#RCU #pointer #lockless #rc #arc

rcu-clean

Smart pointers using RCU with Deref support

7 releases

✓ Uses Rust 2018 edition

0.1.6 Jan 15, 2019
0.1.5 Jan 1, 2019
0.1.3 Dec 31, 2018

#52 in Memory management

Download history 31/week @ 2018-12-24 42/week @ 2018-12-31 18/week @ 2019-01-07 29/week @ 2019-01-14 3/week @ 2019-01-21 2/week @ 2019-01-28 2/week @ 2019-02-04 8/week @ 2019-02-18 7/week @ 2019-02-25

36 downloads per month

Apache-2.0/MIT

23KB
452 lines

rcu-clean   Latest version Documentation Build Status

This crate provides easy to use smart-pointers with interior mutability. These smart pointers use RCU to allow simultaneous reads and updates. They implement Deref for reads, which makes them both convenient (ergonomic) and fast on reads, particularly for the Arc version that would otherwise require taking a Mutex or RwLock in order to read the pointer. The downside is that old versions of the data are only freed when you have called the clean method on each copy of the pointer.


lib.rs:

This crate provides a set (eventually) of smart pointer types that allow read access with no guards (and minimal to no overhead) and no need to call [std::borrow::Borrow]. These smart pointers each allow internal mutability (obtaining mutable references) by a Read-Copy-Update approach, so you get a mutable reference to a private copy of the data, which you can mutate at will. When the mutation is complete the pointer is atomically updated. Old references to the data may still exist, and will still be a valid reference to the old data.

Basically, these smart pointers allow internal mutability through a slow and painful process, while keeping read-only access both fast and easy (in particular, no need to call ptr.borrow() everywhere). Write access is guarded, but read access is not.

The names of the types are based on the standard smart pointer types.

  1. [BoxRcu] is an owned pointer similar to [std::box::Box]. If you like, it is actually closer to Box<RefCell<T>>, or even Box<Mutex<T>>, but without the nuisance of having to call borrow when reading.

  2. [RcRcu] is a reference counted pointer similar to [std::rc::Rc]. If you like, it is actually closer to Rc<RefCell<T>>, but without the nuisance of having to call borrow when reading.

  3. ArcRcu is planned to be a thread-safe reference counted pointer similar to [std::sync::Arc]. It is actually closer to Arc<RwLock<T>>, but without the nuisance of having to call read before reading.

Cleaning

Due to this crate's read-copy-update semantics, old copies of your data are kept until we are confident that there are no longer any references to them. Because we do not have any guards on the read references, this must be done manually. This is the cost we pay for extra convenience (and much improved read speed in the case of ArcRcu) on the read operations. You have two options for how to handle this.

One option is to simply store those extra copies until then entire smart pointer itself is freed. That is what happens if you do nothing, and for small data that is only mutated once, it's a fine option. However, for [ArcRcu] and [RcRcu] there will be a slowdown on reading until you do call clean, since an extra level of pointer redirection will be required.

The other option is to call clean() when convenient. clean takes a &mut self, so when it is called, the compiler will prove to us that there are no other references out there via this smart pointer. For BoxCell that is sufficient to prove that we can free the data. In the case of the reference counted data pointers, we keep track of a count of how many copies have been dereferenced since the last time clean was called. We could probably be more accurate with "epoch" tracking, but I don't know that the complexity will be worthwhile.

Dependencies