Re: [PATCH 3/3] x86/ftrace: Use text_poke()

[Peter Zijlstra]

On Fri, Oct 04, 2019 at 10:10:47AM +0200, Daniel Bristot de Oliveira wrote:
> 1) the enabling/disabling ftrace path
> 2) the int3 path - if a thread/irq is running a kernel function
> 3) the IPI - that affects all CPUs, even those that are not "hitting" trace
> code, e.g., user-space.
> 
> The first one is for sure a cold-path. The second one is a hot-path: any task
> running kernel functions will hit it. But IMHO, the hottest one is the IPIs,
> because it will run on all CPUs, e.g., even isolated CPUs that are running in
> user-space.

Well, we can fix that, just like RCU. In fact, I suspect RCU has all the
bits required to cure this.

For NOHZ_FULL CPUs you can forgo the IPI and delay it until the next kernel
entry, just like RCU.