JfrThreadSampling.cpp (353 - 360):

static void drain_enqueued_cpu_time_requests(const JfrTicks& now, JfrThreadLocal* tl, JavaThread* jt, Thread* current, bool lock) {
  assert(tl != nullptr, "invariant");
  assert(jt != nullptr, "invariant");
  assert(current != nullptr, "invariant");
#ifdef LINUX
  tl->set_do_async_processing_of_cpu_time_jfr_requests(false);
  if (lock) {
    tl->acquire_cpu_time_jfr_dequeue_lock(); <<---

This is your synchronization point on return from native code, which is effectively a spinlock. This can cause problems when a large number of threads are being processed by the "do_async_processing" request call.

We should fix this as a bug after integration (use a proper Monitor as a synchronization point).