Re: [PATCH resend] sched/smt: fix unbalance sched_smt_present dec/inc

From: yangyingliang
Date: Tue Jul 02 2024 - 07:39:05 EST





> -----Original Messages-----
&gt; From: "Peter Zijlstra" <peterz@xxxxxxxxxxxxx>
&gt; Sent Time: 2024-07-02 16:44:18 (Tuesday)
&gt; To: "Yang Yingliang" <yangyingliang@xxxxxxxxxxxxxxx>
&gt; Cc: linux-kernel@xxxxxxxxxxxxxxx, mingo@xxxxxxxxxx, juri.lelli@xxxxxxxxxx, vincent.guittot@xxxxxxxxxx, dietmar.eggemann@xxxxxxx, rostedt@xxxxxxxxxxx, bsegall@xxxxxxxxxx, mgorman@xxxxxxx, bristot@xxxxxxxxxx, vschneid@xxxxxxxxxx, tglx@xxxxxxxxxxxxx, yu.c.chen@xxxxxxxxx, tim.c.chen@xxxxxxxxxxxxxxx, yangyingliang@xxxxxxxxxx, liwei391@xxxxxxxxxx
&gt; Subject: Re: [PATCH resend] sched/smt: fix unbalance sched_smt_present dec/inc
&gt;
&gt; On Tue, Jul 02, 2024 at 04:11:28PM +0800, Yang Yingliang wrote:
&gt; &gt; From: Yang Yingliang <yangyingliang@xxxxxxxxxx>
&gt; &gt;
&gt; &gt; I got the following warn report while doing stress test:
&gt; &gt;
&gt; &gt; jump label: negative count!
&gt; &gt; WARNING: CPU: 3 PID: 38 at kernel/jump_label.c:263 static_key_slow_try_dec+0x9d/0xb0
&gt; &gt; Call Trace:
&gt; &gt; <task>
&gt; &gt; __static_key_slow_dec_cpuslocked+0x16/0x70
&gt; &gt; sched_cpu_deactivate+0x26e/0x2a0
&gt; &gt; cpuhp_invoke_callback+0x3ad/0x10d0
&gt; &gt; cpuhp_thread_fun+0x3f5/0x680
&gt; &gt; smpboot_thread_fn+0x56d/0x8d0
&gt; &gt; kthread+0x309/0x400
&gt; &gt; ret_from_fork+0x41/0x70
&gt; &gt; ret_from_fork_asm+0x1b/0x30
&gt; &gt; </task>
&gt; &gt;
&gt; &gt; Because when cpuset_cpu_inactive() fails in sched_cpu_deactivate(),
&gt; &gt; the cpu offline failed, but sched_smt_present is decremented before
&gt; &gt; calling sched_cpu_deactivate(), it leads to unbalanced dec/inc, so
&gt; &gt; fix it by incrementing sched_smt_present in the error path.
&gt; &gt;
&gt; &gt; Fixes: c5511d03ec09 ("sched/smt: Make sched_smt_present track topology")
&gt; &gt; Reviewed-by: Chen Yu <yu.c.chen@xxxxxxxxx>
&gt; &gt; Reviewed-by: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
&gt; &gt; Signed-off-by: Yang Yingliang <yangyingliang@xxxxxxxxxx>
&gt; &gt; ---
&gt; &gt; kernel/sched/core.c | 4 ++++
&gt; &gt; 1 file changed, 4 insertions(+)
&gt; &gt;
&gt; &gt; diff --git a/kernel/sched/core.c b/kernel/sched/core.c
&gt; &gt; index bcf2c4cc0522..5ab6717b57e0 100644
&gt; &gt; --- a/kernel/sched/core.c
&gt; &gt; +++ b/kernel/sched/core.c
&gt; &gt; @@ -9756,6 +9756,10 @@ int sched_cpu_deactivate(unsigned int cpu)
&gt; &gt; sched_update_numa(cpu, false);
&gt; &gt; ret = cpuset_cpu_inactive(cpu);
&gt; &gt; if (ret) {
&gt; &gt; +#ifdef CONFIG_SCHED_SMT
&gt; &gt; + if (cpumask_weight(cpu_smt_mask(cpu)) == 2)
&gt; &gt; + static_branch_inc_cpuslocked(&amp;sched_smt_present);
&gt; &gt; +#endif
&gt; &gt; balance_push_set(cpu, false);
&gt; &gt; set_cpu_active(cpu, true);
&gt; &gt; sched_update_numa(cpu, true);
&gt;
&gt; Yes, does indeed appear needed, however!, when I look at
&gt; what else goes before this failure, should we not also call
&gt; set_rq_online() and things like that?
Yes, set_rq_online() is needed in the error path. I will send a new patch to add this.

&gt;
&gt; That is, can we rework things to be less fragile by sharing code between
&gt; this error path and sched_cpu_activate() ?
</yangyingliang@xxxxxxxxxx></tim.c.chen@xxxxxxxxxxxxxxx></yu.c.chen@xxxxxxxxx></yangyingliang@xxxxxxxxxx></yangyingliang@xxxxxxxxxxxxxxx></peterz@xxxxxxxxxxxxx>