Re: [PATCH -next v2] mm/hotplug: fix a null-ptr-deref during NUMA boot

[Pingfan Liu <kernelfans@gmail.com>]

On Fri, May 31, 2019 at 5:03 PM Michal Hocko <mhocko@kernel.org> wrote:
>
> On Thu 30-05-19 20:55:32, Pingfan Liu wrote:
> > On Wed, May 29, 2019 at 2:20 AM Michal Hocko <mhocko@kernel.org> wrote:
> > >
> > > [Sorry for a late reply]
> > >
> > > On Thu 23-05-19 11:58:45, Pingfan Liu wrote:
> > > > On Wed, May 22, 2019 at 7:16 PM Michal Hocko <mhocko@kernel.org> wrote:
> > > > >
> > > > > On Wed 22-05-19 15:12:16, Pingfan Liu wrote:
> > > [...]
> > > > > > But in fact, we already have for_each_node_state(nid, N_MEMORY) to
> > > > > > cover this purpose.
> > > > >
> > > > > I do not really think we want to spread N_MEMORY outside of the core MM.
> > > > > It is quite confusing IMHO.
> > > > > .
> > > > But it has already like this. Just git grep N_MEMORY.
> > >
> > > I might be wrong but I suspect a closer review would reveal that the use
> > > will be inconsistent or dubious so following the existing users is not
> > > the best approach.
> > >
> > > > > > Furthermore, changing the definition of online may
> > > > > > break something in the scheduler, e.g. in task_numa_migrate(), where
> > > > > > it calls for_each_online_node.
> > > > >
> > > > > Could you be more specific please? Why should numa balancing consider
> > > > > nodes without any memory?
> > > > >
> > > > As my understanding, the destination cpu can be on a memory less node.
> > > > BTW, there are several functions in the scheduler facing the same
> > > > scenario, task_numa_migrate() is an example.
> > >
> > > Even if the destination node is memoryless then any migration would fail
> > > because there is no memory. Anyway I still do not see how using online
> > > node would break anything.
> > >
> > Suppose we have nodes A, B,C, where C is memory less but has little
> > distance to B, comparing with the one from A to B. Then if a task is
> > running on A, but prefer to run on B due to memory footprint.
> > task_numa_migrate() allows us to migrate the task to node C. Changing
> > for_each_online_node will break this.
>
> That would require the task to have preferred node to be C no? Or do I
> missunderstand the task migration logic?
I think in task_numa_migrate(), the migration logic should looks like:
  env.dst_nid = p->numa_preferred_nid; //Here dst nid is B
But later in
  if (env.best_cpu == -1 || (p->numa_group &&
p->numa_group->active_nodes > 1)) {
    for_each_online_node(nid) {
[...]
       task_numa_find_cpu(&env, taskimp, groupimp); // Here is a
chance to change p->numa_preferred_nid

There are serveral other broken by changing for_each_online_node(),
-1. show_numa_stats()
-2. init_numa_topology_type(), where sched_numa_topology_type may be
mistaken evaluated.
-3. ... can check call to for_each_online_node() one by one in scheduler.

That is my understanding of the code.

Thanks,
  Pingfan