Re: 2.6.33-rc1 unusable due to scheduler issues, circular locking, WARNs and BUGs

[Xiaotian Feng <xtfeng@gmail.com>]

On Tue, Dec 22, 2009 at 3:50 PM, Américo Wang <xiyou.wangcong@gmail.com> wrote:
> On Tue, Dec 22, 2009 at 3:41 PM, Xiaotian Feng <xtfeng@gmail.com> wrote:
>> On Tue, Dec 22, 2009 at 3:19 PM, Américo Wang <xiyou.wangcong@gmail.com> wrote:
>>> [Fix top-posting]
>>>
>>> On Tue, Dec 22, 2009 at 1:42 PM, Xiaotian Feng <xtfeng@gmail.com> wrote:
>>>>
>>>> On Tue, Dec 22, 2009 at 8:17 AM, Eric Paris <eparis@redhat.com> wrote:
>>>>> Trying to build a kernel on a 48 core x86_64 box using make -j 64 and
>>>>> I'm exploding in the scheduler.  I'm running (and building) kernel
>>>>> f7b84a6ba7eaeba4e1df8feddca1473a7db369a5  There are three distinct
>>>>> signatures of problems.  Some boots I'll see all 3 of these failures
>>>>> sometimes only 1 or 2 of them.  That's the reason they are kinda split
>>>>> up in dmesg.
>>>>>
>>>>> 1) gcc/3141 is trying to acquire lock:
>>>>>  (&(&sem->wait_lock)->rlock){......}, at: [<ffffffff81223234>] __down_read_trylock+0x13/0x46
>>>>>
>>>>> but task is already holding lock:
>>>>>  (&rq->lock){-.-.-.}, at: [<ffffffff8103dd2d>] task_rq_lock+0x51/0x83
>>>>>
>>>>> 2) WARN() in kernel/sched_fair.c:1001 hrtick_start_fair()
>>>>>
>>>>> 3) NULL pointer dereference at 0000000000000168 in check_preempt_wakeup
>>>>>      kernel/sched_fair.c
>>>>>
>>>>> Full backtraces are in the attached dmesg.
>>>>>
>>>> Does a revert of cd29fe6f2637cc2ccbda5ac65f5332d6bf5fa3c6 fix this problem?
>>>
>>>
>>> I don't think so...
>>>
>>> I think the most suspicious commit here is ab19cb23. It kicked
>>> "local_irq_save()"
>>> out, which means if the task is selected to run on another cpu which doesn't
>>> disable irq, we will have a page fault, thun we will try to hold mm->mmap_sem
>>> while we are holding rq->lock already.
>>
>> The page fault is from kernel  NULL pointer deref.  You should connect
>> the lockdep warning and kernel BUG together.
>>
>
> Interesting.
>
> 1) Doesn't this NULL ptr def expose that we have a potential problem?
>
> 2) For NULL ptr def problem, commit 3a7e73a2e2 seems more suspicious..

I don't think so,
(gdb) l *check_preempt_wakeup+0x170
0xffffffff8103c815 is in is_same_group (kernel/sched_fair.c:154).
(gdb) assemble check_preempt_wakeup
<snip>
0xffffffff8103c815 <is_same_group+0>:   mov    0x168(%rsi),%rax
<snip>

The panic is from NULL pointer deref at 0000000000000168, so some time
in is_same_group() while loop,
parent_entity(*pse) = NULL, then is_same_group() trying to visit
pse->cfs_rq, NULL pointer deref was triggered.

commit 3a7e73, the behaviour for find_matching_se() is same as before,
this commit should not be the buggy one.

>