Sebastian,
This patch removes the raw spinlock operations when updating cputtime
in the vtime_* functions in kernel/sched/cputime.c.
Based on Frederic's commit b7ce2277f087fd052, there is no need for
the raw spinlocks in vtime_* functions to guard against writer
concurrency and the RT versions of write_seqcount_begin() and
write_seqcount_end() make calls to preempt_disable_rt() and
preempt_enable_rt(), so we'll be in atomic context while updating
cputime.
I've run this patch on x86_64 4.4.6-rt14 and the RHEL-RT kernel,
with 12h rteval runs on systems with both no tuning and systems
with isolcpus/rcu_nocbs/nohz_full cpus. No ill effects seen.
Signed-off-by: Rik van Riel <[email protected]>
Signed-off-by: Clark Williams <[email protected]>
---
kernel/sched/cputime.c | 18 ------------------
1 file changed, 18 deletions(-)
diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
index 0f75a38cff96..9a823ced7e4a 100644
--- a/kernel/sched/cputime.c
+++ b/kernel/sched/cputime.c
@@ -696,45 +696,37 @@ static void __vtime_account_system(struct task_struct *tsk)
void vtime_account_system(struct task_struct *tsk)
{
- raw_spin_lock(&tsk->vtime_lock);
write_seqcount_begin(&tsk->vtime_seq);
__vtime_account_system(tsk);
write_seqcount_end(&tsk->vtime_seq);
- raw_spin_unlock(&tsk->vtime_lock);
}
void vtime_gen_account_irq_exit(struct task_struct *tsk)
{
- raw_spin_lock(&tsk->vtime_lock);
write_seqcount_begin(&tsk->vtime_seq);
__vtime_account_system(tsk);
if (context_tracking_in_user())
tsk->vtime_snap_whence = VTIME_USER;
write_seqcount_end(&tsk->vtime_seq);
- raw_spin_unlock(&tsk->vtime_lock);
}
void vtime_account_user(struct task_struct *tsk)
{
cputime_t delta_cpu;
- raw_spin_lock(&tsk->vtime_lock);
write_seqcount_begin(&tsk->vtime_seq);
delta_cpu = get_vtime_delta(tsk);
tsk->vtime_snap_whence = VTIME_SYS;
account_user_time(tsk, delta_cpu, cputime_to_scaled(delta_cpu));
write_seqcount_end(&tsk->vtime_seq);
- raw_spin_unlock(&tsk->vtime_lock);
}
void vtime_user_enter(struct task_struct *tsk)
{
- raw_spin_lock(&tsk->vtime_lock);
write_seqcount_begin(&tsk->vtime_seq);
__vtime_account_system(tsk);
tsk->vtime_snap_whence = VTIME_USER;
write_seqcount_end(&tsk->vtime_seq);
- raw_spin_unlock(&tsk->vtime_lock);
}
void vtime_guest_enter(struct task_struct *tsk)
@@ -746,23 +738,19 @@ void vtime_guest_enter(struct task_struct *tsk)
* synchronization against the reader (task_gtime())
* that can thus safely catch up with a tickless delta.
*/
- raw_spin_lock(&tsk->vtime_lock);
write_seqcount_begin(&tsk->vtime_seq);
__vtime_account_system(tsk);
current->flags |= PF_VCPU;
write_seqcount_end(&tsk->vtime_seq);
- raw_spin_unlock(&tsk->vtime_lock);
}
EXPORT_SYMBOL_GPL(vtime_guest_enter);
void vtime_guest_exit(struct task_struct *tsk)
{
- raw_spin_lock(&tsk->vtime_lock);
write_seqcount_begin(&tsk->vtime_seq);
__vtime_account_system(tsk);
current->flags &= ~PF_VCPU;
write_seqcount_end(&tsk->vtime_seq);
- raw_spin_unlock(&tsk->vtime_lock);
}
EXPORT_SYMBOL_GPL(vtime_guest_exit);
@@ -775,30 +763,24 @@ void vtime_account_idle(struct task_struct *tsk)
void arch_vtime_task_switch(struct task_struct *prev)
{
- raw_spin_lock(&prev->vtime_lock);
write_seqcount_begin(&prev->vtime_seq);
prev->vtime_snap_whence = VTIME_SLEEPING;
write_seqcount_end(&prev->vtime_seq);
- raw_spin_unlock(&prev->vtime_lock);
- raw_spin_lock(¤t->vtime_lock);
write_seqcount_begin(¤t->vtime_seq);
current->vtime_snap_whence = VTIME_SYS;
current->vtime_snap = sched_clock_cpu(smp_processor_id());
write_seqcount_end(¤t->vtime_seq);
- raw_spin_unlock(¤t->vtime_lock);
}
void vtime_init_idle(struct task_struct *t, int cpu)
{
unsigned long flags;
- raw_spin_lock_irqsave(&t->vtime_lock, flags);
write_seqcount_begin(&t->vtime_seq);
t->vtime_snap_whence = VTIME_SYS;
t->vtime_snap = sched_clock_cpu(cpu);
write_seqcount_end(&t->vtime_seq);
- raw_spin_unlock_irqrestore(&t->vtime_lock, flags);
}
cputime_t task_gtime(struct task_struct *t)
--
2.5.5
On 04/15/2016 05:09 AM, Clark Williams wrote:
> Sebastian,
Hi Clark,
> This patch removes the raw spinlock operations when updating cputtime
> in the vtime_* functions in kernel/sched/cputime.c.
>
> Based on Frederic's commit b7ce2277f087fd052, there is no need for
> the raw spinlocks in vtime_* functions to guard against writer
> concurrency and the RT versions of write_seqcount_begin() and
> write_seqcount_end() make calls to preempt_disable_rt() and
> preempt_enable_rt(), so we'll be in atomic context while updating
> cputime.
>
> I've run this patch on x86_64 4.4.6-rt14 and the RHEL-RT kernel,
> with 12h rteval runs on systems with both no tuning and systems
> with isolcpus/rcu_nocbs/nohz_full cpus. No ill effects seen.
Frederic's patch made it into 4.5-rc1. What I did yesterday was:
* sched/cputime: Convert vtime_seqlock to seqcount
* sched/cputime: Clarify vtime symbols and document them
* Revert "vtime: Split lock and seqcount"
which is what you want, correct?
>
> Signed-off-by: Rik van Riel <[email protected]>
> Signed-off-by: Clark Williams <[email protected]>
Sebastian
On Fri, 15 Apr 2016 09:13:34 +0200
Sebastian Andrzej Siewior <[email protected]> wrote:
> Frederic's patch made it into 4.5-rc1. What I did yesterday was:
>
> * sched/cputime: Convert vtime_seqlock to seqcount
> * sched/cputime: Clarify vtime symbols and document them
> * Revert "vtime: Split lock and seqcount"
>
> which is what you want, correct?
> >
> > Signed-off-by: Rik van Riel <[email protected]>
> > Signed-off-by: Clark Williams <[email protected]>
>
> Sebastian
>
Yes, that's it. I just wanted to get rid of the redundant locking operations and reduce the accounting overhead by a bit.
Thanks!
Clark