1 //! Implementation of a data-race detector using Lamport Timestamps / Vector-clocks
2 //! based on the Dynamic Race Detection for C++:
3 //! https://www.doc.ic.ac.uk/~afd/homepages/papers/pdfs/2017/POPL.pdf
4 //! which does not report false-positives when fences are used, and gives better
5 //! accuracy in presence of read-modify-write operations.
7 //! The implementation contains modifications to correctly model the changes to the memory model in C++20
8 //! regarding the weakening of release sequences: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0982r1.html.
9 //! Relaxed stores now unconditionally block all currently active release sequences and so per-thread tracking of release
10 //! sequences is not needed.
12 //! The implementation also models races with memory allocation and deallocation via treating allocation and
13 //! deallocation as a type of write internally for detecting data-races.
15 //! This does not explore weak memory orders and so can still miss data-races
16 //! but should not report false-positives
18 //! Data-race definition from(https://en.cppreference.com/w/cpp/language/memory_model#Threads_and_data_races):
19 //! a data race occurs between two memory accesses if they are on different threads, at least one operation
20 //! is non-atomic, at least one operation is a write and neither access happens-before the other. Read the link
21 //! for full definition.
23 //! This re-uses vector indexes for threads that are known to be unable to report data-races, this is valid
24 //! because it only re-uses vector indexes once all currently-active (not-terminated) threads have an internal
25 //! vector clock that happens-after the join operation of the candidate thread. Threads that have not been joined
26 //! on are not considered. Since the thread's vector clock will only increase and a data-race implies that
27 //! there is some index x where clock[x] > thread_clock, when this is true clock[candidate-idx] > thread_clock
28 //! can never hold and hence a data-race can never be reported in that vector index again.
29 //! This means that the thread-index can be safely re-used, starting on the next timestamp for the newly created
32 //! The sequentially consistent ordering corresponds to the ordering that the threads
33 //! are currently scheduled, this means that the data-race detector has no additional
34 //! logic for sequentially consistent accesses at the moment since they are indistinguishable
35 //! from acquire/release operations. If weak memory orderings are explored then this
36 //! may need to change or be updated accordingly.
38 //! Per the C++ spec for the memory model a sequentially consistent operation:
39 //! "A load operation with this memory order performs an acquire operation,
40 //! a store performs a release operation, and read-modify-write performs
41 //! both an acquire operation and a release operation, plus a single total
42 //! order exists in which all threads observe all modifications in the same
43 //! order (see Sequentially-consistent ordering below) "
44 //! So in the absence of weak memory effects a seq-cst load & a seq-cst store is identical
45 //! to a acquire load and a release store given the global sequentially consistent order
48 //! The timestamps used in the data-race detector assign each sequence of non-atomic operations
49 //! followed by a single atomic or concurrent operation a single timestamp.
50 //! Write, Read, Write, ThreadJoin will be represented by a single timestamp value on a thread.
51 //! This is because extra increment operations between the operations in the sequence are not
52 //! required for accurate reporting of data-race values.
54 //! As per the paper a threads timestamp is only incremented after a release operation is performed
55 //! so some atomic operations that only perform acquires do not increment the timestamp. Due to shared
56 //! code some atomic operations may increment the timestamp when not necessary but this has no effect
57 //! on the data-race detection code.
60 //! currently we have our own local copy of the currently active thread index and names, this is due
61 //! in part to the inability to access the current location of threads.active_thread inside the AllocExtra
62 //! read, write and deallocate functions and should be cleaned up in the future.
65 cell::{Cell, Ref, RefCell, RefMut},
71 use rustc_data_structures::fx::{FxHashMap, FxHashSet};
72 use rustc_index::vec::{Idx, IndexVec};
73 use rustc_middle::{mir, ty::layout::TyAndLayout};
74 use rustc_target::abi::Size;
77 ImmTy, Immediate, InterpResult, MPlaceTy, MemPlaceMeta, MiriEvalContext, MiriEvalContextExt,
78 OpTy, Pointer, RangeMap, ScalarMaybeUninit, Tag, ThreadId, VClock, VTimestamp,
82 pub type AllocExtra = VClockAlloc;
83 pub type MemoryExtra = Rc<GlobalState>;
85 /// Valid atomic read-write operations, alias of atomic::Ordering (not non-exhaustive).
86 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
95 /// Valid atomic read operations, subset of atomic::Ordering.
96 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
97 pub enum AtomicReadOp {
103 /// Valid atomic write operations, subset of atomic::Ordering.
104 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
105 pub enum AtomicWriteOp {
111 /// Valid atomic fence operations, subset of atomic::Ordering.
112 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
113 pub enum AtomicFenceOp {
120 /// The current set of vector clocks describing the state
121 /// of a thread, contains the happens-before clock and
122 /// additional metadata to model atomic fence operations.
123 #[derive(Clone, Default, Debug)]
124 struct ThreadClockSet {
125 /// The increasing clock representing timestamps
126 /// that happen-before this thread.
129 /// The set of timestamps that will happen-before this
130 /// thread once it performs an acquire fence.
131 fence_acquire: VClock,
133 /// The last timestamp of happens-before relations that
134 /// have been released by this thread by a fence.
135 fence_release: VClock,
138 impl ThreadClockSet {
139 /// Apply the effects of a release fence to this
140 /// set of thread vector clocks.
142 fn apply_release_fence(&mut self) {
143 self.fence_release.clone_from(&self.clock);
146 /// Apply the effects of a acquire fence to this
147 /// set of thread vector clocks.
149 fn apply_acquire_fence(&mut self) {
150 self.clock.join(&self.fence_acquire);
153 /// Increment the happens-before clock at a
156 fn increment_clock(&mut self, index: VectorIdx) {
157 self.clock.increment_index(index);
160 /// Join the happens-before clock with that of
161 /// another thread, used to model thread join
163 fn join_with(&mut self, other: &ThreadClockSet) {
164 self.clock.join(&other.clock);
168 /// Error returned by finding a data race
169 /// should be elaborated upon.
170 #[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug)]
173 /// Externally stored memory cell clocks
174 /// explicitly to reduce memory usage for the
175 /// common case where no atomic operations
176 /// exists on the memory cell.
