[PATCH v2 0/3] Make device links between KFD and GPU device

[PATCH v2 0/3] Make device links between KFD and GPU device

[Mario Limonciello (AMD)]

Discovering which KFD device is associated with a GPU is relatively
awkward right now in userspace.

This series creates sysfs links between the devices to simplify it
for userspace.

v2:
 * Add tag
 * Fix case that systems > 1 GPU wouldn't show links due to rebuilding
   topology.

Mario Limonciello (AMD) (3):
  amdkfd: Only ignore -ENOENT for KFD init failuires
  amdkfd: Don't rebuild node tree when calling
    kfd_topology_update_sysfs()
  amdkfd: Add device links between kfd device and amdgpu device

 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  8 ++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  4 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c       |  6 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_topology.c     | 42 +++++++++++++++++--
 .../gpu/drm/amd/include/kgd_kfd_interface.h   |  2 +
 6 files changed, 57 insertions(+), 6 deletions(-)

-- 
2.43.0

[PATCH v2 1/3] amdkfd: Only ignore -ENOENT for KFD init failuires

[Mario Limonciello (AMD)]

When compiled without CONFIG_HSA_AMD KFD will return -ENOENT.
As other errors will cause KFD functionality issues this is the
only error code that should be ignored at init.

Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
Reviewed-by: Kent Russell <kent.russell@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index 16adeba4d7e68..e804461e5f272 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -3169,8 +3169,10 @@ static int __init amdgpu_init(void)
 	amdgpu_register_atpx_handler();
 	amdgpu_acpi_detect();
 
-	/* Ignore KFD init failures. Normal when CONFIG_HSA_AMD is not set. */
-	amdgpu_amdkfd_init();
+	/* Ignore KFD init failures when CONFIG_HSA_AMD is not set. */
+	r = amdgpu_amdkfd_init();
+	if (r && r != -ENOENT)
+		goto error_fence;
 
 	if (amdgpu_pp_feature_mask & PP_OVERDRIVE_MASK) {
 		add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK);
-- 
2.43.0

[PATCH v2 2/3] amdkfd: Don't rebuild node tree when calling kfd_topology_update_sysfs()

[Mario Limonciello (AMD)]

There is no need to remove all the nodes and rebuild them.  The content
will be the same.  Instead check whether the node was created and skip
the creation.

Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
---
 drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
index a0990dd2378c1..b65f29294e2d6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
@@ -650,8 +650,8 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,
 	uint32_t i, num_attrs;
 	struct attribute **attrs;
 
-	if (WARN_ON(dev->kobj_node))
-		return -EEXIST;
+	if (dev->kobj_node)
+		return 0;
 
 	/*
 	 * Creating the sysfs folders
@@ -888,8 +888,6 @@ static int kfd_topology_update_sysfs(void)
 			return ret;
 	}
 
-	kfd_remove_sysfs_node_tree();
-
 	return kfd_build_sysfs_node_tree();
 }
 
-- 
2.43.0

[PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu device

[Mario Limonciello (AMD)]

Mapping out a KFD device to a GPU can be done manually by looking at the
domain and location properties.  To make it easier to discover which
KFD device goes with what GPU add bidirectional links.

Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  8 +++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  4 +++
 drivers/gpu/drm/amd/amdkfd/kfd_topology.c     | 36 +++++++++++++++++++
 .../gpu/drm/amd/include/kgd_kfd_interface.h   |  2 ++
 5 files changed, 51 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index a2879d2b7c8ec..5d6cf3adfa7b8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -910,3 +910,11 @@ int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id,
 
 	return r;
 }
+
+int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev)
+{
+	if (!adev->kfd.init_complete || !adev->kfd.dev)
+		return 0;
+
+	return kgd2kfd_create_sysfs_links(adev->kfd.dev);
+}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index da4575676335f..fd92b227a674b 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -272,6 +272,7 @@ int amdgpu_amdkfd_stop_sched(struct amdgpu_device *adev, uint32_t node_id);
 int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id,
 	bool core_override_enable, bool reg_override_enable, bool perfmon_override_enable);
 bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t node_id);
+int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev);
 
 
 /* Read user wptr from a specified user address space with page fault
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 7a0213a07023d..44c9320d72a56 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -4947,6 +4947,10 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 	 */
 	r = amdgpu_device_sys_interface_init(adev);
 
