This is true in principle and it is good calling it out, but in practice I've never seen a mutex-based data structure beat an equivalent lock-free data structure, even at low contention, unless the latter is extremely contrived.
A mutex transaction generally requires 2 fences, one on lock and one on unlock. The one on unlock would not be strictly necessary in principle (on x86 archs the implicit acquire-release semantics would be enough) but you generally do a CAS anyway to atomically check whether there are any waiters that need a wake-up, which implies a fence.
Good lock-free data structures OTOH require just one CAS (or other fenced RMW) on the shared state.
Besides, at large scale, no matter how small your critical section is, it will be preempted every once in a while, and when you care about tail latency that is visible. Lock-free data structures have more predictable latency characteristics (even better if wait-free).
A mutex transaction generally requires 2 fences, one on lock and one on unlock. The one on unlock would not be strictly necessary in principle (on x86 archs the implicit acquire-release semantics would be enough) but you generally do a CAS anyway to atomically check whether there are any waiters that need a wake-up, which implies a fence.
Good lock-free data structures OTOH require just one CAS (or other fenced RMW) on the shared state.
Besides, at large scale, no matter how small your critical section is, it will be preempted every once in a while, and when you care about tail latency that is visible. Lock-free data structures have more predictable latency characteristics (even better if wait-free).