Need information on elevator_dispatch_fn

Need information on elevator_dispatch_fn

[Philipp Gruber]

Hello everyone,

I'm just working on an I/O-scheduler that implements some QoS
functionality.

Now I get some weird problems and need to know what triggers my
elevator_dispatch_fn, and how the return value of it is handled.
For now, I found that the dispatch function is called as long as there
are requests in my queue, so probably as long as it returns 1. On the
other hand, only every second dispatch call returns 1, the others 0, but
still it's working. I couldn't find any documentation about that (but
would like to write some, if I understood it). Could someone please
explain me when and why exactly elevator_dispatch_fn is triggered?

A little background:
Now that I'm implementing my QoS, I have one queue for each process, and
prioritized processes get a timeslice, where they can access the device
exclusively.
A timeslice might be 400ms, but if a process is finished after 200ms
and has no more requests, dispatch will return 0 and then it won't be
triggered anymore. So, when the next process has his timeslice, even if
it has requests, dispatch_fn is not triggered, so the process will
starve and hang forever.
So I need to 'reactivate' the whole dispatching process again. The only
way to to this is probably a timer, but I couldn't get this working
since I don't understand what calls dispatch_fn.

thanks in advance,
Philipp

Re: Need information on elevator_dispatch_fn

[Jens Axboe]

On Wed, Oct 31 2007, Philipp Gruber wrote:
> Hello everyone,
> 
> I'm just working on an I/O-scheduler that implements some QoS
> functionality.
> 
> Now I get some weird problems and need to know what triggers my
> elevator_dispatch_fn, and how the return value of it is handled.
> For now, I found that the dispatch function is called as long as there
> are requests in my queue, so probably as long as it returns 1. On the
> other hand, only every second dispatch call returns 1, the others 0, but
> still it's working. I couldn't find any documentation about that (but
> would like to write some, if I understood it). Could someone please
> explain me when and why exactly elevator_dispatch_fn is triggered?

elevator_dispatch_fn() is in charge of putting request on the dispatch
list. So it'll be called, if q->queue_head is empty to refill that. The
core is __elv_next_request(), which is called when a block driver wants
to queue more IO. The logic is essentially:

elv_next_request()
{
        while (1) {
                if (!list_empty(&q->queue_head))
                        return first rq of list
                if (!elevator_dispatch_fn())
                        return NULL;
        }
}

So if you queue_head is empty AND elevator_dispatch_fn returns 0, then
no request is given to the driver. IF elevator_dispatch_fn returns
non-zero, then it MUST have put request on the queue_head list.

-- 
Jens Axboe

Re: Need information on elevator_dispatch_fn

[Aaron Carroll]

Philipp Gruber wrote:
> So I need to 'reactivate' the whole dispatching process again. The only
> way to to this is probably a timer, but I couldn't get this working
> since I don't understand what calls dispatch_fn.

Hi Philipp,

Take a look how the AS and CFQ schedulers implement this.  In fact, the
CFQ scheduler implements policy similar to what you described.
as_work_handler() in block/as-iosched.c might be a good starting point.

-- Aaron