177 #[derive(Clone, PartialEq, Eq, Default, Debug)]
178 struct AtomicMemoryCellClocks {
179 /// The clock-vector of the timestamp of the last atomic
180 /// read operation performed by each thread.
181 /// This detects potential data-races between atomic read
182 /// and non-atomic write operations.
185 /// The clock-vector of the timestamp of the last atomic
186 /// write operation performed by each thread.
187 /// This detects potential data-races between atomic write
188 /// and non-atomic read or write operations.
189 write_vector: VClock,
191 /// Synchronization vector for acquire-release semantics
192 /// contains the vector of timestamps that will
193 /// happen-before a thread if an acquire-load is
194 /// performed on the data.
198 /// Type of write operation: allocating memory
199 /// non-atomic writes and deallocating memory
200 /// are all treated as writes for the purpose
201 /// of the data-race detector.
202 #[derive(Copy, Clone, PartialEq, Eq, Debug)]
207 /// Standard unsynchronized write.
210 /// Deallocate memory.
211 /// Some races with deallocation will be missed and instead
212 /// reported as invalid accesses of freed memory due to
213 /// the order of checks.
217 fn get_descriptor(self) -> &'static str {
219 WriteType::Allocate => "Allocate",
220 WriteType::Write => "Write",
221 WriteType::Deallocate => "Deallocate",
226 /// Memory Cell vector clock metadata
227 /// for data-race detection.
228 #[derive(Clone, PartialEq, Eq, Debug)]
229 struct MemoryCellClocks {
230 /// The vector-clock timestamp of the last write
231 /// corresponding to the writing threads timestamp.
234 /// The identifier of the vector index, corresponding to a thread
235 /// that performed the last write operation.
236 write_index: VectorIdx,
238 /// The type of operation that the write index represents,
239 /// either newly allocated memory, a non-atomic write or
240 /// a deallocation of memory.
241 write_type: WriteType,
243 /// The vector-clock of the timestamp of the last read operation
244 /// performed by a thread since the last write operation occurred.
245 /// It is reset to zero on each write operation.
248 /// Atomic acquire & release sequence tracking clocks.
249 /// For non-atomic memory in the common case this
250 /// value is set to None.
251 atomic_ops: Option<Box<AtomicMemoryCellClocks>>,
254 impl MemoryCellClocks {
255 /// Create a new set of clocks representing memory allocated
256 /// at a given vector timestamp and index.
257 fn new(alloc: VTimestamp, alloc_index: VectorIdx) -> Self {
259 read: VClock::default(),
261 write_index: alloc_index,
262 write_type: WriteType::Allocate,
267 /// Load the internal atomic memory cells if they exist.
269 fn atomic(&self) -> Option<&AtomicMemoryCellClocks> {
270 match &self.atomic_ops {
271 Some(op) => Some(&*op),
276 /// Load or create the internal atomic memory metadata
277 /// if it does not exist.
279 fn atomic_mut(&mut self) -> &mut AtomicMemoryCellClocks {
280 self.atomic_ops.get_or_insert_with(Default::default)
283 /// Update memory cell data-race tracking for atomic
284 /// load acquire semantics, is a no-op if this memory was
285 /// not used previously as atomic memory.
288 clocks: &mut ThreadClockSet,
290 ) -> Result<(), DataRace> {
291 self.atomic_read_detect(clocks, index)?;
292 if let Some(atomic) = self.atomic() {
293 clocks.clock.join(&atomic.sync_vector);
298 /// Update memory cell data-race tracking for atomic
299 /// load relaxed semantics, is a no-op if this memory was
300 /// not used previously as atomic memory.
303 clocks: &mut ThreadClockSet,
305 ) -> Result<(), DataRace> {
306 self.atomic_read_detect(clocks, index)?;
307 if let Some(atomic) = self.atomic() {
308 clocks.fence_acquire.join(&atomic.sync_vector);
313 /// Update the memory cell data-race tracking for atomic
314 /// store release semantics.
315 fn store_release(&mut self, clocks: &ThreadClockSet, index: VectorIdx) -> Result<(), DataRace> {
316 self.atomic_write_detect(clocks, index)?;
317 let atomic = self.atomic_mut();
318 atomic.sync_vector.clone_from(&clocks.clock);
322 /// Update the memory cell data-race tracking for atomic
323 /// store relaxed semantics.
324 fn store_relaxed(&mut self, clocks: &ThreadClockSet, index: VectorIdx) -> Result<(), DataRace> {
325 self.atomic_write_detect(clocks, index)?;
327 // The handling of release sequences was changed in C++20 and so
328 // the code here is different to the paper since now all relaxed
329 // stores block release sequences. The exception for same-thread
330 // relaxed stores has been removed.
331 let atomic = self.atomic_mut();
332 atomic.sync_vector.clone_from(&clocks.fence_release);
336 /// Update the memory cell data-race tracking for atomic
337 /// store release semantics for RMW operations.
338 fn rmw_release(&mut self, clocks: &ThreadClockSet, index: VectorIdx) -> Result<(), DataRace> {
339 self.atomic_write_detect(clocks, index)?;
340 let atomic = self.atomic_mut();
341 atomic.sync_vector.join(&clocks.clock);
345 /// Update the memory cell data-race tracking for atomic
346 /// store relaxed semantics for RMW operations.
347 fn rmw_relaxed(&mut self, clocks: &ThreadClockSet, index: VectorIdx) -> Result<(), DataRace> {
348 self.atomic_write_detect(clocks, index)?;
349 let atomic = self.atomic_mut();
350 atomic.sync_vector.join(&clocks.fence_release);
354 /// Detect data-races with an atomic read, caused by a non-atomic write that does
355 /// not happen-before the atomic-read.
356 fn atomic_read_detect(
358 clocks: &ThreadClockSet,
360 ) -> Result<(), DataRace> {
361 log::trace!("Atomic read with vectors: {:#?} :: {:#?}", self, clocks);
362 if self.write <= clocks.clock[self.write_index] {
363 let atomic = self.atomic_mut();
364 atomic.read_vector.set_at_index(&clocks.clock, index);
371 /// Detect data-races with an atomic write, either with a non-atomic read or with
372 /// a non-atomic write.
