rubicon enables a dangerous form of dynamic linking in Rust through cdylib crates and carefully-enforced invariants.

Webster's Dictionary defines 'rubicon' as:

a bounding or limiting line. especially: one that when crossed, commits a person irrevocably.

In this case, I see it as the limiting line between several shared objects, within the same address space, each including their own copy of the same Rust code.

Dynamic linking concepts have different names on different platforms:

Throughout this document, macOS naming conventions are preferred.

(This section is up-to-date as of Rust 1.79 / 2024-07-18)

cargo and rustc support some form of dynamic linking, through the -C prefer-dynamic compiler flag.

This flag will:

• Link against the pre-built libstd-HASH.dylib, shipped via rustup (assuming you're not using -Z build-std)
• Try to link against libfoobar.dylib, for any crate foobar that includes dylib in its crate-type

rustc has an internal algorithm to decide which linkage to use for which dependency. That algorithm is best-effort, and it can fail.

Regardless, it assumes that rustc has knowledge of the entire dependency graph at link time.

However, one might want to split the dependency graph on purpose:

Let's call Rust's "supported" dynamic linking model "1graph".

rubicon enables (at your own risk), a different model, which we'll call "xgraph".

In the "xgraph" model, every "module" of your application — anything that might make sense to build separately, like "a bunch of tree-sitter grammars", or "a whole JavaScript runtime", is its own dependency graph, rooted at a crate with a crate-type of cdylib.

In the "xgraph" model, your application's "shared object" (Linux executables, macOS executables, etc. are just shared objects — not too different from libraries, except they have an entry point) does not have any references to its modules — by the time main() is executed, none of the modules are loaded yet.

Instead, modules are loaded explicitly through a crate like libloading, which under the hood, uses whatever facilities the platform's dynamic linker-loader exposes. This lets you choose which modules to load and when.

The "xgraph" model is dangerous — we must use discipline to get it to work at all.

In particular, we'll maintain the following invariants:

• A. Modules are NEVER UNLOADED, only loaded.
• B. The EXACT SAME RUSTC VERSION is used to build the app and all modules
• C. The EXACT SAME CARGO FEATURES are enabled for crates that both the app and some modules depend on.

Unloading modules ("A") would break a significant assumption in all Rust programs: that 'static lasts for the entirety of the program's execution. When unloading a module, we can make something 'static disappear.

Although nobody can stop you from unloading modules, what you're writing at this point is no longer safe Rust.

Mixing rustc versions ("B") might result in differences in struct layouts, for example. For a struct like:

struct Blah {
    a: u64,
    b: u32,
}

...there's no guarantee which field will be first, if there will be padding, what order the fields will be in. We pray that struct layouts match across the same compiler version, but even that might not be guaranteed? (citation needed)

Mixing cargo feature sets ("C") might, again, result in differences in struct layouts:

struct Blah {
    #[cfg(feature = "foo")]
    a: u64,
    b: u32
}

// if the app has `foo` enabled, and we pass a &Blah` to
// a module that doesn't have `foo` enabled, then the
// layout won't match.

Or function signatures. Or the (duplicate) code being run at any time.

In the 1graph model, rustc is able to see the entire dependency graph — as a result, it's able to avoid duplicates of a dependency altogether: if the app and some of its modules depend on tokio, then there'll be a single libtokio.dylib that they all depend on — no duplication whatsoever.

In the xgraph model, we're unable to achieve that. By design, the app and all of its modules are built and linked in complete isolation. As long as they agree on a thin FFI (Foreign Function Interface) boundary, which might be provided by a "common" crate everyone depends on, they can be built.

It is possible for the app and its modules to link dynamically against tokio: there will be, for each target (the app is a target, each module is a target), a libtokio.dylib file.

However, that file will not have the same contents for each target, because tokio exposes generic functions.

This code:

tokio::spawn(async move {
    println!("Hello, world!");
});

Will cause the spawn function to be monomorphized, turning from this:

pub fn spawn<F>(future: F) -> JoinHandle<F::Output> ⓘ
where
    F: Future + Send + 'static,
    F::Output: Send + 'static,

Into something like this (the mangling here is not realistic):

pub fn spawn__OpaqueType__FOO(future: OpaqueType__FOO) -> JoinHandle<()> ⓘ

If in another module, we have that code:

let jh = tokio::spawn(async move {
    // make yourself wanted
    tokio::time::sleep(std::time::Duration::from_secs(1)).await;
    println!("Oh hey, you're early!");
    42
});
let answer = jh.await.unwrap();

Then it will cause another monomorphization of tokio's spawn function, which might look something like this:

pub fn spawn__OpaqueType__BAR(future: OpaqueType__BAR) -> JoinHandle<i32> ⓘ

And now, you'll have:

bin/
  app/
    executable
    libtokio.dylib
      (exports spawn__OpaqueType__FOO)
  mod_a/
    libmod_a.dylib
    libtokio.dylib
      (export spawn__OpaqueType__BAR)

At this point, executable refers to its own libtokio.dylib (by absolute path), and libmod_a.dylib, to its own, separate, libtokio.dylib.

