1 //! Check the validity invariant of a given value, and tell the user
2 //! where in the value it got violated.
3 //! In const context, this goes even further and tries to approximate const safety.
4 //! That's useful because it means other passes (e.g. promotion) can rely on `const`s
8 use std::ops::RangeInclusive;
11 use rustc::ty::layout::{self, LayoutOf, TyLayout, VariantIdx};
12 use rustc_data_structures::fx::FxHashSet;
14 use rustc_span::symbol::{sym, Symbol};
19 CheckInAllocMsg, GlobalAlloc, InterpCx, InterpResult, MPlaceTy, Machine, MemPlaceMeta, OpTy,
23 macro_rules! throw_validation_failure {
24 ($what:expr, $where:expr, $details:expr) => {{
25 let mut msg = format!("encountered {}", $what);
27 if !where_.is_empty() {
29 write_path(&mut msg, where_);
31 write!(&mut msg, ", but expected {}", $details).unwrap();
32 throw_unsup!(ValidationFailure(msg))
34 ($what:expr, $where:expr) => {{
35 let mut msg = format!("encountered {}", $what);
37 if !where_.is_empty() {
39 write_path(&mut msg, where_);
41 throw_unsup!(ValidationFailure(msg))
45 macro_rules! try_validation {
46 ($e:expr, $what:expr, $where:expr, $details:expr) => {{
49 Err(_) => throw_validation_failure!($what, $where, $details),
53 ($e:expr, $what:expr, $where:expr) => {{
56 Err(_) => throw_validation_failure!($what, $where),
61 /// We want to show a nice path to the invalid field for diagnostics,
62 /// but avoid string operations in the happy case where no error happens.
63 /// So we track a `Vec<PathElem>` where `PathElem` contains all the data we
64 /// need to later print something for the user.
65 #[derive(Copy, Clone, Debug)]
69 GeneratorState(VariantIdx),
78 /// State for tracking recursive validation of references
79 pub struct RefTracking<T, PATH = ()> {
80 pub seen: FxHashSet<T>,
81 pub todo: Vec<(T, PATH)>,
84 impl<T: Copy + Eq + Hash + std::fmt::Debug, PATH: Default> RefTracking<T, PATH> {
85 pub fn empty() -> Self {
86 RefTracking { seen: FxHashSet::default(), todo: vec![] }
88 pub fn new(op: T) -> Self {
89 let mut ref_tracking_for_consts =
90 RefTracking { seen: FxHashSet::default(), todo: vec![(op, PATH::default())] };
91 ref_tracking_for_consts.seen.insert(op);
92 ref_tracking_for_consts
95 pub fn track(&mut self, op: T, path: impl FnOnce() -> PATH) {
96 if self.seen.insert(op) {
97 trace!("Recursing below ptr {:#?}", op);
99 // Remember to come back to this later.
100 self.todo.push((op, path));
106 fn write_path(out: &mut String, path: &Vec<PathElem>) {
107 use self::PathElem::*;
109 for elem in path.iter() {
111 Field(name) => write!(out, ".{}", name),
112 Variant(name) => write!(out, ".<downcast-variant({})>", name),
113 GeneratorState(idx) => write!(out, ".<generator-state({})>", idx.index()),
114 ClosureVar(name) => write!(out, ".<closure-var({})>", name),
115 TupleElem(idx) => write!(out, ".{}", idx),
116 ArrayElem(idx) => write!(out, "[{}]", idx),
119 // This does not match Rust syntax, but it is more readable for long paths -- and
120 // some of the other items here also are not Rust syntax. Actually we can't
121 // even use the usual syntax because we are just showing the projections,
123 write!(out, ".<deref>")
125 Tag => write!(out, ".<enum-tag>"),
126 DynDowncast => write!(out, ".<dyn-downcast>"),
132 // Test if a range that wraps at overflow contains `test`
133 fn wrapping_range_contains(r: &RangeInclusive<u128>, test: u128) -> bool {
134 let (lo, hi) = r.clone().into_inner();
137 (..=hi).contains(&test) || (lo..).contains(&test)
144 // Formats such that a sentence like "expected something {}" to mean
145 // "expected something <in the given range>" makes sense.
