Rollup merge of #124220 - RalfJung:interpret-wrong-vtable, r=oli-obk

Miri: detect wrong vtables in wide pointers

Fixes https://github.com/rust-lang/miri/issues/3497.
Needed to catch the UB that https://github.com/rust-lang/rust/pull/123572 will start exploiting.

r? `@oli-obk`

compiler/rustc_const_eval/src/errors.rs

@@ -498,6 +498,7 @@ impl<'a> ReportErrorExt for UndefinedBehaviorInfo<'a> {
             InvalidTag(_) => const_eval_invalid_tag,
             InvalidFunctionPointer(_) => const_eval_invalid_function_pointer,
             InvalidVTablePointer(_) => const_eval_invalid_vtable_pointer,
+            InvalidVTableTrait { .. } => const_eval_invalid_vtable_trait,
             InvalidStr(_) => const_eval_invalid_str,
             InvalidUninitBytes(None) => const_eval_invalid_uninit_bytes_unknown,
             InvalidUninitBytes(Some(_)) => const_eval_invalid_uninit_bytes,
@@ -537,6 +538,7 @@ impl<'a> ReportErrorExt for UndefinedBehaviorInfo<'a> {
             | DeadLocal
             | UninhabitedEnumVariantWritten(_)
             | UninhabitedEnumVariantRead(_) => {}
+
             BoundsCheckFailed { len, index } => {
                 diag.arg("len", len);
                 diag.arg("index", index);
@@ -544,6 +546,13 @@ impl<'a> ReportErrorExt for UndefinedBehaviorInfo<'a> {
             UnterminatedCString(ptr) | InvalidFunctionPointer(ptr) | InvalidVTablePointer(ptr) => {
                 diag.arg("pointer", ptr);
             }
+            InvalidVTableTrait { expected_trait, vtable_trait } => {
+                diag.arg("expected_trait", expected_trait.to_string());
+                diag.arg(
+                    "vtable_trait",
+                    vtable_trait.map(|t| t.to_string()).unwrap_or_else(|| format!("<trivial>")),
+                );
+            }
             PointerUseAfterFree(alloc_id, msg) => {
                 diag.arg("alloc_id", alloc_id)
                     .arg("bad_pointer_message", bad_pointer_message(msg, dcx));
@@ -634,6 +643,7 @@ impl<'tcx> ReportErrorExt for ValidationErrorInfo<'tcx> {
             UninhabitedEnumVariant => const_eval_validation_uninhabited_enum_variant,
             Uninit { .. } => const_eval_validation_uninit,
             InvalidVTablePtr { .. } => const_eval_validation_invalid_vtable_ptr,
+            InvalidMetaWrongTrait { .. } => const_eval_validation_invalid_vtable_trait,
             InvalidMetaSliceTooLarge { ptr_kind: PointerKind::Box } => {
                 const_eval_validation_invalid_box_slice_meta
             }
@@ -773,6 +783,13 @@ impl<'tcx> ReportErrorExt for ValidationErrorInfo<'tcx> {
             DanglingPtrNoProvenance { pointer, .. } => {
                 err.arg("pointer", pointer);
             }
+            InvalidMetaWrongTrait { expected_trait: ref_trait, vtable_trait } => {
+                err.arg("ref_trait", ref_trait.to_string());
+                err.arg(
+                    "vtable_trait",
+                    vtable_trait.map(|t| t.to_string()).unwrap_or_else(|| format!("<trivial>")),
+                );
+            }
             NullPtr { .. }
             | PtrToStatic { .. }
             | ConstRefToMutable

compiler/rustc_const_eval/src/interpret/cast.rs

@@ -393,6 +393,7 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
                 let val = self.read_immediate(src)?;
                 if data_a.principal() == data_b.principal() {
                     // A NOP cast that doesn't actually change anything, should be allowed even with mismatching vtables.
+                    // (But currently mismatching vtables violate the validity invariant so UB is triggered anyway.)
                     return self.write_immediate(*val, dest);
                 }
                 let (old_data, old_vptr) = val.to_scalar_pair();
@@ -400,7 +401,10 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
                 let old_vptr = old_vptr.to_pointer(self)?;
                 let (ty, old_trait) = self.get_ptr_vtable(old_vptr)?;
                 if old_trait != data_a.principal() {
-                    throw_ub_custom!(fluent::const_eval_upcast_mismatch);
+                    throw_ub!(InvalidVTableTrait {
+                        expected_trait: data_a,
+                        vtable_trait: old_trait,
+                    });
                 }
                 let new_vptr = self.get_vtable_ptr(ty, data_b.principal())?;
                 self.write_immediate(Immediate::new_dyn_trait(old_data, new_vptr, self), dest)

