Bpftrace for Linux 2018 | Hacker News Reader

Bpftrace for Linux 2018

brendangregg.com

268 pointsdeirdres21 comments

Hi Brendan, thanks for the update and the work on this! Coming from dtrace, the main thing missing in bpftrace when doing kernel debugging is being able to navigate structs.

You've mentioned that we do have BTF now in Linux 4.18. I've tried to find if it was leveraged in bpftrace, but it looks like it isn't yet. I see you've opened an issue in one of the bpftrace forks related to this here: https://github.com/ajor/bpftrace/issues/30.

Do you know if there's been any update on this? Support for BTF/CTF would be amazing.

That's the old repo (we should add a note to it pointing people to https://github.com/iovisor/bpftrace instead!)

Alastair added struct support for kprobes yesterday, based on the functionality in bcc (which bpftrace uses). That was the final missing piece, and why I'm posting about it now. See the last example here:

https://github.com/iovisor/bpftrace/blob/master/docs/referen...

It's the same struct support we had in bcc, using linux kernel headers. Some things are available, some are not. Just like in bcc, we sometimes need to define some of a missing struct.

BTF will mean that all structs are available, plus, bpftrace should be able to automatically figure out the arguments so you don't have to cast them. It'll be great, but in the meantime, I am able to do tons of stuff.

Fair enough. Having to include headers is one thing I found tricky with BCC, as there's sometimes non-trivial dependencies between various header files, especially for kernel modules that aren't included in the mainline kernel. That said I'm very excited to see all the progress that's been done in bpftrace, and looking forward to what's to come.

This is huge. The Linux community should be glad that Netflix hired Brendan Gregg to work on Linux performance.

Is this because they are also using Linux or do they plan to use it on FreeBSD?

Those aren’t mutually exclusive options.

For now they are, and that might hold true in the future as well.

There are some projects to bring BPF features needed (generally called eBPF) to FreeBSD, such as https://github.com/YutaroHayakawa/generic-ebpf - but there's no accepted mainline implementation as of yet.

They use both. Control plane runs on Linux, CDN on FreeBSD.

Seeing this come together makes me super happy.

Now the next time the hot rage of an audio xrun overcomes me I will have a path forward!

In general audio xruns actually is an area that could use a permanent userspace program.

Musicians need low latency audio processing which Jack2 can provide but the random xrun and the internet flimflam/mummery of how to sort it out is waaaaay beyond most of us (I have less of an excuse but yah its a vibe killer when you are trying to make music).

The goal would be "hey you had a/some xrun and its might be mostly this disk or process or..."

Maybe this needs hard ai but maybe it just needs a compass like tool...

This is a great update. I'd wondered how ply differed from bpftrace. And structs support is a huge step in achieving parity with DTrace.

DTrace has really helped me out* on macOS. To be without it is to lose a superpower, so I'm really rooting for bpftrace to bring the same tracing experience to Linux. I've been excitedly following bcc, too. It's just the right tool for HTTP metrics. But my most basic experiments so far have failed, and I don't know how to get support.

* DTrace is a real swiss army-knife. It seems intended for observing production systems, but I've found it indispensible for local development. The mysqld provider is great for tracing queries and perf. Opensnoop answers all sorts of questions. And DTrace saved a C++ project of mine: explaining a deadlock that would've otherwise spelled the end of the software. The most fun I've had is using the pid provider to hook text out of visual novels (in WiNE!).

>But my most basic experiments so far have failed, and I don't know how to get support.

I'd recommend the #iovisor irc channel on oftc. I haven't stopped by in a bit, but the times I did, I usually was able to get assistance from nice people!

Hey Brendan! Super exciting - I'm already updating some internal documentation to point people towards this :)

When are we getting a Linux performance engineering book with all the bpf/bcc/bpftrace stuff in it?

Is there any chance this tool or maybe a wrapper for it could get a higher level, less C-like syntax?

The syntax seems to be inspired by awk more than c.

1. Obviously Linux-only. How is it DTrace 2.0?

2. That’s a staggering amount of NIH, from CTF to syntax. Why?

1. There's work to bring eBPF to other operating systems.

2. bpftrace is built from the ground-up to operate with eBPF, which does much more than DTrace. I made a list in the post, and that's only happened in the last two years. I understand that if I'm narrowing eBPF to just DTrace vs bpftrace comparisons, you might question why we're doing all this, but consider the bigger picture of eBPF.

The 2 major problems (ie the Elefant in the room) compared to DTrace are missing:

1. security: eBPF still supports arrays and and is therefore inherently insecure to be used in the kernel. DTrace didn't support that feature on purpose and not because they couldn't do it. see spectre/meltdown.

2. availability: eBPF probes only work within the kernel. DTrace extends the same interface to the libc level and the user space level. you can trace not only kernel calls, also lib calls and with its high-level language support (Java, python, perl, ruby,...) into userspace. Linux-only eBPF will never be able to support that seemless integration. thanksfully we now have Oracle Linux with proper DTrace support on Linux.

1. I'd say that the other way around: eBPF has already been hardened in a post spectre/meltdown world (see lkml), and I'm not sure DTrace has had the same level of security attention.

2. That's inaccurate, see the bpftrace docs for uprobes and usdt.

Hmm, looks like you are missing some crucial details, eBPF has the same mitigations to the CPU bugs in place as the rest of the kernel (e.g. retpolines on tail calls, index masking for arrays, sanitation of speculative store bypass, etc). Also, eBPF probes do not only work in the kernel but user space as well via uprobes as one example.

Of course they ran after possible array problems post-spectre, but nevertheless it still allows precise timings on anything cache related. it's still highly insecure and needs to be turned off in production. and this is where it makes sense to use.

2nd: do you see uprobes supported anywhere? I still can only use DTrace with high level languages. and I wouldn't use this island solution anyway. DTrace works everywhere, now even on windows. eBPF is Linux-only. And should have stayed netfilter only BTW.

Reply on news.ycombinator.com