Even if you were to edit the DT_NEEDED / LC_LOAD_DYLIB information to have the modules point to executable's version of the dynamic libraries, you would find yourself with a "missing symbol" error at runtime!

None of the libtokio.dylib files you have contain all the symbols required.

To make a libtokio.dylib file that contains ALL THE SYMBOLS required, you would need rustc to be aware of the whole dependency graph: hence, you'd be back to the 1graph model.

Hence, when using the xgraph, we accept the reality that code from dependencies will be duplicated.

That first column corresponds to all functions, types, etc. that are not generic, or that are instantiated the exact same way in each independent depgraph.

There will be a copy of each of these in the application executable AND in each libmod_etc.dylib file. That's unavoidable for now.

Now that we've made our peace with the fact there will be code duplication, and that, as long as that code EXACTLY MATCHES across different copies, it's okay, we need to address the fact that duplicating globals is never okay.

In particular, by globals, we mean:

• thread-locals (declared via the std::thread_local! macro)
• process-locals (more commonly called "statics", declared via the static keyword)

static sample_process_local: AtomicU64 = AtomicU64::new(0);

std::thread_local! {
    static sample_thread_local: u64 = 42;
}

fn blah() {
    let sample_local = 42;
}

Take tracing, for example: it lets you emit "events" that a "subscriber" can process. It's used for structured logging: the event could be of level INFO and include information about some HTTP request, for example.

tracing allows registering a "global" dispatcher, through tracing::dispatcher::set_global_default. This sets a process-global:

static mut GLOBAL_DISPATCH: Dispatch = Dispatch {
    subscriber: Kind::Global(&NO_SUBSCRIBER),
};

The problem is that, since all targets (the app, all its modules) have their own copy of tracing, they also have their own GLOBAL_DISPATCH process-local.

It doesn't matter to mod_a if we've registered a global dispatcher from the app: according to mod_a's copy of GLOBAL_DISPATH — there's no subscriber!

There's only one fix for this: everyone must share the same GLOBAL_DISPATCH: it must be exported from app, and imported from all its modules.

In a perfect world, there'd be a rustc flag like -C globals-linkage=[import,export]: we'd set it to export for our app, so that it would declare those as exported symbols, the kind you can look up with dlsym, and that dynamic libraries you load later can use, because they're part of the set of symbols the dynamic linker-loader searches.

There are, however, two roadblocks we must hop.

The first is that dynamic symbols are not exported for executables. Luckily, there's a linker flag for that: -rdynamic (also known as --export-dynamic).

The second is that there is no such rustc flag at all.

Export a static is easy enough. Instead of:

static MERCHANDISE: u64 = 42;

We can do:

#[used]
static MERCHANDISE: u64 = 42;

And we'll get a mangled symbol:

❯ cargo build --quiet
❯ nm -gp ./target/debug/librubicon.dylib | grep MERCHANDISE
00000000000099f0 S __ZN7rubicon11MERCHANDISE17h03e39e78778de1fdE

The #[no_mangle] attribute implies #[used], and also disables name mangling:

#[no_mangle]
static MERCHANDISE: u64 = 42;

❯ cargo build --quiet
❯ nm -gp ./target/debug/librubicon.dylib | grep MERCHANDISE
00000000000099f0 S _MERCHANDISE

(Just ignore the _ prefix — linkers are cute like that.)

In fact, we can even specify our own export name if we want:

#[export_name = "STILL_MERCHANDISE"]
static PINK_UNICORN: u64 = 42;

❯ cargo build --quiet
❯ nm -gp ./target/debug/librubicon.dylib | grep MERCHANDISE
00000000000099f0 S _STILL_MERCHANDISE

However, when importing, there is no way to opt into mangling.

We can either import it as-is, without mangling:

extern "C" {
    static MERCHANDISE: u64;
}

// (only here to force the linker to import MERCHANDISE)
#[used]
static MERCHANDISE_ADDR: &u64 = unsafe { &MERCHANDISE };

# needed to avoid link errors: `MERCHANDISE` is not present at link time, it's
# only expected to be present at load time.
❯ export RUSTFLAGS="-Clink-arg=-undefined -Clink-arg=dynamic_lookup"

❯ cargo build --quiet
❯ nm -gp ./target/debug/librubicon.dylib | grep MERCHANDISE
00000000000e0210 S __ZN7rubicon16MERCHANDISE_ADDR17h2755f244419dcf79E
                 U _MERCHANDISE

Or we can specify a link_name explicitly:

extern "C" {
    #[link_name = "STILL_MERCHANDISE"]
    static MERCHANDISE: u64;
}

// (only here to force the linker to import MERCHANDISE)
#[used]
static MERCHANDISE_ADDR: &u64 = unsafe { &MERCHANDISE };

00000000000e0210 S __ZN7rubicon16MERCHANDISE_ADDR17h2755f244419dcf79E
                 U _STILL_MERCHANDISE

All these alternatives, quite frankly, suck.