+	r = amdgpu_amdkfd_create_sysfs_links(adev);
+	if (r)
+		dev_err(adev->dev, "Failed to create KFD sysfs link: %d\n", r);
+
 	if (IS_ENABLED(CONFIG_PERF_EVENTS))
 		r = amdgpu_pmu_init(adev);
 	if (r)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
index b65f29294e2d6..796fd411a7dcc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
@@ -79,6 +79,37 @@ struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
 	return device;
 }
 
+int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd)
+{
+	struct kfd_topology_device *top_dev;
+	int ret = -ENODEV;
+
+	if (!kfd)
+		return -EINVAL;
+
+	down_read(&topology_lock);
+
+	list_for_each_entry(top_dev, &topology_device_list, list) {
+		struct kobject *amdgpu_kobj;
+
+		if (!top_dev->gpu || top_dev->gpu->kfd != kfd || !top_dev->kobj_node)
+			continue;
+
+		amdgpu_kobj = &top_dev->gpu->adev->dev->kobj;
+		ret = sysfs_create_link(top_dev->kobj_node, amdgpu_kobj, "device");
+		if (ret)
+			break;
+
+		ret = sysfs_create_link(amdgpu_kobj, top_dev->kobj_node, "kfd");
+		if (ret)
+			sysfs_remove_link(top_dev->kobj_node, "device");
+		break;
+	}
+
+	up_read(&topology_lock);
+	return ret;
+}
+
 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
 {
 	struct kfd_topology_device *top_dev = NULL;
@@ -571,6 +602,11 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
 	struct kfd_mem_properties *mem;
 	struct kfd_perf_properties *perf;
 
+	if (dev->gpu) {
+		sysfs_remove_link(dev->kobj_node, "device");
+		sysfs_remove_link(&dev->gpu->adev->dev->kobj, "kfd");
+	}
+
 	if (dev->kobj_iolink) {
 		list_for_each_entry(iolink, &dev->io_link_props, list)
 			if (iolink->kobj) {
diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
index 9aba8596faa7e..f6db1dc634399 100644
--- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
+++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
@@ -335,4 +335,6 @@ struct kfd2kgd_calls {
 					  int engine, int queue);
 };
 
+int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd);
+
 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
-- 
2.43.0

RE: [PATCH v2 2/3] amdkfd: Don't rebuild node tree when calling kfd_topology_update_sysfs()

[Russell, Kent]

[Public]

Seems reasonable to me, but I'd want Felix (unlikely due to PTO) or Harish to weigh in to make sure I am not missing something obvious.

From what I can tell, it was called way back when KFD Topology was first written, back when it was for APUs-only. So they'd do 2 calls to update sysfs, once for the CPU and once for the GPU. I think the logic just never got touched because it still works, even though we've changed things a bit now.

 Kent

> -----Original Message-----
> From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> On Behalf Of Mario
> Limonciello (AMD)
> Sent: Wednesday, December 10, 2025 3:15 PM
> To: amd-gfx@lists.freedesktop.org
> Cc: Mario Limonciello (AMD) <superm1@kernel.org>
> Subject: [PATCH v2 2/3] amdkfd: Don't rebuild node tree when calling
> kfd_topology_update_sysfs()
>
> There is no need to remove all the nodes and rebuild them.  The content
> will be the same.  Instead check whether the node was created and skip
> the creation.
>
> Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
> ---
>  drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> index a0990dd2378c1..b65f29294e2d6 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> @@ -650,8 +650,8 @@ static int kfd_build_sysfs_node_entry(struct
> kfd_topology_device *dev,
>       uint32_t i, num_attrs;
>       struct attribute **attrs;
>
> -     if (WARN_ON(dev->kobj_node))
> -             return -EEXIST;
> +     if (dev->kobj_node)
> +             return 0;
>
>       /*
>        * Creating the sysfs folders
> @@ -888,8 +888,6 @@ static int kfd_topology_update_sysfs(void)
>                       return ret;
>       }
>
> -     kfd_remove_sysfs_node_tree();
> -
>       return kfd_build_sysfs_node_tree();
>  }
>
> --
> 2.43.0

RE: [PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu device

[Russell, Kent]

[Public]

