[PATCH] sched/cpuacct: Use __this_cpu_add() instead of this_cpu_ptr()

[PATCH] sched/cpuacct: Use __this_cpu_add() instead of this_cpu_ptr()

[Muchun Song]

There seems to be no difference between the two, but on some
architectures(e.g. x86_64), there will be optimizations for
__this_cpu_add(). We can disassemble the code for you to see
the difference between them on x86_64.

  1) this_cpu_ptr(ca->cpuusage)->usages[index] += cputime;

     ffffffff810d7227: add %gs:0x7ef37fa9(%rip),%rax # f1d8 <this_cpu_off>
     ffffffff810d722f: add %rsi,(%rax)               # %rsi is @cputime

This result in two add instructions emitted by the compiler.

  2) __this_cpu_add(ca->cpuusage->usages[index], cputime);

     ffffffff810d7227: add %rsi,%gs:(%rax)           # %rsi is @cputime

This result in only one add instruction emitted by the compiler.

So we have enough reasons to use the __this_cpu_add().

Signed-off-by: Muchun Song <songmuchun@bytedance.com>
---
 kernel/sched/cpuacct.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
index 9fbb103834345..6448b0438ffb2 100644
--- a/kernel/sched/cpuacct.c
+++ b/kernel/sched/cpuacct.c
@@ -347,7 +347,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
 	rcu_read_lock();
 
 	for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
-		this_cpu_ptr(ca->cpuusage)->usages[index] += cputime;
+		__this_cpu_add(ca->cpuusage->usages[index], cputime);
 
 	rcu_read_unlock();
 }
@@ -363,7 +363,7 @@ void cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
 
 	rcu_read_lock();
 	for (ca = task_ca(tsk); ca != &root_cpuacct; ca = parent_ca(ca))
-		this_cpu_ptr(ca->cpustat)->cpustat[index] += val;
+		__this_cpu_add(ca->cpustat->cpustat[index], val);
 	rcu_read_unlock();
 }
 
-- 
2.11.0

Re: [PATCH] sched/cpuacct: Use __this_cpu_add() instead of this_cpu_ptr()

[Peter Zijlstra]

On Thu, Apr 16, 2020 at 02:53:10PM +0800, Muchun Song wrote:
> There seems to be no difference between the two, but on some
> architectures(e.g. x86_64), there will be optimizations for
> __this_cpu_add(). We can disassemble the code for you to see
> the difference between them on x86_64.
> 
>   1) this_cpu_ptr(ca->cpuusage)->usages[index] += cputime;
> 
>      ffffffff810d7227: add %gs:0x7ef37fa9(%rip),%rax # f1d8 <this_cpu_off>
>      ffffffff810d722f: add %rsi,(%rax)               # %rsi is @cputime
> 
> This result in two add instructions emitted by the compiler.
> 
>   2) __this_cpu_add(ca->cpuusage->usages[index], cputime);
> 
>      ffffffff810d7227: add %rsi,%gs:(%rax)           # %rsi is @cputime
> 
> This result in only one add instruction emitted by the compiler.
> 
> So we have enough reasons to use the __this_cpu_add().

The patch is OK, but I can't take it with such complete nonsense for a
Changelog.

The reason this_cpu_add() and __this_cpu_add() exist and are different
is for different calling context. this_cpu_*() is always safe and
correct, but as you notice, not always optimal. __this_cpu_*() relies on
the caller already having preemption (and or IRQs disabled) to allow for
better code-gen.

Now, the below call-sites have rq->lock taken, and this means preemption
(and IRQs) are indeed disabled, so it is safe to use __this_cpu_*().

> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> ---
>  kernel/sched/cpuacct.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
> index 9fbb103834345..6448b0438ffb2 100644
> --- a/kernel/sched/cpuacct.c
> +++ b/kernel/sched/cpuacct.c
> @@ -347,7 +347,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
>  	rcu_read_lock();
>  
>  	for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
> -		this_cpu_ptr(ca->cpuusage)->usages[index] += cputime;
> +		__this_cpu_add(ca->cpuusage->usages[index], cputime);
>  
>  	rcu_read_unlock();
>  }
> @@ -363,7 +363,7 @@ void cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
>  
>  	rcu_read_lock();
>  	for (ca = task_ca(tsk); ca != &root_cpuacct; ca = parent_ca(ca))
> -		this_cpu_ptr(ca->cpustat)->cpustat[index] += val;
> +		__this_cpu_add(ca->cpustat->cpustat[index], val);
>  	rcu_read_unlock();
>  }
>  
> -- 
> 2.11.0
> 

Re: [External] Re: [PATCH] sched/cpuacct: Use __this_cpu_add() instead of this_cpu_ptr()

[Muchun Song]

On Thu, Apr 16, 2020 at 3:27 PM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Thu, Apr 16, 2020 at 02:53:10PM +0800, Muchun Song wrote:
> > There seems to be no difference between the two, but on some
> > architectures(e.g. x86_64), there will be optimizations for
> > __this_cpu_add(). We can disassemble the code for you to see
> > the difference between them on x86_64.
> >
> >   1) this_cpu_ptr(ca->cpuusage)->usages[index] += cputime;
> >
> >      ffffffff810d7227: add %gs:0x7ef37fa9(%rip),%rax # f1d8 <this_cpu_off>
> >      ffffffff810d722f: add %rsi,(%rax)               # %rsi is @cputime
> >
> > This result in two add instructions emitted by the compiler.
> >
> >   2) __this_cpu_add(ca->cpuusage->usages[index], cputime);
> >
> >      ffffffff810d7227: add %rsi,%gs:(%rax)           # %rsi is @cputime
> >
> > This result in only one add instruction emitted by the compiler.
> >
> > So we have enough reasons to use the __this_cpu_add().
>
> The patch is OK, but I can't take it with such complete nonsense for a
> Changelog.
>
> The reason this_cpu_add() and __this_cpu_add() exist and are different
> is for different calling context. this_cpu_*() is always safe and
> correct, but as you notice, not always optimal. __this_cpu_*() relies on
> the caller already having preemption (and or IRQs disabled) to allow for
> better code-gen.
>
> Now, the below call-sites have rq->lock taken, and this means preemption
> (and IRQs) are indeed disabled, so it is safe to use __this_cpu_*().

Thanks Peter. I will update the changelog.

>
> > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > ---
> >  kernel/sched/cpuacct.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/sched/cpuacct.c b/kernel/sched/cpuacct.c
> > index 9fbb103834345..6448b0438ffb2 100644
> > --- a/kernel/sched/cpuacct.c
> > +++ b/kernel/sched/cpuacct.c
> > @@ -347,7 +347,7 @@ void cpuacct_charge(struct task_struct *tsk, u64 cputime)
> >       rcu_read_lock();
> >
> >       for (ca = task_ca(tsk); ca; ca = parent_ca(ca))
> > -             this_cpu_ptr(ca->cpuusage)->usages[index] += cputime;
> > +             __this_cpu_add(ca->cpuusage->usages[index], cputime);
> >
> >       rcu_read_unlock();
> >  }
> > @@ -363,7 +363,7 @@ void cpuacct_account_field(struct task_struct *tsk, int index, u64 val)
> >
> >       rcu_read_lock();
> >       for (ca = task_ca(tsk); ca != &root_cpuacct; ca = parent_ca(ca))
> > -             this_cpu_ptr(ca->cpustat)->cpustat[index] += val;
> > +             __this_cpu_add(ca->cpustat->cpustat[index], val);
> >       rcu_read_unlock();
> >  }
> >
> > --
> > 2.11.0
> >



-- 
Yours,
Muchun