PSA Meeting Minutes Dec 6, 2017

Attendees:

• AndyF, Antonin, Samar, Han, Vladimir, Calin
• online: AndyK, Chris James Coole, Joe Tardo

Agenda

• Schedule going forward

• Recirculation flag (Han, Andy) 15 min:

  • https://github.com/p4lang/p4-spec/issues/504

• Naming (Calin) 15 min:

  • PSA table property names https://github.com/p4lang/p4-spec/pull/481
  • Consistent naming conventions https://github.com/p4lang/p4-spec/issues/483

• Miscellaneous (all) 60 min:

  • Should there be a clone_port field for egress-to-egress clones? https://github.com/p4lang/p4-spec/issues/492
  • Limits on parameter values of Hash extern, and others https://github.com/p4lang/p4-spec/issues/478
  • Interaction between direct counters/meters and action profile/selector https://github.com/p4lang/p4-spec/issues/466
  • LAG resolution in PRE - https://github.com/p4lang/p4-spec/issues/508
  • State of direct meters and counters when entries are added and removed https://github.com/p4lang/p4-spec/issues/359
  • Should InternetChecksum() extern be guaranteed initialized as if clear() was called? https://github.com/p4lang/p4-spec/issues/458
  • Is drop allowed in egress? (only remaining issue is I think creating an example suggested by Samar Abdi) https://github.com/p4lang/p4-spec/issues/444
  • Using virtual functions to specify action chosen in action_selector https://github.com/p4lang/p4-spec/issues/446
  • Document effect of multiple pipelines on PSA https://github.com/p4lang/p4-spec/issues/353

Discussion

Schedule in the next few weeks

• meeting on Dec 20 is on
• meeting on Jan 3 is canceled

Recirculation

• Decision: recirculate port selected only in ingress
• AI (Han): finish the PR. Make sure egress port to get into egress deparser

Naming

• AI (Vladimir): idletime - think whether we need to bring up for discussion
• AI (Andy): size to be reserved in the language
• AI (Calin): stick with the current names for externs, add prefixes to constants, add prefixes to psa_implementation.

clone_port field for egress-to-egress

• need a mechanism to assign the port
• options:
  • metadata field
  • use a session, with a special id for CPU, which is set automatically
    • is there additional info that we want to carry with a session? QoS, priorities.
• session id may contain anything that goes into an egress port
• AI (Andy): put together a PR using session

LAG resolution in PRE

• multicast ports must be singletons
• use case scenario: l2 vlan
• fail over combinatorial explosion
  • avoid by reprogramming the selector in case of failure
• lag to mcast group resolution is delegated to P4Runtime
• Decision: no LAG in mcast group (PRE2 on whiteboard)
• AI (Andy): to write down the restriction

What is in a multicast group

• flat set of (port, replication id)
• pairs need to be unique: globally, per mcast group, port per rid.
• Decision: pair unique per mcast group
• AI (Andy): we also need to document based on the P4Runtime PR

InternetChecksum()

• AI (Andy): extern is automatically cleared when constructed and it is constructed at the beginning of each packet processing (thread)

Limits on param values for Hash

• min requirement: power of 2
• implementations are free to support other sizes

Dynamic field selection

• how do we supply inputs to hash at runtime
  • will need P4Runtime support
  • contain fields from multiple structs/headers
    • give the global list and specify what to drop off
• AI: put on the agenda for next time.