Messages in this thread | | | Date | Sat, 11 Dec 2021 12:12:15 +0100 | From | Peter Zijlstra <> | Subject | Re: [PATCH] sched/fair: Again ignore percpu threads for imbalance pulls |
| |
On Sat, Dec 11, 2021 at 05:48:08PM +0800, Yihao Wu wrote: > commit 2f5f4cce496e ("sched/fair: Ignore percpu threads for imbalance > pulls") was meant to fix a performance issue, when load balance tries to > migrate pinned kernel threads at MC domain level. This was destined to > fail. After it fails, it further makes wakeup balance at NUMA domain level > messed up. The most severe case that I noticed and frequently occurs: > |sum_nr_running(node1) - sum_nr_running(node2)| > 100 > > However the original bugfix failed, because it covers only case 1) below. > 1) Created by create_kthread > 2) Created by kernel_thread > No kthread is assigned to task_struct in case 2 (Please refer to comments > in free_kthread_struct) so it simply won't work. > > The easist way to cover both cases is to check nr_cpus_allowed, just as > discussed in the mailing list of the v1 version of the original fix. > > * lmbench3.lat_proc -P 104 fork (2 NUMA, and 26 cores, 2 threads) > > w/out patch w/ patch > fork+exit latency 1660 ms 1520 ms ( 8.4%) > > Fixes: 2f5f4cce496e ("sched/fair: Ignore percpu threads for imbalance pulls") > Signed-off-by: Yihao Wu <wuyihao@linux.alibaba.com> > --- > kernel/kthread.c | 6 +----- > 1 file changed, 1 insertion(+), 5 deletions(-) > > diff --git a/kernel/kthread.c b/kernel/kthread.c > index 4a4d7092a2d8..cb05d3ff2de4 100644 > --- a/kernel/kthread.c > +++ b/kernel/kthread.c > @@ -543,11 +543,7 @@ void kthread_set_per_cpu(struct task_struct *k, int cpu) > > bool kthread_is_per_cpu(struct task_struct *p) > { > - struct kthread *kthread = __to_kthread(p); > - if (!kthread) > - return false; > - > - return test_bit(KTHREAD_IS_PER_CPU, &kthread->flags); > + return (p->flags & PF_KTHREAD) && p->nr_cpus_allowed == 1; > }
NAK, this will break lots of things.
| |