Re: [PATCH v4] sched/numa: fix unsafe get_task_struct() in task_numa_assign()

[Kirill Tkhai <tkhai@yandex.ru>]

10.11.2014, 19:45, "Sasha Levin" <sasha.levin@oracle.com>:
> On 11/10/2014 11:36 AM, Kirill Tkhai wrote:
>>  I mean task_numa_find_cpu(). If a garbage is in cpumask_of_node(env->dst_nid)
>>  and cpu is bigger than mask, the check
>>
>>  cpumask_test_cpu(cpu, tsk_cpus_allowed(env->p)
>>
>>  may be true.
>>
>>  So, we dereference wrong rq in task_numa_compare(). It's not rq at all.
>>  Strange cpu may be from here. It's just a int number in a wrong memory.
>
> But the odds of the spinlock magic and owner pointer matching up are slim
> to none in that case. The memory is also likely to be valid since KASAN didn't
> complain about the access, so I don't believe it to be an access to freed memory.

I'm not good with lockdep checks, so I can't judge right now...
Just a hypothesis.

>>  A hypothesis that below may help:
>>
>>  diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>  index 826fdf3..a2b4a8a 100644
>>  --- a/kernel/sched/fair.c
>>  +++ b/kernel/sched/fair.c
>>  @@ -1376,6 +1376,9 @@ static void task_numa_find_cpu(struct task_numa_env *env,
>>   {
>>           int cpu;
>>
>>  + if (!node_online(env->dst_nid))
>>  + return;
>
> I've changed that to BUG_ON(!node_online(env->dst_nid)) and will run it for a
> bit.

I've looked one more time, and it looks like it's better to check for
BUG_ON(env->dst_nid > MAX_NUMNODES). node_online() may do not work for
insane nids.

Anyway, even if it's not connected, we need to initialize numa_preferred_nid
of init_task, because it's inherited by kernel_init() (and /sbin/init too).
I'll send the patch tomorrow.

Kirill