Vector is empty after cloning struct with uninitialized member

Question

In Rust 1.29.0 one of my tests has started failing. I managed to get the strange bug down to this example:

#[derive(Clone, Debug)]
struct CountDrop<'a>(&'a std::cell::RefCell<usize>);

struct MayContainValue<T> {
    value: std::mem::ManuallyDrop<T>,
    has_value: u32,
}

impl<T: Clone> Clone for MayContainValue<T> {
    fn clone(&self) -> Self {
        Self {
            value: if self.has_value > 0 {
                self.value.clone()
            } else {
                unsafe { std::mem::uninitialized() }
            },
            has_value: self.has_value,
        }
    }
}

impl<T> Drop for MayContainValue<T> {
    fn drop(&mut self) {
        if self.has_value > 0 {
            unsafe {
                std::mem::ManuallyDrop::drop(&mut self.value);
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    #[test]
    fn check_drops() {
        let n = 2000;
        let drops = std::cell::RefCell::new(0usize);

        let mut slots = Vec::new();
        for _ in 0..n {
            slots.push(MayContainValue {
                value: std::mem::ManuallyDrop::new(CountDrop(&drops)),
                has_value: 1,
            });
        }

        unsafe { std::mem::ManuallyDrop::drop(&mut slots[0].value); }
        slots[0].has_value = 0;

        assert_eq!(slots.len(), slots.clone().len());
    }
}

I know the code looks strange; it is all ripped out of context. I reproduced this problem with cargo test on 64-bit Ubuntu on Rust 1.29.0. A friend could not reproduce on Windows with the same Rust version.

Other things that stop reproduction:

Lowering n below ~900.
Not running the example from within cargo test.
Replacing CountDrop's member with u64.
Using a Rust version before 1.29.0.

What's going on here? Yes, MayContainValue can have an uninitialized member, but this is never used in any way.

I also managed to reproduce this on play.rust-lang.org.

I'm not interested in 'solutions' that involve re-engineering MayContainValue in some safe way with Option or enum, I'm using manual storage and occupied/vacant discrimination for a good reason.

Shepmaster · Accepted Answer · 2018-09-16T00:07:06.263

TL;DR: Yes, creating an uninitialized reference is always undefined behavior. You cannot use mem::uninitialized safely with generics. There is not currently a good workaround for your specific case.

Running your code in valgrind reports 3 errors, each with the same stack trace:

==741== Conditional jump or move depends on uninitialised value(s)
==741==    at 0x11907F: <alloc::vec::Vec<T> as alloc::vec::SpecExtend<T, I>>::spec_extend (vec.rs:1892)
==741==    by 0x11861C: <alloc::vec::Vec<T> as alloc::vec::SpecExtend<&'a T, I>>::spec_extend (vec.rs:1942)
==741==    by 0x11895C: <alloc::vec::Vec<T>>::extend_from_slice (vec.rs:1396)
==741==    by 0x11C1A2: alloc::slice::hack::to_vec (slice.rs:168)
==741==    by 0x11C643: alloc::slice::<impl [T]>::to_vec (slice.rs:369)
==741==    by 0x118C1E: <alloc::vec::Vec<T> as core::clone::Clone>::clone (vec.rs:1676)
==741==    by 0x11AF89: md::tests::check_drops (main.rs:51)
==741==    by 0x119D39: md::__test::TESTS::{{closure}} (main.rs:36)
==741==    by 0x11935D: core::ops::function::FnOnce::call_once (function.rs:223)
==741==    by 0x11F09E: {{closure}} (lib.rs:1451)
==741==    by 0x11F09E: call_once<closure,()> (function.rs:223)
==741==    by 0x11F09E: <F as alloc::boxed::FnBox<A>>::call_box (boxed.rs:642)
==741==    by 0x17B469: __rust_maybe_catch_panic (lib.rs:105)
==741==    by 0x14044F: try<(),std::panic::AssertUnwindSafe<alloc::boxed::Box<FnBox<()>>>> (panicking.rs:289)
==741==    by 0x14044F: catch_unwind<std::panic::AssertUnwindSafe<alloc::boxed::Box<FnBox<()>>>,()> (panic.rs:392)
==741==    by 0x14044F: {{closure}} (lib.rs:1406)
==741==    by 0x14044F: std::sys_common::backtrace::__rust_begin_short_backtrace (backtrace.rs:136)

