[patch RFC 00/15] mm/highmem: Provide a preemptible variant of kmap_atomic & friends

[Peter Zijlstra]

On Thu, Sep 24, 2020 at 08:32:41AM -0400, Steven Rostedt wrote:
> Anyway, instead of blocking. What about having a counter of number of
> migrate disabled tasks per cpu, and when taking a migrate_disable(), and there's
> already another task with migrate_disabled() set, and the current task has
> an affinity greater than 1, it tries to migrate to another CPU?

That doesn't solve the problem. On wakeup we should already prefer an
idle CPU over one running a (RT) task, but you can always wake more
tasks than there's CPUs around and you'll _have_ to stack at some point.

The trick is how to unstack them correctly. We need to detect when a
migrate_disable() task _should_ start running again, and migrate away
whoever is in the way at that point.

It turns out, that getting selected for pull-balance is exactly that
condition, and clearly a migrate_disable() task cannot be pulled, but we
can use that signal to try and pull away the running task that's in the
way.