On main, attempting to apply settings to port qsfp31 fails. Little dummy test with curl that hits the dpd stub binary built for Linux:
# request body
% cat req.json
{"links":{"0":{"addrs":[],"params":{"autoneg":false,"kr":false,"lane":0,"speed":"Speed0G"}}}}
# posting to qsfp0, ..., qsfp30 is fine
% curl -X POST -H 'content-type: application/json' -H 'api-version: 12.0.0' --data @req.json 'http://[::1]:12224/port/qsfp0/settings'
{"links":{"0":{"params":{"lane":0,"speed":"Speed0G","fec":null,"autoneg":false,"kr":false,"tx_eq":null},"addrs":[]}}}
% curl -X POST -H 'content-type: application/json' -H 'api-version: 12.0.0' --data @req.json 'http://[::1]:12224/port/qsfp30/settings'
{"links":{"0":{"params":{"lane":0,"speed":"Speed0G","fec":null,"autoneg":false,"kr":false,"tx_eq":null},"addrs":[]}}}
# posting to qsfp31 fails
% curl -X POST -H 'content-type: application/json' -H 'api-version: 12.0.0' --data @req.json 'http://[::1]:12224/port/qsfp31/settings'
{
"request_id": "038b79cd-d798-41d8-8a73-90dbef511cfc",
"error_code": "invalid data: no such port",
"message": "no such port"
}
qsfp31 definitely works in the product (@sion42x has tested it on a racklette), so this appears to be a stub-only bug. I tried tracing through what's going on, and I think there's an off-by-one disagreement in Connector numbering between PortMap (which maps an API-level PortId to a Connector) and PortData (which maps a (Connector, channel) tuple to an ASIC ID):
PortMap maps PortId::Rear(0..32) to connectors QSFP(1..33) (but the order is scrambled) and then maps PortId::Qsfp(0..32) to connectors Qsfp(33..65) (not scrambled: PortId::Qsfp(0) becomes connector QSFP(33), up through PortId::Qsfp(31) mapping to connector QSFP(64)).
PortData's map of connectors to ASIC IDs contains keys for connectors QSFP(0..64); this does not have an entry for QSFP(64) itself.
When we try to map PortId::Qsfp(31) to an ASIC ID, PortMap tells us it's connector QSFP(64), and then we get the not found error when trying to look up QSFP(64) in PortData.
I'm very unfamiliar with all of this, but AFAICT, PortMap is shared between the stub and real ASIC, so the problem is presumably in the stub's PortData. If I change this line in asic::tofino_stub::ports::init():
+++ b/asic/src/tofino_stub/ports.rs
@@ -249,12 +249,13 @@ pub fn init() -> AsicResult<PortData> {
let eth_port = Some(CPU_PORT as u16);
connectors.insert(Connector::CPU, PhysPort::new(Connector::CPU)?);
- for id in 0..QSFP_PORT_COUNT {
+ for id in 1..=QSFP_PORT_COUNT {
let connector = Connector::QSFP(id);
let mut phys_port = PhysPort::new(connector)?;
phys_port.media = PortMedia::Optical;
I'm able to apply settings to qsfp31. But I don't know if shifting the connector indexing like this may cause other problems somewhere else in the stub.
On main, attempting to apply settings to port
qsfp31fails. Little dummy test with curl that hits thedpdstub binary built for Linux:qsfp31definitely works in the product (@sion42x has tested it on a racklette), so this appears to be a stub-only bug. I tried tracing through what's going on, and I think there's an off-by-one disagreement inConnectornumbering betweenPortMap(which maps an API-levelPortIdto aConnector) andPortData(which maps a(Connector, channel)tuple to an ASIC ID):PortMapmapsPortId::Rear(0..32)to connectorsQSFP(1..33)(but the order is scrambled) and then mapsPortId::Qsfp(0..32)to connectorsQsfp(33..65)(not scrambled:PortId::Qsfp(0)becomes connectorQSFP(33), up throughPortId::Qsfp(31)mapping to connectorQSFP(64)).PortData's map of connectors to ASIC IDs contains keys for connectorsQSFP(0..64); this does not have an entry forQSFP(64)itself.When we try to map
PortId::Qsfp(31)to an ASIC ID,PortMaptells us it's connectorQSFP(64), and then we get the not found error when trying to look upQSFP(64)inPortData.I'm very unfamiliar with all of this, but AFAICT,
PortMapis shared between the stub and real ASIC, so the problem is presumably in the stub'sPortData. If I change this line inasic::tofino_stub::ports::init():I'm able to apply settings to
qsfp31. But I don't know if shifting the connector indexing like this may cause other problems somewhere else in the stub.