Re: blkiocg_update_io_add_stats(): INFO: trying to register non-static key

[Jens Axboe <jaxboe@fusionio.com>]

On 2010-06-11 09:15, Peter Zijlstra wrote:
> On Wed, 2010-06-02 at 15:04 +0200, Jens Axboe wrote:
>> On Wed, Jun 02 2010, Ingo Molnar wrote:
>>>
>>> btw., there's another warning triggered by the new blk-cgroups stats code:
>>>
>>> sda:INFO: trying to register non-static key.
>>> the code is fine but needs lockdep annotation.
>>> turning off the locking correctness validator.
>>> Pid: 81, comm: async/2 Not tainted 2.6.35-rc1-tip-01073-gd2f7698-dirty #6765
>>> Call Trace:
>>>  [<ffffffff81065d6d>] register_lock_class+0x15f/0x365
>>>  [<ffffffff8105c426>] ? sched_clock_local+0x1d/0x83
>>>  [<ffffffff8105c557>] ? sched_clock_cpu+0xcb/0xd9
>>>  [<ffffffff81067184>] __lock_acquire+0x97/0x481
>>>  [<ffffffff810088d4>] ? native_sched_clock+0x37/0x6d
>>>  [<ffffffff81321b74>] ? blkiocg_update_io_add_stats+0x32/0x12e
>>>  [<ffffffff81067644>] lock_acquire+0xd6/0xfd
>>>  [<ffffffff81321b74>] ? blkiocg_update_io_add_stats+0x32/0x12e
>>>  [<ffffffff81850859>] _raw_spin_lock_irqsave+0x47/0x82
>>>  [<ffffffff81321b74>] ? blkiocg_update_io_add_stats+0x32/0x12e
>>>  [<ffffffff81321b74>] blkiocg_update_io_add_stats+0x32/0x12e
>>
>> So that's blkg->stats_lock - help me out, what is lockdep complaining
>> about? The lock is initialized, what kind of lockdep annotation magic do
>> we need to sprinkle on it?
> 
> It basically says its lock instance isn't properly initialised. Usually
> spin_lock_init() will set lock->dep_map->key, for static locks,
> initialised with __SPIN_LOCK_UNLOCKED() ->key doesn't get set and we use
> the lock's address as key.
> 
> Now lockdep requires the key to be in static storage, so if you try to
> used __SPIN_LOCK_UNLOCKED() on dynamically allocated locks (the most
> common form is using static forms like DEFINE_foo() on stack variables),
> things go bang.
> 
> That said, the block_cgroup.c code seems to use spin_lock_init() so it
> _should_ all work out. Use before init/after free perhaps?

Exactly, I did double check that. But yes, could be a bug where it's
used before being initialized, though it seems to do that when the
struct is allocated. So perhaps user-after-free indeed, but it happened
at boot.

-- 
Jens Axboe