Skip to content

Latest commit

 

History

History
243 lines (186 loc) · 10.9 KB

1183-swap-out-jemalloc.md

File metadata and controls

243 lines (186 loc) · 10.9 KB
Raw

Summary

Add support to the compiler to override the default allocator, allowing a different allocator to be used by default in Rust programs. Additionally, also switch the default allocator for dynamic libraries and static libraries to using the system malloc instead of jemalloc.

Note: this RFC has been superseded by RFC 1974.

Motivation

Note that this issue was discussed quite a bit in the past, and the meat of this RFC draws from Niko's post.

Currently all Rust programs by default use jemalloc for an allocator because it is a fairly reasonable default as it is commonly much faster than the default system allocator. This is not desirable, however, when embedding Rust code into other runtimes. Using jemalloc implies that Rust will be using one allocator while the host application (e.g. Ruby, Firefox, etc) will be using a separate allocator. Having two allocators in one process generally hurts performance and is not recommended, so the Rust toolchain needs to provide a method to configure the allocator.

In addition to using an entirely separate allocator altogether, some Rust programs may want to simply instrument allocations or shim in additional functionality (such as memory tracking statistics). This is currently quite difficult to do, and would be accommodated with a custom allocation scheme.

Detailed design

The high level design can be found in this gist, but this RFC intends to expound on the idea to make it more concrete in terms of what the compiler implementation will look like. A sample implementation is available of this section.

High level design

The design of this RFC from 10,000 feet (referred to below), which was previously outlined looks like:

  1. Define a set of symbols which correspond to the APIs specified in alloc::heap. The liballoc library will call these symbols directly. Note that this means that each of the symbols take information like the size of allocations and such.
  2. Create two shim libraries which implement these allocation-related functions. Each shim is shipped with the compiler in the form of a static library. One shim will redirect to the system allocator, the other shim will bundle a jemalloc build along with Rust shims to redirect to jemalloc.
  3. Intermediate artifacts (rlibs) do not resolve this dependency, they're just left dangling.
  4. When producing a "final artifact", rustc by default links in one of two shims:
    • If we're producing a staticlib or a dylib, link the system shim.
    • If we're producing an exe and all dependencies are rlibs link the jemalloc shim.

The final link step will be optional, and one could link in any compliant allocator at that time if so desired.

New Attributes

Two new unstable attributes will be added to the compiler:

  • #![needs_allocator] indicates that a library requires the "allocation symbols" to link successfully. This attribute will be attached to liballoc and no other library should need to be tagged as such. Additionally, most crates don't need to worry about this attribute as they'll transitively link to liballoc.
  • #![allocator] indicates that a crate is an allocator crate. This is currently also used for tagging FFI functions as an "allocation function" to leverage more LLVM optimizations as well.

All crates implementing the Rust allocation API must be tagged with #![allocator] to get properly recognized and handled.

New Crates

Two new unstable crates will be added to the standard distribution:

  • alloc_system is a crate that will be tagged with #![allocator] and will redirect allocation requests to the system allocator.
  • alloc_jemalloc is another allocator crate that will bundle a static copy of jemalloc to redirect allocations to.

Both crates will be available to link to manually, but they will not be available in stable Rust to start out.

Allocation functions

Each crate tagged #![allocator] is expected to provide the full suite of allocation functions used by Rust, defined as:

extern {
    fn __rust_allocate(size: usize, align: usize) -> *mut u8;
    fn __rust_deallocate(ptr: *mut u8, old_size: usize, align: usize);
    fn __rust_reallocate(ptr: *mut u8, old_size: usize, size: usize,
                         align: usize) -> *mut u8;
    fn __rust_reallocate_inplace(ptr: *mut u8, old_size: usize, size: usize,
                                 align: usize) -> usize;
    fn __rust_usable_size(size: usize, align: usize) -> usize;
}

The exact API of all these symbols is considered unstable (hence the leading __). This otherwise currently maps to what liballoc expects today. The compiler will not currently typecheck #![allocator] crates to ensure these symbols are defined and have the correct signature.

Also note that to define the above API in a Rust crate it would look something like:

#[no_mangle]
pub extern fn __rust_allocate(size: usize, align: usize) -> *mut u8 {
    /* ... */
}

Limitations of #![allocator]

Allocator crates (those tagged with #![allocator]) are not allowed to transitively depend on a crate which is tagged with #![needs_allocator]. This would introduce a circular dependency which is difficult to link and is highly likely to otherwise just lead to infinite recursion.

