Re: CFQ: async queue blocks the whole system

[Jens Axboe <axboe@kernel.dk>]

On 2011-06-10 11:29, Vivek Goyal wrote:
> On Fri, Jun 10, 2011 at 11:20:43AM +0200, Jens Axboe wrote:
>> On 2011-06-10 11:17, Vivek Goyal wrote:
>>> On Fri, Jun 10, 2011 at 09:19:12AM +0800, Shaohua Li wrote:
>>>
>>> [..]
>>>>> If there is no major advantage of draining sync requests before async
>>>>> is dispatched, I think this should be an easy fix.
>>>> I thought this is to avoid sync latency if we switch from an async
>>>> queue to sync queue later.
>>>
>>> Is it about the sync request latency which has already been dispatched? I 
>>> really wish that driver and disk should do some prioritazation for reads
>>> here and CFQ does not have to jump through hoops like drain sync requests
>>> before async requests are dispatched.
>>
>> That would never work. Are you suggesting putting that logic in all
>> drivers? Or relying on hardware to get the fairness right? Not going to
>> happen.
> 
> I was hoping that hardware does some prioritization. Well, in this case
> even if hardware maintains FIFO behavior it should be good enough.
> 
> But I would not claim anything in this regard as I have never experimented
> with it and have no idea that how sync latencies are impacted if we don't
> drain the queue before dispathing WRITEs.
> 
> I was just wondering that with current generation hardware is it bad
> enough that we need to keep this logic around?

The logic for draining around sync/async switch isn't that old, in fact
it dates only back to the last round of interactiveness fury. So yes,
it's definitely needed. It made a big difference on hardware that was
just out-of-the-store back then. I think you are putting way too much
faith into the sanity and goals of companies making this hardware.

-- 
Jens Axboe