[PATCH 09/11] drm/xe: Allow prefetch-only VM bind IOCTLs to use VM read lock
Matthew Brost
matthew.brost at intel.com
Wed Aug 6 06:22:40 UTC 2025
Prefetch-only VM bind IOCTLs do not modify VMAs after pinning userptr
pages. Downgrade vm->lock to read mode once pinning is complete.
Lays the groundwork for prefetch IOCTLs to use threaded migration.
Signed-off-by: Matthew Brost <matthew.brost at intel.com>
---
drivers/gpu/drm/xe/xe_vm.c | 41 ++++++++++++++++++++++++++++----
drivers/gpu/drm/xe/xe_vm_types.h | 4 +++-
2 files changed, 39 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
index 2498cff58fe7..3211827ef6d7 100644
--- a/drivers/gpu/drm/xe/xe_vm.c
+++ b/drivers/gpu/drm/xe/xe_vm.c
@@ -1763,6 +1763,12 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags)
err = xe_svm_init(vm);
if (err)
goto err_no_resv;
+ } else {
+ /*
+ * Avoid lockdep explosions in
+ * xe_vm_assert_write_mode_or_garbage_collector
+ */
+ mutex_init(&vm->svm.garbage_collector.lock);
}
vm_resv_obj = drm_gpuvm_resv_object_alloc(&xe->drm);
@@ -1996,6 +2002,8 @@ void xe_vm_close_and_put(struct xe_vm *vm)
if (xe_vm_in_fault_mode(vm))
xe_svm_fini(vm);
+ else
+ mutex_destroy(&vm->svm.garbage_collector.lock);
up_write(&vm->lock);
@@ -2365,10 +2373,12 @@ vm_bind_ioctl_ops_create(struct xe_vm *vm, struct xe_vma_ops *vops,
switch (operation) {
case DRM_XE_VM_BIND_OP_MAP:
case DRM_XE_VM_BIND_OP_MAP_USERPTR:
+ vops->flags |= XE_VMA_OPS_FLAG_MODIFIES_GPUVA;
ops = drm_gpuvm_sm_map_ops_create(&vm->gpuvm, addr, range,
obj, bo_offset_or_userptr);
break;
case DRM_XE_VM_BIND_OP_UNMAP:
+ vops->flags |= XE_VMA_OPS_FLAG_MODIFIES_GPUVA;
ops = drm_gpuvm_sm_unmap_ops_create(&vm->gpuvm, addr, range);
break;
case DRM_XE_VM_BIND_OP_PREFETCH:
@@ -2377,6 +2387,7 @@ vm_bind_ioctl_ops_create(struct xe_vm *vm, struct xe_vma_ops *vops,
case DRM_XE_VM_BIND_OP_UNMAP_ALL:
xe_assert(vm->xe, bo);
+ vops->flags |= XE_VMA_OPS_FLAG_MODIFIES_GPUVA;
err = xe_bo_lock(bo, true);
if (err)
return ERR_PTR(err);
@@ -2584,10 +2595,12 @@ static int xe_vma_op_commit(struct xe_vm *vm, struct xe_vma_op *op)
{
int err = 0;
- xe_vm_assert_write_mode_or_garbage_collector(vm);
+ lockdep_assert_held(&vm->lock);
switch (op->base.op) {
case DRM_GPUVA_OP_MAP:
+ xe_vm_assert_write_mode_or_garbage_collector(vm);
+
err |= xe_vm_insert_vma(vm, op->map.vma);
if (!err)
op->flags |= XE_VMA_OP_COMMITTED;
@@ -2597,6 +2610,8 @@ static int xe_vma_op_commit(struct xe_vm *vm, struct xe_vma_op *op)
u8 tile_present =
gpuva_to_vma(op->base.remap.unmap->va)->tile_present;
+ xe_vm_assert_write_mode_or_garbage_collector(vm);
+
prep_vma_destroy(vm, gpuva_to_vma(op->base.remap.unmap->va),
true);
op->flags |= XE_VMA_OP_COMMITTED;
@@ -2630,6 +2645,8 @@ static int xe_vma_op_commit(struct xe_vm *vm, struct xe_vma_op *op)
break;
}
case DRM_GPUVA_OP_UNMAP:
+ xe_vm_assert_write_mode_or_garbage_collector(vm);
+
prep_vma_destroy(vm, gpuva_to_vma(op->base.unmap.va), true);
op->flags |= XE_VMA_OP_COMMITTED;
break;
