9 unstable releases (3 breaking)

0.7.2 Oct 22, 2023
0.7.1 Oct 5, 2023
0.7.0 Jul 25, 2023
0.6.3 May 30, 2023
0.4.8 Apr 20, 2023

#165 in Unix APIs

Download history 57/week @ 2023-10-29 34/week @ 2023-11-05 194/week @ 2023-11-12 86/week @ 2023-11-19 218/week @ 2023-11-26 50/week @ 2023-12-03 50/week @ 2023-12-10 46/week @ 2023-12-17 42/week @ 2023-12-24 73/week @ 2023-12-31 49/week @ 2024-01-07 61/week @ 2024-01-14 96/week @ 2024-01-21 44/week @ 2024-01-28 109/week @ 2024-02-04 47/week @ 2024-02-11

297 downloads per month

MIT/Apache

470KB
9K SLoC

perf-event: a Rust interface to Linux performance monitoring

This crate is a wrapper around the Linux perf_event_open API. It allows you to access a wide variety of the performance monitoring counters that are available in Linux.

This crate is a fork of Jim Blandy's perf-event crate that has been updated to support more features of the perf_event_open API.

Getting Started

Add the following to your Cargo.toml

perf-event2 = "0.7"

Use Builder to create a perf counter, then use enable and disable to stop and start countinng. Call read to get your count. If you need to use multiple counters at once you can use Group. If you want to sample events from the kernel, check out Sampler.

Examples

For example, this counts the number of cycles used by the call to println!. Try adjusting the length of the vector to see the cycle count change.

use perf_event::Builder;
use perf_event::events::Hardware;

fn main() -> std::io::Result<()> {
    let mut counter = Builder::new(Hardware::INSTRUCTIONS).build()?;

    let vec = (0..=51).collect::<Vec<_>>();

    counter.enable()?;
    println!("{:?}", vec);
    counter.disable()?;

    println!("{} instructions retired", counter.read()?);

    Ok(())
}

The examples directory includes programs that count other sorts of events.

Differences between perf-event2 and perf-event

perf-event2 supports all the same features that perf-event does but it also supports the following:

  • All counter options introduced up to linux kernel 6.0.
  • Creating a Group to monitor anything other than all threads in the current process.
  • Sampled events via the kernel (e.g. gathering stack traces, etc.). Parsing the records emitted by the kernel is supported via the perf-event-data crate.
  • Direct access to the underlying perf_event_attr struct exposed by the kernel.

Migrating From perf-event

perf-event2 v0.4.8 is exactly the same as perf-event v0.4.8. You should be to just replace perf-event -> perf_event2 in your Cargo.toml and continue going with no changes. To get the new features, however, you will need to upgrade to v0.5 or later.

The main change to be aware of is that Builder::new now takes an event directly. Where you would previously do this

let counter = Builder::new()
    .kind(some_event)
    // ...
    .build()?

you will now need to do this instead

let counter = Builder::new(some_event)
    // ...
    .build()?;

Note that if you didn't call kind the default event type was Hardware::INSTRUCTIONS.

See also

  • The original perf-event crate is still perfectly workable depending on your use case.
  • Markus Stange's linux-perf-event-reader allows you to parse records emitted by perf and should allow you to parse those emitted directly by the kernel as well.
  • The perfcnt crate provides similar coverage of the perf_event_open API and appears to support parsing samples emitted by the kernel as well.
  • The not-perf project is a rewrite of perf in Rust, and has a bunch of code for dealing with the Linux perf API.

Dependencies

~515KB
~11K SLoC