Re: Oops on Power8 (was Re: [PATCH v2 1/7] workqueue: make workqueue available early during boot)

From: Michael Ellerman
Date: Tue Oct 11 2016 - 07:22:24 EST

Tejun Heo <tj@xxxxxxxxxx> writes:

> Hello, Michael.
> On Mon, Oct 10, 2016 at 09:22:55PM +1100, Michael Ellerman wrote:
>> This patch seems to be causing one of my Power8 boxes not to boot.
>> Specifically commit 3347fa092821 ("workqueue: make workqueue available
>> early during boot") in linux-next.
>> If I revert this on top of next-20161005 then the machine boots again.
>> I've attached the oops below. It looks like the cfs_rq of p->se is NULL?
> Hah, weird that it's arch dependent, or maybe it's just different
> config options. Most likely, it's caused by workqueue_init() call
> being moved too early. Can you please try the following patch and see
> whether the problem goes away?

No that doesn't help.

What does is this:

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 94732d1ab00a..4e79549d242f 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -1614,7 +1614,8 @@ int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
* [ this allows ->select_task() to simply return task_cpu(p) and
* not worry about this generic constraint ]
- if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
+ if (unlikely(cpu >= nr_cpu_ids ||
+ !cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
cpu = select_fallback_rq(task_cpu(p), p);

The oops happens because we're in enqueue_task_fair() and p->se->cfs_rq
is NULL.

The cfs_rq is NULL because we did set_task_rq(p, 2048), where 2048 is
NR_CPUS. That causes us to index past the end of the tg->cfs_rq array in
set_task_rq() and happen to get NULL.

We never should have done set_task_rq(p, 2048), because 2048 is >=
nr_cpu_ids, which means it's not a valid CPU number, and set_task_rq()
doesn't cope with that.

The reason we're calling set_task_rq() with CPU 2048 is because
in select_task_rq() we had tsk_nr_cpus_allowed() = 0, because
tsk_cpus_allowed(p) is an empty cpu mask.

That means we do in select_task_rq():
cpu = cpumask_any(tsk_cpus_allowed(p));

And when tsk_cpus_allowed(p) is empty cpumask_any() returns nr_cpu_ids,
causing cpu to be set to 2048 in my case.

select_task_rq() then does the check to see if it should use a fallback

if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
cpu = select_fallback_rq(task_cpu(p), p);

But in both those checks we end up indexing off the end of the cpu mask,
because cpu is >= nr_cpu_ids. At least on my system they both return
true and so we return cpu == 2048.

The patch above is pretty clearly not the right fix, though maybe it's a
good safety measure.

Presumably we shouldn't be ending up with tsk_cpus_allowed() being
empty, but I haven't had time to track down why that's happening.