Re: [PATCH 3/3] cgroup/cpuset: Allow only one active attach operation per cpuset

[Christian Brauner]

On Fri, Apr 07, 2023 at 11:30:14AM -0400, Waiman Long wrote:
> On 4/6/23 05:44, Christian Brauner wrote:
> > On Fri, Mar 31, 2023 at 10:50:45AM -0400, Waiman Long wrote:
> > > The current cpuset code uses the global cpuset_attach_old_cs variable
> > > to store the old cpuset value between consecutive cpuset_can_attach()
> > > and cpuset_attach() calls. Since a caller of cpuset_can_attach() may
> > > not need to hold the global cgroup_threadgroup_rwsem, parallel cpuset
> > > attach operations are possible.
> > > 
> > > When there are concurrent cpuset attach operations in progress,
> > > cpuset_attach() may fetch the wrong value from cpuset_attach_old_cs
> > > causing incorrect result.  To avoid this problem while still allowing
> > > certain level of parallelism, drop cpuset_attach_old_cs and use a
> > > per-cpuset attach_old_cs value. Also restrict to at most one active
> > > attach operation per cpuset to avoid corrupting the value of the
> > > per-cpuset attach_old_cs value.
> > > 
> > > Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
> > > ---
> > >   kernel/cgroup/cpuset.c | 19 ++++++++++++++-----
> > >   1 file changed, 14 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
> > > index 2367de611c42..3f925c261513 100644
> > > --- a/kernel/cgroup/cpuset.c
> > > +++ b/kernel/cgroup/cpuset.c
> > > @@ -198,6 +198,8 @@ struct cpuset {
> > >   	/* Handle for cpuset.cpus.partition */
> > >   	struct cgroup_file partition_file;
> > > +
> > > +	struct cpuset *attach_old_cs;
> > >   };
> > >   /*
> > > @@ -2456,22 +2458,27 @@ static int fmeter_getrate(struct fmeter *fmp)
> > >   	return val;
> > >   }
> > > -static struct cpuset *cpuset_attach_old_cs;
> > > -
> > >   /* Called by cgroups to determine if a cpuset is usable; cpuset_rwsem held */
> > >   static int cpuset_can_attach(struct cgroup_taskset *tset)
> > >   {
> > >   	struct cgroup_subsys_state *css;
> > > -	struct cpuset *cs;
> > > +	struct cpuset *cs, *oldcs;
> > >   	struct task_struct *task;
> > >   	int ret;
> > >   	/* used later by cpuset_attach() */
> > > -	cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
> > > +	oldcs = task_cs(cgroup_taskset_first(tset, &css));
> > >   	cs = css_cs(css);
> > >   	percpu_down_write(&cpuset_rwsem);
> > > +	/*
> > > +	 * Only one cpuset attach operation is allowed for each cpuset.
> > > +	 */
> > > +	ret = -EBUSY;
> > > +	if (cs->attach_in_progress)
> > > +		goto out_unlock;
> > That'll mean CLONE_INTO_CGROUP becomes even more interestig because it
> > isn't subject to this restriction in contrast to fork()+migrate, right?
> 
> This patch is not related to the CLONE_INTO_CGROUP. In fact, I have dropped

I know that. My point had been that this patch introduced a new
restriction affecting parallelism in the migration path while the
CLONE_INTO_CGROUP path remained unaffected.