| Message ID | 1561996022-28829-1-git-send-email-vincent.guittot@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | [v2] sched/fair: fix imbalance due to CPU affinity | expand |
On Mon, Jul 01, 2019 at 05:47:02PM +0200, Vincent Guittot wrote: > The load_balance() has a dedicated mecanism to detect when an imbalance > is due to CPU affinity and must be handled at parent level. In this case, > the imbalance field of the parent's sched_group is set. > > The description of sg_imbalanced() gives a typical example of two groups > of 4 CPUs each and 4 tasks each with a cpumask covering 1 CPU of the first > group and 3 CPUs of the second group. Something like: > > { 0 1 2 3 } { 4 5 6 7 } > * * * * > > But the load_balance fails to fix this UC on my octo cores system > made of 2 clusters of quad cores. > > Whereas the load_balance is able to detect that the imbalanced is due to > CPU affinity, it fails to fix it because the imbalance field is cleared > before letting parent level a chance to run. In fact, when the imbalance is > detected, the load_balance reruns without the CPU with pinned tasks. But > there is no other running tasks in the situation described above and > everything looks balanced this time so the imbalance field is immediately > cleared. > > The imbalance field should not be cleared if there is no other task to move > when the imbalance is detected. > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Thanks!
On 01/07/2019 16:47, Vincent Guittot wrote: > The load_balance() has a dedicated mecanism to detect when an imbalance > is due to CPU affinity and must be handled at parent level. In this case, > the imbalance field of the parent's sched_group is set. > > The description of sg_imbalanced() gives a typical example of two groups > of 4 CPUs each and 4 tasks each with a cpumask covering 1 CPU of the first > group and 3 CPUs of the second group. Something like: > > { 0 1 2 3 } { 4 5 6 7 } > * * * * > > But the load_balance fails to fix this UC on my octo cores system > made of 2 clusters of quad cores. > > Whereas the load_balance is able to detect that the imbalanced is due to > CPU affinity, it fails to fix it because the imbalance field is cleared > before letting parent level a chance to run. In fact, when the imbalance is > detected, the load_balance reruns without the CPU with pinned tasks. But > there is no other running tasks in the situation described above and > everything looks balanced this time so the imbalance field is immediately > cleared. > > The imbalance field should not be cleared if there is no other task to move > when the imbalance is detected. > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Does that want a Cc: stable@vger.kernel.org Fixes: afdeee0510db ("sched: Fix imbalance flag reset") ?
On Tue, 2 Jul 2019 at 11:34, Valentin Schneider <valentin.schneider@arm.com> wrote: > > On 01/07/2019 16:47, Vincent Guittot wrote: > > The load_balance() has a dedicated mecanism to detect when an imbalance > > is due to CPU affinity and must be handled at parent level. In this case, > > the imbalance field of the parent's sched_group is set. > > > > The description of sg_imbalanced() gives a typical example of two groups > > of 4 CPUs each and 4 tasks each with a cpumask covering 1 CPU of the first > > group and 3 CPUs of the second group. Something like: > > > > { 0 1 2 3 } { 4 5 6 7 } > > * * * * > > > > But the load_balance fails to fix this UC on my octo cores system > > made of 2 clusters of quad cores. > > > > Whereas the load_balance is able to detect that the imbalanced is due to > > CPU affinity, it fails to fix it because the imbalance field is cleared > > before letting parent level a chance to run. In fact, when the imbalance is > > detected, the load_balance reruns without the CPU with pinned tasks. But > > there is no other running tasks in the situation described above and > > everything looks balanced this time so the imbalance field is immediately > > cleared. > > > > The imbalance field should not be cleared if there is no other task to move > > when the imbalance is detected. > > > > Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> > > Does that want a > > Cc: stable@vger.kernel.org > Fixes: afdeee0510db ("sched: Fix imbalance flag reset") I was not sure that this has been introduced by this patch or following changes. I haven't been able to test it on such old kernel with my platform > > ? >
