Re: [PATCH 05/10] writeback: support > 1 flusher thread per bdi

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

On Mon, Jul 06 2009, Artem Bityutskiy wrote:
> Jamie Lokier wrote:
>> Artem Bityutskiy wrote:
>>> Jens Axboe wrote:
>>>> +static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work
>>>> *work)
>>>> +{
>>>> +	if (work) {
>>>> +		work->seen = bdi->wb_mask;
>>>> +		BUG_ON(!work->seen);
>>>> +		atomic_set(&work->pending, bdi->wb_cnt);
>>>> +		BUG_ON(!bdi->wb_cnt);
>>>> +
>>>> +		/*
>>>> +		 * Make sure stores are seen before it appears on the list
>>>> +		 */
>>>> +		smp_mb();
>>>> +
>>>> +		spin_lock(&bdi->wb_lock);
>>>> +		list_add_tail_rcu(&work->list, &bdi->work_list);
>>>> +		spin_unlock(&bdi->wb_lock);
>>>> +	}
>>> Doesn't spin_lock() include an implicit memory barrier?
>>> After &bdi->wb_lock is acquired, it is guaranteed that all
>>> memory operations are finished.
>>
>> I'm pretty sure spin_lock() is an "acquire" barrier, which just guarantees
>> loads/stores after the spin_lock() are done after taking the lock.
>>
>> It doesn't guarantee anything about loads/stores before the spin_lock().
>
> Right, but comment says memops should be flushed before the
> list is changed.

The comment says that the _above_ stores should be seen before it
appears on the list, it doesn't say anything about the list itself. What
matters is that the ->seen/pending must be fully visible before it
appears on the list. A spin_lock() doesn't guarentee that, and the bdi
thread could even see the work before the spin_unlock() is started.

-- 
Jens Axboe