Re: [PATCH v1 4/5] drm/xe: implement VM_BIND decompression in vm_bind_ioctl

[Matthew Auld <matthew.auld@intel.com>]

On 18/09/2025 15:25, Nitin Gote wrote:
> Implement handling of VM_BIND(..., DECOMPRESS) in xe_vm_bind_ioctl.
> 
> - Validate decompression preconditions (VRAM buffer, flat CCS support,
>    XE2+ hardware, and an uncompressed PAT index).
> - Mark the BO with DRM_XE_VM_BIND_FLAG_DECOMPRESS.
> - Invalidate any overlapping VMA when required.
> - Use helper xe_vm_exec_lock_vm_and_bo(..., wait_bookkeep=true)
>    to atomically acquire the VM reservation and optional BO reservation and
>    handle BOOKKEEP wait/retry.
>    * Drop vm->lock before calling the helper.
>    * Do VMA invalidation or schedule the migrate resolve while reservations
>      are held.
>    * Call drm_exec_fini() to release reservations, then wait on the
>      migrate/decompression fence without holding reservations and
>      re-acquire vm->lock to continue.
> - Defer decompression when the VM is in fault mode.
> 
> This schedules an in-place GPU resolve (xe_migrate_resolve) for
> decompression and waits for completion before proceeding with the bind.
> The change centralises drm_exec + BOOKKEEP retry logic and avoids
> sleeping while holding resv locks.
> 
> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
> Signed-off-by: Nitin Gote <nitin.r.gote@intel.com>
> ---
>   drivers/gpu/drm/xe/xe_vm.c | 134 +++++++++++++++++++++++++++++++++++--
>   1 file changed, 130 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index fae88c4c981e..cdeb7995eab1 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -3266,7 +3266,8 @@ ALLOW_ERROR_INJECTION(vm_bind_ioctl_ops_execute, ERRNO);
>   	 DRM_XE_VM_BIND_FLAG_NULL | \
>   	 DRM_XE_VM_BIND_FLAG_DUMPABLE | \
>   	 DRM_XE_VM_BIND_FLAG_CHECK_PXP | \
> -	 DRM_XE_VM_BIND_FLAG_CPU_ADDR_MIRROR)
> +	 DRM_XE_VM_BIND_FLAG_CPU_ADDR_MIRROR | \
> +	 DRM_XE_VM_BIND_FLAG_DECOMPRESS)
>   
>   #ifdef TEST_VM_OPS_ERROR
>   #define SUPPORTED_FLAGS	(SUPPORTED_FLAGS_STUB | FORCE_OP_ERROR)
> @@ -3324,6 +3325,7 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
>   		bool is_null = flags & DRM_XE_VM_BIND_FLAG_NULL;
>   		bool is_cpu_addr_mirror = flags &
>   			DRM_XE_VM_BIND_FLAG_CPU_ADDR_MIRROR;
> +		bool is_decompress = flags & DRM_XE_VM_BIND_FLAG_DECOMPRESS;
>   		u16 pat_index = (*bind_ops)[i].pat_index;
>   		u16 coh_mode;
>   
> @@ -3361,7 +3363,7 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
>   				 (is_null || is_cpu_addr_mirror)) ||
>   		    XE_IOCTL_DBG(xe, !obj &&
>   				 op == DRM_XE_VM_BIND_OP_MAP &&
> -				 !is_null && !is_cpu_addr_mirror) ||
> +				 is_decompress && !is_null && !is_cpu_addr_mirror) ||
>   		    XE_IOCTL_DBG(xe, !obj &&
>   				 op == DRM_XE_VM_BIND_OP_UNMAP_ALL) ||
>   		    XE_IOCTL_DBG(xe, addr &&
> @@ -3378,8 +3380,8 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
>   				 op == DRM_XE_VM_BIND_OP_PREFETCH) ||
>   		    XE_IOCTL_DBG(xe, prefetch_region &&
>   				 op != DRM_XE_VM_BIND_OP_PREFETCH) ||
> -		    XE_IOCTL_DBG(xe,  (prefetch_region != DRM_XE_CONSULT_MEM_ADVISE_PREF_LOC &&
> -				       !(BIT(prefetch_region) & xe->info.mem_region_mask))) ||
> +		    XE_IOCTL_DBG(xe, (prefetch_region != DRM_XE_CONSULT_MEM_ADVISE_PREF_LOC &&
> +				      !(BIT(prefetch_region) & xe->info.mem_region_mask))) ||
>   		    XE_IOCTL_DBG(xe, obj &&
>   				 op == DRM_XE_VM_BIND_OP_UNMAP)) {
>   			err = -EINVAL;
> @@ -3698,6 +3700,130 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   		u64 obj_offset = bind_ops[i].obj_offset;
>   		u32 prefetch_region = bind_ops[i].prefetch_mem_region_instance;
>   		u16 pat_index = bind_ops[i].pat_index;
> +		bool do_decompress = flags & DRM_XE_VM_BIND_FLAG_DECOMPRESS;
> +
> +		/* Handle decompression process */
> +		if (do_decompress && bos[i]) {
> +			struct xe_vma *existing_vma;
> +			struct dma_fence *decomp_fence;
> +			struct xe_tile *tile = xe_device_get_root_tile(xe);
> +
> +			bos[i]->flags |= DRM_XE_VM_BIND_FLAG_DECOMPRESS;
> +
> +			/* Validate VRAM buffer suitable for decompression or not */
> +			if (!mem_type_is_vram(bos[i]->ttm.resource->mem_type)) {

I think this can crash with NPD or other nasties since ttm.resource is 
dynamic state and we need to hold dma-resv lock if touching it (only 
held below AFAICT). Also memory can get evicted or can be deferred until 
first use, so this would return an error for those cases?

