Re: [PATCH] mm: reduce spinlock contention in release_pages()

[Matthew Wilcox]

On Mon, Nov 29, 2021 at 09:39:16AM +0100, Michal Hocko wrote:
> On Fri 26-11-21 16:26:23, Hao Lee wrote:
> [...]
> > I will try Matthew's idea to use semaphore or mutex to limit the number of BE
> > jobs that are in the exiting path. This sounds like a feasible approach for
> > our scenario...
> 
> I am not really sure this is something that would be acceptable. Your
> problem is resource partitioning. Papering that over by a lock is not
> the right way to go. Besides that you will likely hit a hard question on
> how many tasks to allow to run concurrently. Whatever the value some
> workload will very likely going to suffer. We cannot assume admin to
> chose the right value because there is no clear answer for that. Not to
> mention other potential problems - e.g. even more priority inversions
> etc.

I don't see how we get priority inversions.  These tasks are exiting; at
the point they take the semaphore, they should not be holding any locks.
They're holding a resource (memory) that needs to be released, but a
task wanting to acquire memory must already be prepared to sleep.

I see this as being a thundering herd problem.  We have dozens, maybe
hundreds of tasks all trying to free their memory at once.  If we force
the herd to go through a narrow gap, they arrive at the spinlock in an
orderly manner.