk;
+		adev->pm.current_sclk = adev->clock.sclk.def;
+		adev->pm.current_mclk = adev->clock.mclk.def;
 
 		ret = 0;
 	}
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7 -0
index 3fb02f5b91c9..03417e7e8961 100644
@@ -4422,6 +4422,13 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 
 	amdgpu_device_check_iommu_direct_map(adev);
 
+	if (adev->pm.dpm_enabled) {
+		adev->clock.sclk.min = amdgpu_dpm_get_sclk(adev, true);
+		adev->clock.sclk.max = amdgpu_dpm_get_sclk(adev, false);
+		adev->clock.mclk.min = amdgpu_dpm_get_mclk(adev, true);
+		adev->clock.mclk.max = amdgpu_dpm_get_mclk(adev, false);
+	}
+
 	return 0;
 
 release_ras_con:
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -14
index dbb05d51682b..781851cf8dc9 100644
@@ -827,19 +827,10 @@ static int amdgpu_info(struct drm_device *dev, void *data, struct drm_file *filp
 		dev_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
 		/* return all clocks in KHz */
 		dev_info->gpu_counter_freq = amdgpu_asic_get_xclk(adev) * 10;
-		if (adev->pm.dpm_enabled) {
-			dev_info->max_engine_clock = amdgpu_dpm_get_sclk(adev, false) * 10;
-			dev_info->max_memory_clock = amdgpu_dpm_get_mclk(adev, false) * 10;
-			dev_info->min_engine_clock = amdgpu_dpm_get_sclk(adev, true) * 10;
-			dev_info->min_memory_clock = amdgpu_dpm_get_mclk(adev, true) * 10;
-		} else {
-			dev_info->max_engine_clock =
-				dev_info->min_engine_clock =
-					adev->clock.default_sclk * 10;
-			dev_info->max_memory_clock =
-				dev_info->min_memory_clock =
-					adev->clock.default_mclk * 10;
-		}
+		dev_info->min_engine_clock = adev->clock.sclk.min * 10;
+		dev_info->max_engine_clock = adev->clock.sclk.max * 10;
+		dev_info->min_memory_clock = adev->clock.mclk.min * 10;
+		dev_info->max_memory_clock = adev->clock.mclk.max * 10;
 		dev_info->enabled_rb_pipes_mask = adev->gfx.config.backend_enable_mask;
 		dev_info->num_rb_pipes = adev->gfx.config.max_backends_per_se *
 			adev->gfx.config.max_shader_engines;
@@ -1315,12 +1306,12 @@ int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
 	case AMDGPU_INFO_MAX_IBS:
 	case AMDGPU_INFO_GPUVM_FAULT:
 	case AMDGPU_INFO_GB_ADDR_CONFIG:
+	case AMDGPU_INFO_DEV_INFO:
 		need_runtime_pm = false;
 		break;
 
 	case AMDGPU_INFO_TIMESTAMP:
 	case AMDGPU_INFO_READ_MMR_REG:
-	case AMDGPU_INFO_DEV_INFO:
 	case AMDGPU_INFO_SENSOR:
 		need_runtime_pm = true;
 		break;
drivers/gpu/drm/amd/pm/legacy-dpm/kv_dpm.c +4 -4
index 6bb42d04b247..75c262acb704 100644
@@ -2996,10 +2996,10 @@ static int kv_dpm_sw_init(void *handle)
 	adev->pm.dpm.state = POWER_STATE_TYPE_BALANCED;
 	adev->pm.dpm.user_state = POWER_STATE_TYPE_BALANCED;
 	adev->pm.dpm.forced_level = AMD_DPM_FORCED_LEVEL_AUTO;
-	adev->pm.default_sclk = adev->clock.default_sclk;
-	adev->pm.default_mclk = adev->clock.default_mclk;
-	adev->pm.current_sclk = adev->clock.default_sclk;
-	adev->pm.current_mclk = adev->clock.default_mclk;
+	adev->pm.default_sclk = adev->clock.sclk.def;
+	adev->pm.default_mclk = adev->clock.mclk.def;
+	adev->pm.current_sclk = adev->clock.sclk.def;
+	adev->pm.current_mclk = adev->clock.mclk.def;
 	adev->pm.int_thermal_type = THERMAL_TYPE_NONE;
 