Reducing while keeping the Valgrind error (or one extremely similar) leads to

use std::{iter, mem};

fn main() {
    let a = unsafe { mem::uninitialized::<&()>() };
    let mut b = iter::once(a);
    let c = b.next();
    let _d = match c {
        Some(_) => 1,
        None => 2,
    };
}

Running this smaller reproduction in Miri in the playground leads to this error:

error[E0080]: constant evaluation error: attempted to read undefined bytes
 --> src/main.rs:7:20
  |
7 |     let _d = match c {
  |                    ^ attempted to read undefined bytes
  |
note: inside call to `main`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/rt.rs:74:34
  |
74|     lang_start_internal(&move || main().report(), argc, argv)
  |                                  ^^^^^^
note: inside call to `closure`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/rt.rs:59:75
  |
59|             ::sys_common::backtrace::__rust_begin_short_backtrace(move || main())
  |                                                                           ^^^^^^
note: inside call to `closure`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/sys_common/backtrace.rs:136:5
  |
13|     f()
  |     ^^^
note: inside call to `std::sys_common::backtrace::__rust_begin_short_backtrace::<[closure@DefId(1/1:1823 ~ std[82ff]::rt[0]::lang_start_internal[0]::{{closure}}[0]::{{closure}}[0]) 0:&dyn std::ops::Fn() -> i32 + std::marker::Sync + std::panic::RefUnwindSafe], i32>`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/rt.rs:59:13
  |
59|             ::sys_common::backtrace::__rust_begin_short_backtrace(move || main())
  |             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `closure`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/panicking.rs:310:40
  |
31|             ptr::write(&mut (*data).r, f());
  |                                        ^^^
note: inside call to `std::panicking::try::do_call::<[closure@DefId(1/1:1822 ~ std[82ff]::rt[0]::lang_start_internal[0]::{{closure}}[0]) 0:&&dyn std::ops::Fn() -> i32 + std::marker::Sync + std::panic::RefUnwindSafe], i32>`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/panicking.rs:306:5
  |
30| /     fn do_call<F: FnOnce() -> R, R>(data: *mut u8) {
30| |         unsafe {
30| |             let data = data as *mut Data<F, R>;
30| |             let f = ptr::read(&mut (*data).f);
31| |             ptr::write(&mut (*data).r, f());
31| |         }
31| |     }
  | |_____^
note: inside call to `std::panicking::try::<i32, [closure@DefId(1/1:1822 ~ std[82ff]::rt[0]::lang_start_internal[0]::{{closure}}[0]) 0:&&dyn std::ops::Fn() -> i32 + std::marker::Sync + std::panic::RefUnwindSafe]>`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/panic.rs:392:9
  |
39|         panicking::try(f)
  |         ^^^^^^^^^^^^^^^^^
note: inside call to `std::panic::catch_unwind::<[closure@DefId(1/1:1822 ~ std[82ff]::rt[0]::lang_start_internal[0]::{{closure}}[0]) 0:&&dyn std::ops::Fn() -> i32 + std::marker::Sync + std::panic::RefUnwindSafe], i32>`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/rt.rs:58:25
  |
58|           let exit_code = panic::catch_unwind(|| {
  |  _________________________^
59| |             ::sys_common::backtrace::__rust_begin_short_backtrace(move || main())
60| |         });
  | |__________^
note: inside call to `std::rt::lang_start_internal`
 --> /root/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/rt.rs:74:5
  |
74|     lang_start_internal(&move || main().report(), argc, argv)
  |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

The short version is that mem::uninitialized creates a null pointer, which is being treated as a reference. That's the undefined behavior.

In your original code, the Vec::clone is implemented by iterating over an iterator. Iterator::next returns an Option<T>, so you have an option of a reference, which causes the null pointer optimization to kick in. This counts as a None, which terminates the iteration early, resulting in your empty second vector.

It turns out that having mem::uninitialized, a piece of code that gives you C-like semantics, is a giant footgun and is frequently misused (surprise!), so you aren't alone here. The main things you should follow as replacements are:

Francis Gagné · Answer 2 · 2018-09-16T00:41:50.287

Rust 1.29.0 changed the definition of ManuallyDrop. It used to be a union (with a single member), but now it's a struct and a lang item. The role of the lang item in the compiler is to force the type to not have a destructor, even if it wraps a type that has once.

I tried copying the old definition of ManuallyDrop (which requires nightly, unless a T: Copy bound is added) and using that instead of the one from std, and it avoids the issue (at least on the Playground). I also tried dropping the second slot (slots[1]) instead of the first (slots[0]) and that also happens to work.