Reviewed-by: Kent Russell <kent.russell@amd.com>



> -----Original Message-----
> From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> On Behalf Of Mario
> Limonciello (AMD)
> Sent: Wednesday, December 10, 2025 3:15 PM
> To: amd-gfx@lists.freedesktop.org
> Cc: Mario Limonciello (AMD) <superm1@kernel.org>
> Subject: [PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu
> device
>
> Mapping out a KFD device to a GPU can be done manually by looking at the
> domain and location properties.  To make it easier to discover which
> KFD device goes with what GPU add bidirectional links.
>
> Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  8 +++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  4 +++
>  drivers/gpu/drm/amd/amdkfd/kfd_topology.c     | 36 +++++++++++++++++++
>  .../gpu/drm/amd/include/kgd_kfd_interface.h   |  2 ++
>  5 files changed, 51 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> index a2879d2b7c8ec..5d6cf3adfa7b8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> @@ -910,3 +910,11 @@ int amdgpu_amdkfd_config_sq_perfmon(struct
> amdgpu_device *adev, uint32_t xcp_id,
>
>       return r;
>  }
> +
> +int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev)
> +{
> +     if (!adev->kfd.init_complete || !adev->kfd.dev)
> +             return 0;
> +
> +     return kgd2kfd_create_sysfs_links(adev->kfd.dev);
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> index da4575676335f..fd92b227a674b 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> @@ -272,6 +272,7 @@ int amdgpu_amdkfd_stop_sched(struct amdgpu_device
> *adev, uint32_t node_id);
>  int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t
> xcp_id,
>       bool core_override_enable, bool reg_override_enable, bool
> perfmon_override_enable);
>  bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t
> node_id);
> +int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev);
>
>
>  /* Read user wptr from a specified user address space with page fault
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 7a0213a07023d..44c9320d72a56 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -4947,6 +4947,10 @@ int amdgpu_device_init(struct amdgpu_device *adev,
>        */
>       r = amdgpu_device_sys_interface_init(adev);
>
> +     r = amdgpu_amdkfd_create_sysfs_links(adev);
> +     if (r)
> +             dev_err(adev->dev, "Failed to create KFD sysfs link: %d\n", r);
> +
>       if (IS_ENABLED(CONFIG_PERF_EVENTS))
>               r = amdgpu_pmu_init(adev);
>       if (r)
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> index b65f29294e2d6..796fd411a7dcc 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> @@ -79,6 +79,37 @@ struct kfd_topology_device
> *kfd_topology_device_by_proximity_domain(
>       return device;
>  }
>
> +int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd)
> +{
> +     struct kfd_topology_device *top_dev;
> +     int ret = -ENODEV;
> +
> +     if (!kfd)
> +             return -EINVAL;
> +
> +     down_read(&topology_lock);
> +
> +     list_for_each_entry(top_dev, &topology_device_list, list) {
> +             struct kobject *amdgpu_kobj;
> +
> +             if (!top_dev->gpu || top_dev->gpu->kfd != kfd || !top_dev->kobj_node)
> +                     continue;
> +
> +             amdgpu_kobj = &top_dev->gpu->adev->dev->kobj;
> +             ret = sysfs_create_link(top_dev->kobj_node, amdgpu_kobj, "device");
> +             if (ret)
> +                     break;
> +
> +             ret = sysfs_create_link(amdgpu_kobj, top_dev->kobj_node, "kfd");
> +             if (ret)
> +                     sysfs_remove_link(top_dev->kobj_node, "device");
> +             break;
> +     }
> +
> +     up_read(&topology_lock);
> +     return ret;
> +}
> +
>  struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
>  {
>       struct kfd_topology_device *top_dev = NULL;
> @@ -571,6 +602,11 @@ static void kfd_remove_sysfs_node_entry(struct
> kfd_topology_device *dev)
>       struct kfd_mem_properties *mem;
>       struct kfd_perf_properties *perf;
>
> +     if (dev->gpu) {
> +             sysfs_remove_link(dev->kobj_node, "device");
> +             sysfs_remove_link(&dev->gpu->adev->dev->kobj, "kfd");
> +     }
> +
>       if (dev->kobj_iolink) {
>               list_for_each_entry(iolink, &dev->io_link_props, list)
>                       if (iolink->kobj) {
> diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> index 9aba8596faa7e..f6db1dc634399 100644
> --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> @@ -335,4 +335,6 @@ struct kfd2kgd_calls {
>                                         int engine, int queue);
>  };
>
> +int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd);
> +
>  #endif       /* KGD_KFD_INTERFACE_H_INCLUDED */
> --
> 2.43.0