> +				drm_err(&vm->xe->drm,
> +					"Decompression requires VRAM buffer\n");
> +				err = -EINVAL;
> +				goto unwind_ops;
> +			}
> +
> +			/* Check hardware support for decompression */
> +			if (!xe_device_has_flat_ccs(vm->xe)) {
> +				drm_err(&vm->xe->drm,
> +					"Decompression requires flat CCS support\n");
> +				err = -EOPNOTSUPP;
> +				goto unwind_ops;
> +			}
> +
> +			if (GRAPHICS_VER(vm->xe) < 20) {
> +				drm_err(&vm->xe->drm,
> +					"Decompression requires XE2+ hardware\n");
> +				err = -EOPNOTSUPP;
> +				goto unwind_ops;
> +			}
> +
> +			/* Validate that user provided an uncompressed PAT index */
> +			if (pat_index == vm->xe->pat.idx[XE_CACHE_NONE_COMPRESSION]) {

XE_CACHE_NONE_COMPRESSION is one of many possible compression enabled 
indexes. Why do we only check this one? Also for all of these uAPI 
sanity checks I think it's best practice to use XE_IOCTL_DBG.

> +				drm_err(&vm->xe->drm,
> +					"Decompression requires uncompressed PAT index\n");
> +				err = -EINVAL;
> +				goto unwind_ops;
> +			}
> +
> +			/* Invalidate VMA so GPU mappings are destroyed */
> +			existing_vma = xe_vm_find_overlapping_vma(vm, addr, range);
> +			if (existing_vma) {
> +				struct xe_bo *existing_bo = xe_vma_bo(existing_vma);
> +
> +				drm_dbg(&vm->xe->drm,
> +					"Found overlapping VMA - automatic invalidation will occur\n");
> +
> +				/* Drop vm->lock and atomically take VM+BO resvs via drm_exec */
> +				up_write(&vm->lock);
> +				struct drm_exec exec;
> +
> +				err = xe_vm_exec_lock_vm_and_bo(vm, existing_bo, &exec, true, true);
> +				if (err) {
> +					down_write(&vm->lock);
> +					goto put_obj;
> +				}
> +
> +				/* Invalidate the VMA while reservations are held. */
> +				err = xe_vm_invalidate_vma(existing_vma);
> +
> +				/* release reservations (drm_exec_fini releases VM+BO resvs) */
> +				drm_exec_fini(&exec);
> +
> +				/* Re-acquire vm->lock */
> +				down_write(&vm->lock);
> +
> +				if (err)
> +					goto put_obj;
> +			}
> +
> +			/* Handle decompression differently for fault mode */
> +			if (xe_vm_in_fault_mode(vm)) {
> +				drm_dbg(&vm->xe->drm,
> +					"Deferring decompression for fault mode\n");
> +				/* Skip immediate decompression */
> +				goto unwind_ops;
> +			}
> +
> +			/* Drop vm->lock and atomically acquire VM+BO reservations via drm_exec.
> +			 * Hold the reservations while we wait for BOOKKEEP and schedule the
> +			 * in-place resolve. Release the reservations (drm_exec_fini) before
> +			 * waiting on the migrate fence so we don't hold resv locks while
> +			 * sleeping.
> +			 */
> +			up_write(&vm->lock);
> +			struct drm_exec exec;
> +
> +			err = xe_vm_exec_lock_vm_and_bo(vm, bos[i], &exec, true, true);
> +			if (err) {
> +				drm_err(&vm->xe->drm,
> +					"Failed to acquire reservations for decompression: %d\n",
> +					err);
> +				down_write(&vm->lock);
> +				goto unwind_ops;
> +			}
> +
> +			/* With exec holding reservations, schedule the in-place decompression */
> +			decomp_fence = xe_migrate_resolve(tile->migrate,
> +							  bos[i], bos[i],
> +							  bos[i]->ttm.resource,
> +							  bos[i]->ttm.resource,
> +							  false);
> +
> +			/* release the reservations so scheduler can observe BO fences */
> +			drm_exec_fini(&exec);
> +
> +			if (IS_ERR(decomp_fence)) {
> +				err = PTR_ERR(decomp_fence);
> +				drm_err(&vm->xe->drm,
> +					"In-place decompression failed: %d\n", err);
> +				/* vm->lock must be held for unwind_ops cleanup */
> +				down_write(&vm->lock);
> +				goto unwind_ops;
> +			}
> +
> +			/* Wait for in-place decompression to complete without holding vm->lock */
> +			dma_fence_wait(decomp_fence, false);
> +			dma_fence_put(decomp_fence);
> +
> +			/* Now re-acquire vm->lock and continue */
> +			down_write(&vm->lock);
> +		}
>   
>   		ops[i] = vm_bind_ioctl_ops_create(vm, &vops, bos[i], obj_offset,
>   						  addr, range, op, flags,