} while (!state_.compare_exchange_strong(
    oldState, newState,
    std::memory_order_seq_cst,    // success
    std::memory_order_relaxed));  // failure

The success ordering seq_cst looks stronger than what is required. modifyState has exactly three callers in the Folly tree:

- add() at L83
- resume() at L112
- worker() at L152

In all three sites, every atomic access related to state_ is on the same address (the packed state_ itself; pause() uses fetch_or(kPausedBit, relaxed) on the same word). Cross-address publication is handled by other primitives:

So the state_ CAS appears to be purely a counter update with no cross-address publication role. Single-location coherence + RMW atomicity already linearize the three callers' updates and guarantee no lost increments/decrements.


Is the seq_cst required  for a reason I'm missing?