Re: namespace: deadlock in dec_pid_namespaces

[Eric W. Biederman]

Nikolay Borisov <n.borisov.lkml@xxxxxxxxx> writes:

> On 20.01.2017 20:05, Eric W. Biederman wrote:
>> Nikolay Borisov <n.borisov.lkml@xxxxxxxxx> writes:
>> 
>>> On 20.01.2017 15:07, Dmitry Vyukov wrote:
>>>> Hello,
>>>>
>>>> I've got the following deadlock report while running syzkaller fuzzer
>>>> on eec0d3d065bfcdf9cd5f56dd2a36b94d12d32297 of linux-next (on odroid
>>>> device if it matters):
>> 
>> I am puzzled I thought we had fixed this with:
>>   add7c65ca426 ("pid: fix lockdep deadlock warning due to ucount_lock")
>> But apparently not.  We  just moved it from hardirq to softirq context.  Bah.
>> 
>> Thank you very much for the report.
>> 
>> Nikolay can you make your change use spinlock_irq?  And have put_ucounts
>> do spin_lock_irqsave?  That way we just don't care where we call this.
>
> Like the one attached?

Exactly thank you.  Dmitry if you have time to test that patch and
verify it fixes your issue I would appreciate it.

> I haven't really taken careful look as to whether
> the function where _irq versions do fiddle with irq state, since this
> might cause a problem if we unconditionally enable them.

In code paths where we can sleep irqs must come in enabled or it's a
bug.

spin_lock_irq which unconditionally disables irqs is thus safe on the
allocation path.

Similary spin_unlock_irq which unconditionally enables irqs is also safe
on the allocation path.

Eric