compiler/rustc_const_eval/src/interpret/place.rs

@@ -1020,16 +1020,20 @@ where
     pub(super) fn unpack_dyn_trait(
         &self,
         mplace: &MPlaceTy<'tcx, M::Provenance>,
+        expected_trait: &'tcx ty::List<ty::PolyExistentialPredicate<'tcx>>,
     ) -> InterpResult<'tcx, (MPlaceTy<'tcx, M::Provenance>, Pointer<Option<M::Provenance>>)> {
         assert!(
             matches!(mplace.layout.ty.kind(), ty::Dynamic(_, _, ty::Dyn)),
             "`unpack_dyn_trait` only makes sense on `dyn*` types"
         );
         let vtable = mplace.meta().unwrap_meta().to_pointer(self)?;
-        let (ty, _) = self.get_ptr_vtable(vtable)?;
-        let layout = self.layout_of(ty)?;
+        let (ty, vtable_trait) = self.get_ptr_vtable(vtable)?;
+        if expected_trait.principal() != vtable_trait {
+            throw_ub!(InvalidVTableTrait { expected_trait, vtable_trait });
+        }
         // This is a kind of transmute, from a place with unsized type and metadata to
         // a place with sized type and no metadata.
+        let layout = self.layout_of(ty)?;
         let mplace =
             MPlaceTy { mplace: MemPlace { meta: MemPlaceMeta::None, ..mplace.mplace }, layout };
         Ok((mplace, vtable))
@@ -1040,6 +1044,7 @@ where
     pub(super) fn unpack_dyn_star<P: Projectable<'tcx, M::Provenance>>(
         &self,
         val: &P,
+        expected_trait: &'tcx ty::List<ty::PolyExistentialPredicate<'tcx>>,
     ) -> InterpResult<'tcx, (P, Pointer<Option<M::Provenance>>)> {
         assert!(
             matches!(val.layout().ty.kind(), ty::Dynamic(_, _, ty::DynStar)),
@@ -1048,10 +1053,12 @@ where
         let data = self.project_field(val, 0)?;
         let vtable = self.project_field(val, 1)?;
         let vtable = self.read_pointer(&vtable.to_op(self)?)?;
-        let (ty, _) = self.get_ptr_vtable(vtable)?;
+        let (ty, vtable_trait) = self.get_ptr_vtable(vtable)?;
+        if expected_trait.principal() != vtable_trait {
+            throw_ub!(InvalidVTableTrait { expected_trait, vtable_trait });
+        }
+        // `data` is already the right thing but has the wrong type. So we transmute it.
         let layout = self.layout_of(ty)?;
-        // `data` is already the right thing but has the wrong type. So we transmute it, by
-        // projecting with offset 0.
         let data = data.transmute(layout, self)?;
         Ok((data, vtable))
     }

compiler/rustc_const_eval/src/interpret/terminator.rs

@@ -803,11 +803,8 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
                 let (vptr, dyn_ty, adjusted_receiver) = if let ty::Dynamic(data, _, ty::DynStar) =
                     receiver_place.layout.ty.kind()
                 {
-                    let (recv, vptr) = self.unpack_dyn_star(&receiver_place)?;
-                    let (dyn_ty, dyn_trait) = self.get_ptr_vtable(vptr)?;
-                    if dyn_trait != data.principal() {
-                        throw_ub_custom!(fluent::const_eval_dyn_star_call_vtable_mismatch);
-                    }
+                    let (recv, vptr) = self.unpack_dyn_star(&receiver_place, data)?;
+                    let (dyn_ty, _dyn_trait) = self.get_ptr_vtable(vptr)?;
 
                     (vptr, dyn_ty, recv.ptr())
                 } else {
@@ -829,7 +826,10 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
                     let vptr = receiver_place.meta().unwrap_meta().to_pointer(self)?;
                     let (dyn_ty, dyn_trait) = self.get_ptr_vtable(vptr)?;
                     if dyn_trait != data.principal() {
-                        throw_ub_custom!(fluent::const_eval_dyn_call_vtable_mismatch);
+                        throw_ub!(InvalidVTableTrait {
+                            expected_trait: data,
+                            vtable_trait: dyn_trait,
+                        });
                     }
 