373 fn atomic_write_detect(
375 clocks: &ThreadClockSet,
377 ) -> Result<(), DataRace> {
378 log::trace!("Atomic write with vectors: {:#?} :: {:#?}", self, clocks);
379 if self.write <= clocks.clock[self.write_index] && self.read <= clocks.clock {
380 let atomic = self.atomic_mut();
381 atomic.write_vector.set_at_index(&clocks.clock, index);
388 /// Detect races for non-atomic read operations at the current memory cell
389 /// returns true if a data-race is detected.
392 clocks: &ThreadClockSet,
394 ) -> Result<(), DataRace> {
395 log::trace!("Unsynchronized read with vectors: {:#?} :: {:#?}", self, clocks);
396 if self.write <= clocks.clock[self.write_index] {
397 let race_free = if let Some(atomic) = self.atomic() {
398 atomic.write_vector <= clocks.clock
403 self.read.set_at_index(&clocks.clock, index);
413 /// Detect races for non-atomic write operations at the current memory cell
414 /// returns true if a data-race is detected.
415 fn write_race_detect(
417 clocks: &ThreadClockSet,
419 write_type: WriteType,
420 ) -> Result<(), DataRace> {
421 log::trace!("Unsynchronized write with vectors: {:#?} :: {:#?}", self, clocks);
422 if self.write <= clocks.clock[self.write_index] && self.read <= clocks.clock {
423 let race_free = if let Some(atomic) = self.atomic() {
424 atomic.write_vector <= clocks.clock && atomic.read_vector <= clocks.clock
429 self.write = clocks.clock[index];
430 self.write_index = index;
431 self.write_type = write_type;
432 self.read.set_zero_vector();
443 /// Evaluation context extensions.
444 impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for MiriEvalContext<'mir, 'tcx> {}
445 pub trait EvalContextExt<'mir, 'tcx: 'mir>: MiriEvalContextExt<'mir, 'tcx> {
446 /// Atomic variant of read_scalar_at_offset.
447 fn read_scalar_at_offset_atomic(
451 layout: TyAndLayout<'tcx>,
452 atomic: AtomicReadOp,
453 ) -> InterpResult<'tcx, ScalarMaybeUninit<Tag>> {
454 let this = self.eval_context_ref();
455 let op_place = this.deref_operand(op)?;
456 let offset = Size::from_bytes(offset);
458 // Ensure that the following read at an offset is within bounds.
459 assert!(op_place.layout.size >= offset + layout.size);
460 let value_place = op_place.offset(offset, MemPlaceMeta::None, layout, this)?;
461 this.read_scalar_atomic(value_place, atomic)
464 /// Atomic variant of write_scalar_at_offset.
465 fn write_scalar_at_offset_atomic(
469 value: impl Into<ScalarMaybeUninit<Tag>>,
470 layout: TyAndLayout<'tcx>,
471 atomic: AtomicWriteOp,
472 ) -> InterpResult<'tcx> {
473 let this = self.eval_context_mut();
474 let op_place = this.deref_operand(op)?;
475 let offset = Size::from_bytes(offset);
477 // Ensure that the following read at an offset is within bounds.
478 assert!(op_place.layout.size >= offset + layout.size);
479 let value_place = op_place.offset(offset, MemPlaceMeta::None, layout, this)?;
480 this.write_scalar_atomic(value.into(), value_place, atomic)
483 /// Perform an atomic read operation at the memory location.
484 fn read_scalar_atomic(
486 place: MPlaceTy<'tcx, Tag>,
487 atomic: AtomicReadOp,
488 ) -> InterpResult<'tcx, ScalarMaybeUninit<Tag>> {
489 let this = self.eval_context_ref();
490 let scalar = this.allow_data_races_ref(move |this| this.read_scalar(place.into()))?;
491 self.validate_atomic_load(place, atomic)?;
495 /// Perform an atomic write operation at the memory location.
496 fn write_scalar_atomic(
498 val: ScalarMaybeUninit<Tag>,
499 dest: MPlaceTy<'tcx, Tag>,
500 atomic: AtomicWriteOp,
501 ) -> InterpResult<'tcx> {
502 let this = self.eval_context_mut();
503 this.allow_data_races_mut(move |this| this.write_scalar(val, dest.into()))?;
504 self.validate_atomic_store(dest, atomic)
507 /// Perform a atomic operation on a memory location.
508 fn atomic_op_immediate(
510 place: MPlaceTy<'tcx, Tag>,
511 rhs: ImmTy<'tcx, Tag>,
515 ) -> InterpResult<'tcx, ImmTy<'tcx, Tag>> {
516 let this = self.eval_context_mut();
518 let old = this.allow_data_races_mut(|this| this.read_immediate(place.into()))?;
520 // Atomics wrap around on overflow.
521 let val = this.binary_op(op, old, rhs)?;
522 let val = if neg { this.unary_op(mir::UnOp::Not, val)? } else { val };
523 this.allow_data_races_mut(|this| this.write_immediate(*val, place.into()))?;
525 this.validate_atomic_rmw(place, atomic)?;
529 /// Perform an atomic exchange with a memory place and a new
530 /// scalar value, the old value is returned.
531 fn atomic_exchange_scalar(
533 place: MPlaceTy<'tcx, Tag>,
534 new: ScalarMaybeUninit<Tag>,
536 ) -> InterpResult<'tcx, ScalarMaybeUninit<Tag>> {
537 let this = self.eval_context_mut();
539 let old = this.allow_data_races_mut(|this| this.read_scalar(place.into()))?;
540 this.allow_data_races_mut(|this| this.write_scalar(new, place.into()))?;
541 this.validate_atomic_rmw(place, atomic)?;
545 /// Perform an atomic compare and exchange at a given memory location.
546 /// On success an atomic RMW operation is performed and on failure
547 /// only an atomic read occurs.