Although I haven't been able to reproduce the problem natively on my system (running Arch Linux x86_64), I found something interesting by using miri:

francis@francis-arch /data/git/miri master
$ MIRI_SYSROOT=~/.xargo/HOST cargo run -- /data/src/rust/so-manually-drop-1_29/src/main.rs
    Finished dev [unoptimized + debuginfo] target(s) in 0.03s                                                                                                                                                                                
     Running `target/debug/miri /data/src/rust/so-manually-drop-1_29/src/main.rs`
error[E0080]: constant evaluation error: attempted to read undefined bytes
    --> /home/francis/.rustup/toolchains/nightly-2018-09-15-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/liballoc/vec.rs:1903:32
     |
1903 |                 for element in iterator {
     |                                ^^^^^^^^ attempted to read undefined bytes
     |
note: inside call to `<std::vec::Vec<T> as std::vec::SpecExtend<T, I>><MayContainValue<CountDrop>, std::iter::Cloned<std::slice::Iter<MayContainValue<CountDrop>>>>::spec_extend`
    --> /home/francis/.rustup/toolchains/nightly-2018-09-15-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/liballoc/vec.rs:1953:9
     |
1953 |         self.spec_extend(iterator.cloned())
     |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `<std::vec::Vec<T> as std::vec::SpecExtend<&'a T, I>><MayContainValue<CountDrop>, std::slice::Iter<MayContainValue<CountDrop>>>::spec_extend`
    --> /home/francis/.rustup/toolchains/nightly-2018-09-15-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/liballoc/vec.rs:1402:9
     |
1402 |         self.spec_extend(other.iter())
     |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `<std::vec::Vec<T>><MayContainValue<CountDrop>>::extend_from_slice`
    --> /home/francis/.rustup/toolchains/nightly-2018-09-15-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/liballoc/slice.rs:168:9
     |
168  |         vector.extend_from_slice(s);
     |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `std::slice::hack::to_vec::<MayContainValue<CountDrop>>`
    --> /home/francis/.rustup/toolchains/nightly-2018-09-15-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/liballoc/slice.rs:369:9
     |
369  |         hack::to_vec(self)
     |         ^^^^^^^^^^^^^^^^^^
note: inside call to `std::slice::<impl [T]><MayContainValue<CountDrop>>::to_vec`
    --> /home/francis/.rustup/toolchains/nightly-2018-09-15-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/liballoc/vec.rs:1687:9
     |
1687 |         <[T]>::to_vec(&**self)
     |         ^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `<std::vec::Vec<T> as std::clone::Clone><MayContainValue<CountDrop>>::clone`
    --> /data/src/rust/so-manually-drop-1_29/src/main.rs:54:33
     |
54   |         assert_eq!(slots.len(), slots.clone().len());
     |                                 ^^^^^^^^^^^^^
note: inside call to `tests::check_drops`
    --> /data/src/rust/so-manually-drop-1_29/src/main.rs:33:5
     |
33   |     tests::check_drops();
     |     ^^^^^^^^^^^^^^^^^^^^

error: aborting due to previous error

For more information about this error, try `rustc --explain E0080`.

(Note: I can get the same error without using Xargo, but then miri doesn't show the source code for the stack frames in std.)

If I do this again with the original definition of ManuallyDrop, then miri doesn't report any issue. This confirms that the new definition of ManuallyDrop causes your program to have undefined behavior.

When I change std::mem::uninitialized() to std::mem::zeroed(), I can reliably reproduce the issue. When running natively, if it happens that the uninitialized memory is all zeroes, then you'll get the issue, otherwise you won't.

By calling std::mem::zeroed(), I've made the program generate null references, which are documented as undefined behavior in Rust. When the vector is cloned, an iterator is used (as shown in miri's output above). Iterator::next returns an Option<T>; that T here has a reference in it (coming from CountDrops), which causes Option's memory layout to be optimized: instead of having a discrete discriminant, it uses a null reference to represent its None value. Since I am generating null references, the iterator returns None on the first item and thus the vector ends up empty.

What's interesting is that when ManuallyDrop was defined as a union, Option's memory layout was not optimized.

println!("{}", std::mem::size_of::<Option<std::mem::ManuallyDrop<CountDrop<'static>>>>());
// prints 16 in Rust 1.28, but 8 in Rust 1.29

There is a discussion about this situation in #52898.

Vector is empty after cloning struct with uninitialized member

2 Answers2