PSA Meeting Minutes Jan 31, 2018

Attendees:

Postponed to later releases:

mental model: table with match key the values in the set, actions that set the next state
static vs. dynamic entries
- tables have typically dynamic entries
- in parser, most entries are static entries. Exception is the dynamic entries that are in the value set
parser sub-language is not sufficient to express dynamic entries
first proposal with an extern is the most clear code
externs do not have semantics in the language
- so using the extern as a label is outside the language
- will require changing the language
is_member does not require language changes -- the only virtue
- group votes that this is a kludge
Tom: proposes to have the select as a control plane entity that is addressable
AI: Han to bring the proposal to the language WG
AI: Andy move the PVS to rip it out and mention open issues
AI: Han to tag the feature as experimental, but keep in the compiler

sending to CPU is identical to sending to any other port and any metadata needs to be added in explicit headers
AI: Andy to clarify in the text

PR available
AI: people agree that this is a reasonable recommendation. Andy to merge the PR.

PR available
requirements not recommendations. control plane implementations rely on these operations being atomic.
implementation of Ranges -- do we accept piece-wise atomic behavior for Range?
AI: Andy to add exception to ranges in the text. Implementations can add annotations to side-step atomicity. Andy to add to PR explicit note that if a PSA implementation supports doing an apply operation on the same table multiple times per packet (there is no requirement that this be supported, though), then one or more control plane updates may occur on the table between those apply operations. It is individual control plane updates that are atomic relative to individual apply calls.

Andy gave overview of changes, which include notes on digest message relative ordering from data plane to control plane. No requirements there on PSA implementations, more warnings to control plane authors on there being no ordering guarantees here.
Han: Should Digest pack calls be restricted to IngressParser and EgressDeparser control blocks only?
AI: Andy to update PR to propose that PSA implementations only need to support Digest instances and pack calls in IngressParser and EgressParser.

Andy gave overview of the additions and their motivation
No issues raised during discussion.
It proposes specific numerical bit widths intended to be large enough to hold values of types like PortId_t, MulticastGroup_t, ClassOfService_t, etc. across all PSA implementations. Everyone interested please review those numbers to see if they look too small, or much wider than necessary.
AI: Samar to review and provide comments after meeting

Samar to create proposed addition of 1 ‘congestion experienced’ bit filled in by PRE.
No expectation that this be in PSA spec 1.0.