Re: [PATCH v2 2/3] amdkfd: Don't rebuild node tree when calling kfd_topology_update_sysfs()

[Kasiviswanathan, Harish]

Hi Mario,

This wouldn't work when we change compute and/or memory partitions. After a partition switch we need to recreate all the nodes. If you have an instinct machine you can try changing partitions 

$ rocm-smi --setcomputepartition CPX

Best Regards,
Harish





________________________________________
From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> on behalf of Mario Limonciello (AMD) <superm1@kernel.org>
Sent: Wednesday, December 10, 2025 3:15 PM
To: amd-gfx@lists.freedesktop.org <amd-gfx@lists.freedesktop.org>
Cc: Mario Limonciello (AMD) <superm1@kernel.org>
Subject: [PATCH v2 2/3] amdkfd: Don't rebuild node tree when calling kfd_topology_update_sysfs()


There is no need to remove all the nodes and rebuild them.  The content

will be the same.  Instead check whether the node was created and skip

the creation.



Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>

---

 drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 6 ++----

 1 file changed, 2 insertions(+), 4 deletions(-)



diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c

index a0990dd2378c1..b65f29294e2d6 100644

--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c

+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c

@@ -650,8 +650,8 @@ static int kfd_build_sysfs_node_entry(struct kfd_topology_device *dev,

         uint32_t i, num_attrs;

         struct attribute **attrs;

 

-       if (WARN_ON(dev->kobj_node))

-               return -EEXIST;

+       if (dev->kobj_node)

+               return 0;

 

         /*

          * Creating the sysfs folders

@@ -888,8 +888,6 @@ static int kfd_topology_update_sysfs(void)

                         return ret;

         }

 

-       kfd_remove_sysfs_node_tree();

-

         return kfd_build_sysfs_node_tree();

 }

 

--

2.43.0

Re: [PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu device

[Kasiviswanathan, Harish]

Similar comment to previous patch. Once you do spatial partitioning (like QPX or CPX), there is no one-to-one correspondence between drm node and kfd node. Partitions don't change device node however, you could have multiple (4 or 8) kfd nodes. 

Best Regards,
Harish


________________________________________
From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> on behalf of Mario Limonciello (AMD) <superm1@kernel.org>
Sent: Wednesday, December 10, 2025 3:15 PM
To: amd-gfx@lists.freedesktop.org <amd-gfx@lists.freedesktop.org>
Cc: Mario Limonciello (AMD) <superm1@kernel.org>
Subject: [PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu device


Mapping out a KFD device to a GPU can be done manually by looking at the

domain and location properties.  To make it easier to discover which

KFD device goes with what GPU add bidirectional links.



Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>

---

 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  8 +++++

 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +

 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  4 +++

 drivers/gpu/drm/amd/amdkfd/kfd_topology.c     | 36 +++++++++++++++++++

 .../gpu/drm/amd/include/kgd_kfd_interface.h   |  2 ++

 5 files changed, 51 insertions(+)



diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c

index a2879d2b7c8ec..5d6cf3adfa7b8 100644

--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c

@@ -910,3 +910,11 @@ int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id,

 

         return r;

 }

+

+int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev)

+{

+       if (!adev->kfd.init_complete || !adev->kfd.dev)

+               return 0;

+

+       return kgd2kfd_create_sysfs_links(adev->kfd.dev);

+}

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h

index da4575676335f..fd92b227a674b 100644

--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h

@@ -272,6 +272,7 @@ int amdgpu_amdkfd_stop_sched(struct amdgpu_device *adev, uint32_t node_id);

 int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id,

         bool core_override_enable, bool reg_override_enable, bool perfmon_override_enable);

 bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t node_id);

+int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev);

 

 

 /* Read user wptr from a specified user address space with page fault

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

index 7a0213a07023d..44c9320d72a56 100644

--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

@@ -4947,6 +4947,10 @@ int amdgpu_device_init(struct amdgpu_device *adev,

          */

         r = amdgpu_device_sys_interface_init(adev);

 

+       r = amdgpu_amdkfd_create_sysfs_links(adev);

+       if (r)

+               dev_err(adev->dev, "Failed to create KFD sysfs link: %d\n", r);

