Safepoint spin time latency

A JVM service has periodic p99 pauses of 200-600ms. You assume it's GC, but the GC logs show the actual collection 'stop-the-world' portion is tiny (sub-10ms), yet the GC log line's total pause TIME is large, and -XX:+PrintSafepointStatistics shows long 'time to safepoint' (the 'spin'+'block'+'sync' phases) before collections and other safepoint operations. The service has a few very hot tight loops doing big array crunching. p50 is great; CPU is fine; no allocation spike. How do you triage and what's actually happening?