[LSF/MM TOPIC] BPF: Extending eBPF as a generic tracing engine

[Kris Van Hees <kris.van.hees@oracle.com>]

Topic: Extending eBPF as a generic tracing engine

My presentation covers a few new features to the eBPF implementation in
support of tracing tools.  The main goal is to allow tracing tools the
ability to encode actions with probes/events, beyond the context of what
the current probes/events provide.

The features to discuss are:

    - Allowing BPF-to-BPF tail-calling between BPF programs of different
      program types (ussing a context conversion/casting mechanism that
      BPF program types can provide.  This patch is being posted to the
      netdev group this weekend, along with an example on how to use it.
    - Using the context conversion mechanism to associate task info with
      probe/event context, and have a BPF program operate on that, which
      reduces the dependency (or need) for helper functions.
    - Consider the option of having BPF program sub-types (for some program
      types), which could even provide for loadable modules to implement
      some specific extensions to a more generic BPF program type.
    - Extending the context handling code and access checking to allow
      for pointer references to auxilliary structs (to an arbitrary depth).
      This makes it possible to do a lot more with contexts, such as
      having it store a pointer to a trace output buffer so you can have
      BPF instructions store to it directly.

The talk will focus on why this will help tracing a lot, and how it will
advance BPF as a tracing facility.  I want to solicit feedback on these
features so there can be common road forward to enhancing tracing based on
BPF.