146 fn wrapping_range_format(r: &RangeInclusive<u128>, max_hi: u128) -> String {
147 let (lo, hi) = r.clone().into_inner();
148 debug_assert!(hi <= max_hi);
150 format!("less or equal to {}, or greater or equal to {}", hi, lo)
152 format!("equal to {}", lo)
154 debug_assert!(hi < max_hi, "should not be printing if the range covers everything");
155 format!("less or equal to {}", hi)
156 } else if hi == max_hi {
157 debug_assert!(lo > 0, "should not be printing if the range covers everything");
158 format!("greater or equal to {}", lo)
160 format!("in the range {:?}", r)
164 struct ValidityVisitor<'rt, 'mir, 'tcx, M: Machine<'mir, 'tcx>> {
165 /// The `path` may be pushed to, but the part that is present when a function
166 /// starts must not be changed! `visit_fields` and `visit_array` rely on
167 /// this stack discipline.
169 ref_tracking_for_consts:
170 Option<&'rt mut RefTracking<MPlaceTy<'tcx, M::PointerTag>, Vec<PathElem>>>,
171 ecx: &'rt InterpCx<'mir, 'tcx, M>,
174 impl<'rt, 'mir, 'tcx, M: Machine<'mir, 'tcx>> ValidityVisitor<'rt, 'mir, 'tcx, M> {
175 fn aggregate_field_path_elem(&mut self, layout: TyLayout<'tcx>, field: usize) -> PathElem {
176 match layout.ty.kind {
177 // generators and closures.
178 ty::Closure(def_id, _) | ty::Generator(def_id, _, _) => {
180 if def_id.is_local() {
181 let tables = self.ecx.tcx.typeck_tables_of(def_id);
182 if let Some(upvars) = tables.upvar_list.get(&def_id) {
183 // Sometimes the index is beyond the number of upvars (seen
185 if let Some((&var_hir_id, _)) = upvars.get_index(field) {
186 let node = self.ecx.tcx.hir().get(var_hir_id);
187 if let hir::Node::Binding(pat) = node {
188 if let hir::PatKind::Binding(_, _, ident, _) = pat.kind {
189 name = Some(ident.name);
196 PathElem::ClosureVar(name.unwrap_or_else(|| {
197 // Fall back to showing the field index.
203 ty::Tuple(_) => PathElem::TupleElem(field),
206 ty::Adt(def, ..) if def.is_enum() => {
207 // we might be projecting *to* a variant, or to a field *in*a variant.
208 match layout.variants {
209 layout::Variants::Single { index } =>
212 PathElem::Field(def.variants[index].fields[field].ident.name)
219 ty::Adt(def, _) => PathElem::Field(def.non_enum_variant().fields[field].ident.name),
222 ty::Array(..) | ty::Slice(..) => PathElem::ArrayElem(field),
225 ty::Dynamic(..) => PathElem::DynDowncast,
227 // nothing else has an aggregate layout
228 _ => bug!("aggregate_field_path_elem: got non-aggregate type {:?}", layout.ty),
234 new_op: OpTy<'tcx, M::PointerTag>,
236 ) -> InterpResult<'tcx> {
237 // Remember the old state
238 let path_len = self.path.len();
240 self.path.push(elem);
241 self.visit_value(new_op)?;
243 self.path.truncate(path_len);
247 fn check_wide_ptr_meta(
249 meta: MemPlaceMeta<M::PointerTag>,
250 pointee: TyLayout<'tcx>,
251 ) -> InterpResult<'tcx> {
252 let tail = self.ecx.tcx.struct_tail_erasing_lifetimes(pointee.ty, self.ecx.param_env);
255 let vtable = meta.unwrap_meta();
257 self.ecx.memory.check_ptr_access(
259 3 * self.ecx.tcx.data_layout.pointer_size, // drop, size, align
260 self.ecx.tcx.data_layout.pointer_align.abi,
262 "dangling or unaligned vtable pointer in wide pointer or too small vtable",
266 self.ecx.read_drop_type_from_vtable(vtable),
267 "invalid drop fn in vtable",
271 self.ecx.read_size_and_align_from_vtable(vtable),
272 "invalid size or align in vtable",
275 // FIXME: More checks for the vtable.
