Re: [QUESTION] Are these issues PREEMPT_RT-specific?

[Steven Rostedt <rostedt@goodmis.org>]

On Wed, 9 Sep 2026 11:05:05 +0800
Ran Hongyun <ranhongyun1@huawei.com> wrote:

> Florian,
> 
> Thank you for your reply.
> 
> Now I understand that the issues are particularly critical on PREEMPT_RT. 
> However, I’m still not entirely clear about the "performance or throughput
> drop" you mentioned for !RT systems. Since !RT kernels are
> non-preemptible, unbounded priority inversion and stalls should not occur.

That is incorrect. !RT kernels have been preemptible for decades. There are
different levels:

  SERVER - non preemptible (but no distributions use that)
  VOLUNTARY - Has preemptible points in the kernel (might_sleep()) - most distros used this
  FULL - Preempts everywhere that doesn't turn off preemption
  RT - Turns spin locks into mutexes and preempts there too

Regardless of that, unbound priority inversion can happen even in SERVER
mode. All it takes is three tasks sharing a mutex. Only RT implements
priority inheritance (for all locks but reader locks) in the kernel which
would prevent unbounded priority inversion with normal locks.

> Are there any other factors that could still cause such a performance or
> throughput drop?
> 
> In other words, if I am running a !RT system, what benefit can I actually
> expect from switching from rwlock to spinlock?

Lack of priority inheritance is true for both RT and !RT in rwlocks. Hence
that does affect both of them.

-- Steve