548 fn atomic_compare_exchange_scalar(
550 place: MPlaceTy<'tcx, Tag>,
551 expect_old: ImmTy<'tcx, Tag>,
552 new: ScalarMaybeUninit<Tag>,
555 ) -> InterpResult<'tcx, Immediate<Tag>> {
556 let this = self.eval_context_mut();
558 // Failure ordering cannot be stronger than success ordering, therefore first attempt
559 // to read with the failure ordering and if successful then try again with the success
560 // read ordering and write in the success case.
561 // Read as immediate for the sake of `binary_op()`
562 let old = this.allow_data_races_mut(|this| this.read_immediate(place.into()))?;
564 // `binary_op` will bail if either of them is not a scalar.
565 let eq = this.overflowing_binary_op(mir::BinOp::Eq, old, expect_old)?.0;
566 let res = Immediate::ScalarPair(old.to_scalar_or_uninit(), eq.into());
568 // Update ptr depending on comparison.
569 // if successful, perform a full rw-atomic validation
570 // otherwise treat this as an atomic load with the fail ordering.
572 this.allow_data_races_mut(|this| this.write_scalar(new, place.into()))?;
573 this.validate_atomic_rmw(place, success)?;
575 this.validate_atomic_load(place, fail)?;
578 // Return the old value.
582 /// Update the data-race detector for an atomic read occurring at the
583 /// associated memory-place and on the current thread.
584 fn validate_atomic_load(
586 place: MPlaceTy<'tcx, Tag>,
587 atomic: AtomicReadOp,
588 ) -> InterpResult<'tcx> {
589 let this = self.eval_context_ref();
590 this.validate_atomic_op(
594 move |memory, clocks, index, atomic| {
595 if atomic == AtomicReadOp::Relaxed {
596 memory.load_relaxed(&mut *clocks, index)
598 memory.load_acquire(&mut *clocks, index)
604 /// Update the data-race detector for an atomic write occurring at the
605 /// associated memory-place and on the current thread.
606 fn validate_atomic_store(
608 place: MPlaceTy<'tcx, Tag>,
609 atomic: AtomicWriteOp,
610 ) -> InterpResult<'tcx> {
611 let this = self.eval_context_ref();
612 this.validate_atomic_op(
616 move |memory, clocks, index, atomic| {
617 if atomic == AtomicWriteOp::Relaxed {
618 memory.store_relaxed(clocks, index)
620 memory.store_release(clocks, index)
626 /// Update the data-race detector for an atomic read-modify-write occurring
627 /// at the associated memory place and on the current thread.
628 fn validate_atomic_rmw(
630 place: MPlaceTy<'tcx, Tag>,
632 ) -> InterpResult<'tcx> {
634 let acquire = matches!(atomic, Acquire | AcqRel | SeqCst);
635 let release = matches!(atomic, Release | AcqRel | SeqCst);
636 let this = self.eval_context_ref();
637 this.validate_atomic_op(place, atomic, "Atomic RMW", move |memory, clocks, index, _| {
639 memory.load_acquire(clocks, index)?;
641 memory.load_relaxed(clocks, index)?;
644 memory.rmw_release(clocks, index)
646 memory.rmw_relaxed(clocks, index)
651 /// Update the data-race detector for an atomic fence on the current thread.
652 fn validate_atomic_fence(&mut self, atomic: AtomicFenceOp) -> InterpResult<'tcx> {
653 let this = self.eval_context_mut();
654 if let Some(data_race) = &this.memory.extra.data_race {
655 data_race.maybe_perform_sync_operation(move |index, mut clocks| {
656 log::trace!("Atomic fence on {:?} with ordering {:?}", index, atomic);
658 // Apply data-race detection for the current fences
659 // this treats AcqRel and SeqCst as the same as a acquire
660 // and release fence applied in the same timestamp.
661 if atomic != AtomicFenceOp::Release {
662 // Either Acquire | AcqRel | SeqCst
663 clocks.apply_acquire_fence();
665 if atomic != AtomicFenceOp::Acquire {
666 // Either Release | AcqRel | SeqCst
667 clocks.apply_release_fence();
670 // Increment timestamp in case of release semantics.
671 Ok(atomic != AtomicFenceOp::Acquire)
679 /// Vector clock metadata for a logical memory allocation.
680 #[derive(Debug, Clone)]
681 pub struct VClockAlloc {
682 /// Assigning each byte a MemoryCellClocks.
683 alloc_ranges: RefCell<RangeMap<MemoryCellClocks>>,
685 /// Pointer to global state.
690 /// Create a new data-race detector for newly allocated memory.
691 pub fn new_allocation(global: &MemoryExtra, len: Size, track_alloc: bool) -> VClockAlloc {
692 let (alloc_timestamp, alloc_index) = if track_alloc {
693 let (alloc_index, clocks) = global.current_thread_state();
694 let alloc_timestamp = clocks.clock[alloc_index];
695 (alloc_timestamp, alloc_index)
697 (0, VectorIdx::MAX_INDEX)
700 global: Rc::clone(global),
701 alloc_ranges: RefCell::new(RangeMap::new(
702 len, MemoryCellClocks::new(alloc_timestamp, alloc_index)
707 // Find an index, if one exists where the value
708 // in `l` is greater than the value in `r`.
709 fn find_gt_index(l: &VClock, r: &VClock) -> Option<VectorIdx> {
710 log::trace!("Find index where not {:?} <= {:?}", l, r);
711 let l_slice = l.as_slice();
712 let r_slice = r.as_slice();
717 .find_map(|(idx, (&l, &r))| if l > r { Some(idx) } else { None })
719 if l_slice.len() > r_slice.len() {
720 // By invariant, if l_slice is longer
721 // then one element must be larger.
722 // This just validates that this is true
723 // and reports earlier elements first.
724 let l_remainder_slice = &l_slice[r_slice.len()..];
725 let idx = l_remainder_slice
728 .find_map(|(idx, &r)| if r == 0 { None } else { Some(idx) })
729 .expect("Invalid VClock Invariant");
730 Some(idx + r_slice.len())
735 .map(|idx| VectorIdx::new(idx))
738 /// Report a data-race found in the program.
