Messages in this thread Patch in this message | ![/](/images/icornerl.gif) | | From | lizhe.67@bytedanc ... | Subject | [RFC] sched/isolation: Fix CPU affinity issues for several task | Date | Tue, 30 Apr 2024 11:39:48 +0800 |
| |
From: Li Zhe <lizhe.67@bytedance.com>
If the parameter of cmdline "nohz_full=" contains cpu 0, the cpu affinity of the kernel thread "kthreadd", "rcu_sched", "rcuos%d", "rcuog%d" will always be 0x01, that is, these threads can only run on cpu 0. This is obviously not in line with the original design.
The root cause of this problem is that variables 'cpu_valid_mask' in functions __set_cpus_allowed_ptr_locked only contain cpu 0 before smp initialization is completed. If we call set_cpus_allowed_ptr and pass in a cpumask that does not contain cpu 0, the function call will return failure. Thread "kthreadd" and "rcu_sched" call the function set_cpus_allowed_ptr early in the system startup. Thread "rcuos%d" and "rcuog%d" inherit the wrong cpu affinity of "kthreadd".
I tried to fix this problem by adapting the function set_cpus_allowed_ptr, but the variable task_struct->cpus_ptr will be referenced or modified in the scheduled process, which seems to make it more difficult to fix this problem by adapting the function set_cpus_allowed_ptr. So this patch clear cpu 0 from nohz_full range to fix this problem.
Signed-off-by: Li Zhe <lizhe.67@bytedance.com> --- kernel/sched/isolation.c | 7 +++++++ 1 file changed, 7 insertions(+)
diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c index 5891e715f00d..7b9bcfcd3c55 100644 --- a/kernel/sched/isolation.c +++ b/kernel/sched/isolation.c @@ -152,6 +152,13 @@ static int __init housekeeping_setup(char *str, unsigned long flags) if (cpumask_empty(non_housekeeping_mask)) goto free_housekeeping_staging; + if ((flags & HK_FLAG_KTHREAD) && + cpumask_test_cpu(smp_processor_id(), non_housekeeping_mask)) { + pr_warn("Housekeeping: Clearing cpu %d from nohz_full range\n", smp_processor_id()); + __cpumask_set_cpu(smp_processor_id(), housekeeping_staging); + __cpumask_clear_cpu(smp_processor_id(), non_housekeeping_mask); + } + if (!housekeeping.flags) { /* First setup call ("nohz_full=" or "isolcpus=") */ enum hk_type type; -- 2.20.1
| ![\](/images/icornerr.gif) |