Re: [PATCH 4/7] libata: use lazy workqueues for the pio task

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

On Mon, Aug 24 2009, Jeff Garzik wrote:
> On 08/24/2009 12:42 PM, Jens Axboe wrote:
>> On Mon, Aug 24 2009, Jeff Garzik wrote:
>>> No objections to the code, operationally...
>>>
>>> But it is disappointing that the "1 thread on UP" problem is not solved
>>> while changing this libata area.  Is there no way to specify a minimum
>>> lazy-thread count?
>>>
>>> A key problem continues to be tying to the number of CPUs, which is
>>> quite inappropriate for libata.
>>
>> We'll solve that next, the first problem is reducing the per-cpu
>> threads. Lots of places use per-cpu workqueues because that is what is
>> available, not necessarily because it's an appropriate choice. Like the
>> ata_wq above, it's not even a good fit.
>
> Agreed + sounds great.
>
> Thanks -- both for hacking libata for this, and more generally, for  
> attacking the too-many-kthreads problem!  :)  It's just sad how many  
> unused workqueue threads hang about, on every modern Linux box.

It is, it's one of those problems that's gotten totally out of hand. A
handful of wasted threads is easily ignored, but once you are safely
into the three digits it's just too much.

I took a quick look at converting libata to slow-work, and it's an easy
fit (and would solve the UP problem too). The remaining piece is a
slow_work_enqueue_delayed(), since we do use pio task queue with a small
delay from one path.

So I hope that we can get by with slow-work with a few tweaks here and
there, and just retain workqueues for the true performance (or
persistent) case. The lazy workqueues is still a nice addition I think,
since they don't hang around forever when things go idle.

-- 
Jens Axboe