relib
is a framework for reloadable dynamic libraries written in Rust.
relib_demo.mp4
Currently Linux has the best support, Windows is supported partially, macOS is not supported (not tested), see support matrix.
relib
tries to be a safe (as much as possible) runtime of native, almost normal Rust programs. Programs that can be safely unloaded (without memory leaks and crashes) without closing the whole OS process.
Since it's not possible to make this completely safe: memory leaks, UB can still happen (for example, due to some unsafe call to C library), you should only use unloading for development (see live reload example). relib
can also be used without unloading, see "Usage without unloading".
See feature support matrix for what relib
offers to improve unloading of dynamic libraries in Rust. And for what not, check out limitations.
See examples.
See docs
of relib
crate.
Currently, relib
doesn't check in runtime that function signatures (arguments, return types) specified in imports and exports traits (main
and before_unload
as well) are exactly the same for host and module.
To ensure at least something about ABI relib
checks and requires that host and module are compiled with the same rustc and relib
version.
For ABI stable types, you can use abi_stable or stabby crate for it, see abi_stable
usage example.
Currently, relib
knows nothing about file descriptors or network sockets (unlike background threads) so, for example, if your program stores them in static items and does not properly close them they will leak after unloading.
note: relib provides before_unload
callback API when you need to cleanup something manually (similar to Rust Drop).
If your program deadlocks unloading won't work and you will have to kill the whole process.
Non-Copy
types (for example, a heap allocated string) are always implicitly cloned (and must implement Clone
trait) on host-module boundary when returned from an export or import. Since host and module can use different global allocators and dealloc
expects a pointer allocated exactly via this global allocator.
For example:
// a type that is common for host and module
#[repr(C)]
#[derive(Debug)]
struct MemoryChunk {
ptr: *const u8,
len: usize,
}
// allocates new chunk of memory using global allocator (will be called in generated bindings)
impl Clone for MemoryChunk { ... }
// deallocates it (will be called in generated bindings)
impl Drop for MemoryChunk { ... }
// host:
impl Imports for ModuleImportsImpl {
fn example() -> MemoryChunk {
MemoryChunk { ... }
}
}
// module:
// returned value will be implicitly cloned by using Clone trait
let chunk: MemoryChunk = unsafe { gen_imports::example() }; // gen_imports is defined by relib_interface::include_imports!()
-
Reference-counting pointers don't allocate new memory when cloned, but reuse old one, so they must not be moved through module-host boundary (Rc or Arc in std, but keep in mind that these std types don't have stable ABI).
-
It's still possible to use raw pointers to avoid cloning if you're sure of what you're doing.
Parameters are limited to Copy
types, moving non-Copy
types is not possible.
For example:
// a type that is common for host and module
#[repr(C)]
#[derive(Debug)]
struct MemoryChunk {
ptr: *const u8,
len: usize,
}
// allocates new chunk of memory using global allocator
impl Clone for MemoryChunk { ... }
// deallocates it
impl Drop for MemoryChunk { ... }
// host:
impl Imports for ModuleImportsImpl {
fn example(chunk: &MemoryChunk) {
// if owned value is needed just call .to_owned() explicitly:
let chunk = chunk.to_owned();
}
}
// module:
let chunk = MemoryChunk { ... };
unsafe { gen_imports::example(&chunk) }; // gen_imports is defined by relib_interface::include_imports!()
Because when you can you should pass values by reference to avoid cost of the cloning allocations.
It is the same reason as with return values: host and module can use different global allocators and dealloc
expects a pointer allocated exactly via this global allocator. So if moving non-Copy
types would be possible relib
would still clone parameters implicitly.
Due to the code generation this code may not compile: (RStr
is FFI-safe equivalent of &str
from abi_stable)
// shared:
pub trait Exports {
fn ret_ref(str: RStr) -> RStr;
}
// module:
impl Exports for ModuleExportsImpl {
fn ret_ref(str: RStr) -> RStr {
str.slice(..) // equal to str[..]
}
}
Will result in:
error[E0621]: explicit lifetime required in the type of `str`
--> .../generated_module_exports.rs:234:9
In order to fix it you need add explicit lifetime to trait
pub trait Exports {
fn ret_ref<'a>(str: RStr<'a>) -> RStr<'a>;
}
And suppress clippy lint if you use it:
pub trait Exports {
#[expect(clippy::needless_lifetimes)]
fn ret_ref<'a>(str: RStr<'a>) -> RStr<'a>;
}
It's not possible specify lifetime bounds for imports and exports as it's too complex to implement (there is no for<'a, 'b: 'a> fn(...)
syntax).
Example: (RStr
is FFI-safe equivalent of &str
from abi_stable)
pub trait Exports {
fn returns_b<'a, 'b: 'a>(a: RStr<'a>, b: RStr<'b>) -> RStr<'a> {
b
}
}
If you can you should use WebAssembly since it's much more memory-safe approach. But what if WASM is not enough for you for some of these reasons: (some of which may be resolved in the future)
- you need to communicate with C++ or C
- you want to use all features of Rust (for example, multi threading, panics, backtraces may not be supported really well in WASM ecosystem)
- you've already written something in normal Rust and don't want to rewrite it to work in WASM
- you don't need sandboxing/isolation
- performance
- bugs in WASM runtimes
Awesome fasterthanlime's article ❤️ https://fasterthanli.me/articles/so-you-want-to-live-reload-rust