@@ -2831,10 +2848,12 @@ static void xe_vma_op_unwind(struct xe_vm *vm, struct xe_vma_op *op,
bool post_commit, bool prev_post_commit,
bool next_post_commit)
{
- xe_vm_assert_write_mode_or_garbage_collector(vm);
+ lockdep_assert_held(&vm->lock);
switch (op->base.op) {
case DRM_GPUVA_OP_MAP:
+ xe_vm_assert_write_mode_or_garbage_collector(vm);
+
if (op->map.vma) {
prep_vma_destroy(vm, op->map.vma, post_commit);
xe_vma_destroy_unlocked(op->map.vma);
@@ -2844,6 +2863,8 @@ static void xe_vma_op_unwind(struct xe_vm *vm, struct xe_vma_op *op,
{
struct xe_vma *vma = gpuva_to_vma(op->base.unmap.va);
+ xe_vm_assert_write_mode_or_garbage_collector(vm);
+
if (vma) {
down_read(&vm->userptr.notifier_lock);
vma->gpuva.flags &= ~XE_VMA_DESTROYED;
@@ -2857,6 +2878,8 @@ static void xe_vma_op_unwind(struct xe_vm *vm, struct xe_vma_op *op,
{
struct xe_vma *vma = gpuva_to_vma(op->base.remap.unmap->va);
+ xe_vm_assert_write_mode_or_garbage_collector(vm);
+
if (op->remap.prev) {
prep_vma_destroy(vm, op->remap.prev, prev_post_commit);
xe_vma_destroy_unlocked(op->remap.prev);
@@ -3313,7 +3336,7 @@ static struct dma_fence *vm_bind_ioctl_ops_execute(struct xe_vm *vm,
struct dma_fence *fence;
int err;
- lockdep_assert_held_write(&vm->lock);
+ lockdep_assert_held(&vm->lock);
drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
DRM_EXEC_IGNORE_DUPLICATES, 0);
@@ -3587,7 +3610,7 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
u32 num_syncs, num_ufence = 0;
struct xe_sync_entry *syncs = NULL;
struct drm_xe_vm_bind_op *bind_ops;
- struct xe_vma_ops vops;
+ struct xe_vma_ops vops = { .flags = 0 };
struct dma_fence *fence;
int err;
int i;
@@ -3753,6 +3776,11 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
goto unwind_ops;
}
+ if (!(vops.flags & XE_VMA_OPS_FLAG_MODIFIES_GPUVA)) {
+ vops.flags |= XE_VMA_OPS_FLAG_DOWNGRADE_LOCK;
+ downgrade_write(&vm->lock);
+ }
+
err = xe_vma_ops_alloc(&vops, args->num_binds > 1);
if (err)
goto unwind_ops;
@@ -3785,7 +3813,10 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
for (i = 0; i < args->num_binds; ++i)
xe_bo_put(bos[i]);
release_vm_lock:
- up_write(&vm->lock);
+ if (vops.flags & XE_VMA_OPS_FLAG_DOWNGRADE_LOCK)
+ up_read(&vm->lock);
+ else
+ up_write(&vm->lock);
put_exec_queue:
if (q)
xe_exec_queue_put(q);
diff --git a/drivers/gpu/drm/xe/xe_vm_types.h b/drivers/gpu/drm/xe/xe_vm_types.h
index 1aabdedbfa92..332822e6ee7f 100644
--- a/drivers/gpu/drm/xe/xe_vm_types.h
+++ b/drivers/gpu/drm/xe/xe_vm_types.h
@@ -481,7 +481,9 @@ struct xe_vma_ops {
/** @pt_update_ops: page table update operations */
struct xe_vm_pgtable_update_ops pt_update_ops[XE_MAX_TILES_PER_DEVICE];
/** @flag: signify the properties within xe_vma_ops*/
-#define XE_VMA_OPS_FLAG_HAS_SVM_PREFETCH BIT(0)
+#define XE_VMA_OPS_FLAG_HAS_SVM_PREFETCH BIT(0)
+#define XE_VMA_OPS_FLAG_MODIFIES_GPUVA BIT(1)
+#define XE_VMA_OPS_FLAG_DOWNGRADE_LOCK BIT(2)
u32 flags;
#ifdef TEST_VM_OPS_ERROR
/** @inject_error: inject error to test error handling */
--
2.34.1
More information about the Intel-xe
mailing list