Re: [PATCH 1/2] sched/topology: Annotate RCU pointers properly

[Pierre Gondois]

Hello Valentin,

On 1/4/24 10:54, Valentin Schneider wrote:
> On 03/01/24 13:56, Pierre Gondois wrote:
> > Cleanup RCU-related spare errors by annotating RCU pointers.
> >
> > sched_domains_numa_distance:
> >    error: incompatible types in comparison expression
> >    (different address spaces):
> >        int [noderef] __rcu *
> >        int *
> >
> > sched_domains_numa_masks:
> >    error: incompatible types in comparison expression
> >    (different address spaces):
> >        struct cpumask **[noderef] __rcu *
> >        struct cpumask ***
> >
> > Signed-off-by: Pierre Gondois <pierre.gondois@xxxxxxx>
>
> That's from when the NUMA topologies were made dynamic, which should be:
> Fixes: 0fb3978b0aac ("sched/numa: Fix NUMA topology for systems with CPU-less nodes")

Ok yes

> > ---
> >   kernel/sched/topology.c | 4 ++--
> >   1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> > index 10d1391e7416..0342a4f41f09 100644
> > --- a/kernel/sched/topology.c
> > +++ b/kernel/sched/topology.c
> > @@ -1542,8 +1542,8 @@ static int			sched_domains_numa_levels;
> >   static int			sched_domains_curr_level;
> >
> >   int				sched_max_numa_distance;
> > -static int			*sched_domains_numa_distance;
> > -static struct cpumask		***sched_domains_numa_masks;
> > +static int			__rcu *sched_domains_numa_distance;
> > +static struct cpumask		** __rcu *sched_domains_numa_masks;
>
> I understand that's what sparse is asking for, but that looks odd to me. We
> use it as:
>
>    rcu_assign_pointer(sched_domains_numa_masks, foo);
>
> so why isn't it
>
>    __rcu ***sched_domains_numa_masks;
>
> ?
>
> This isn't a pointer to an RCU-protected array of masks, this is an
> RCU-protected double array of masks.

I think:
  static struct cpumask ** __rcu *sched_domains_numa_masks;
should denote an RCU-protected array^3 of 'struct cpumask', when
  static struct cpumask __rcu ***sched_domains_numa_masks;
would denote an array^2 of RCU-protected 'struct cpumask*',
and assignments would look like:
  rcu_assign_pointer(**sched_domains_numa_masks, foo);

Meaning that, when taking as a better example:
  static int __rcu *sched_domains_numa_distance;
Here we would like to avoid having 'access after free' to the array of
integer allocated to sched_domains_numa_distance.

For sched_domains_numa_masks:
  static struct cpumask	** __rcu *sched_domains_numa_masks;
  rcu_assign_pointer(sched_domains_numa_masks, foo);
  bar = rcu_dereference(sched_domains_numa_masks);
once the first array pointed by sched_domains_numa_masks is accessed/assigned,
we know the RCU-framework makes it safe against 'accesses after free' when
accessing the level-2 array of sched_domains_numa_masks, or accessing the mask
(level 3).

Please let me know if the reasoning seems dodgy.

About the kernel test robot:
  kernel/sched/topology.c:1998:19: sparse: sparse: incorrect type in assignment (different address spaces) @@
       expected int *distances @@
       got int [noderef] __rcu *static [addressable] [assigned] [toplevel] sched_domains_numa_distance
  kernel/sched/topology.c:2000:15: sparse: sparse: incorrect type in assignment (different address spaces) @@
       expected struct cpumask ***masks @@
       got struct cpumask **[noderef] __rcu *static [addressable] [assigned] [toplevel] sched_domains_numa_masks @@
I think the warnings can be ignored since the two pointers (distances, masks)
are only dereferenced after a 'synchronize_rcu()'

Regards,
Pierre

> >   #endif
> >
> >   /*
> > --
> > 2.25.1