739 /// This finds the two racing threads and the type
740 /// of data-race that occurred. This will also
741 /// return info about the memory location the data-race
745 fn report_data_race<'tcx>(
746 global: &MemoryExtra,
747 range: &MemoryCellClocks,
750 pointer: Pointer<Tag>,
752 ) -> InterpResult<'tcx> {
753 let (current_index, current_clocks) = global.current_thread_state();
755 let (other_action, other_thread, other_clock) = if range.write
756 > current_clocks.clock[range.write_index]
758 // Convert the write action into the vector clock it
759 // represents for diagnostic purposes.
760 write_clock = VClock::new_with_index(range.write_index, range.write);
761 (range.write_type.get_descriptor(), range.write_index, &write_clock)
762 } else if let Some(idx) = Self::find_gt_index(&range.read, ¤t_clocks.clock) {
763 ("Read", idx, &range.read)
764 } else if !is_atomic {
765 if let Some(atomic) = range.atomic() {
766 if let Some(idx) = Self::find_gt_index(&atomic.write_vector, ¤t_clocks.clock)
768 ("Atomic Store", idx, &atomic.write_vector)
769 } else if let Some(idx) =
770 Self::find_gt_index(&atomic.read_vector, ¤t_clocks.clock)
772 ("Atomic Load", idx, &atomic.read_vector)
775 "Failed to report data-race for non-atomic operation: no race found"
780 "Failed to report data-race for non-atomic operation: no atomic component"
784 unreachable!("Failed to report data-race for atomic operation")
787 // Load elaborated thread information about the racing thread actions.
788 let current_thread_info = global.print_thread_metadata(current_index);
789 let other_thread_info = global.print_thread_metadata(other_thread);
791 // Throw the data-race detection.
793 "Data race detected between {} on {} and {} on {}, memory({:?},offset={},size={})\
794 \n\t\t -current vector clock = {:?}\
795 \n\t\t -conflicting timestamp = {:?}",
801 pointer.offset.bytes(),
803 current_clocks.clock,
808 /// Detect data-races for an unsynchronized read operation, will not perform
809 /// data-race detection if `multi-threaded` is false, either due to no threads
810 /// being created or if it is temporarily disabled during a racy read or write
811 /// operation for which data-race detection is handled separately, for example
812 /// atomic read operations.
813 pub fn read<'tcx>(&self, pointer: Pointer<Tag>, len: Size) -> InterpResult<'tcx> {
814 if self.global.multi_threaded.get() {
815 let (index, clocks) = self.global.current_thread_state();
816 let mut alloc_ranges = self.alloc_ranges.borrow_mut();
817 for (_, range) in alloc_ranges.iter_mut(pointer.offset, len) {
818 if let Err(DataRace) = range.read_race_detect(&*clocks, index) {
820 return Self::report_data_race(
836 // Shared code for detecting data-races on unique access to a section of memory
837 fn unique_access<'tcx>(
839 pointer: Pointer<Tag>,
841 write_type: WriteType,
842 ) -> InterpResult<'tcx> {
843 if self.global.multi_threaded.get() {
844 let (index, clocks) = self.global.current_thread_state();
845 for (_, range) in self.alloc_ranges.get_mut().iter_mut(pointer.offset, len) {
846 if let Err(DataRace) = range.write_race_detect(&*clocks, index, write_type) {
848 return Self::report_data_race(
851 write_type.get_descriptor(),
864 /// Detect data-races for an unsynchronized write operation, will not perform
865 /// data-race threads if `multi-threaded` is false, either due to no threads
866 /// being created or if it is temporarily disabled during a racy read or write
868 pub fn write<'tcx>(&mut self, pointer: Pointer<Tag>, len: Size) -> InterpResult<'tcx> {
869 self.unique_access(pointer, len, WriteType::Write)
872 /// Detect data-races for an unsynchronized deallocate operation, will not perform
873 /// data-race threads if `multi-threaded` is false, either due to no threads
874 /// being created or if it is temporarily disabled during a racy read or write
876 pub fn deallocate<'tcx>(&mut self, pointer: Pointer<Tag>, len: Size) -> InterpResult<'tcx> {
877 self.unique_access(pointer, len, WriteType::Deallocate)
881 impl<'mir, 'tcx: 'mir> EvalContextPrivExt<'mir, 'tcx> for MiriEvalContext<'mir, 'tcx> {}
882 trait EvalContextPrivExt<'mir, 'tcx: 'mir>: MiriEvalContextExt<'mir, 'tcx> {
883 // Temporarily allow data-races to occur, this should only be
884 // used if either one of the appropriate `validate_atomic` functions
885 // will be called to treat a memory access as atomic or if the memory
886 // being accessed should be treated as internal state, that cannot be
887 // accessed by the interpreted program.
889 fn allow_data_races_ref<R>(&self, op: impl FnOnce(&MiriEvalContext<'mir, 'tcx>) -> R) -> R {
890 let this = self.eval_context_ref();
891 let old = if let Some(data_race) = &this.memory.extra.data_race {
892 data_race.multi_threaded.replace(false)
896 let result = op(this);
897 if let Some(data_race) = &this.memory.extra.data_race {
898 data_race.multi_threaded.set(old);
903 /// Same as `allow_data_races_ref`, this temporarily disables any data-race detection and
904 /// so should only be used for atomic operations or internal state that the program cannot
907 fn allow_data_races_mut<R>(
909 op: impl FnOnce(&mut MiriEvalContext<'mir, 'tcx>) -> R,
911 let this = self.eval_context_mut();
912 let old = if let Some(data_race) = &this.memory.extra.data_race {
913 data_race.multi_threaded.replace(false)
917 let result = op(this);
918 if let Some(data_race) = &this.memory.extra.data_race {
919 data_race.multi_threaded.set(old);
924 /// Generic atomic operation implementation,
925 /// this accesses memory via get_raw instead of
926 /// get_raw_mut, due to issues calling get_raw_mut
927 /// for atomic loads from read-only memory.
928 /// FIXME: is this valid, or should get_raw_mut be used for
929 /// atomic-stores/atomic-rmw?
