11 releases (7 breaking)

0.9.0 May 24, 2024
0.8.0 Aug 29, 2023
0.7.0 Apr 21, 2022
0.6.0 Sep 18, 2021
0.1.1 Jul 26, 2017

#307 in Filesystem

Download history 917/week @ 2024-07-20 546/week @ 2024-07-27 513/week @ 2024-08-03 703/week @ 2024-08-10 372/week @ 2024-08-17 515/week @ 2024-08-24 777/week @ 2024-08-31 428/week @ 2024-09-07 347/week @ 2024-09-14 446/week @ 2024-09-21 463/week @ 2024-09-28 821/week @ 2024-10-05 761/week @ 2024-10-12 535/week @ 2024-10-19 449/week @ 2024-10-26 387/week @ 2024-11-02

2,284 downloads per month
Used in 5 crates

MIT/Apache

13KB
165 lines

mio-aio

A library for integrating file I/O with mio, using POSIX AIO and kqueue. File I/O can be seamlessly mixed with network I/O and timers in the same event loop.

Build Status

# Cargo.toml
[dependencies]
mio-aio = "0.7.0"
mio = "0.8.0"

Usage

Usage of this crate is based on the mio_aio::AioCb type, which is a wrapper around nix::AioCb. You can create one using constructors that are similar to what nix provides. Registration is the same as any mio type, except that AioCb must be individually registered. The underlying file does not get registered with mio. Once registered, you can issue the AioCb using methods that wrap the nix type: read, write, etc. After mio's poll method returns the event, call AioCb::aio_return to get the final status. At this point, the kernel has forgotten about the AioCb. There is no need to deregister it (though deregistration does not hurt).

Platforms

mio-aio works on FreeBSD. It will probably also work on DragonflyBSD. It does not work on Linux or MacOS.

Unfortunately, Linux includes a poor implementation of POSIX AIO that emulates asynchronous I/O in glibc using userland threads. Worse, epoll(2) can't deliver completion notifications for POSIX AIO. That means that it can't be supported by mio-aio. But there's still hope for Linux users! Linux has a non-standard asynchronous file I/O API called libaio. Libaio has better performance than Linux's POSIX AIO. It still can't deliver completion notification throuh epoll(2), however. What it can do is deliver completion notification through an eventfd(2). And epoll can poll an eventfd. So a Linux programmer wishing to use mio with files could theoretically write a mio-libaio crate that uses one eventfd per reactor to poll all libaio operations . Then he could implement a portability layer above mio, for example in tokio.

On MacOS AIO only supports notification using signals, not kqueue. On MacOS mio-aio could theoretically run aio-suspend in a separate thread, which would send completion notification to the main thread's reactor. Performance would suffer, however.

License

mio-aio is primarily distributed under the terms of both the MIT license and the Apache License (Version 2.0).

See LICENSE-APACHE, and LICENSE-MIT for details.

Dependencies

~2–12MB
~86K SLoC