The compiler will also not immediately verify that crates tagged with #![allocator] do indeed define an appropriate allocation API, and vice versa if a crate defines an allocation API the compiler will not verify that it is tagged with #![allocator]. This means that the only meaning #![allocator] has to the compiler is to signal that the default allocator should not be linked.

Default allocator specifications

Target specifications will be extended with two keys: lib_allocation_crate and exe_allocation_crate, describing the default allocator crate for these two kinds of artifacts for each target. The compiler will by default have all targets redirect to alloc_system for both scenarios, but alloc_jemalloc will be used for binaries on OSX, Bitrig, DragonFly, FreeBSD, Linux, OpenBSD, and GNU Windows. MSVC will notably not use jemalloc by default for binaries (we don't currently build jemalloc on MSVC).

Injecting an allocator

As described above, the compiler will inject an allocator if necessary into the current compilation. The compiler, however, cannot blindly do so as it can easily lead to link errors (or worse, two allocators), so it will have some heuristics for only injecting an allocator when necessary. The steps taken by the compiler for any particular compilation will be:

  • If no crate in the dependency graph is tagged with #![needs_allocator], then the compiler does not inject an allocator.
  • If only an rlib is being produced, no allocator is injected.
  • If any crate tagged with #[allocator] has been explicitly linked to (e.g. via an extern crate statement directly or transitively) then no allocator is injected.
  • If two allocators have been linked to explicitly an error is generated.
  • If only a binary is being produced, then the target's exe_allocation_crate value is injected, otherwise the lib_allocation_crate is injected.

The compiler will also record that the injected crate is injected, so later compilations know that rlibs don't actually require the injected crate at runtime (allowing it to be overridden).

Allocators in practice

Most libraries written in Rust wouldn't interact with the scheme proposed in this RFC at all as they wouldn't explicitly link with an allocator and generally are compiled as rlibs. If a Rust dynamic library is used as a dependency, then its original choice of allocator is propagated throughout the crate graph, but this rarely happens (except for the compiler itself, which will continue to use jemalloc).

Authors of crates which are embedded into other runtimes will start using the system allocator by default with no extra annotation needed. If they wish to funnel Rust allocations to the same source as the host application's allocations then a crate can be written and linked in.

Finally, providers of allocators will simply provide a crate to do so, and then applications and/or libraries can make explicit use of the allocator by depending on it as usual.

Drawbacks

A significant amount of API surface area is being added to the compiler and standard distribution as part of this RFC, but it is possible for it to all enter as #[unstable], so we can take our time stabilizing it and perhaps only stabilize a subset over time.

The limitation of an allocator crate not being able to link to the standard library (or libcollections) may be a somewhat significant hit to the ergonomics of defining an allocator, but allocators are traditionally a very niche class of library and end up defining their own data structures regardless.

Libraries on crates.io may accidentally link to an allocator and not actually use any specific API from it (other than the standard allocation symbols), forcing transitive dependants to silently use that allocator.

This RFC does not specify the ability to swap out the allocator via the command line, which is certainly possible and sometimes more convenient than modifying the source itself.

It's possible to define an allocator API (e.g. define the symbols) but then forget the #![allocator] annotation, causing the compiler to wind up linking two allocators, which may cause link errors that are difficult to debug.

Alternatives

The compiler's knowledge about allocators could be simplified quite a bit to the point where a compiler flag is used to just turn injection on/off, and then it's the responsibility of the application to define the necessary symbols if the flag is turned off. The current implementation of this RFC, however, is not seen as overly invasive and the benefits of "everything's just a crate" seems worth it for the mild amount of complexity in the compiler.

Many of the names (such as alloc_system) have a number of alternatives, and the naming of attributes and functions could perhaps follow a stronger convention.

Unresolved questions

Does this enable jemalloc to be built without a prefix on Linux? This would enable us to direct LLVM allocations to jemalloc, which would be quite nice!

Should BSD-like systems use Rust's jemalloc by default? Many of them have jemalloc as the system allocator and even the special APIs we use from jemalloc.

Updates since being accepted

Note: this RFC has been superseded by RFC 1974.