277 ty::Slice(..) | ty::Str => {
278 let _len = try_validation!(
279 meta.unwrap_meta().to_machine_usize(self.ecx),
280 "non-integer slice length in wide pointer",
283 // We do not check that `len * elem_size <= isize::MAX`:
284 // that is only required for references, and there it falls out of the
285 // "dereferenceable" check performed by Stacked Borrows.
288 // Unsized, but not wide.
290 _ => bug!("Unexpected unsized type tail: {:?}", tail),
297 impl<'rt, 'mir, 'tcx, M: Machine<'mir, 'tcx>> ValueVisitor<'mir, 'tcx, M>
298 for ValidityVisitor<'rt, 'mir, 'tcx, M>
300 type V = OpTy<'tcx, M::PointerTag>;
303 fn ecx(&self) -> &InterpCx<'mir, 'tcx, M> {
310 old_op: OpTy<'tcx, M::PointerTag>,
312 new_op: OpTy<'tcx, M::PointerTag>,
313 ) -> InterpResult<'tcx> {
314 let elem = self.aggregate_field_path_elem(old_op.layout, field);
315 self.visit_elem(new_op, elem)
321 old_op: OpTy<'tcx, M::PointerTag>,
322 variant_id: VariantIdx,
323 new_op: OpTy<'tcx, M::PointerTag>,
324 ) -> InterpResult<'tcx> {
325 let name = match old_op.layout.ty.kind {
326 ty::Adt(adt, _) => PathElem::Variant(adt.variants[variant_id].ident.name),
327 // Generators also have variants
328 ty::Generator(..) => PathElem::GeneratorState(variant_id),
329 _ => bug!("Unexpected type with variant: {:?}", old_op.layout.ty),
331 self.visit_elem(new_op, name)
335 fn visit_value(&mut self, op: OpTy<'tcx, M::PointerTag>) -> InterpResult<'tcx> {
336 trace!("visit_value: {:?}, {:?}", *op, op.layout);
337 // Translate some possible errors to something nicer.
338 match self.walk_value(op) {
340 Err(err) => match err.kind {
341 err_ub!(InvalidDiscriminant(val)) => {
342 throw_validation_failure!(val, self.path, "a valid enum discriminant")
344 err_unsup!(ReadPointerAsBytes) => {
345 throw_validation_failure!("a pointer", self.path, "plain (non-pointer) bytes")
352 fn visit_primitive(&mut self, value: OpTy<'tcx, M::PointerTag>) -> InterpResult<'tcx> {
353 let value = self.ecx.read_immediate(value)?;
354 // Go over all the primitive types
355 let ty = value.layout.ty;
358 let value = value.to_scalar_or_undef();
359 try_validation!(value.to_bool(), value, self.path, "a boolean");
362 let value = value.to_scalar_or_undef();
363 try_validation!(value.to_char(), value, self.path, "a valid unicode codepoint");
365 ty::Float(_) | ty::Int(_) | ty::Uint(_) => {
366 // NOTE: Keep this in sync with the array optimization for int/float
368 let size = value.layout.size;
369 let value = value.to_scalar_or_undef();
370 if self.ref_tracking_for_consts.is_some() {
371 // Integers/floats in CTFE: Must be scalar bits, pointers are dangerous
376 "initialized plain (non-pointer) bytes"
379 // At run-time, for now, we accept *anything* for these types, including
380 // undef. We should fix that, but let's start low.
384 // We are conservative with undef for integers, but try to
385 // actually enforce our current rules for raw pointers.
387 try_validation!(self.ecx.ref_to_mplace(value), "undefined pointer", self.path);
388 if place.layout.is_unsized() {
389 self.check_wide_ptr_meta(place.meta, place.layout)?;
392 _ if ty.is_box() || ty.is_region_ptr() => {
393 // Handle wide pointers.
394 // Check metadata early, for better diagnostics
396 try_validation!(self.ecx.ref_to_mplace(value), "undefined pointer", self.path);
397 if place.layout.is_unsized() {
398 self.check_wide_ptr_meta(place.meta, place.layout)?;
400 // Make sure this is dereferenceable and all.
401 let (size, align) = self
403 .size_and_align_of(place.meta, place.layout)?
404 // for the purpose of validity, consider foreign types to have
405 // alignment and size determined by the layout (size will be 0,
406 // alignment should take attributes into account).
