On Tue, 10 May, at 07:43:16PM, Peter Zijlstra wrote:
> Mike reported that the recent commit 3a47d5124a95 ("sched/fair: Fix
> fairness issue on migration") broke interactivity and the signal
> starve test.
>
> The problem is that I assumed ENQUEUE_WAKING was only set when we do a
> cross-cpu wakeup (migration), which isn't true. This means we now
> destroy the vruntime history of tasks and wakeup-preemption suffers.
>
> Cure this by making my assumption true, only call
> sched_class::task_waking() when we do a cross-cpu wakeup. This avoids
> the indirect call in the case we do a local wakeup.
>
> Cc: Pavan Kondeti <pkond...@codeaurora.org>
> Cc: Ben Segall <bseg...@google.com>
> Cc: Matt Fleming <m...@codeblueprint.co.uk>
> Cc: Morten Rasmussen <morten.rasmus...@arm.com>
> Cc: Paul Turner <p...@google.com>
> Cc: Thomas Gleixner <t...@linutronix.de>
> Cc: byungchul.p...@lge.com
> Cc: Andrew Hunter <a...@google.com>
> Fixes: 3a47d5124a95 ("sched/fair: Fix fairness issue on migration")
> Reported-by: Mike Galbraith <mgalbra...@suse.de>
> Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org>
> ---
> kernel/sched/core.c | 29 +++++++++++++++++++++--------
> kernel/sched/fair.c | 41 ++++++++++++++++++++++++++++++++++++++++-
> 2 files changed, 61 insertions(+), 9 deletions(-)
This patch appears to cause a regression for hackbench -pipe of
between ~8% and ~10% with groups >= NR_CPU.
I haven't probed much yet, but it looks like the vruntime of tasks has
gone nuts.