930 fn validate_atomic_op<A: Debug + Copy>(
932 place: MPlaceTy<'tcx, Tag>,
936 &mut MemoryCellClocks,
940 ) -> Result<(), DataRace>,
941 ) -> InterpResult<'tcx> {
942 let this = self.eval_context_ref();
943 if let Some(data_race) = &this.memory.extra.data_race {
944 if data_race.multi_threaded.get() {
945 // Load and log the atomic operation.
946 let place_ptr = place.ptr.assert_ptr();
947 let size = place.layout.size;
949 &this.memory.get_raw(place_ptr.alloc_id)?.extra.data_race.as_ref().unwrap();
951 "Atomic op({}) with ordering {:?} on memory({:?}, offset={}, size={})",
955 place_ptr.offset.bytes(),
959 // Perform the atomic operation.
960 let data_race = &alloc_meta.global;
961 data_race.maybe_perform_sync_operation(|index, mut clocks| {
963 alloc_meta.alloc_ranges.borrow_mut().iter_mut(place_ptr.offset, size)
965 if let Err(DataRace) = op(range, &mut *clocks, index, atomic) {
967 return VClockAlloc::report_data_race(
978 // This conservatively assumes all operations have release semantics
982 // Log changes to atomic memory.
983 if log::log_enabled!(log::Level::Trace) {
984 for (_, range) in alloc_meta.alloc_ranges.borrow().iter(place_ptr.offset, size)
987 "Updated atomic memory({:?}, offset={}, size={}) to {:#?}",
988 place.ptr.assert_ptr().alloc_id,
989 place_ptr.offset.bytes(),
1001 /// Extra metadata associated with a thread.
1002 #[derive(Debug, Clone, Default)]
1003 struct ThreadExtraState {
1004 /// The current vector index in use by the
1005 /// thread currently, this is set to None
1006 /// after the vector index has been re-used
1007 /// and hence the value will never need to be
1008 /// read during data-race reporting.
1009 vector_index: Option<VectorIdx>,
1011 /// The name of the thread, updated for better
1012 /// diagnostics when reporting detected data
1014 thread_name: Option<Box<str>>,
1016 /// Thread termination vector clock, this
1017 /// is set on thread termination and is used
1018 /// for joining on threads since the vector_index
1019 /// may be re-used when the join operation occurs.
1020 termination_vector_clock: Option<VClock>,
1023 /// Global data-race detection state, contains the currently
1024 /// executing thread as well as the vector-clocks associated
1025 /// with each of the threads.
1026 #[derive(Debug, Clone)]
1027 pub struct GlobalState {
1028 /// Set to true once the first additional
1029 /// thread has launched, due to the dependency
1030 /// between before and after a thread launch.
1031 /// Any data-races must be recorded after this
1032 /// so concurrent execution can ignore recording
1034 multi_threaded: Cell<bool>,
1036 /// Mapping of a vector index to a known set of thread
1037 /// clocks, this is not directly mapping from a thread id
1038 /// since it may refer to multiple threads.
1039 vector_clocks: RefCell<IndexVec<VectorIdx, ThreadClockSet>>,
1041 /// Mapping of a given vector index to the current thread
1042 /// that the execution is representing, this may change
1043 /// if a vector index is re-assigned to a new thread.
1044 vector_info: RefCell<IndexVec<VectorIdx, ThreadId>>,
1046 /// The mapping of a given thread to associated thread metadata.
1047 thread_info: RefCell<IndexVec<ThreadId, ThreadExtraState>>,
1049 /// The current vector index being executed.
1050 current_index: Cell<VectorIdx>,
1052 /// Potential vector indices that could be re-used on thread creation
1053 /// values are inserted here on after the thread has terminated and
1054 /// been joined with, and hence may potentially become free
1055 /// for use as the index for a new thread.
1056 /// Elements in this set may still require the vector index to
1057 /// report data-races, and can only be re-used after all
1058 /// active vector-clocks catch up with the threads timestamp.
1059 reuse_candidates: RefCell<FxHashSet<VectorIdx>>,
1061 /// Counts the number of threads that are currently active
1062 /// if the number of active threads reduces to 1 and then
1063 /// a join operation occurs with the remaining main thread
1064 /// then multi-threaded execution may be disabled.
1065 active_thread_count: Cell<usize>,
1067 /// This contains threads that have terminated, but not yet joined
1068 /// and so cannot become re-use candidates until a join operation
1070 /// The associated vector index will be moved into re-use candidates
1071 /// after the join operation occurs.
1072 terminated_threads: RefCell<FxHashMap<ThreadId, VectorIdx>>,
1076 /// Create a new global state, setup with just thread-id=0
1077 /// advanced to timestamp = 1.
1078 pub fn new() -> Self {
1079 let global_state = GlobalState {
1080 multi_threaded: Cell::new(false),
1081 vector_clocks: RefCell::new(IndexVec::new()),
1082 vector_info: RefCell::new(IndexVec::new()),
1083 thread_info: RefCell::new(IndexVec::new()),
1084 current_index: Cell::new(VectorIdx::new(0)),
1085 active_thread_count: Cell::new(1),
1086 reuse_candidates: RefCell::new(FxHashSet::default()),
1087 terminated_threads: RefCell::new(FxHashMap::default()),
1090 // Setup the main-thread since it is not explicitly created:
1091 // uses vector index and thread-id 0, also the rust runtime gives
1092 // the main-thread a name of "main".
1093 let index = global_state.vector_clocks.borrow_mut().push(ThreadClockSet::default());
1094 global_state.vector_info.borrow_mut().push(ThreadId::new(0));
1095 global_state.thread_info.borrow_mut().push(ThreadExtraState {
1096 vector_index: Some(index),
1097 thread_name: Some("main".to_string().into_boxed_str()),
1098 termination_vector_clock: None,
1104 // Try to find vector index values that can potentially be re-used
1105 // by a new thread instead of a new vector index being created.
1106 fn find_vector_index_reuse_candidate(&self) -> Option<VectorIdx> {
1107 let mut reuse = self.reuse_candidates.borrow_mut();
1108 let vector_clocks = self.vector_clocks.borrow();
1109 let vector_info = self.vector_info.borrow();
1110 let terminated_threads = self.terminated_threads.borrow();
1111 for &candidate in reuse.iter() {
1112 let target_timestamp = vector_clocks[candidate].clock[candidate];
1113 if vector_clocks.iter_enumerated().all(|(clock_idx, clock)| {
1114 // The thread happens before the clock, and hence cannot report
1115 // a data-race with this the candidate index.
