On 10/12/26 11:24, 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, exclude it from both the kernel-queue pool and
> vmid_uq_mask_gfxhub, so neither the round-robin kernel-queue allocator
> nor MES/Userqueues ever maps another context onto it.
>
> Assisted-by: Claude:claude-sonnet-5
> Signed-off-by: Sunil Khatri <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c | 46 ++++++++++++++++---------
> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 2 ++
> 2 files changed, 31 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> index 392a18b1f2b4b..ed1efde29a9a5 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ids.c
> @@ -638,26 +638,45 @@ void amdgpu_vmid_reset_all(struct amdgpu_device *adev)
> * @vmid_kq_mask_gfxhub: kernel-queue VMID pool mask for GFXHUB
> * @vmid_kq_mask_mmhub: kernel-queue VMID pool mask for MMHUB
> *
> - * Sets both hubs' kernel-queue VMID pools (tracked least-recently-used
> - * via id_mgr[].ids_lru), and both hubs' KFD/MES user-queue masks.
> - * vmid_uq_mask_gfxhub is set to the
> - * complement of @vmid_kq_mask_gfxhub, since unlike MMHUB's it is always a
> - * clean partition of the full VMID range; vmid_uq_mask_mmhub is the
> - * fixed pool set here instead, since MMHUB's kq and uq pools are
> - * allowed to overlap. A VMID dedicated to a single fixed purpose on
> - * MMHUB (npa_vmid) is excluded from both MMHUB pools; it is the
> - * caller's responsibility to have already set npa_vmid.
> + * Sets both GFXHUB and MMHUB kernel-queue VMID pools
> + * via id_mgr[].ids_lru, and KFD/MES user-queue masks too.
> */
> void amdgpu_vmid_mgr_set_vmid_mask(struct amdgpu_device *adev,
> unsigned long vmid_kq_mask_gfxhub,
> unsigned long vmid_kq_mask_mmhub)
> {
> + unsigned long uq_mask_gfxhub, uq_mask_mmhub;
> unsigned int i;
>
> BUILD_BUG_ON(AMDGPU_NUM_VMID > BITS_PER_LONG);
>
> - if (adev->vm_manager.npa_vmid)
> + /*
> + * Derive the uq masks from the kq masks as given, then carve each
> + * hub's dedicated out-of-pool VMID out of both of that hub's pools.
> + */
> + uq_mask_gfxhub = GENMASK(AMDGPU_NUM_VMID - 1, 1) & ~vmid_kq_mask_gfxhub;
> + uq_mask_mmhub = 0xFF00;
That looks like it was derived from the original first_kfd_vmid or some fixed
MES value.
For now the patch is Reviewed-by: Christian König <[email protected]>,
but I think we need to take a look at that value again.
Regards,
Christian.
> +
> + /*
> + * SPM uses a dedicated vmid that is reserved from gfxhub.
> + * Unmask the reserved vmids from both user and kernel queues
> + */
> + if (adev->vm_manager.spm_vmid) {
> + vmid_kq_mask_gfxhub &= ~BIT(adev->vm_manager.spm_vmid);
> + uq_mask_gfxhub &= ~BIT(adev->vm_manager.spm_vmid);
> + }
> +
> + /*
> + * NPA uses a dedicated vmid that is reserved from mmhub.
> + * Unmask the reserved VMIDs from both user and kernel queues
> + */
> + if (adev->vm_manager.npa_vmid) {
> vmid_kq_mask_mmhub &= ~BIT(adev->vm_manager.npa_vmid);
> + uq_mask_mmhub &= ~BIT(adev->vm_manager.npa_vmid);
> + }
> +
> + adev->vm_manager.vmid_uq_mask_gfxhub = uq_mask_gfxhub;
> + adev->vm_manager.vmid_uq_mask_mmhub = uq_mask_mmhub;
>
> for (i = 0; i < AMDGPU_MAX_VMHUBS; i++) {
> bool is_mmhub = AMDGPU_IS_MMHUB0(i) || AMDGPU_IS_MMHUB1(i);
> @@ -666,13 +685,6 @@ void amdgpu_vmid_mgr_set_vmid_mask(struct amdgpu_device
> *adev,
> is_mmhub ? &vmid_kq_mask_mmhub :
> &vmid_kq_mask_gfxhub,
> AMDGPU_NUM_VMID);
> }
> -
> - adev->vm_manager.vmid_uq_mask_gfxhub =
> - GENMASK(AMDGPU_NUM_VMID - 1, 1) & ~vmid_kq_mask_gfxhub;
> -
> - adev->vm_manager.vmid_uq_mask_mmhub = 0xFF00;
> - if (adev->vm_manager.npa_vmid)
> - adev->vm_manager.vmid_uq_mask_mmhub &=
> ~BIT(adev->vm_manager.npa_vmid);
> }
>
> /**
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index d99e42ca308ef..ddcdc7256ef07 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -411,7 +411,9 @@ struct amdgpu_vm_manager {
>
> /* Global registration of recent page fault information */
> struct amdgpu_vm_fault_info fault_info;
> + /* VMIDs dedicated to a single fixed purpose, 0 if unused */
> unsigned int npa_vmid;
> + unsigned int spm_vmid;
>
> /* VMID masks for GFXHUB and MMHUB */
> uint32_t vmid_uq_mask_gfxhub;
