Re: [PATCH -mm -V2] mm, swap: Fix race between swapoff and some swap operations

[Andrew Morton]

On Fri, 15 Dec 2017 11:04:43 +0100 Michal Hocko <mhocko@xxxxxxxxxx> wrote:

> On Thu 14-12-17 12:42:46, Andrew Morton wrote:
> > On Thu, 14 Dec 2017 16:17:18 +0100 Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > 
> > > > as fast as possible, SRCU instead of reference count is used to
> > > > implement get/put_swap_device().  From get_swap_device() to
> > > > put_swap_device(), the reader side of SRCU is locked, so
> > > > synchronize_srcu() in swapoff() will wait until put_swap_device() is
> > > > called.
> > > 
> > > It is quite unfortunate to pull SRCU as a dependency to the core kernel.
> > > Different attempts to do this have failed in the past. This one is
> > > slightly different though because I would suspect that those tiny
> > > systems do not configure swap. But who knows, maybe they do.
> > > 
> > > Anyway, if you are worried about performance then I would expect some
> > > numbers to back that worry. So why don't simply start with simpler
> > > ref count based and then optimize it later based on some actual numbers.
> > > Btw. have you considered pcp refcount framework. I would suspect that
> > > this would give you close to SRCU performance.
> > 
> > <squeaky-wheel>Or use stop_kernel() ;)</squeaky-wheel>
> 
> well, stop_kernel is a _huge_ hammer.

But it's very simple and requires zero code changes on the fast path. 
This makes it appropriate for swapoff!

> I think we can do much better
> without a large complexity. A simple ref counting (or pcp refcounting if
> the former has measurable complexity) should do just fine.

I'd like to be able to compare the implementations ;)