1116 let no_data_race = clock.clock[candidate] >= target_timestamp;
1118 // The vector represents a thread that has terminated and hence cannot
1119 // report a data-race with the candidate index.
1120 let thread_id = vector_info[clock_idx];
1121 let vector_terminated =
1122 reuse.contains(&clock_idx) || terminated_threads.contains_key(&thread_id);
1124 // The vector index cannot report a race with the candidate index
1125 // and hence allows the candidate index to be re-used.
1126 no_data_race || vector_terminated
1128 // All vector clocks for each vector index are equal to
1129 // the target timestamp, and the thread is known to have
1130 // terminated, therefore this vector clock index cannot
1131 // report any more data-races.
1132 assert!(reuse.remove(&candidate));
1133 return Some(candidate);
1139 // Hook for thread creation, enabled multi-threaded execution and marks
1140 // the current thread timestamp as happening-before the current thread.
1142 pub fn thread_created(&self, thread: ThreadId) {
1143 let current_index = self.current_index();
1145 // Increment the number of active threads.
1146 let active_threads = self.active_thread_count.get();
1147 self.active_thread_count.set(active_threads + 1);
1149 // Enable multi-threaded execution, there are now two threads
1150 // so data-races are now possible.
1151 self.multi_threaded.set(true);
1153 // Load and setup the associated thread metadata
1154 let mut thread_info = self.thread_info.borrow_mut();
1155 thread_info.ensure_contains_elem(thread, Default::default);
1157 // Assign a vector index for the thread, attempting to re-use an old
1158 // vector index that can no longer report any data-races if possible.
1159 let created_index = if let Some(reuse_index) = self.find_vector_index_reuse_candidate() {
1160 // Now re-configure the re-use candidate, increment the clock
1161 // for the new sync use of the vector.
1162 let mut vector_clocks = self.vector_clocks.borrow_mut();
1163 vector_clocks[reuse_index].increment_clock(reuse_index);
1165 // Locate the old thread the vector was associated with and update
1166 // it to represent the new thread instead.
1167 let mut vector_info = self.vector_info.borrow_mut();
1168 let old_thread = vector_info[reuse_index];
1169 vector_info[reuse_index] = thread;
1171 // Mark the thread the vector index was associated with as no longer
1172 // representing a thread index.
1173 thread_info[old_thread].vector_index = None;
1177 // No vector re-use candidates available, instead create
1178 // a new vector index.
1179 let mut vector_info = self.vector_info.borrow_mut();
1180 vector_info.push(thread)
1183 log::trace!("Creating thread = {:?} with vector index = {:?}", thread, created_index);
1185 // Mark the chosen vector index as in use by the thread.
1186 thread_info[thread].vector_index = Some(created_index);
1188 // Create a thread clock set if applicable.
1189 let mut vector_clocks = self.vector_clocks.borrow_mut();
1190 if created_index == vector_clocks.next_index() {
1191 vector_clocks.push(ThreadClockSet::default());
1194 // Now load the two clocks and configure the initial state.
1195 let (current, created) = vector_clocks.pick2_mut(current_index, created_index);
1197 // Join the created with current, since the current threads
1198 // previous actions happen-before the created thread.
1199 created.join_with(current);
1201 // Advance both threads after the synchronized operation.
1202 // Both operations are considered to have release semantics.
1203 current.increment_clock(current_index);
1204 created.increment_clock(created_index);
1207 /// Hook on a thread join to update the implicit happens-before relation
1208 /// between the joined thread and the current thread.
1210 pub fn thread_joined(&self, current_thread: ThreadId, join_thread: ThreadId) {
1211 let mut clocks_vec = self.vector_clocks.borrow_mut();
1212 let thread_info = self.thread_info.borrow();
1214 // Load the vector clock of the current thread.
1215 let current_index = thread_info[current_thread]
1217 .expect("Performed thread join on thread with no assigned vector");
1218 let current = &mut clocks_vec[current_index];
1220 // Load the associated vector clock for the terminated thread.
1221 let join_clock = thread_info[join_thread]
1222 .termination_vector_clock
1224 .expect("Joined with thread but thread has not terminated");
1227 // The join thread happens-before the current thread
1228 // so update the current vector clock.
1229 // Is not a release operation so the clock is not incremented.
1230 current.clock.join(join_clock);
1232 // Check the number of active threads, if the value is 1
1233 // then test for potentially disabling multi-threaded execution.
1234 let active_threads = self.active_thread_count.get();
1235 if active_threads == 1 {
1236 // May potentially be able to disable multi-threaded execution.
1237 let current_clock = &clocks_vec[current_index];
1240 .all(|(idx, clocks)| clocks.clock[idx] <= current_clock.clock[idx])
1242 // All thread terminations happen-before the current clock
1243 // therefore no data-races can be reported until a new thread
1244 // is created, so disable multi-threaded execution.
1245 self.multi_threaded.set(false);
1249 // If the thread is marked as terminated but not joined
1250 // then move the thread to the re-use set.
1251 let mut termination = self.terminated_threads.borrow_mut();
1252 if let Some(index) = termination.remove(&join_thread) {
1253 let mut reuse = self.reuse_candidates.borrow_mut();
1254 reuse.insert(index);
1258 /// On thread termination, the vector-clock may re-used
1259 /// in the future once all remaining thread-clocks catch
1260 /// up with the time index of the terminated thread.
1261 /// This assigns thread termination with a unique index
1262 /// which will be used to join the thread
1263 /// This should be called strictly before any calls to
1264 /// `thread_joined`.
1266 pub fn thread_terminated(&self) {
1267 let current_index = self.current_index();
1269 // Increment the clock to a unique termination timestamp.
1270 let mut vector_clocks = self.vector_clocks.borrow_mut();
1271 let current_clocks = &mut vector_clocks[current_index];
1272 current_clocks.increment_clock(current_index);
1274 // Load the current thread id for the executing vector.
