fn start_control_channel_capacity(cfg: &SessionManagerConfig) -> usizeExpand description
Capacity of the control Start protocol lane (every variant but StartSession).
Fed by SessionManager::dispatch_message with try_send, and an overflow drops the
message. For most Start messages that is recoverable, but a dropped SsaCommit is not: there is
no NACK or retransmission, so the coefficient cell stays empty forever, the commitment never
completes, and the cycle dies on the supervisor’s max_deposit_wait deadline. It therefore keeps
the whole PIX commitment term.
PIX load is the entire commitment set of an SSA cycle, chunked into packet-sized messages.
Batching multiplies it: an Exit may ask for up to
ssas_per_request SSAs at once, so the per-cycle
term is scaled by the configured maximum even when dynamic admission selects less.
The per-cycle burst is bounded by two independent limits, and the capacity takes the smaller:
quota_range.end() / PIX_QUOTA_BYTES_PER_SHAREispolys × (threshold + surplus), an over-estimate by that whole second factor, since a cycle carries one constant term per polynomial and nothing else. The quota alone does not reveal how the product splits, so the over-estimate cannot be undone from it.MAX_POLYS_PER_SSAis the number of polynomialsSessionManager::check_pix_paramswill accept, whatever the quota says. It therefore bounds the commitments a cycle can ever deliver.
Clamping to the second matters because this capacity is reserved: an operator-settable
quota_range feeding an unclamped derivation is an allocation with no upper bound — at
quota_range.end() = 1e13 it asks for 77 GB. At the default dimensions the commitments reach
the MAX_POLYS_PER_SSA ceiling, so the capacity term is the resulting burst plus
START_PROTOCOL_CHANNEL_RESERVE: the margin over the burst is the reserve itself.
The batch factor is bounded by MAX_SSA_BATCH_SIZE for the same allocation reason, and is
clamped here rather than taken on trust from the config, so callers that build a
SessionManagerConfig without going through SessionManager::new cannot inflate it.
Deliberately not sized by maximum_sessions: that quantity sizes the session lane, where a
StartSession flood’s own drops are taken, and this lane is the one a flood must not touch.