+

         if (IS_ENABLED(CONFIG_PERF_EVENTS))

                 r = amdgpu_pmu_init(adev);

         if (r)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c

index b65f29294e2d6..796fd411a7dcc 100644

--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c

+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c

@@ -79,6 +79,37 @@ struct kfd_topology_device *kfd_topology_device_by_proximity_domain(

         return device;

 }

 

+int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd)

+{

+       struct kfd_topology_device *top_dev;

+       int ret = -ENODEV;

+

+       if (!kfd)

+               return -EINVAL;

+

+       down_read(&topology_lock);

+

+       list_for_each_entry(top_dev, &topology_device_list, list) {

+               struct kobject *amdgpu_kobj;

+

+               if (!top_dev->gpu || top_dev->gpu->kfd != kfd || !top_dev->kobj_node)

+                       continue;

+

+               amdgpu_kobj = &top_dev->gpu->adev->dev->kobj;

+               ret = sysfs_create_link(top_dev->kobj_node, amdgpu_kobj, "device");

+               if (ret)

+                       break;

+

+               ret = sysfs_create_link(amdgpu_kobj, top_dev->kobj_node, "kfd");

+               if (ret)

+                       sysfs_remove_link(top_dev->kobj_node, "device");

+               break;

+       }

+

+       up_read(&topology_lock);

+       return ret;

+}

+

 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)

 {

         struct kfd_topology_device *top_dev = NULL;

@@ -571,6 +602,11 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)

         struct kfd_mem_properties *mem;

         struct kfd_perf_properties *perf;

 

+       if (dev->gpu) {

+               sysfs_remove_link(dev->kobj_node, "device");

+               sysfs_remove_link(&dev->gpu->adev->dev->kobj, "kfd");

+       }

+

         if (dev->kobj_iolink) {

                 list_for_each_entry(iolink, &dev->io_link_props, list)

                         if (iolink->kobj) {

diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h

index 9aba8596faa7e..f6db1dc634399 100644

--- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h

+++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h

@@ -335,4 +335,6 @@ struct kfd2kgd_calls {

                                           int engine, int queue);

 };

 

+int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd);

+

 #endif  /* KGD_KFD_INTERFACE_H_INCLUDED */

--

2.43.0

Re: [PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu device

[Mario Limonciello]

Thanks.  I've posted a v3 that hopefully addresses this.