1275 let vector_info = self.vector_info.borrow();
1276 let current_thread = vector_info[current_index];
1278 // Load the current thread metadata, and move to a terminated
1279 // vector state. Setting up the vector clock all join operations
1281 let mut thread_info = self.thread_info.borrow_mut();
1282 let current = &mut thread_info[current_thread];
1283 current.termination_vector_clock = Some(current_clocks.clock.clone());
1285 // Add this thread as a candidate for re-use after a thread join
1287 let mut termination = self.terminated_threads.borrow_mut();
1288 termination.insert(current_thread, current_index);
1290 // Reduce the number of active threads, now that a thread has
1292 let mut active_threads = self.active_thread_count.get();
1293 active_threads -= 1;
1294 self.active_thread_count.set(active_threads);
1297 /// Hook for updating the local tracker of the currently
1298 /// enabled thread, should always be updated whenever
1299 /// `active_thread` in thread.rs is updated.
1301 pub fn thread_set_active(&self, thread: ThreadId) {
1302 let thread_info = self.thread_info.borrow();
1303 let vector_idx = thread_info[thread]
1305 .expect("Setting thread active with no assigned vector");
1306 self.current_index.set(vector_idx);
1309 /// Hook for updating the local tracker of the threads name
1310 /// this should always mirror the local value in thread.rs
1311 /// the thread name is used for improved diagnostics
1312 /// during a data-race.
1314 pub fn thread_set_name(&self, thread: ThreadId, name: String) {
1315 let name = name.into_boxed_str();
1316 let mut thread_info = self.thread_info.borrow_mut();
1317 thread_info[thread].thread_name = Some(name);
1320 /// Attempt to perform a synchronized operation, this
1321 /// will perform no operation if multi-threading is
1322 /// not currently enabled.
1323 /// Otherwise it will increment the clock for the current
1324 /// vector before and after the operation for data-race
1325 /// detection between any happens-before edges the
1326 /// operation may create.
1327 fn maybe_perform_sync_operation<'tcx>(
1329 op: impl FnOnce(VectorIdx, RefMut<'_, ThreadClockSet>) -> InterpResult<'tcx, bool>,
1330 ) -> InterpResult<'tcx> {
1331 if self.multi_threaded.get() {
1332 let (index, clocks) = self.current_thread_state_mut();
1333 if op(index, clocks)? {
1334 let (_, mut clocks) = self.current_thread_state_mut();
1335 clocks.increment_clock(index);
1341 /// Internal utility to identify a thread stored internally
1342 /// returns the id and the name for better diagnostics.
1343 fn print_thread_metadata(&self, vector: VectorIdx) -> String {
1344 let thread = self.vector_info.borrow()[vector];
1345 let thread_name = &self.thread_info.borrow()[thread].thread_name;
1346 if let Some(name) = thread_name {
1347 let name: &str = name;
1348 format!("Thread(id = {:?}, name = {:?})", thread.to_u32(), &*name)
1350 format!("Thread(id = {:?})", thread.to_u32())
1354 /// Acquire a lock, express that the previous call of
1355 /// `validate_lock_release` must happen before this.
1356 /// As this is an acquire operation, the thread timestamp is not
1358 pub fn validate_lock_acquire(&self, lock: &VClock, thread: ThreadId) {
1359 let (_, mut clocks) = self.load_thread_state_mut(thread);
1360 clocks.clock.join(&lock);
1363 /// Release a lock handle, express that this happens-before
1364 /// any subsequent calls to `validate_lock_acquire`.
1365 /// For normal locks this should be equivalent to `validate_lock_release_shared`
1366 /// since an acquire operation should have occurred before, however
1367 /// for futex & condvar operations this is not the case and this
1368 /// operation must be used.
1369 pub fn validate_lock_release(&self, lock: &mut VClock, thread: ThreadId) {
1370 let (index, mut clocks) = self.load_thread_state_mut(thread);
1371 lock.clone_from(&clocks.clock);
1372 clocks.increment_clock(index);
1375 /// Release a lock handle, express that this happens-before
1376 /// any subsequent calls to `validate_lock_acquire` as well
1377 /// as any previous calls to this function after any
1378 /// `validate_lock_release` calls.
1379 /// For normal locks this should be equivalent to `validate_lock_release`.
1380 /// This function only exists for joining over the set of concurrent readers
1381 /// in a read-write lock and should not be used for anything else.
1382 pub fn validate_lock_release_shared(&self, lock: &mut VClock, thread: ThreadId) {
1383 let (index, mut clocks) = self.load_thread_state_mut(thread);
1384 lock.join(&clocks.clock);
1385 clocks.increment_clock(index);
1388 /// Load the vector index used by the given thread as well as the set of vector clocks
1389 /// used by the thread.
1391 fn load_thread_state_mut(&self, thread: ThreadId) -> (VectorIdx, RefMut<'_, ThreadClockSet>) {
1392 let index = self.thread_info.borrow()[thread]
1394 .expect("Loading thread state for thread with no assigned vector");
1395 let ref_vector = self.vector_clocks.borrow_mut();
1396 let clocks = RefMut::map(ref_vector, |vec| &mut vec[index]);
1400 /// Load the current vector clock in use and the current set of thread clocks
1401 /// in use for the vector.
1403 fn current_thread_state(&self) -> (VectorIdx, Ref<'_, ThreadClockSet>) {
1404 let index = self.current_index();
1405 let ref_vector = self.vector_clocks.borrow();
1406 let clocks = Ref::map(ref_vector, |vec| &vec[index]);
1410 /// Load the current vector clock in use and the current set of thread clocks
1411 /// in use for the vector mutably for modification.
1413 fn current_thread_state_mut(&self) -> (VectorIdx, RefMut<'_, ThreadClockSet>) {
1414 let index = self.current_index();
1415 let ref_vector = self.vector_clocks.borrow_mut();
1416 let clocks = RefMut::map(ref_vector, |vec| &mut vec[index]);
1420 /// Return the current thread, should be the same
1421 /// as the data-race active thread.
1423 fn current_index(&self) -> VectorIdx {
1424 self.current_index.get()