PFC headroom and cyclic waiting

PFC operates on individual priorities of a full-duplex Ethernet link, unlike a pause that stops all ordinary traffic on that link. Several unrelated flows can share a paused priority, so one congested destination may delay another flow that would otherwise have a clear path. Classification, queue mapping, and trust settings must agree across endpoints and switches. A DSCP or VLAN-priority label is not by itself proof that all hops place packets in the intended protected buffer.

The pause takes time to affect arrivals. Packets already transmitted, the pause frame's travel, reaction time, and serialization contribute to the bytes that can still arrive. Reserve headroom for the declared link and implementation rather than equating total shared switch memory with guaranteed space for every paused port and priority. Multiple protected priorities and simultaneous congested ports compete for real resources. Threshold design must account for those allocations and their failure or reconfiguration behavior.

A queue can be congested yet still make progress as downstream service drains it. Deadlock requires a stronger failure of progress: resources wait on one another in a dependency cycle with no enabled drain or independent escape. PFC can participate in such cycles when traffic occupying one buffer depends on another paused buffer that eventually depends back on the first. A momentary pause or a high queue watermark alone does not prove a deadlock.

Analyze the channel and buffer dependency graph together with routing, priority mapping, and actual escape behavior. A directed cycle is a warning, not an unconditional proof that every traffic trace deadlocks; occupancy and blocking conditions must materialize. Watchdogs may break a stall by dropping or disabling protection, changing the losslessness promise. End-to-end congestion control and disciplined topology or buffer design reduce risks, but no single checkbox substitutes for a progress argument and failure tests.

Correlate pause duration by priority with queue occupancy, receiver progress, drops, and application completion. A device may report frames, time, or transitions under a counter containing the word pause; these are different units. Persistent upstream waiting can originate from a slow receiver or local PCIe pressure rather than a purely fabric-wide load increase. Test whether unrelated flows in the same priority stall and whether independent priorities keep progressing under the same event.

The queue lab illustrates only fixed-rate backlog. Reset and set both λ=4, μ=2, predict growth, then Step; for a separate case Reset and re-set λ=2, μ=4 before stepping. It has no pause frames, finite shared pools, dependency graph, or deadlock detector. Even λ=μ is merely constant fluid backlog, not evidence of a cyclic wait. Explain which downstream release conditions and dependencies would be needed before calling a real paused system deadlocked.