The basic JFR data structure used to keep memory buffers are currently doubly-linked and are accessed mutually exclusive on writes. This can lead to stalls in certain situations where the lists are heavily traversed for space. This is especially the case when writing larger-than thread local sized events, which are rare in practice but there are currently some tests that provoke this situation. This can cause the VM to have to wait a long time to reach a safepoint.

A mostly concurrent, although not necessarily non-blocking, implementation could avoid this bottleneck and improve overall levels of concurrency.