From: Nitin Gote Date: Wed, 4 Mar 2026 12:38:02 +0000 (+0530) Subject: drm/xe: implement VM_BIND decompression in vm_bind_ioctl X-Git-Url: http://git.ipfire.org/gitweb.cgi?a=commitdiff_plain;h=2b484419700a0f563c695312374eb8cd5264b82c;p=thirdparty%2Fkernel%2Flinux.git drm/xe: implement VM_BIND decompression in vm_bind_ioctl Implement handling of VM_BIND(..., DECOMPRESS) in xe_vm_bind_ioctl. Key changes: - Parse and record per-op intent (op->map.request_decompress) when the DECOMPRESS flag is present. - Use xe_pat_index_get_comp_en() helper to check if a PAT index has compression enabled via the XE2_COMP_EN bit. - Validate DECOMPRESS preconditions in the ioctl path: - Only valid for MAP ops. - The provided pat_index must select the device's "no-compression" PAT. - Only meaningful on devices with flat CCS and the required XE2+ otherwise return -EOPNOTSUPP. - Use XE_IOCTL_DBG for uAPI sanity checks. - Implement xe_bo_decompress(): For VRAM BOs run xe_bo_move_notify(), reserve one fence slot, schedule xe_migrate_resolve(), and attach the returned fence with DMA_RESV_USAGE_KERNEL. Non-VRAM cases are silent no-ops. - Wire scheduling into vma_lock_and_validate() so VM_BIND will schedule decompression when request_decompress is set. - Handle fault-mode VMs by performing decompression synchronously during the bind process, ensuring that the resolve is completed before the bind finishes. This schedules an in-place GPU resolve (xe_migrate_resolve) for decompression. Compute PR: https://github.com/intel/compute-runtime/pull/898 IGT PR: https://patchwork.freedesktop.org/series/157553/ v7: Rebase on latest drm-tip and add compute and igt pr info v6: (Matt Auld) - Rebase as xe_pat_index_get_comp_en() is added in separate patch - Drop vm param from xe_bo_decompress(), instead of it extract tile from bo - Reject decompression on igpu instead of silent skipping to avoid any failure on Xe2+igpu as xe_device_has_flat_ccs() can sometimes be false on igpu due some setting in the BIOS to turn off compression on igpu. - Nits v5: (Matt) - Correct the condition check of xe_pat_index_get_comp_en v4: (Matt) - Introduce xe_pat_index_get_comp_en(), which checks XE2_COMP_EN for the pat_index - .interruptible should be true, everything else false v3: (Matt) - s/xe_bo_schedule_decompress/xe_bo_decompress - skip the decrompress step if the BO isn't in VRAM - start/size not required in xe_bo_schedule_decompress - Use xe_bo_move_notify instead of xe_vm_invalidate_vma with respect to invalidation. - Nits v2: - Move decompression work out of vm_bind ioctl. (Matt) - Put that work in a small helper at the BO/migrate layer invoke it from vma_lock_and_validate which already runs under drm_exec. - Move lightweight checks to vm_bind_ioctl_check_args (Matthew Auld) Cc: Matthew Brost Cc: Matthew Auld Reviewed-by: Matthew Auld Acked-by: Michal Mrozek Signed-off-by: Nitin Gote Signed-off-by: Matthew Auld Link: https://patch.msgid.link/20260304123758.3050386-8-nitin.r.gote@intel.com --- diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c index d6c2cb959cdd5..2bcdf75bc3d8f 100644 --- a/drivers/gpu/drm/xe/xe_bo.c +++ b/drivers/gpu/drm/xe/xe_bo.c @@ -3330,6 +3330,61 @@ int xe_gem_mmap_offset_ioctl(struct drm_device *dev, void *data, return 0; } +/** + * xe_bo_decompress - schedule in-place decompress and install fence + * @bo: buffer object (caller should hold drm_exec reservations for VM+BO) + * + * Schedules an in-place resolve via the migrate layer and installs the + * returned dma_fence into the BO kernel reservation slot (DMA_RESV_USAGE_KERNEL). + * In preempt fence mode, this operation interrupts hardware execution + * which is expensive. Page fault mode is recommended for better performance. + * + * The resolve path only runs for VRAM-backed buffers (currently dGPU-only); + * iGPU/system-memory objects fail the resource check and bypass the resolve. + * + * Returns 0 on success, negative errno on error. + */ +int xe_bo_decompress(struct xe_bo *bo) +{ + struct xe_device *xe = xe_bo_device(bo); + struct xe_tile *tile = xe_device_get_root_tile(xe); + struct dma_fence *decomp_fence = NULL; + struct ttm_operation_ctx op_ctx = { + .interruptible = true, + .no_wait_gpu = false, + .gfp_retry_mayfail = false, + }; + int err = 0; + + /* Silently skip decompression for non-VRAM buffers */ + if (!bo->ttm.resource || !mem_type_is_vram(bo->ttm.resource->mem_type)) + return 0; + + /* Notify before scheduling resolve */ + err = xe_bo_move_notify(bo, &op_ctx); + if (err) + return err; + + /* Reserve fence slot before scheduling */ + err = dma_resv_reserve_fences(bo->ttm.base.resv, 1); + if (err) + return err; + + /* Schedule the in-place decompression */ + decomp_fence = xe_migrate_resolve(tile->migrate, + bo, + bo->ttm.resource); + + if (IS_ERR(decomp_fence)) + return PTR_ERR(decomp_fence); + + /* Install kernel-usage fence */ + dma_resv_add_fence(bo->ttm.base.resv, decomp_fence, DMA_RESV_USAGE_KERNEL); + dma_fence_put(decomp_fence); + + return 0; +} + /** * xe_bo_lock() - Lock the buffer object's dma_resv object * @bo: The struct xe_bo whose lock is to be taken diff --git a/drivers/gpu/drm/xe/xe_bo.h b/drivers/gpu/drm/xe/xe_bo.h index c914ab719f20e..2cbac16f7db75 100644 --- a/drivers/gpu/drm/xe/xe_bo.h +++ b/drivers/gpu/drm/xe/xe_bo.h @@ -312,6 +312,8 @@ int xe_bo_dumb_create(struct drm_file *file_priv, bool xe_bo_needs_ccs_pages(struct xe_bo *bo); +int xe_bo_decompress(struct xe_bo *bo); + static inline size_t xe_bo_ccs_pages_start(struct xe_bo *bo) { return PAGE_ALIGN(xe_bo_size(bo)); diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c index dea25aae2928d..20515ec9fcaea 100644 --- a/drivers/gpu/drm/xe/xe_vm.c +++ b/drivers/gpu/drm/xe/xe_vm.c @@ -2362,6 +2362,7 @@ vm_bind_ioctl_ops_create(struct xe_vm *vm, struct xe_vma_ops *vops, op->map.vma_flags |= XE_VMA_DUMPABLE; if (flags & DRM_XE_VM_BIND_FLAG_MADVISE_AUTORESET) op->map.vma_flags |= XE_VMA_MADV_AUTORESET; + op->map.request_decompress = flags & DRM_XE_VM_BIND_FLAG_DECOMPRESS; op->map.pat_index = pat_index; op->map.invalidate_on_bind = __xe_vm_needs_clear_scratch_pages(vm, flags); @@ -2902,7 +2903,7 @@ static void vm_bind_ioctl_ops_unwind(struct xe_vm *vm, } static int vma_lock_and_validate(struct drm_exec *exec, struct xe_vma *vma, - bool res_evict, bool validate) + bool res_evict, bool validate, bool request_decompress) { struct xe_bo *bo = xe_vma_bo(vma); struct xe_vm *vm = xe_vma_vm(vma); @@ -2915,6 +2916,12 @@ static int vma_lock_and_validate(struct drm_exec *exec, struct xe_vma *vma, err = xe_bo_validate(bo, vm, xe_vm_allow_vm_eviction(vm) && res_evict, exec); + + if (err) + return err; + + if (request_decompress) + err = xe_bo_decompress(bo); } return err; @@ -3009,7 +3016,8 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, err = vma_lock_and_validate(exec, op->map.vma, res_evict, !xe_vm_in_fault_mode(vm) || - op->map.immediate); + op->map.immediate, + op->map.request_decompress); break; case DRM_GPUVA_OP_REMAP: err = check_ufence(gpuva_to_vma(op->base.remap.unmap->va)); @@ -3018,13 +3026,13 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, err = vma_lock_and_validate(exec, gpuva_to_vma(op->base.remap.unmap->va), - res_evict, false); + res_evict, false, false); if (!err && op->remap.prev) err = vma_lock_and_validate(exec, op->remap.prev, - res_evict, true); + res_evict, true, false); if (!err && op->remap.next) err = vma_lock_and_validate(exec, op->remap.next, - res_evict, true); + res_evict, true, false); break; case DRM_GPUVA_OP_UNMAP: err = check_ufence(gpuva_to_vma(op->base.unmap.va)); @@ -3033,7 +3041,7 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, err = vma_lock_and_validate(exec, gpuva_to_vma(op->base.unmap.va), - res_evict, false); + res_evict, false, false); break; case DRM_GPUVA_OP_PREFETCH: { @@ -3048,7 +3056,7 @@ static int op_lock_and_prep(struct drm_exec *exec, struct xe_vm *vm, err = vma_lock_and_validate(exec, gpuva_to_vma(op->base.prefetch.va), - res_evict, false); + res_evict, false, false); if (!err && !xe_vma_has_no_bo(vma)) err = xe_bo_migrate(xe_vma_bo(vma), region_to_mem_type[region], @@ -3370,7 +3378,8 @@ ALLOW_ERROR_INJECTION(vm_bind_ioctl_ops_execute, ERRNO); 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_MADVISE_AUTORESET) + DRM_XE_VM_BIND_FLAG_MADVISE_AUTORESET | \ + DRM_XE_VM_BIND_FLAG_DECOMPRESS) #ifdef TEST_VM_OPS_ERROR #define SUPPORTED_FLAGS (SUPPORTED_FLAGS_STUB | FORCE_OP_ERROR) @@ -3431,6 +3440,7 @@ 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; bool comp_en; @@ -3467,7 +3477,9 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm, XE_IOCTL_DBG(xe, obj_offset && (is_null || is_cpu_addr_mirror)) || XE_IOCTL_DBG(xe, 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, is_decompress && + xe_pat_index_get_comp_en(xe, pat_index)) || XE_IOCTL_DBG(xe, !obj && op == DRM_XE_VM_BIND_OP_MAP && !is_null && !is_cpu_addr_mirror) || @@ -3509,6 +3521,13 @@ static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm, err = -EINVAL; goto free_bind_ops; } + + if (is_decompress && (XE_IOCTL_DBG(xe, !xe_device_has_flat_ccs(xe)) || + XE_IOCTL_DBG(xe, GRAPHICS_VER(xe) < 20) || + XE_IOCTL_DBG(xe, !IS_DGFX(xe)))) { + err = -EOPNOTSUPP; + goto free_bind_ops; + } } return 0; diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h index de6544165cfad..69e80c94138a2 100644 --- a/drivers/gpu/drm/xe/xe_vm_types.h +++ b/drivers/gpu/drm/xe/xe_vm_types.h @@ -377,6 +377,8 @@ struct xe_vma_op_map { bool immediate; /** @read_only: Read only */ bool invalidate_on_bind; + /** @request_decompress: schedule decompression for GPU map */ + bool request_decompress; /** @pat_index: The pat index to use for this operation. */ u16 pat_index; };