On 22 February 2013 13:32, Frederic Weisbecker fweisbec@gmail.com wrote:
On Thu, Feb 21, 2013 at 09:29:16AM +0100, Vincent Guittot wrote:
On my smp platform which is made of 5 cores in 2 clusters, I have the nr_busy_cpu field of sched_group_power struct that is not null when the platform is fully idle. The root cause seems to be: During the boot sequence, some CPUs reach the idle loop and set their NOHZ_IDLE flag while waiting for others CPUs to boot. But the nr_busy_cpus field is initialized later with the assumption that all CPUs are in the busy state whereas some CPUs have already set their NOHZ_IDLE flag. During the initialization of the sched_domain, we set the NOHZ_IDLE flag when nr_busy_cpus is initialized to 0 in order to have a coherent configuration. If a CPU enters idle and call set_cpu_sd_state_idle during the build of the new sched_domain it will not corrupt the initial state set_cpu_sd_state_busy is modified and clears the NOHZ_IDLE only if a non NULL sched_domain is attached to the CPU (which is the case during the rebuild)
Change since V3;
- NOHZ flag is not cleared if a NULL domain is attached to the CPU
- Remove patch 2/2 which becomes useless with latest modifications
Change since V2:
- change the initialization to idle state instead of busy state so a CPU that enters idle during the build of the sched_domain will not corrupt the initialization state
Change since V1:
- remove the patch for SCHED softirq on an idle core use case as it was a side effect of the other use cases.
Signed-off-by: Vincent Guittot vincent.guittot@linaro.org
kernel/sched/core.c | 4 +++- kernel/sched/fair.c | 9 +++++++-- 2 files changed, 10 insertions(+), 3 deletions(-)
diff --git a/kernel/sched/core.c b/kernel/sched/core.c index 26058d0..c730a4e 100644 --- a/kernel/sched/core.c +++ b/kernel/sched/core.c @@ -5884,7 +5884,9 @@ static void init_sched_groups_power(int cpu, struct sched_domain *sd) return;
update_group_power(sd, cpu);
atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
atomic_set(&sg->sgp->nr_busy_cpus, 0);set_bit(NOHZ_IDLE, nohz_flags(cpu));}
int __weak arch_sd_sibling_asym_packing(void) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 81fa536..2701a92 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -5403,15 +5403,20 @@ static inline void set_cpu_sd_state_busy(void) { struct sched_domain *sd; int cpu = smp_processor_id();
int clear = 0; if (!test_bit(NOHZ_IDLE, nohz_flags(cpu))) return;
clear_bit(NOHZ_IDLE, nohz_flags(cpu)); rcu_read_lock();for_each_domain(cpu, sd)
for_each_domain(cpu, sd) { atomic_inc(&sd->groups->sgp->nr_busy_cpus);clear = 1;} rcu_read_unlock();if (likely(clear))clear_bit(NOHZ_IDLE, nohz_flags(cpu));I fear there is still a race window:
= CPU 0 = = CPU 1 = // NOHZ_IDLE is set set_cpu_sd_state_busy() { dom1 = rcu_dereference(dom1); inc(dom1->nr_busy_cpus) rcu_assign_pointer(dom 1, NULL) // create new domain init_sched_group_power() { atomic_set(&tmp->nr_busy_cpus, 0); set_bit(NOHZ_IDLE, nohz_flags(cpu 1)); rcu_assign_pointer(dom 1, tmp) clear_bit(NOHZ_IDLE, nohz_flags(cpu)); }I don't know if there is any sane way to deal with this issue other than having nr_busy_cpus and nohz_flags in the same object sharing the same lifecycle.
I wanted to avoid having to use the sd pointer for testing NOHZ_IDLE flag because it occurs each time we go into idle but it seems to be not easily feasible. Another solution could be to add a synchronization step between rcu_assign_pointer(dom 1, NULL) and create new domain to ensure that all pending access to old sd values, has finished but this will imply a potential delay in the rebuild of sched_domain and i'm not sure that it's acceptable
Vincent