On 31-Oct 19:45, Peter Zijlstra wrote:
> On Tue, Oct 30, 2018 at 04:09:47PM +0000, Patrick Bellasi wrote:
>
> > Let's fix this by ensuring to always discount the task estimated
> > utilization from the CPU's estimated utilization when the task is also
> > the current one. The same benchmark of the bug report, executed on a
> > dual socket 40 CPUs Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz machine,
> > reports these "Execl Throughput" figures (higher the better):
>
> Before this we have:
>
> /* Discount task's blocked util from CPU's util */
> util -= min_t(unsigned int, util, task_util(p));
>
> at the very least that comment is now inaccurate, since @p might not be
> blocked.
Right... will fix this too.
> > @@ -6258,8 +6267,17 @@ static unsigned long cpu_util_wake(int cpu, struct
> > task_struct *p)
> > * covered by the following code when estimated utilization is
> > * enabled.
> > */
> > - if (sched_feat(UTIL_EST))
> > - util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
> > + if (sched_feat(UTIL_EST)) {
> > + unsigned int estimated =
> > + READ_ONCE(cfs_rq->avg.util_est.enqueued);
> > +
> > + if (unlikely(current == p || task_on_rq_queued(p))) {
>
> I'm confused by the need for 'current == p', afaict task_on_rq_queued(p)
> is sufficient -- we've already established task_cpu(p) == cpu earlier.
Mmm... you right, I've got confused by the fact that current is
removed from the RBTree, but we keep tracking it as:
on_rq = TASK_ON_RQ_QUEUED
... unless, select_task_rq_fair() races with LB's:
detach_task()
p->on_rq = TASK_ON_RQ_MIGRATING;
-----------------------------------A
deactivate_task() \
dequeue_task() +- RaceTime
util_est_dequeue() /
-----------------------------------B
set_task_cpu()
migrate_task_rq{_fair}()
detach_entity_cfs_rq()
where, in [A..B] we will still avoid to discount *p's estimated
utilization. :/
Do you think we can live with that for the time being, maybe by just
adding a comment, or should we try to close that too ?
Eventually, the (current == p) check, maybe moved to the right of the
OR condition above, should certainly close the race window for the
specific UnixBench's execl case. Assuming for example the execl is
executed by a misfit task which is target of an active load balance...
> > + estimated -= min_t(unsigned int, estimated,
> > + (_task_util_est(p) | UTIL_AVG_UNCHANGED));
> > + }
> > +
> > + util = max(util, estimated);
> > + }
>
> Also, I think it is about time we find a suitable name for:
>
> #define xxx(_var, _val) do { \
remove_contrib(_var, _val) ?
> typeof(_var) var = (_var); \
> typeof(_var) val = (_val); \
> typeof(_var) res = var - val; \
> if (res > var) \
> res = 0; \
> (_var) = res; \
> } while (0)
>
> Which is basically sub_positive() but without the READ_ONCE/WRITE_ONCE
> stuff.
Perhaps there are still some paths in where sub_positive() can be
recycled... will look better into that and see what we can do on that
polishing side. However, I'll keep all that in a different patch.
> We do that:
>
> var -= min_t(typeof(var), var, val);
>
> pattern _all_ over.
Cheers Patrick
--
#include <best/regards.h>
Patrick Bellasi