407 .unwrap_or_else(|| (place.layout.size, place.layout.align.abi));
408 let ptr: Option<_> = match self.ecx.memory.check_ptr_access_align(
412 CheckInAllocMsg::InboundsTest,
417 "{:?} did not pass access check for size {:?}, align {:?}",
418 place.ptr, size, align
421 err_unsup!(InvalidNullPointerUsage) => {
422 throw_validation_failure!("NULL reference", self.path)
424 err_unsup!(AlignmentCheckFailed { required, has }) => {
425 throw_validation_failure!(
427 "unaligned reference \
428 (required {} byte alignment but found {})",
435 err_unsup!(ReadBytesAsPointer) => throw_validation_failure!(
436 "dangling reference (created from integer)",
439 _ => throw_validation_failure!(
440 "dangling reference (not entirely in bounds)",
446 // Recursive checking
447 if let Some(ref mut ref_tracking) = self.ref_tracking_for_consts {
448 if let Some(ptr) = ptr {
450 // Skip validation entirely for some external statics
451 let alloc_kind = self.ecx.tcx.alloc_map.lock().get(ptr.alloc_id);
452 if let Some(GlobalAlloc::Static(did)) = alloc_kind {
453 // `extern static` cannot be validated as they have no body.
454 // FIXME: Statics from other crates are also skipped.
455 // They might be checked at a different type, but for now we
456 // want to avoid recursing too deeply. This is not sound!
457 if !did.is_local() || self.ecx.tcx.is_foreign_item(did) {
462 // Proceed recursively even for ZST, no reason to skip them!
463 // `!` is a ZST and we want to validate it.
464 // Normalize before handing `place` to tracking because that will
465 // check for duplicates.
466 let place = if size.bytes() > 0 {
467 self.ecx.force_mplace_ptr(place).expect("we already bounds-checked")
471 let path = &self.path;
472 ref_tracking.track(place, || {
473 // We need to clone the path anyway, make sure it gets created
474 // with enough space for the additional `Deref`.
475 let mut new_path = Vec::with_capacity(path.len() + 1);
476 new_path.clone_from(path);
477 new_path.push(PathElem::Deref);
483 let value = value.to_scalar_or_undef();
484 let _fn = try_validation!(
485 value.not_undef().and_then(|ptr| self.ecx.memory.get_fn(ptr)),
490 // FIXME: Check if the signature matches
492 // This should be all the primitive types
493 _ => bug!("Unexpected primitive type {}", value.layout.ty),
498 fn visit_uninhabited(&mut self) -> InterpResult<'tcx> {
499 throw_validation_failure!("a value of an uninhabited type", self.path)
504 op: OpTy<'tcx, M::PointerTag>,
505 layout: &layout::Scalar,
506 ) -> InterpResult<'tcx> {
507 let value = self.ecx.read_scalar(op)?;
508 // Determine the allowed range
509 let (lo, hi) = layout.valid_range.clone().into_inner();
510 // `max_hi` is as big as the size fits
511 let max_hi = u128::max_value() >> (128 - op.layout.size.bits());
512 assert!(hi <= max_hi);
513 // We could also write `(hi + 1) % (max_hi + 1) == lo` but `max_hi + 1` overflows for `u128`
514 if (lo == 0 && hi == max_hi) || (hi + 1 == lo) {
518 // At least one value is excluded. Get the bits.
519 let value = try_validation!(
523 format_args!("something {}", wrapping_range_format(&layout.valid_range, max_hi),)
525 let bits = match value.to_bits_or_ptr(op.layout.size, self.ecx) {
527 if lo == 1 && hi == max_hi {
528 // Only NULL is the niche. So make sure the ptr is NOT NULL.
529 if self.ecx.memory.ptr_may_be_null(ptr) {
530 throw_validation_failure!(
531 "a potentially NULL pointer",
534 "something that cannot possibly fail to be {}",
535 wrapping_range_format(&layout.valid_range, max_hi)
541 // Conservatively, we reject, because the pointer *could* have a bad
543 throw_validation_failure!(
547 "something that cannot possibly fail to be {}",
548 wrapping_range_format(&layout.valid_range, max_hi)
555 // Now compare. This is slightly subtle because this is a special "wrap-around" range.