 	if (amdgpu_dpm == 0)
drivers/gpu/drm/amd/pm/legacy-dpm/si_dpm.c +6 -6
index a1baa13ab2c2..e72eb59af6a5 100644
@@ -7232,8 +7232,8 @@ static void si_parse_pplib_clock_info(struct amdgpu_device *adev,
 	if (rps->class & ATOM_PPLIB_CLASSIFICATION_BOOT) {
 		u16 vddc, vddci, mvdd;
 		amdgpu_atombios_get_default_voltages(adev, &vddc, &vddci, &mvdd);
-		pl->mclk = adev->clock.default_mclk;
-		pl->sclk = adev->clock.default_sclk;
+		pl->mclk = adev->clock.mclk.def;
+		pl->sclk = adev->clock.sclk.def;
 		pl->vddc = vddc;
 		pl->vddci = vddci;
 		si_pi->mvdd_bootup_value = mvdd;
@@ -7733,10 +7733,10 @@ static int si_dpm_sw_init(void *handle)
 	adev->pm.dpm.state = POWER_STATE_TYPE_BALANCED;
 	adev->pm.dpm.user_state = POWER_STATE_TYPE_BALANCED;
 	adev->pm.dpm.forced_level = AMD_DPM_FORCED_LEVEL_AUTO;
-	adev->pm.default_sclk = adev->clock.default_sclk;
-	adev->pm.default_mclk = adev->clock.default_mclk;
-	adev->pm.current_sclk = adev->clock.default_sclk;
-	adev->pm.current_mclk = adev->clock.default_mclk;
+	adev->pm.default_sclk = adev->clock.sclk.def;
+	adev->pm.default_mclk = adev->clock.mclk.def;
+	adev->pm.current_sclk = adev->clock.sclk.def;
+	adev->pm.current_mclk = adev->clock.mclk.def;
 	adev->pm.int_thermal_type = THERMAL_TYPE_NONE;
 
 	if (amdgpu_dpm == 0)
Signature · 1 line
2.40.1
#15 lore raw

PATCH 6/6 drm/amdgpu: resume the device from amdgpu_gem_fault

In reply to Pierre-Eric Pelloux-Prayer

The fault handler may push some work to the GPU through amdgpu_bo_move so use the pm_runtime functions before that.

Since we're in an interrupt context, we can't use the sync version, so pm_runtime_get is called.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +11 -2
index 1f22b4208729..ec120e33536d 100644
@@ -30,6 +30,7 @@
 #include <linux/pagemap.h>
 #include <linux/pci.h>
 #include <linux/dma-buf.h>
+#include <linux/pm_runtime.h>
 
 #include <drm/amdgpu_drm.h>
 #include <drm/drm_drv.h>
@@ -52,9 +53,13 @@ static vm_fault_t amdgpu_gem_fault(struct vm_fault *vmf)
 	vm_fault_t ret;
 	int idx;
 
+	ret = pm_runtime_get(ddev->dev);
+	if (ret < 0)
+		return ret;
+
 	ret = ttm_bo_vm_reserve(bo, vmf);
 	if (ret)
-		return ret;
+		goto put_pm;
 
 	if (drm_dev_enter(ddev, &idx)) {
 		ret = amdgpu_bo_fault_reserve_notify(bo);
@@ -71,10 +76,14 @@ static vm_fault_t amdgpu_gem_fault(struct vm_fault *vmf)
 		ret = ttm_bo_vm_dummy_page(vmf, vmf->vma->vm_page_prot);
 	}
 	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
-		return ret;
+		goto put_pm;
 
 unlock:
 	dma_resv_unlock(bo->base.resv);
+put_pm:
+	pm_runtime_mark_last_busy(ddev->dev);
+	pm_runtime_put_autosuspend(ddev->dev);
+
 	return ret;
 }
 
Signature · 1 line
2.40.1