PSA Meeting Minutes Oct 18, 2017

Attendees:

AIs from last time (30 min):
- Action profiles (Samar) - https://github.com/p4lang/p4-spec/pull/410
- Checksum (Nate) - https://github.com/p4lang/p4-spec/pull/421 - merged without init method
- Digest (Han) - https://github.com/p4lang/p4-spec/pull/419 -- need example(s)
- Clone, resubmit, recirculate (Han) -- need more examplesd
Make queue lengths readable (15 min) https://github.com/p4lang/p4-spec/issues/335
Default metadata values for all packet types (20 min) -- share spreadsheet?
Bridged metadata (20 min)

use of selectors as virtual functions require language support.
action selector
- Vlad: all members in an action group can only have the same action? different action data, but same action
- Antonin: in P4_16 the actions are properties of the table implementation
P4runtime does not support changing the hash algo, or the fields used in the selector
- tables sharing the same action selector implies that they use the match keys for selector (and the same order) AI: Samar to add this constraint
- definition of the hash algorithm is part of the selector definition
Action selectors will be revisited to make them easier to use with potential language chages
AI: Samar to update the text to restrict to one action for all members in the group
AI: Vlad review the current example and propose more examples
AI: Han to add a virtual function example

example: mirror on drop
- uses metadata: clone,clone_portand drop. Separate reason in user metadata
- programmer can identify cloned packets and she is in charge of parsing
- header tag may not be reordered, but the rest of the header may be optimized -- if properly annotated
issues with the transitions in the example, because it was not parsing the rest of the packet AI: Han to fix
Vlad: clone_in ? Can we isolate the parsing for cloning from the rest of the packet parsing
AndyF/Tom: fields list for clone and autogenerate packing and parsing (just as for bridged metadata)
- issue: naming of these fields in the different controls -- use global names
AndyK: weird to have headers that implicitly have wire order get reordered by the compiler
- serializabe data should preserve the order
- non-serializable -- compiler can do whatever optimizations it can
- enforce the distinction and provide guidelines to the compiler on what's allowed
A particular implementation will have to mark where the clone_in is in the packet (start, len)
Tom: each pipeline will have to have an out param that is a union of all clone/recirc types, and a corresponding in param that has these members valid only when the packet has clones.
AI: Han to work on a proposal

Egress-only has access to queues after the processing
If available in ingress, useful for lag
units: packets, bytes, packets and bytes?
can we support total bytes in the queue?
- that should be the goal -- sync up with the INT WG
AI: review the INT spec and decide whether we have a matching specification or we just punt to INT