Re: [PATCH 1/5] xen/sched: rtds: add global-EDF utilization admission control

[Jürgen Groß <jgross@xxxxxxxx>]

On 14.09.26 11:13, Jan Beulich wrote:
> On 14.09.2026 11:10, Juergen Gross wrote:
> > On 26.08.26 06:57, Furkan Caliskan wrote:
> > > @@ -635,6 +656,54 @@ replq_reinsert(const struct scheduler *ops, struct rt_unit 
> > > *svc)
> > >            set_timer(&rt_priv(ops)->repl_timer, rearm_svc->cur_deadline);
> > >    }
> > >    
> > > +/*
> > > + * budget << RTDS_UTIL_SHIFT can't overflow: rt_validate_params()
> > > + * caps budget at RTDS_MAX_BUDGET. period == 0 means "no
> > > + * reservation" (a unit being removed), not an error.
> > > + */
> > > +static uint64_t
> > > +rt_unit_utilization(s_time_t period, s_time_t budget)
> > > +{
> > > +    if ( period <= 0 )
> > > +        return 0;
> > > +
> > > +    return ((uint64_t)budget << RTDS_UTIL_SHIFT) / (uint64_t)period;
> > > +}
> > > +
> > > +/*
> > > + * Utilization capacity of the cpupool domain d resides in.
> > > + */
> > > +static uint64_t
> > > +rt_utilization_cap(const struct domain *d)
> > > +{
> > > +    unsigned int cpus = cpumask_weight(cpupool_domain_master_cpumask(d));
> > > +
> > > +    return (uint64_t)cpus * RTDS_UTIL_SCALE * RTDS_UTIL_CAP_PCT / 100;
> > > +}
> > > +
> > > +/*
> > > + * Replaces a unit's reservation and updates prv->utilization
> > > + * to match. Growth that would push utilization over the
> > > + * cpupool's cap is refused. Removing a unit or shrinking
> > > + * a unit's reservation always succeed.
> > > + */
> > > +static int
> > > +rt_admission_test(struct rt_private *prv, const struct domain *d,
> > > +                  s_time_t old_period, s_time_t old_budget,
> > > +                  s_time_t new_period, s_time_t new_budget)
> > I think the function name isn't appropriate. The function doesn't test only, it
> > is setting the utilization, too.
> >
> > What about rt_try_set_utilization() instead? And make the return type bool
> > (true on success).
> >
> > > +{
> > > +    uint64_t old_util = rt_unit_utilization(old_period, old_budget);
> > > +    uint64_t new_util = rt_unit_utilization(new_period, new_budget);
> > > +    uint64_t total    = prv->utilization - old_util + new_util;
> > > +
> > > +    if ( new_util > old_util && total > rt_utilization_cap(d) )
> > > +        return -EINVAL;
> > > +
> > > +    prv->utilization = total;
> > > +
> > > +    return 0;
> > > +}
> > > +
> > >    /*
> > >     * Pick a valid resource for the unit vc
> > >     * Valid resource of an unit is intesection of unit's affinity
> > > @@ -892,8 +983,19 @@ rt_alloc_udata(const struct scheduler *ops, struct 
> > > sched_unit *unit, void *dd)
> > >    static void cf_check
> > >    rt_free_udata(const struct scheduler *ops, void *priv)
> > >    {
> > > +    struct rt_private *prv = rt_priv(ops);
> > >        struct rt_unit *svc = priv;
> > >    
> > > +    if ( svc && !is_idle_unit(svc->unit) )
> > > +    {
> > > +        unsigned long flags;
> > > +
> > > +        spin_lock_irqsave(&prv->lock, flags);
> > > +        rt_admission_test(prv, svc->unit->domain,
> > > +                           svc->period, svc->budget, 0, 0);
> > This use case clearly shows you are not only testing. :-)
> Yet at the same time ignoring possible errors.
I don't see how this could happen, as new_period and new_budget are both 0 here.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature.asc
Description: OpenPGP digital signature