Even if an element is popped from a queue, crossbeam would run its destructor inside the epoch-based garbage collector. This is a source of double frees.

The flaw was corrected by wrapping elements inside queues in a ManuallyDrop.

Thanks to @c0gent for reporting the issue.

Affected versions of this crate entered a corrupted state if mem::size_of::<T>() % allocation_granularity() != 0 and a specific allocation pattern was used: sufficiently shifting the deque elements over the mirrored page boundary.

This allows an attacker that controls controls both element insertion and removal to corrupt the deque, such that reading elements from it would read bytes corresponding to other elements in the deque. (e.g. a read of T could read some bytes from one value and some bytes from an adjacent one, resulting in a T whose value representation is not meaningful). This is undefined behavior.

The flaw was corrected by using a pair of pointers to track the head and tail of the deque instead of a pair of indices. This pair of pointers are represented using a Rust slice.

Affected versions of the crate incremented the current index of the drain filter iterator before calling the predicate function self.pred.

If the predicate function panics, it is possible for the last element in the iterator to be dropped twice.