Re: 2.6.19-rc3-git7: scsi_device_unbusy: inconsistent lock state

[Jens Axboe <jens.axboe@oracle.com>]

On Mon, Oct 30 2006, Arjan van de Ven wrote:
> 
> > which has always been considered safe, while not very pretty.
> 
> 
> actually it's different I think (based on a brief inspection of the
> code, I could well be wrong): 
> get_request_wait() causes a get_request() call with a GFP_NOIO gfp_mask
> which perculates upto cfq_set_request() as argument.
> cfq_set_request() then calls the inline cfq_get_queue() (which isn't in
> the backtrace due to inlining) which does
>                 } else if (gfp_mask & __GFP_WAIT) {
>                         /* 
>                          * Inform the allocator of the fact that we will
>                          * just repeat this allocation if it fails, to allow
>                          * the allocator to do whatever it needs to attempt to
>                          * free memory.
>                          */
>                         spin_unlock_irq(cfqd->queue->queue_lock);
> 
> which enables interrupts right smack in the middle of holding a whole
> bunch of locks.....

Where do you get 'a bunch' from? If you call get_request() with a
gfp_mask that includes __GFP_WAIT with a spinlock held, it's a bug. Just
as if you had called kmalloc() or similar with __GFP_WAIT set and
holding a lock. cfq even includes a warning check:

        might_sleep_if(gfp_mask & __GFP_WAIT);

So there's no bug there, cfq even grabbed the lock on its own before
calling cfq_get_queue().

> so to me it looks like lockdep at least has the appearance of moaning
> about a reasonably fishy situation...

To me it looks more about lockdep complaining because it doesn't grok
the full picture. The question is how to shut it up.

-- 
Jens Axboe