On 02/07/2019 11:00, Vincent Guittot wrote: >> Does that want a >> >> Cc: stable@vger.kernel.org >> Fixes: afdeee0510db ("sched: Fix imbalance flag reset") > > I was not sure that this has been introduced by this patch or > following changes. I haven't been able to test it on such old kernel > with my platform > Right, seems like 65a4433aebe3 ("sched/fair: Fix load_balance() affinity redo path") also played in this area. From surface level it looks like it only reduced the amount of CPUs the load_balance() redo can use (and interestingly it mentions the exact same bug as you observed, through triggered slightly differently). I'd be inclined to say that the issue was introduced by afdeee0510db, since from looking at the code from that time I can see the issue happening: - try to pull from a CPU with only tasks pinned to itself - set sgc->imbalance - redo with a CPU that sees no big imbalance - goto out_balanced - env.LBF_ALL_PINNED is still set but we clear sgc->imbalance >> >> ? >>
On Tue, 2 Jul 2019 at 16:29, Valentin Schneider <valentin.schneider@arm.com> wrote: > > > > On 02/07/2019 11:00, Vincent Guittot wrote: > >> Does that want a > >> > >> Cc: stable@vger.kernel.org > >> Fixes: afdeee0510db ("sched: Fix imbalance flag reset") > > > > I was not sure that this has been introduced by this patch or > > following changes. I haven't been able to test it on such old kernel > > with my platform > > > > Right, seems like > > 65a4433aebe3 ("sched/fair: Fix load_balance() affinity redo path") > > also played in this area. From surface level it looks like it only reduced > the amount of CPUs the load_balance() redo can use (and interestingly it > mentions the exact same bug as you observed, through triggered slightly > differently). > > I'd be inclined to say that the issue was introduced by afdeee0510db, since > from looking at the code from that time I can see the issue happening: I agree that the patch seems to be the root cause when reading code. But it also means that the bug is there for almost 5 years and has never been seen before I did some functional tests on my rework of the load balance That's why a real test would have confirmed that nothing else happens in the meantime > > - try to pull from a CPU with only tasks pinned to itself > - set sgc->imbalance > - redo with a CPU that sees no big imbalance > - goto out_balanced > - env.LBF_ALL_PINNED is still set but we clear sgc->imbalance > > >> > >> ? > >>
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index b798fe7..fff5632 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -8992,9 +8992,10 @@ static int load_balance(int this_cpu, struct rq *this_rq, out_balanced: /* * We reach balance although we may have faced some affinity - * constraints. Clear the imbalance flag if it was set. + * constraints. Clear the imbalance flag only if other tasks got + * a chance to move and fix the imbalance. */ - if (sd_parent) { + if (sd_parent && !(env.flags & LBF_ALL_PINNED)) { int *group_imbalance = &sd_parent->groups->sgc->imbalance; if (*group_imbalance)
The load_balance() has a dedicated mecanism to detect when an imbalance is due to CPU affinity and must be handled at parent level. In this case, the imbalance field of the parent's sched_group is set. The description of sg_imbalanced() gives a typical example of two groups of 4 CPUs each and 4 tasks each with a cpumask covering 1 CPU of the first group and 3 CPUs of the second group. Something like: { 0 1 2 3 } { 4 5 6 7 } * * * * But the load_balance fails to fix this UC on my octo cores system made of 2 clusters of quad cores. Whereas the load_balance is able to detect that the imbalanced is due to CPU affinity, it fails to fix it because the imbalance field is cleared before letting parent level a chance to run. In fact, when the imbalance is detected, the load_balance reruns without the CPU with pinned tasks. But there is no other running tasks in the situation described above and everything looks balanced this time so the imbalance field is immediately cleared. The imbalance field should not be cleared if there is no other task to move when the imbalance is detected. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> --- Sorry, I sent the patch before rebasing it on top of sched-tip and it might conlfict when applying because it was on top of my ongoing rework of load_balance This version is rebased on top of latest shced-tip/sched/core kernel/sched/fair.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) -- 2.7.4