Re: [PATCH percpu/for-4.7-fixes 1/2] percpu: fix synchronization between chunk->map_extend_work and chunk destruction

[Vlastimil Babka]

On 26.5.2016 21:21, Tejun Heo wrote:
> Hello,
> 
> On Thu, May 26, 2016 at 11:19:06AM +0200, Vlastimil Babka wrote:
>>>  	if (is_atomic) {
>>>  		margin = 3;
>>>
>>>  		if (chunk->map_alloc <
>>> -		    chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW &&
>>> -		    pcpu_async_enabled)
>>> -			schedule_work(&chunk->map_extend_work);
>>> +		    chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) {
>>> +			if (list_empty(&chunk->map_extend_list)) {
> 
>> So why this list_empty condition? Doesn't it deserve a comment then? And
> 
> Because doing list_add() twice corrupts the list.  I'm not sure that
> deserves a comment.  We can do list_move() instead but that isn't
> necessarily better.

Ugh, right, somehow I thought it was testing &pcpu_map_extend_chunks.
My second question was based on the assumption that the list can have only one
item. Sorry about the noise.

>> isn't using a list an overkill in that case?
> 
> That would require rebalance work to scan all chunks whenever it's
> scheduled and if a lot of atomic allocations are taking place, it has
> some possibility to become expensive with a lot of chunks.
> 
> Thanks.
>