RUSTSEC-2026-0143

When the vmem feature is enabled, VmemStorage<T>::new(Box<[UnsafeSyncCell<T>]>) (and every public constructor that funnels through it — ConcurrentHeapRB::default(cap), ConcurrentHeapRB::from(Vec<T>), From<Box<[T]>>, etc.) bit-copies the input buffer into a freshly mmap'd region with ptr::copy_nonoverlapping, then lets the source Box<[UnsafeSyncCell<T>]> drop normally.

Because UnsafeSyncCell<T> has a Drop impl that runs assume_init_drop on its inner MaybeUninit<T>, the source-side T values are dropped at the end of new, while bitwise duplicates (carrying the same heap pointers for String, Box, Vec, Arc, …) remain inside the mmap region. When the ring buffer is later destructed, the destructor's drop_in_place over the slice runs UnsafeSyncCell::drop a second time on every cell — a deterministic double-free of every heap-owning element. Reachable from 100% safe Rust.

Trigger

let v: Vec<Vec<u32>> = (0..1024).map(|i| vec![i, i+1, i+2]).collect();
let rb: oneringbuf::SharedVmemRB<Vec<u32>> = oneringbuf::SharedVmemRB::from(v);
drop(rb);
// glibc: free(): double free detected in tcache 2 -> abort
// ASan: AddressSanitizer: attempting double-free

Any T with a non-trivial Drop reproduces deterministically.

Fix

Fixed in 0.7.1 via upstream PR skilvingr/rust-oneringbuf#3, which deallocates the source Box<[UnsafeSyncCell<T>]> without running per-element Drop after the bytes have been copied into the mmap region. Vulnerable versions (< 0.7.1) have been yanked from crates.io.