Re: [PATCH v4 5/5] cgroup/cpuset: Optimize out unneeded cpuset_can_fork/cpuset_cancel_fork calls

[Tejun Heo <tj@kernel.org>]

Hello,

On Wed, Apr 12, 2023 at 02:40:53PM -0400, Waiman Long wrote:
> On 4/12/23 14:27, Tejun Heo wrote:
> > On Tue, Apr 11, 2023 at 09:36:01AM -0400, Waiman Long wrote:
> > > The newly introduced cpuset_can_fork() and cpuset_cancel_fork() calls
> > > are only needed when the CLONE_INTO_CGROUP flag is set which is not
> > > likely. Adding an extra cpuset_can_fork() call does introduce a bit
> > > of performance overhead in the fork/clone fastpath. To reduce this
> > > performance overhead, introduce a new clone_into_cgroup_can_fork flag
> > > into the cgroup_subsys structure. This flag, when set, will call the
> > > can_fork and cancel_fork methods only if the CLONE_INTO_CGROUP flag
> > > is set.
> > > 
> > > The cpuset code is now modified to set this flag. The same cpuset
> > > checking code in cpuset_can_fork() and cpuset_cancel_fork() will have
> > > to stay as the cgroups can be different, but the cpusets may still be
> > > the same. So the same check must be present in both cpuset_fork() and
> > > cpuset_can_fork() to make sure that attach_in_progress is correctly set.
> > > 
> > > Signed-off-by: Waiman Long <longman@redhat.com>
> > Waiman, I'm not necessarily against this optimization but can we at least
> > have some performance numbers to show that this is actually meaningful?
> > Given how heavy our fork path is, I'm not too sure this would show up in any
> > meaningful way.
> 
> That make sense to me. I am OK to leave it for now as it is an optimization
> patch anyway.
> 
> BTW, another question that I have is about the cgroup_threadgroup_rwsem. It
> is currently a percpu rwsem. Is it possible to change it into a regular
> rwsem instead? It is causing quite a bit of latency for workloads that
> require rather frequent changes to cgroups. I know we have a "favordynmods"
> mount option to disable the percpu operation. This will still be less
> performant than a normal rwsem. Of course the downside is that the fork/exit
> fastpaths will be slowed down a bit.

I don't know. Maybe? A rwsem actually has a scalability factor in that the
more CPUs are forking, the more expensive the rwsem becomes, so it is a bit
more of a concern. Another factor is that in majority of use cases we're
almost completely bypassing write-locking percpu_rwsem, so it feel a bit sad
to convert it to a regular rwsem. So, if favordynmods is good enough, I'd
like to keep it that way.

Thanks.

-- 
tejun