SRCU latency?

A few years ago each SRCU instance was effectively independent and synchronize_srcu_expedited (which could busy wait but is nowhere as bad as its _rcu counterpart) was guaranteed to be basically instantaneous if there was no long running usage of srcu_read_lock/unlock. This was a bit counter to the "sleepable" RCU name, but in fact I always thought of it more as a "subsystem" RCU.

Since then, SRCU has been rewritten to be more scalable in the tracking of grace periods, but the current code is... a bit impenetrable. How true is the above guarantee these days?