556 if wrapping_range_contains(&layout.valid_range, bits) {
559 throw_validation_failure!(
562 format_args!("something {}", wrapping_range_format(&layout.valid_range, max_hi))
569 op: OpTy<'tcx, M::PointerTag>,
570 fields: impl Iterator<Item = InterpResult<'tcx, Self::V>>,
571 ) -> InterpResult<'tcx> {
572 match op.layout.ty.kind {
574 let mplace = op.assert_mem_place(self.ecx); // strings are never immediate
576 self.ecx.read_str(mplace),
577 "uninitialized or non-UTF-8 data in str",
581 ty::Array(tys, ..) | ty::Slice(tys)
583 // This optimization applies for types that can hold arbitrary bytes (such as
584 // integer and floating point types) or for structs or tuples with no fields.
585 // FIXME(wesleywiser) This logic could be extended further to arbitrary structs
586 // or tuples made up of integer/floating point types or inhabited ZSTs with no
589 ty::Int(..) | ty::Uint(..) | ty::Float(..) => true,
590 ty::Tuple(tys) if tys.len() == 0 => true,
592 if adt_def.is_struct() && adt_def.all_fields().next().is_none() =>
600 // Optimized handling for arrays of integer/float type.
602 // Arrays cannot be immediate, slices are never immediate.
603 let mplace = op.assert_mem_place(self.ecx);
604 // This is the length of the array/slice.
605 let len = mplace.len(self.ecx)?;
606 // Zero length slices have nothing to be checked.
610 // This is the element type size.
611 let layout = self.ecx.layout_of(tys)?;
612 // Empty tuples and fieldless structs (the only ZSTs that allow reaching this code)
613 // have no data to be checked.
617 // This is the size in bytes of the whole array.
618 let size = layout.size * len;
619 // Size is not 0, get a pointer.
620 let ptr = self.ecx.force_ptr(mplace.ptr)?;
622 // Optimization: we just check the entire range at once.
623 // NOTE: Keep this in sync with the handling of integer and float
624 // types above, in `visit_primitive`.
625 // In run-time mode, we accept pointers in here. This is actually more
626 // permissive than a per-element check would be, e.g., we accept
627 // an &[u8] that contains a pointer even though bytewise checking would
628 // reject it. However, that's good: We don't inherently want
629 // to reject those pointers, we just do not have the machinery to
630 // talk about parts of a pointer.
631 // We also accept undef, for consistency with the slow path.
632 match self.ecx.memory.get_raw(ptr.alloc_id)?.check_bytes(
636 /*allow_ptr_and_undef*/ self.ref_tracking_for_consts.is_none(),
638 // In the happy case, we needn't check anything else.
640 // Some error happened, try to provide a more detailed description.
642 // For some errors we might be able to provide extra information
644 err_unsup!(ReadUndefBytes(offset)) => {
645 // Some byte was undefined, determine which
646 // element that byte belongs to so we can
648 let i = (offset.bytes() / layout.size.bytes()) as usize;
649 self.path.push(PathElem::ArrayElem(i));
651 throw_validation_failure!("undefined bytes", self.path)
653 // Other errors shouldn't be possible
654 _ => return Err(err),
660 self.walk_aggregate(op, fields)? // default handler
667 impl<'mir, 'tcx, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
668 /// This function checks the data at `op`. `op` is assumed to cover valid memory if it
669 /// is an indirect operand.
670 /// It will error if the bits at the destination do not match the ones described by the layout.
672 /// `ref_tracking_for_consts` can be `None` to avoid recursive checking below references.
673 /// This also toggles between "run-time" (no recursion) and "compile-time" (with recursion)
674 /// validation (e.g., pointer values are fine in integers at runtime) and various other const
675 /// specific validation checks.
676 pub fn validate_operand(
678 op: OpTy<'tcx, M::PointerTag>,
680 ref_tracking_for_consts: Option<
681 &mut RefTracking<MPlaceTy<'tcx, M::PointerTag>, Vec<PathElem>>,
683 ) -> InterpResult<'tcx> {
684 trace!("validate_operand: {:?}, {:?}", *op, op.layout.ty);
686 // Construct a visitor
687 let mut visitor = ValidityVisitor { path, ref_tracking_for_consts, ecx: self };
689 // Try to cast to ptr *once* instead of all the time.
690 let op = self.force_op_ptr(op).unwrap_or(op);
693 visitor.visit_value(op)