3 releases

new 0.1.2 Oct 21, 2024
0.1.1 Oct 14, 2024
0.1.0 Oct 12, 2024

#124 in Concurrency

Download history 313/week @ 2024-10-09 187/week @ 2024-10-16

500 downloads per month

GPL-3.0-or-later

175KB
3.5K SLoC

Repository Latest Version Latest Documentation

Userspace RCU

This crate provides safe Rust API to liburcu for Linux systems.

Goals

The goal is to provide traits and primitives where RCU guarantees are always respected.

  • Enforce RCU read locks when accessing RCU protected references.
  • Enforce RCU syncronization when taking ownership of a RCU reference.
  • Enforce memory cleanups in the exposed RCU data structures.

Warnings

Even though liburcu is well tested and used in many applications, this crate is still experimental. It works well in toy applications and stress tests, but I cannot guarantee it's bug free. There may be hidden race conditions or type unsoundness that may lead to undefined behaviors.

Features

This crate offers optional features. By default, all flavors are included.

  • flavor-bp: Enable liburcu-bp flavor.
  • flavor-mb: Enable liburcu-mb flavor.
  • flavor-memb: Enable liburcu-memb flavor.
  • flavor-qsbr: Enable liburcu-qsbr flavor.
  • static: Build liburcu and link statically.
    • This feature requires that liburcu build dependencies are installed.
    • Without this feature, you need to install liburcu our your system.

Types

RCU Context

Every thread that does RCU operations needs to be registered. This is enforced through the RcuContext trait. Depending on the RCU flavor, the implementator will be different. In all cases, a context can be created using RcuContext::rcu_register.

Currently, all contexts register the thread for read and defer operations. Even if your thread don't execute RCU critical sections, it will still be registered for it. It's not optimal, but it currenly simplifies things. In the future, it will be possible for threads to register only on features it needs.

RCU Guard

When accessing RCU protected data, every data structure will require a RCU read guard. It is obtained from RcuContext::rcu_read_lock. The lifetime of the references will be the same as this guard. That way, the Rust compiler guarantees that the RCU read lock is taken.

RCU Reference

When RCU protected data is removed from a container, it returns a RcuRef. This trait defines a RCU protected reference that might still have RCU readers accessing it. To get ownership of this reference, you need to wait for a RCU grace period. It is enforced by calling RcuRef::take_ownership. Dropping a RcuRef without taking ownership will still cleanup safely.

Data Structures

All data structures, except RcuBox<T>, are a wrapper around liburcu-cds API. They all supports RCU read traversal.

Type Description
RcuBox<T> RCU Box<T> with wait-free updates.
[RcuHashMap<K, V>] RCU hashmap with lock-free updates.
RcuList<T> RCU linked list with mutual exclusion on updates.
RcuQueue<T> RCU queue with lock-free updates.
RcuStack<T> RCU stack with wait-free updates.

Example

// register the current thread for RCU operations
let mut context = RcuContextMemb::rcu_register().unwrap();

// create a RCU queue (could be sent to other threads)
let queue = RcuQueue::<u32>::new();

// push/pop operations requires a RCU critical section
let guard = context.rcu_read_lock();

// push data into the queue
queue.push(Job { ... }, &guard);
queue.push(Job { ... }, &guard);
queue.push(Job { ... }, &guard);

// pop data from the queue
let job = queue.pop(&guard).unwrap();

// exit RCU critical section
drop(guard);

// wait for RCU grace period and get ownership
let mut job = job.take_ownership(&mut context);

// do something with the data
job.execute();

Performance

Althought most of the API should have low-overhead on the existing C library, we are currently linking liburcu dynamically, meaning that all the inlined functions are not used. This will have an overhead.

Unlike liburcu, we do not expose an intrusive API to store data in the data structures. This means you don't have to add a special head node in your types. Intrusive containers are more efficient. Althought it's feasible, it is currently not a goal to offer this.

Performance can be improved by enabling link-time optimization (LTO). To do so, we need to build liburcu with LTO and link it statically into the final binary.

  • Install clang compiler.
  • Enable feature flag static.
  • Enable lto = true in your build profile.
  • Execute Cargo with RUSTFLAGS="-Clinker-plugin-lto".

Dependencies

~0.3–2.7MB
~52K SLoC