Re: [RFC][PATCH 6/7] context_tracking: Provide SMP ordering using RCU

[Peter Zijlstra]

On Wed, Sep 22, 2021 at 12:53:50PM -0700, Paul E. McKenney wrote:

> I wasn't saying that the patch doesn't work.  But doesn't it add an IPI?
> Or was I looking at it too early this morning?

Ah, no. The patch allows a user-bound NOHZ_FULL task to be transitioned
remotely. Unlike today, where they'll eventually poke it with a signal
to force a kernel entry, which is bad m'kay :-)

The code in question skips transitioning running tasks, seeing as you
can't tell what they're doing etc.. Howver, with context tracking on
you're supposedly able to tell they're in userspace without disturbing
them -- except you really can't today.

So if you can tell that a current running task is in userspace (hence my
patch) you can allow the task to transition without any further ado,
userspace is a safe state vs kernel text patching.