[Intel-xe] [PATCH v2 20/31] drm/xe: Optimize size of xe_vma allocation
Thomas Hellström
thomas.hellstrom at linux.intel.com
Thu May 11 09:05:42 UTC 2023
On 5/2/23 02:17, Matthew Brost wrote:
> Reduce gt_mask to a u8 from a u64, only allocate userptr state if VMA is
> a userptr, and union of destroy callback and worker.
>
> Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> ---
> drivers/gpu/drm/xe/xe_vm.c | 14 +++--
> drivers/gpu/drm/xe/xe_vm_types.h | 88 +++++++++++++++++---------------
> 2 files changed, 57 insertions(+), 45 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index e5f2fffb2aec..e8d9939ee535 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -814,7 +814,7 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm,
> u64 bo_offset_or_userptr,
> u64 start, u64 end,
> bool read_only, bool null,
> - u64 gt_mask)
> + u8 gt_mask)
> {
> struct xe_vma *vma;
> struct xe_gt *gt;
> @@ -823,7 +823,11 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm,
> XE_BUG_ON(start >= end);
> XE_BUG_ON(end >= vm->size);
>
> - vma = kzalloc(sizeof(*vma), GFP_KERNEL);
> + if (!bo && !null) /* userptr */
> + vma = kzalloc(sizeof(*vma), GFP_KERNEL);
> + else
> + vma = kzalloc(sizeof(*vma) - sizeof(struct xe_userptr),
> + GFP_KERNEL);
> if (!vma) {
> vma = ERR_PTR(-ENOMEM);
> return vma;
> @@ -2149,7 +2153,7 @@ static void print_op(struct xe_device *xe, struct drm_gpuva_op *op)
> static struct drm_gpuva_ops *
> vm_bind_ioctl_ops_create(struct xe_vm *vm, struct xe_bo *bo,
> u64 bo_offset_or_userptr, u64 addr, u64 range,
> - u32 operation, u64 gt_mask, u32 region)
> + u32 operation, u8 gt_mask, u32 region)
> {
> struct drm_gem_object *obj = bo ? &bo->ttm.base : NULL;
> struct ww_acquire_ctx ww;
> @@ -2234,7 +2238,7 @@ vm_bind_ioctl_ops_create(struct xe_vm *vm, struct xe_bo *bo,
> }
>
> static struct xe_vma *new_vma(struct xe_vm *vm, struct drm_gpuva_op_map *op,
> - u64 gt_mask, bool read_only, bool null)
> + u8 gt_mask, bool read_only, bool null)
> {
> struct xe_bo *bo = op->gem.obj ? gem_to_xe_bo(op->gem.obj) : NULL;
> struct xe_vma *vma;
> @@ -3217,8 +3221,8 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> u64 addr = bind_ops[i].addr;
> u32 op = bind_ops[i].op;
> u64 obj_offset = bind_ops[i].obj_offset;
> - u64 gt_mask = bind_ops[i].gt_mask;
> u32 region = bind_ops[i].region;
> + u8 gt_mask = bind_ops[i].gt_mask;
>
> ops[i] = vm_bind_ioctl_ops_create(vm, bos[i], obj_offset,
> addr, range, op, gt_mask,
> diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
> index 22def5483c12..df4797ec4d7f 100644
> --- a/drivers/gpu/drm/xe/xe_vm_types.h
> +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> @@ -34,22 +34,34 @@ struct xe_vm;
> #define XE_VMA_PTE_2M (DRM_GPUVA_USERBITS << 7)
> #define XE_VMA_PTE_1G (DRM_GPUVA_USERBITS << 8)
>
> +/** struct xe_userptr - User pointer */
> +struct xe_userptr {
> + /**
> + * @notifier: MMU notifier for user pointer (invalidation call back)
> + */
> + struct mmu_interval_notifier notifier;
> + /** @sgt: storage for a scatter gather table */
> + struct sg_table sgt;
> + /** @sg: allocated scatter gather table */
> + struct sg_table *sg;
> + /** @notifier_seq: notifier sequence number */
> + unsigned long notifier_seq;
> + /**
> + * @initial_bind: user pointer has been bound at least once.