On 12/15/25 8:45 PM, Kasiviswanathan, Harish wrote:
> Similar comment to previous patch. Once you do spatial partitioning (like QPX or CPX), there is no one-to-one correspondence between drm node and kfd node. Partitions don't change device node however, you could have multiple (4 or 8) kfd nodes.
> 
> Best Regards,
> Harish
> 
> 
> ________________________________________
> From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> on behalf of Mario Limonciello (AMD) <superm1@kernel.org>
> Sent: Wednesday, December 10, 2025 3:15 PM
> To: amd-gfx@lists.freedesktop.org <amd-gfx@lists.freedesktop.org>
> Cc: Mario Limonciello (AMD) <superm1@kernel.org>
> Subject: [PATCH v2 3/3] amdkfd: Add device links between kfd device and amdgpu device
> 
> 
> Mapping out a KFD device to a GPU can be done manually by looking at the
> 
> domain and location properties.  To make it easier to discover which
> 
> KFD device goes with what GPU add bidirectional links.
> 
> 
> 
> Signed-off-by: Mario Limonciello (AMD) <superm1@kernel.org>
> 
> ---
> 
>   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c    |  8 +++++
> 
>   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h    |  1 +
> 
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c    |  4 +++
> 
>   drivers/gpu/drm/amd/amdkfd/kfd_topology.c     | 36 +++++++++++++++++++
> 
>   .../gpu/drm/amd/include/kgd_kfd_interface.h   |  2 ++
> 
>   5 files changed, 51 insertions(+)
> 
> 
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> 
> index a2879d2b7c8ec..5d6cf3adfa7b8 100644
> 
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> 
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> 
> @@ -910,3 +910,11 @@ int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id,
> 
>   
> 
>           return r;
> 
>   }
> 
> +
> 
> +int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev)
> 
> +{
> 
> +       if (!adev->kfd.init_complete || !adev->kfd.dev)
> 
> +               return 0;
> 
> +
> 
> +       return kgd2kfd_create_sysfs_links(adev->kfd.dev);
> 
> +}
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> 
> index da4575676335f..fd92b227a674b 100644
> 
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> 
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
> 
> @@ -272,6 +272,7 @@ int amdgpu_amdkfd_stop_sched(struct amdgpu_device *adev, uint32_t node_id);
> 
>   int amdgpu_amdkfd_config_sq_perfmon(struct amdgpu_device *adev, uint32_t xcp_id,
> 
>           bool core_override_enable, bool reg_override_enable, bool perfmon_override_enable);
> 
>   bool amdgpu_amdkfd_compute_active(struct amdgpu_device *adev, uint32_t node_id);
> 
> +int amdgpu_amdkfd_create_sysfs_links(struct amdgpu_device *adev);
> 
>   
> 
>   
> 
>   /* Read user wptr from a specified user address space with page fault
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> 
> index 7a0213a07023d..44c9320d72a56 100644
> 
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> 
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> 
> @@ -4947,6 +4947,10 @@ int amdgpu_device_init(struct amdgpu_device *adev,
> 
>            */
> 
>           r = amdgpu_device_sys_interface_init(adev);
> 
>   
> 
> +       r = amdgpu_amdkfd_create_sysfs_links(adev);
> 
> +       if (r)
> 
> +               dev_err(adev->dev, "Failed to create KFD sysfs link: %d\n", r);
> 
> +
> 
>           if (IS_ENABLED(CONFIG_PERF_EVENTS))
> 
>                   r = amdgpu_pmu_init(adev);
> 
>           if (r)
> 
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> 
> index b65f29294e2d6..796fd411a7dcc 100644
> 
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> 
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
> 
> @@ -79,6 +79,37 @@ struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
> 
>           return device;
> 
>   }
> 
>   
> 
> +int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd)
> 
> +{
> 
> +       struct kfd_topology_device *top_dev;
> 
> +       int ret = -ENODEV;
> 
> +
> 
> +       if (!kfd)
> 
> +               return -EINVAL;
> 
> +
> 
> +       down_read(&topology_lock);
> 
> +
> 
> +       list_for_each_entry(top_dev, &topology_device_list, list) {
> 
> +               struct kobject *amdgpu_kobj;
> 
> +
> 
> +               if (!top_dev->gpu || top_dev->gpu->kfd != kfd || !top_dev->kobj_node)
> 
> +                       continue;
> 
> +
> 
> +               amdgpu_kobj = &top_dev->gpu->adev->dev->kobj;
> 
> +               ret = sysfs_create_link(top_dev->kobj_node, amdgpu_kobj, "device");
> 
> +               if (ret)
> 
> +                       break;
> 
> +
> 
> +               ret = sysfs_create_link(amdgpu_kobj, top_dev->kobj_node, "kfd");
> 
> +               if (ret)
> 
> +                       sysfs_remove_link(top_dev->kobj_node, "device");
> 
> +               break;
> 
> +       }
> 
> +
> 
> +       up_read(&topology_lock);
> 
> +       return ret;
> 
> +}
> 
> +
> 
>   struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id)
> 
>   {
> 
>           struct kfd_topology_device *top_dev = NULL;
> 
> @@ -571,6 +602,11 @@ static void kfd_remove_sysfs_node_entry(struct kfd_topology_device *dev)
> 
>           struct kfd_mem_properties *mem;
> 
>           struct kfd_perf_properties *perf;
> 
>   
> 
> +       if (dev->gpu) {
> 
> +               sysfs_remove_link(dev->kobj_node, "device");
> 
> +               sysfs_remove_link(&dev->gpu->adev->dev->kobj, "kfd");
> 
> +       }
> 
> +
> 
>           if (dev->kobj_iolink) {
> 
>                   list_for_each_entry(iolink, &dev->io_link_props, list)
> 
>                           if (iolink->kobj) {
> 
> diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> 
> index 9aba8596faa7e..f6db1dc634399 100644
> 
> --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> 
> +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
> 
> @@ -335,4 +335,6 @@ struct kfd2kgd_calls {
> 
>                                             int engine, int queue);
> 
>   };
> 
>   
> 
> +int kgd2kfd_create_sysfs_links(struct kfd_dev *kfd);
> 
> +
> 
>   #endif  /* KGD_KFD_INTERFACE_H_INCLUDED */
> 
> --
> 
> 2.43.0
> 
> 
>