On 9/28/26 20:57, Sunil Khatri wrote:
> The SPM (Streaming Performance Monitor) path needs a VMID that is
> isolated from other GPU contexts. Introduce vm_manager.spm_vmid to let a
> GFXHUB dedicate a VMID to SPM.
>
> When spm_vmid is set, remove that VMID from the MES user-queue mask and
> from the KFD compute pool so that neither MES firmware nor KFD ever maps
> another context onto it. KFD derives its VMID range from the shared
> compute_vmid_bitmap via ffs()/fls(), so clearing the bit there is
> sufficient and needs no change inside the KFD subsystem.
>
> The field defaults to 0 (unused), so this is a no-op until a generation
> opts in by setting spm_vmid.
>
> Signed-off-by: Sunil Khatri <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 4 ++++
> drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 3 +++
> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 2 ++
> 3 files changed, 9 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> index afcfb92fb9f5f..71e47bdf11393 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
> @@ -207,6 +207,10 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device
> *adev)
> for (i = last_valid_bit; i < AMDGPU_MAX_QUEUES; ++i)
> clear_bit(i, gpu_resources.cp_queue_bitmap);
>
> + if (adev->vm_manager.spm_vmid)
> + gpu_resources.compute_vmid_bitmap &=
> + ~BIT(adev->vm_manager.spm_vmid);
> +
That is quite a mess.
> amdgpu_doorbell_get_kfd_info(adev,
> &gpu_resources.doorbell_physical_address,
> &gpu_resources.doorbell_aperture_size,
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
> index 22ecb70fec2cd..afb56a2f9672b 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
> @@ -134,6 +134,9 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
> adev->mes.vmid_mask_mmhub = 0xFF00;
> adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
>
> + if (adev->vm_manager.spm_vmid)
> + adev->mes.vmid_mask_gfxhub &= ~BIT(adev->vm_manager.spm_vmid);
Please remove the vmid_mask_gfxhub field from the MES structure if possible and
move all of that into adev->vm_manager or adev->vm_manager.id_mgr instead.
Regards,
Christian.
> +
> num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
> if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
> dev_warn(adev->dev, "more gfx pipes than supported by MES! (%d
> vs %d)\n",
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index 0f6634b749746..57070da337750 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -414,6 +414,8 @@ struct amdgpu_vm_manager {
> /* Global registration of recent page fault information */
> struct amdgpu_vm_fault_info fault_info;
> unsigned int npa_vmid;
> + /* VMID reserved for SPM, 0 if unused */
> + unsigned int spm_vmid;
> };
>
> struct amdgpu_bo_va_mapping;