> + * write: vm->userptr.notifier_lock in read mode and vm->resv held.
> + * read: vm->userptr.notifier_lock in write mode or vm->resv held.
> + */
> + bool initial_bind;
> +#if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
> + u32 divisor;
> +#endif
> +};
> +
> +/** xe_vma - Virtual memory address */
> struct xe_vma {
> /** @gpuva: Base GPUVA object */
> struct drm_gpuva gpuva;
>
> - /** @gt_mask: GT mask of where to create binding for this VMA */
> - u64 gt_mask;
> -
> - /**
> - * @gt_present: GT mask of binding are present for this VMA.
> - * protected by vm->lock, vm->resv and for userptrs,
> - * vm->userptr.notifier_lock for writing. Needs either for reading,
> - * but if reading is done under the vm->lock only, it needs to be held
> - * in write mode.
> - */
> - u64 gt_present;
> -
> union {
> /** @userptr_link: link into VM repin list if userptr */
> struct list_head userptr_link;
> @@ -77,16 +89,29 @@ struct xe_vma {
> } notifier;
> };
>
> - /** @destroy_cb: callback to destroy VMA when unbind job is done */
> - struct dma_fence_cb destroy_cb;
> + union {
> + /** @destroy_cb: callback to destroy VMA when unbind job is done */
> + struct dma_fence_cb destroy_cb;
> + /** @destroy_work: worker to destroy this BO */
> + struct work_struct destroy_work;
> + };
>
> - /** @destroy_work: worker to destroy this BO */
> - struct work_struct destroy_work;
> + /** @gt_mask: GT mask of where to create binding for this VMA */
> + u8 gt_mask;
> +
> + /**
> + * @gt_present: GT mask of binding are present for this VMA.
> + * protected by vm->lock, vm->resv and for userptrs,
> + * vm->userptr.notifier_lock for writing. Needs either for reading,
> + * but if reading is done under the vm->lock only, it needs to be held
> + * in write mode.
> + */
> + u8 gt_present;
>
> /** @usm: unified shared memory state */
> struct {
> /** @gt_invalidated: VMA has been invalidated */
> - u64 gt_invalidated;
> + u8 gt_invalidated;
> } usm;
>
> struct {
> @@ -97,28 +122,11 @@ struct xe_vma {
> struct list_head link;
> } extobj;
>
> - /** @userptr: user pointer state */
> - struct {
> - /**
> - * @notifier: MMU notifier for user pointer (invalidation call back)
> - */
> - struct mmu_interval_notifier notifier;
> - /** @sgt: storage for a scatter gather table */
> - struct sg_table sgt;
> - /** @sg: allocated scatter gather table */
> - struct sg_table *sg;
> - /** @notifier_seq: notifier sequence number */
> - unsigned long notifier_seq;
> - /**
> - * @initial_bind: user pointer has been bound at least once.
> - * write: vm->userptr.notifier_lock in read mode and vm->resv held.
> - * read: vm->userptr.notifier_lock in write mode or vm->resv held.
> - */
> - bool initial_bind;
> -#if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
> - u32 divisor;
> -#endif
> - } userptr;
> + /**
> + * @userptr: user pointer state, only allocated for VMAs that are
> + * user pointers
> + */
> + struct xe_userptr userptr;
I think this is very fragile, What happens when someone doesn't read the
code and simply adds a member after @userptr, or generic code accidently
dereferences a field in @userptr?
Wouldn't the proper way to do this, if at all, to subclass xe_vma into
an xe_vma_usertptr to guard against such things happening?
For the u8 space optimizations, also a pahole layout before and after
the change would be beneficial in the commit message.
/Thomas
> };
>
> struct xe_device;
> @@ -387,7 +395,7 @@ struct xe_vma_op {
> */
> struct async_op_fence *fence;
> /** @gt_mask: gt mask for this operation */
> - u64 gt_mask;
> + u8 gt_mask;
> /** @flags: operation flags */
> enum xe_vma_op_flags flags;
>
More information about the Intel-xe
mailing list