News from PREEMPT_RT
summary
Local bottom halves get an explicit lock implemented as lock_nested_bh (merged in 6.18) which removes implicit PCPU locking in PREEMPT_RT and eliminates some softirq-induced PI inflation and a coincidental use‑after‑free regression fix. Futexes move from a single global hash to a per-task resizable bucket set using RCU and per-CPU counters to avoid 6.16 slowdowns and enable NUMA-aware futex buckets via new futex ops. Lazy preempt (since 6.13) enables a user-space pin lock prototype with ~23% DB benchmark win. Workqueue and RT docs are also being cleaned up.
AI-written summary · about
chapters (8)
- 0:00Introduction and overview
- 5:06Unlocking bottom halves on PREEMPT_RT
- 12:42Futex optimizations and fixes
- 25:36Lazy preempt in mainline
- 28:47User space pin locks
- 31:35Work queues refactoring
- 36:14Documentation updates
- 38:09Q&A
related talks
53mPREEMPT_RT - How Not to Break It.
24mPREEMPT_RT Q&A with tglx- 53mPREEMPT_RT - How Not to Break It.
50mInside the RT Patch
42mBoF: Realtime Linux
47mReal-Time Linux: What Is Next?
34mDo Existing Drivers Conflict with PREEMPT_RT? a 3-Year Survey
36mJourney of improving RT performance on Single CPU equipped PREEMPT_RT Linux