Re: [RFC PATCH] blk-mq: Clean up references when freeing rqs

[John Garry <john.garry@huawei.com>]

On 03/12/2020 00:55, Ming Lei wrote:

Hi Ming,

>> Yeah, so I said that was another problem which you mentioned there, which
>> I'm not addressing, but I don't think that I'm making thing worse here.
> The thing is that this patch does not fix the issue completely.
> 
>> So AFAICS, the blk-mq/sched code doesn't wait for any "readers" to be
>> finished, such as those running blk_mq_queue_tag_busy_iter or
>> blk_mq_tagset_busy_iter() in another context.
>>
>> So how about the idea of introducing some synchronization primitive, such as
>> semaphore, which those "readers" must grab and release at start and end (of
>> iter), to ensure the requests are not freed during the iteration?
> It looks good, however devil is in details, please make into patch for
> review.

OK, but another thing to say is that I need to find a somewhat reliable 
reproducer for the potential problem you mention. So far this patch 
solves the issue I see (in that kasan stops warning). Let me analyze 
this a bit further.

Thanks,
John