If we opt into mangling, we're safe from name collisions, but we cannot import that symbol again (I'm not counting "manually copying and pasting the mangled name into Rust source code").

If we opt out of mangling, two crates that export CURRENT_STATE will clash.

In practice, we have no choice but to opt out of mangling, and make sure there's no collision between the unmangled globals of various crates in the dependency graph — which means, that's right, we're back to manually prefixing things, like in C.

We've just covered process-locals. The situation for thread-locals is much the same, except we have to do some more trickery because the internals of LocalKey are, well, internal, and cannot be accessed from stable Rust.

Getting all these just right is tricky — that's why rubicon ships macros, which are meant to be used by any crate that has global state, such as tokio, tracing, parking_lot, etc.

This is not as good as a rustc flag, but it's all we got right now. In time, the hope is that rubicon will disappear.

If you maintain a crate that has global state, you might want to make it rubicon-compatible.

You'll need to add a non-optional dependency to it:

cargo add rubicon

Without any features added, it has zero dependencies.

When rubicon/import-globals or rubicon/export-globals is enabled, it will pull in paste, which is a proc-macro: I'm not fond of the idea, but I've explored alternatives and token pasting is the best I can do right now.

Enabling both features at the same time will yield a compile error, and enabling neither will act as if your crate wasn't using rubicon's macros at all (so most users of your crate should be completely unaffected).

Users are in charge of adding their own dependency to rubicon and enabling either feature — this avoids feature proliferation. Provided that there's only one copy of rubicon in the entire depgraph (e.g. everyone is on 3.x), then the scheme works.

rubicon::thread_local! is a drop-in replacement for std::thread_local!.

Before:

std::thread_local! {
    static BUF: RefCell<String> = RefCell::new(String::new());
}

After:

rubicon::thread_local! {
    static BUF: RefCell<String> = RefCell::new(String::new());
}

However, keep in mind that, whenever import/export is enabled, mangling will be disabled for your static. Thus, it might be a good idea to preemptively prefix it:

rubicon::thread_local! {
    static MY_CRATE_BUF: RefCell<String> = RefCell::new(String::new());
}

Before:

static DISPATCHERS: Dispatchers = Dispatchers::new();
static CALLSITES: Callsites = Callsites {
    list_head: AtomicPtr::new(ptr::null_mut()),
    has_locked_callsites: AtomicBool::new(false),
};
static DISPATCHERS: Dispatchers = Dispatchers::new();
static LOCKED_CALLSITES: Lazy<Mutex<Vec<&'static dyn Callsite>>> = Lazy::new(Default::default);

After:

rubicon::process_local! {
    static DISPATCHERS: Dispatchers = Dispatchers::new();
    static CALLSITES: Callsites = Callsites {
        list_head: AtomicPtr::new(ptr::null_mut()),
        has_locked_callsites: AtomicBool::new(false),
    };
    static DISPATCHERS: Dispatchers = Dispatchers::new();
    static LOCKED_CALLSITES: Lazy<Mutex<Vec<&'static dyn Callsite>>> = Lazy::new(Default::default);
}

Both thread_local! and process_local! support multiple definitions.

In addition, process_local! supports static mut, should you really need it (looking at you tracing-core).

Sometimes thread-locals and statics hide in the darndest of places.

For example, tokio depends on parking_lot which has global state (did you know?)

/// Holds the pointer to the currently active `HashTable`.
///
/// # Safety
///
/// Except for the initial value of null, it must always point to a valid `HashTable` instance.
/// Any `HashTable` this global static has ever pointed to must never be freed.
static PARKING_LOT_HASHTABLE: AtomicPtr<HashTable> = AtomicPtr::new(ptr::null_mut());

Assuming all your dependencies are rubicon-compatible, you can implement the xgraph model!

In terms of crates, you'll need

• bin, a bin crate, depends on exports, and libloading
• exports, a lib crate, crate-type=["dylib"] (that's just "dye lib")
  • depends on all your rubicon-compatible dependencies
  • depends on rubicon with feature export-globals enabled
• mod_a, a lib crate, crate-type=["cdylib"] (that's "see dye lib")
  • depends on rubicon with feature import-globals enabled
• mod_b, like mod_a
• mod_c, like mod_a
• etc.

The exports crate is needed to bring all globals in the address space in a way that the dynamic linker can understand.

Technically -rdynamic should help there, but I couldn't get it to work.

That's about it. Don't forget the invariants!

• A. Modules are NEVER UNLOADED, only loaded.
• B. The EXACT SAME RUSTC VERSION is used to build the app and all modules
• C. The EXACT SAME CARGO FEATURES are enabled for crates that both the app and some modules depend on.

You can find a full example in test-crates/ in the rubicon repository.