BPF/EBPF

Chairs ask for and encourage participation and new participants to join
in.

BPF ISA Specification Document Update

Presenter: Dave Thaler
WGLC closes April 5 - ACTION (all Participants) please review the
draft
Reminder, github link is in
https://datatracker.ietf.org/doc/draft-ietf-bpf-isa/
Dave covered changes in the draft since the last meeting (details in
presentation)

Implementation status:

• BPF conformance suite
• eBPF verifier (PREVAIL)

ACTION (Dave): Suresh suggests a policy change to differentiate
depricated vs historical - specificaion required allows flexibility and
simplicity for depricating things (like a text file in the linux kernel
vs an RFC) - clarity is required for the designated expert on acceptable
specifications

Conformance groups discussion around cumulative instructions and
deprication:

• Jagan asked and Dave noted that groups are cumulative. New group
  includes previous group.
• Groups may be depricated to remove all instructions and a new group
  added to include subsets of the depricated group.

Dave asked for clarification around pointer/address clarification from
the working group.

• Discussion at the mic

  • address and index are not accurate
  • ID may be more accurate
  • It was called address since the ID is converted to an address at
    some point by the JIT.
  • A summary could not be stated clearly

• ACTION (Dave): take this discussion to the list to continue
  discussion

ACTION (Dave): code pointer should change to address or code address -
discussion to be held on this

ACTION (Christoph): please review the draft :)

BPF callx instruction discussion

Presenter: Dave Thaler
Dave set up the problem via an example (see slides)
How/where/do we document callx used by clang and gcc in the context of
IETF?

• Alexei: when we support callx r1 - r1 will mean full 64bit addr, and
  addr length will be dependent on where addr is pointing to.
• Christoph: the compilers are doing the wrong thing by using callx.
  0-- - Dave: So should we document the opcodes as dangerous and
  should not be used?

• Alexei:

  1. • using -O0 cannot expect it to work
  3. • Discussion around when the address is replaced by a real
       address, the jit will convert it.

Implementation and performance evaluation of PDM using BPF

Presentation of performance of eBPF

• Documentation was good for helper functions
• It would be good to be able to split packets and for header
  insertion to be able to do the right thing with MTU.