On 6/3/26 18:33, Srinivasan Shanmugam wrote:
> The current EVENTFD UAPI describes subscriptions using a userspace-defined
> event_id and reserves queue_id for future use.
>
> That no longer matches the intended render-node event model. Userspace
> should subscribe to kernel-defined event types, while the kernel keeps
> control of event semantics and signaling sources.
>
> Add a shared UAPI enum for AMDGPU EVENTFD event types and update the
> EVENTFD ioctl documentation to describe subscriptions as notification-only.
>
> queue_id remains part of the UAPI and is now used to distinguish
> queue-scoped subscriptions from GPU/device-scoped subscriptions.
>
> USERQ_EOP, QUEUE_RESET, and SCRATCH may use queue_id for queue-specific
> subscriptions. MEMORY_EXCEPTION is currently GPU-scoped and requires
> queue_id = 0.
>
> Cc: Alex Deucher <[email protected]>
> Cc: Christian König <[email protected]>
> Signed-off-by: Srinivasan Shanmugam <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 6 +++---
> include/uapi/drm/amdgpu_drm.h | 25 +++++++++++++++++++------
> 2 files changed, 22 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
> index cd127a271c19..75fc3a74db28 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
> @@ -643,7 +643,7 @@ int amdgpu_eventfd_ioctl(struct drm_device *dev, void
> *data,
> if (!fpriv)
> return -EINVAL;
>
> - if (args->flags || !args->event_id || args->eventfd < 0)
> + if (args->flags || !args->event_type || args->eventfd < 0)
> return -EINVAL;
>
> /*
> @@ -656,11 +656,11 @@ int amdgpu_eventfd_ioctl(struct drm_device *dev, void
> *data,
> switch (args->op) {
> case DRM_AMDGPU_EVENTFD_OP_BIND:
> return amdgpu_eventfd_bind(&fpriv->eventfd_mgr,
> - args->event_id,
> + args->event_type,
> args->eventfd);
> case DRM_AMDGPU_EVENTFD_OP_UNBIND:
> return amdgpu_eventfd_unbind(&fpriv->eventfd_mgr,
> - args->event_id,
> + args->event_type,
> args->eventfd);
> default:
> return -EINVAL;
> diff --git a/include/uapi/drm/amdgpu_drm.h b/include/uapi/drm/amdgpu_drm.h
> index 63cde4487c47..3dffd9e581a1 100644
> --- a/include/uapi/drm/amdgpu_drm.h
> +++ b/include/uapi/drm/amdgpu_drm.h
> @@ -207,28 +207,41 @@ union drm_amdgpu_gem_create {
> struct drm_amdgpu_gem_create_out out;
> };
>
> +enum drm_amdgpu_event_type {
> + DRM_AMDGPU_EVENT_TYPE_INVALID = 0,
What exactly is the invalid event_type?
> + DRM_AMDGPU_EVENT_TYPE_USERQ_EOP = 1,
> + DRM_AMDGPU_EVENT_TYPE_QUEUE_RESET = 2,
> + DRM_AMDGPU_EVENT_TYPE_MEMORY_EXCEPTION = 3,
> + DRM_AMDGPU_EVENT_TYPE_SCRATCH = 4,
> + DRM_AMDGPU_EVENT_TYPE_GPU_RESET = 5,
> +};
> +
> enum drm_amdgpu_eventfd_op {
> DRM_AMDGPU_EVENTFD_OP_BIND = 0,
> DRM_AMDGPU_EVENTFD_OP_UNBIND = 1,
> };
>
> /**
> - * struct drm_amdgpu_eventfd - bind or unbind an eventfd to an event_id
> + * struct drm_amdgpu_eventfd - bind or unbind an eventfd to an AMDGPU event
> * @op: operation type, see &enum drm_amdgpu_eventfd_op
> - * @event_id: userspace-defined event identifier
> + * @event_type: kernel-defined AMDGPU event type
> * @eventfd: eventfd file descriptor
> - * @queue_id: queue identifier for future queue-specific subscriptions
> + * @queue_id: queue identifier for queue-scoped subscriptions, or 0 for
> + * device/GPU-scoped subscriptions
> * @flags: must be 0
> *
> * This ioctl lets userspace register or unregister eventfd notifications
> * for a render-node event.
> *
> - * For now, @queue_id is reserved for future queue-specific subscriptions
> - * and must be 0.
> + * Eventfd signaling is notification-only.
> + *
> + * USERQ_EOP, QUEUE_RESET, and SCRATCH may use @queue_id for queue-scoped
> + * subscriptions. MEMORY_EXCEPTION is currently GPU-scoped and requires
> + * @queue_id = 0.
> */
> struct drm_amdgpu_eventfd {
> __u32 op;
> - __u32 event_id;
> + __u32 event_type;
That rename should be in the original patch adding the interface.
Maybe best to squash the two patches together.
Regards,
Christian.
> __s32 eventfd;
> __u32 queue_id;
> __u32 flags;
