Expand description

Direct, unsafe bindings for Linux perf_event_open and friends.

Linux’s perf_event_open system call provides access to the processor’s performance measurement counters (things like instructions retired, cache misses, and so on), kernel counters (context switches, page faults), and many other sources of performance information.

You can’t get the perf_event_open function from the libc crate, as you would any other system call. The Linux standard C library does not provide a binding for this function or its associated types and constants.

Rust analogs to the C types and constants from <linux/perf_event.h> and <linux/hw_breakpoint.h>, generated with bindgen, are available in the bindings module.

There are several ioctls for use with perf_event_open file descriptors; see the ioctls module for those.

For a safe and convenient interface to this functionality, see the perf_event crate.

Using the raw API

As the kernel interface evolves, the struct and union types from the bindings module may acquire new fields. To ensure that your code will continue to compile against newer versions of this crate, you should construct values of these types by calling their Default implementations, which return zero-filled values, and then assigning to the fields you care about. For example:

use perf_event_open_sys as sys;

// Construct a zero-filled `perf_event_attr`.
let mut attrs = sys::bindings::perf_event_attr::default();

// Populate the fields we need.
attrs.size = std::mem::size_of::<sys::bindings::perf_event_attr>() as u32;
attrs.type_ = sys::bindings::PERF_TYPE_HARDWARE;
attrs.config = sys::bindings::PERF_COUNT_HW_INSTRUCTIONS as u64;
attrs.set_disabled(1);
attrs.set_exclude_kernel(1);
attrs.set_exclude_hv(1);

// Make the system call.
let result = unsafe {
    sys::perf_event_open(&mut attrs, 0, -1, -1, 0)
};

if result < 0 {
    // ... handle error
}

// ... use `result` as a raw file descriptor

It is not necessary to adjust size to what the running kernel expects: older kernels can accept newer perf_event_attr structs, and vice versa. As long as the size field was properly initialized, an error result of E2BIG indicates that the attrs structure has requested behavior the kernel is too old to support.

When E2BIG is returned, the kernel writes the size it expected back to the size field of the attrs struct. Again, if you want to retry the call, it is not necessary to adjust the size you pass to match what the kernel passed back. The size from the kernel just indicates which version of the API the kernel supports; see the documentation for the PERF_EVENT_ATTR_SIZE_VER... constants for details.

Kernel versions

The bindings in this crate are generated from the Linux kernel headers packaged by Fedora as:

  • x86_64: kernel-headers-5.19.4-200.fc36.x86_64 (PERF_ATTR_SIZE_VER7)
  • aarch64: kernel-headers-5.18.4-201.fc36.aarch64 (PERF_ATTR_SIZE_VER7)

As explained above, bugs aside, it is not necessary to use the version of these structures that matches the kernel you want to run under, so it should always be acceptable to use the latest version of this crate, even if you want to support older kernels.

This crate’s README.md file includes instructions on regenerating the bindings from newer kernel headers. However, this can be a breaking change for users that have not followed the advice above, so regeneration should cause a major version increment.

If you need features that are available only in a more recent version of the types than this crate provides, please file an issue.

Linux API Backward/Forward Compatibility Strategy

(This is more detail than necessary if you just want to use the crate. I want to write this down somewhere so that I have something to refer to when I forget the details.)

It is an important principle of Linux kernel development that new versions of the kernel should not break userspace. If upgrading your kernel breaks a user program, then that’s a bug in the kernel. (This refers to the run-time interface. I don’t know what the stability rules are for the kernel headers: can new headers cause old code to fail to compile? Anyway, run time is our concern here.)

But when you have an open-ended, complex system call like perf_event_open, it’s really important for the interface to be able to evolve. Certainly, old programs must run properly on new kernels, but ideally, it should work the other way, too: a program built against a newer version of the kernel headers should run on an older kernel, as long as it only requests features the old kernel actually supports. That is, simply compiling against newer headers should not be disqualifying - only using those new headers to request new features the running kernel can’t provide should cause an error.

Consider the specific case of passing a struct like perf_event_attr to a system call like perf_event_open. In general, there are two versions of the struct in play: the version the user program was compiled against, and the version the running kernel was compiled against. How can we let old programs call perf_event_open on new kernels, and vice versa?

Linux has a neat strategy for making this work. There are four rules:

  • Every system call that passes a struct to the kernel includes some indication of how large userspace thinks that struct is. For perf_event_open, it’s the size field of the perf_event_attr struct. For ioctls that pass a struct, it’s a bitfield of the request value.

  • Fields are never deleted from structs. At most, newer kernel headers may rename them to __reserved_foo or something like that, but once a field has been placed, its layout in the struct never changes.

  • New fields are added to the end of structs.

  • New fields’ semantics are chosen such that filling them with zeros preserves the old behavior. That is, turning an old struct into a new struct by extending it with zero bytes should always give you a new struct with the same meaning as the old struct.

Then, the kernel’s strategy for receiving structs from userspace is as follows (according to the comments for copy_struct_from_user in the kernel source include/linux/uaccess.h):

  • If the kernel’s struct is larger than the one passed from userspace, then that means the kernel is newer than the userspace program. The kernel copies the userspace data into the initial bytes of its own struct, and zeros the remaining bytes. Since zeroed fields have no effect, the resulting struct properly reflects the user’s intent.

  • If the kernel’s struct is smaller than the one passed from userspace, then that means that a userspace program compiled against newer kernel headers is running on an older kernel. The kernel checks that the excess bytes in the userspace struct are all zero; if they are not, the system call returns E2BIG, indicating that userspace has requested a feature the kernel doesn’t support. If they are all zero, then the kernel initializes its own struct with the bytes from the start of the userspace struct, and drops the rest. Since the dropped bytes were all zero, they did not affect the requested behavior, and the resulting struct reflects the user’s intent.

  • In either case, the kernel verifies that any __reserved_foo fields in its own version of the struct are zero.

This covers both the old-on-new and new-on-old cases, and returns an error only when the call requests functionality the kernel doesn’t support.

You can find one example of using perf_event_open in the perf_event crate, which provides a safe interface to a subset of perf_event_open’s functionality.

Using perf types on other platforms

Although the functions in this crate are only available on Linux and Android, the crate itself should build on Windows and Mac as well. On those platforms, only the bindings module is available: the types are useful to code that needs to parse perf data produced on Linux.

Modules

Types and constants used with perf_event_open.
Ioctls for use with perf_event_open file descriptors.

Functions

The perf_event_open system call.