                     // It might be surprising that we use a pointer as the receiver even if this
@@ -939,13 +939,13 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
         let place = self.force_allocation(place)?;
 
         let place = match place.layout.ty.kind() {
-            ty::Dynamic(_, _, ty::Dyn) => {
+            ty::Dynamic(data, _, ty::Dyn) => {
                 // Dropping a trait object. Need to find actual drop fn.
-                self.unpack_dyn_trait(&place)?.0
+                self.unpack_dyn_trait(&place, data)?.0
             }
-            ty::Dynamic(_, _, ty::DynStar) => {
+            ty::Dynamic(data, _, ty::DynStar) => {
                 // Dropping a `dyn*`. Need to find actual drop fn.
-                self.unpack_dyn_star(&place)?.0
+                self.unpack_dyn_star(&place, data)?.0
             }
             _ => {
                 debug_assert_eq!(

compiler/rustc_const_eval/src/interpret/validity.rs

@@ -339,16 +339,22 @@ impl<'rt, 'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> ValidityVisitor<'rt, 'mir, '
     ) -> InterpResult<'tcx> {
         let tail = self.ecx.tcx.struct_tail_erasing_lifetimes(pointee.ty, self.ecx.param_env);
         match tail.kind() {
-            ty::Dynamic(_, _, ty::Dyn) => {
+            ty::Dynamic(data, _, ty::Dyn) => {
                 let vtable = meta.unwrap_meta().to_pointer(self.ecx)?;
                 // Make sure it is a genuine vtable pointer.
-                let (_ty, _trait) = try_validation!(
+                let (_dyn_ty, dyn_trait) = try_validation!(
                     self.ecx.get_ptr_vtable(vtable),
                     self.path,
                     Ub(DanglingIntPointer(..) | InvalidVTablePointer(..)) =>
                         InvalidVTablePtr { value: format!("{vtable}") }
                 );
-                // FIXME: check if the type/trait match what ty::Dynamic says?
+                // Make sure it is for the right trait.
+                if dyn_trait != data.principal() {
+                    throw_validation_failure!(
+                        self.path,
+                        InvalidMetaWrongTrait { expected_trait: data, vtable_trait: dyn_trait }
+                    );
+                }
             }
             ty::Slice(..) | ty::Str => {
                 let _len = meta.unwrap_meta().to_target_usize(self.ecx)?;
@@ -933,7 +939,16 @@ impl<'rt, 'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> ValueVisitor<'mir, 'tcx, M>
                 }
             }
             _ => {
-                self.walk_value(op)?; // default handler
+                // default handler
+                try_validation!(
+                    self.walk_value(op),
+                    self.path,
+                    // It's not great to catch errors here, since we can't give a very good path,
+                    // but it's better than ICEing.
+                    Ub(InvalidVTableTrait { expected_trait, vtable_trait }) => {
+                        InvalidMetaWrongTrait { expected_trait, vtable_trait: *vtable_trait }
+                    },
+                );
             }
         }
 

compiler/rustc_const_eval/src/interpret/visitor.rs

@@ -88,22 +88,22 @@ pub trait ValueVisitor<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>>: Sized {
         // Special treatment for special types, where the (static) layout is not sufficient.
         match *ty.kind() {
             // If it is a trait object, switch to the real type that was used to create it.
-            ty::Dynamic(_, _, ty::Dyn) => {
+            ty::Dynamic(data, _, ty::Dyn) => {
                 // Dyn types. This is unsized, and the actual dynamic type of the data is given by the
                 // vtable stored in the place metadata.
                 // unsized values are never immediate, so we can assert_mem_place
                 let op = v.to_op(self.ecx())?;
                 let dest = op.assert_mem_place();
-                let inner_mplace = self.ecx().unpack_dyn_trait(&dest)?.0;
+                let inner_mplace = self.ecx().unpack_dyn_trait(&dest, data)?.0;
                 trace!("walk_value: dyn object layout: {:#?}", inner_mplace.layout);
                 // recurse with the inner type
                 return self.visit_field(v, 0, &inner_mplace.into());
             }
-            ty::Dynamic(_, _, ty::DynStar) => {
+            ty::Dynamic(data, _, ty::DynStar) => {
                 // DynStar types. Very different from a dyn type (but strangely part of the
                 // same variant in `TyKind`): These are pairs where the 2nd component is the
                 // vtable, and the first component is the data (which must be ptr-sized).
-                let data = self.ecx().unpack_dyn_star(v)?.0;
+                let data = self.ecx().unpack_dyn_star(v, data)?.0;
                 return self.visit_field(v, 0, &data);
             }
             // Slices do not need special handling here: they have `Array` field