CFQ: Preemption/timeout logic reversed?

CFQ: Preemption/timeout logic reversed?

[Carl Henrik Lunde]

Hi!

It seems the preemption "bonus" logic in CFQ is reversed, a preempted
process is given an additional delay in start time instead of a bonus.
 This seems unfair.  I'm not sure if it's a good idea to let
slice_resid grow without limit as shown below, but isn't this more
like the way it was intended to work?  Or did I misunderstand
something?

PS! The comment above cfq_preempt_queue seems outdated too.

Code not tested, just showing what I mean:

diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
index 664ebfd..ea18d45 100644
--- a/block/cfq-iosched.c
+++ b/block/cfq-iosched.c
@@ -292,7 +292,8 @@ cfq_prio_to_slice(struct cfq_data *cfqd, struct
cfq_queue *cfqq)
 static inline void
 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {
-       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
+       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) +
cfqq->slice_resid + jiffies;
+       cfqq->slice_resid = 0;
        cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
 }

@@ -489,8 +490,6 @@ static void cfq_service_tree_add(struct cfq_data *cfqd,
                        rb_key += jiffies;
        } else if (!add_front) {
                rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
-               rb_key += cfqq->slice_resid;
-               cfqq->slice_resid = 0;
        } else
                rb_key = 0;

@@ -1851,8 +1850,7 @@ cfq_should_preempt(struct cfq_data *cfqd, struct
cfq_queue *new_cfqq,
 }

 /*
- * cfqq preempts the active queue. if we allowed preempt with no slice left,
- * let it have half of its nominal slice.
+ * cfqq preempts the active queue.  start a new time slice.
  */
 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
 {


-- 
mvh
Carl Henrik

Re: CFQ: Preemption/timeout logic reversed?

[Jens Axboe]

On Mon, Apr 20 2009, Carl Henrik Lunde wrote:
> Hi!
> 
> It seems the preemption "bonus" logic in CFQ is reversed, a preempted
> process is given an additional delay in start time instead of a bonus.
>  This seems unfair.  I'm not sure if it's a good idea to let

Hmm? ->slice_resid is a long, so if we preempt the process 10 jiffies
before it was supposed to end, the resid will be -10. So it'll not
increase the rb_key, it'll decrease it.

> slice_resid grow without limit as shown below, but isn't this more
> like the way it was intended to work?  Or did I misunderstand
> something?

->slice_resid is reset when it gets repositioned in the rb tree. The
intent was not to increase the slice length, but instead allow it sooner
service again.

> PS! The comment above cfq_preempt_queue seems outdated too.

Yep, the slice length comment is out dated indeed.

> Code not tested, just showing what I mean:
> 
> diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
> index 664ebfd..ea18d45 100644
> --- a/block/cfq-iosched.c
> +++ b/block/cfq-iosched.c
> @@ -292,7 +292,8 @@ cfq_prio_to_slice(struct cfq_data *cfqd, struct
> cfq_queue *cfqq)
>  static inline void
>  cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
>  {
> -       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
> +       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) +
> cfqq->slice_resid + jiffies;

So if ->slice_resid is negative because we preempted this queue, it'll
now get a shorter slice. That's not very nice :-)

-- 
Jens Axboe

Re: CFQ: Preemption/timeout logic reversed?

[Carl Henrik Lunde]

On Mon, Apr 20, 2009 at 08:16, Jens Axboe <jens.axboe@oracle.com> wrote:
> On Mon, Apr 20 2009, Carl Henrik Lunde wrote:
>> Hi!
>>
>> It seems the preemption "bonus" logic in CFQ is reversed, a preempted
>> process is given an additional delay in start time instead of a bonus.
>>  This seems unfair.  I'm not sure if it's a good idea to let
>
> Hmm? ->slice_resid is a long, so if we preempt the process 10 jiffies
> before it was supposed to end, the resid will be -10. So it'll not
> increase the rb_key, it'll decrease it.

OK, so maybe I'm tired (I am!), but I don't get it. :)

{
        if (... busy_rt_queues)
               cfq_slice_expire(timed_out=1)
}

cfq_slice_expire() {
{
        if (timed_out)
                slice_resid = slice_end - jiffies;
}

if preempted it would be 100 - 90 = +10?

>
>> slice_resid grow without limit as shown below, but isn't this more
>> like the way it was intended to work?  Or did I misunderstand
>> something?
>
> ->slice_resid is reset when it gets repositioned in the rb tree. The
> intent was not to increase the slice length, but instead allow it sooner
> service again.

Fair enough

>> PS! The comment above cfq_preempt_queue seems outdated too.
>
> Yep, the slice length comment is out dated indeed.
>
>> Code not tested, just showing what I mean:
>>
>> diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
>> index 664ebfd..ea18d45 100644
>> --- a/block/cfq-iosched.c
>> +++ b/block/cfq-iosched.c
>> @@ -292,7 +292,8 @@ cfq_prio_to_slice(struct cfq_data *cfqd, struct
>> cfq_queue *cfqq)
>>  static inline void
>>  cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
>>  {
>> -       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
>> +       cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) +
>> cfqq->slice_resid + jiffies;
>
> So if ->slice_resid is negative because we preempted this queue, it'll
> now get a shorter slice. That's not very nice :-)
>
> --
> Jens Axboe
>
>



-- 
mvh
Carl Henrik

Re: CFQ: Preemption/timeout logic reversed?

[Jens Axboe]

On Mon, Apr 20 2009, Carl Henrik Lunde wrote:
> On Mon, Apr 20, 2009 at 08:16, Jens Axboe <jens.axboe@oracle.com> wrote:
> > On Mon, Apr 20 2009, Carl Henrik Lunde wrote:
> >> Hi!
> >>
> >> It seems the preemption "bonus" logic in CFQ is reversed, a preempted
> >> process is given an additional delay in start time instead of a bonus.
> >>  This seems unfair.  I'm not sure if it's a good idea to let
> >
> > Hmm? ->slice_resid is a long, so if we preempt the process 10 jiffies
> > before it was supposed to end, the resid will be -10. So it'll not
> > increase the rb_key, it'll decrease it.
> 
> OK, so maybe I'm tired (I am!), but I don't get it. :)
> 
> {
>         if (... busy_rt_queues)
>                cfq_slice_expire(timed_out=1)
> }
> 
> cfq_slice_expire() {
> {
>         if (timed_out)
>                 slice_resid = slice_end - jiffies;
> }
> 
> if preempted it would be 100 - 90 = +10?

Hrmpf, that is buggy, it should be jiffies - slice_end! Presumably that
has been buggy since the shift to rb service tree, since (IIRC) that was
when it was changed from "add to slice length" to "adjust in tree".

-- 
Jens Axboe