[RFC] DTrace based on eBPF and other tracing facilities

[RFC] DTrace based on eBPF and other tracing facilities

[Kris Van Hees]

A lot of work has been done on various aspects of the tracing infrastructure
in Linux in the past years and with the further development of BPF a quite
powerful execution engine has become available as well.

One of the difficulties we have experienced in furthering DTrace on Linux is
that we have to duplicate functionality already available in the kernel
because that functionality is not easy to make use of.

In the past year or so we have been working towards changing that.  There is
no point in having multiple projects reinvent the same wheel a couple of times
over, especially when there are ways where everyone can benefit from actually
cooperating.  Our current (lofty) goal is to rework the DTrace implementation 
that we currently have to make it more modular and less self-sufficient.  We
are envisioning a future for DTrace where we can leverage its strengths in the
areas where it matters most (e.g. very efficient handling of large amounts of
kernel probes, well defined and understood D language, user familiarity with
existing providers, ...) while building on the existing tracing infrastructure
in Linux.  That also means that we can contribute better to existing pieces
in the infrastructure and work together with other tracing projects to continue
to improve tracing on Linux.

Ideally we would like to see an infrastructure where any tracers can attach
actions to any kind of probe source, and have data generated according to the
actions the tracer associated with the probe source when a specific probe
fires.  The execution of those actions would be done using BPF.

We believe that this proposal would be a benefit to all because it allows us
to pool resources in the areas that really need it.  E.g. if we all depend on
BPF as execution engine we invariably work together to make it as solid as can
be.

Obviously we cannot do this work on our own, and we cannot do it behind closed
doors.  We've created a github repository for the kernel with DTrace added in
at:

  https://github.com/ezannoni/dtrace-linux-kernel/tree/master

We also have a branch there with the most recent BPF-based work:

  https://github.com/ezannoni/dtrace-linux-kernel/tree/nix/bpf/4.19/helpers

Since most (if not all) tracing tools have similar requirements for what may
need to be done when a probe fires, we really want to join forces.

	Cheers,
	Kris

Re: [RFC] DTrace based on eBPF and other tracing facilities

[Aleksa Sarai]

[-- Attachment #1: Type: text/plain, Size: 3422 bytes --]

On 2018-11-15, Kris Van Hees <kris.van.hees@oracle.com> wrote:
> A lot of work has been done on various aspects of the tracing infrastructure
> in Linux in the past years and with the further development of BPF a quite
> powerful execution engine has become available as well.
> 
> One of the difficulties we have experienced in furthering DTrace on Linux is
> that we have to duplicate functionality already available in the kernel
> because that functionality is not easy to make use of.
> 
> In the past year or so we have been working towards changing that.  There is
> no point in having multiple projects reinvent the same wheel a couple of times
> over, especially when there are ways where everyone can benefit from actually
> cooperating.  Our current (lofty) goal is to rework the DTrace implementation 
> that we currently have to make it more modular and less self-sufficient.  We
> are envisioning a future for DTrace where we can leverage its strengths in the
> areas where it matters most (e.g. very efficient handling of large amounts of
> kernel probes, well defined and understood D language, user familiarity with
> existing providers, ...) while building on the existing tracing infrastructure
> in Linux.  That also means that we can contribute better to existing pieces
> in the infrastructure and work together with other tracing projects to continue
> to improve tracing on Linux.
> 
> Ideally we would like to see an infrastructure where any tracers can attach
> actions to any kind of probe source, and have data generated according to the
> actions the tracer associated with the probe source when a specific probe
> fires.  The execution of those actions would be done using BPF.
> 
> We believe that this proposal would be a benefit to all because it allows us
> to pool resources in the areas that really need it.  E.g. if we all depend on
> BPF as execution engine we invariably work together to make it as solid as can
> be.
> 
> Obviously we cannot do this work on our own, and we cannot do it behind closed
> doors.  We've created a github repository for the kernel with DTrace added in
> at:
> 
>   https://github.com/ezannoni/dtrace-linux-kernel/tree/master
> 
> We also have a branch there with the most recent BPF-based work:
> 
>   https://github.com/ezannoni/dtrace-linux-kernel/tree/nix/bpf/4.19/helpers
> 
> Since most (if not all) tracing tools have similar requirements for what may
> need to be done when a probe fires, we really want to join forces.

Have you taken a look at bpftrace[1]? It uses LLVM to compile to BPF
bytecode and has a DTrace-like syntax. Yes, it's a from-scratch
reimplementation, but as a user of tracing I don't really mind either
way (I think most people like DTrace because of how seamlessly all
aspects of probing work -- but I believe that most of this is a result
of the DTrace kernel infrastructure more than the user-space tooling).

I'm not entirely sure how USDT interacts with the current Linux tracing
tools, but it would be nice if USDT worked on Linux the same way it
worked on Solaris (though I will admit I haven't played with it much on
Linux, I'm sure the other tracing folks will be able to comment on this
far more than myself).

[1]: https://github.com/iovisor/bpftrace

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]