Re: sched/deadline: Server stops while fair tasks are still runnable

[Juri Lelli <juri.lelli@redhat.com>]

Hi Gabriele,

On 12/01/26 10:02, Gabriele Monaco wrote:
> The boost model in [1] spotted cases with fair tasks running after the server
> stopped. I believe this means that it's equally possible for those tasks to
> starve without the server even trying to start.
> 
> There are 2 related situation that can result in this:
> 
> 1. dl_server_idle is set to true when the CPU is really idle
> 2. a fair task wakes up (nothing happens since the server is still active)
> 3. dl_server_timer stops the server since it's idle, although it isn't really
> 
> ```
>     <idle>-0    d.h3.  2.521220: event_boost:     -12: idle x dl_replenish -> idle
>     <idle>-0    d.h3.  2.521227: event_laxity:    -12: zero_laxity_wait x dl_replenish_idle -> idle_wait
>  thread3-3-580  d..4.  3.338067: event_boost:     -12: idle x dl_server_resume_throttled -> throttled
>  thread3-3-580  d..3.  3.338076: sched_wakeup:    ksoftirqd/12:118 [120] CPU:012
>  thread5-5-582  d.h2.  3.471237: event_boost:     -12: throttled x dl_server_stop -> stopped (final)
>  thread5-5-582  d.h2.  3.471240: event_laxity:    -12: idle_wait x dl_server_stop -> stopped (final)
> ktimers/12-117  d..3.  3.538088: error_boost:     -12: event sched_switch_in not expected in the state stopped
> ktimers/12-117  d..2.  3.538089: sched_switch:    ktimers/12:117 [98] S ==> ksoftirqd/12:118 [120]
> ```
> 
> 
> 1. dl_server_timer runs when the CPU is idle but after a fair task wakes up
> 2. the call to update_curr sets dl_server_idle since it's only checking rq->curr
> (which is indeed idle)
> 3. the server is stopped within the same timer call although a fair task just
> woke up
> 
> ```
> ktimers/13-125  d.s52  7.309878: event_boost:    -12: stopped x dl_server_start -> ready
> ktimers/13-125  d.s52  7.309878: event_laxity:   -12: stopped x dl_server_start -> zero_laxity_wait
> ktimers/13-125  d.s42  7.309879: sched_wakeup:   kworker/u519:2:5559 [120] CPU:012
>     <idle>-0    d.h3.  7.309889: event_boost:    -12: ready x dl_replenish -> ready
>     <idle>-0    d.h3.  7.309889: event_laxity:   -12: zero_laxity_wait x dl_replenish_idle -> idle_wait
>     <idle>-0    d.h3.  7.309890: event_boost:    -12: ready x dl_server_stop -> stopped (final)
>     <idle>-0    d.h3.  7.309891: event_laxity:   -12: idle_wait x dl_server_stop -> stopped (final)
>     <idle>-0    d..3.  7.309895: error_boost:    -12: event sched_switch_in not expected in the state stopped
>     <idle>-0    d..2.  7.309896: sched_switch:   swapper/12:0 [120] R ==> kworker/u519:2:5559 [120]
> 
> ```
> 
> In both cases, if no new fair task wakes up, the one that woke up right before
> stopping the server could starve. That's a problem isn't it?

Yeah, think so. :/

Looks like we clear dl_defer_idle only while updating the fair task
entity (which didn't happen yet in both cases above).

> I tried a quick solution to the first case by clearing dl_server_idle every time
> the server started (even if it's still active, that is at every fair wakeup) and
> I think the second case could be avoided by relying on idle_cpu() instead of
> looking only on rq->curr, taking also waking tasks into account (not tried).

Not sure if you want to wait for Peter to chime in, but if you have
patches we can definitely take a look. :)

Thanks,
Juri