[PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Jisheng Zhang]

cpuidle_ops is initialized once by arm_cpuidle_read_ops() during
initialization, and thereafter is mostly read in arm_cpuidle_suspend()

The fact that it is mostly read and not written to makes it candidates
for __read_mostly declarations.

Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
---
 arch/arm/kernel/cpuidle.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
index 7dccc96..762e0929 100644
--- a/arch/arm/kernel/cpuidle.c
+++ b/arch/arm/kernel/cpuidle.c
@@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
 static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
 	__used __section(__cpuidle_method_of_table_end);
 
-static struct cpuidle_ops cpuidle_ops[NR_CPUS];
+static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;
 
 /**
  * arm_cpuidle_simple_enter() - a wrapper to cpu_do_idle()
-- 
2.8.1

Re: [PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Arnd Bergmann]

On Wednesday, August 10, 2016 4:49:57 PM CEST Jisheng Zhang wrote:
> diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
> index 7dccc96..762e0929 100644
> --- a/arch/arm/kernel/cpuidle.c
> +++ b/arch/arm/kernel/cpuidle.c
> @@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
>  static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
>         __used __section(__cpuidle_method_of_table_end);
>  
> -static struct cpuidle_ops cpuidle_ops[NR_CPUS];
> +static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;

Should this perhaps be percpu data instead?

	Arnd

Re: [PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Jisheng Zhang]

Dear Arnd,

On Wed, 10 Aug 2016 10:57:57 +0200 Arnd Bergmann wrote:

> On Wednesday, August 10, 2016 4:49:57 PM CEST Jisheng Zhang wrote:
> > diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
> > index 7dccc96..762e0929 100644
> > --- a/arch/arm/kernel/cpuidle.c
> > +++ b/arch/arm/kernel/cpuidle.c
> > @@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
> >  static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
> >         __used __section(__cpuidle_method_of_table_end);
> >  
> > -static struct cpuidle_ops cpuidle_ops[NR_CPUS];
> > +static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;  
> 
> Should this perhaps be percpu data instead?
> 

Per my understanding, percpu is used for those vars with normal read/write
frequency, while the cpuidle_ops is read mostly, so IMHO, __read_mostly
is suitable, what do you think?

Thanks,
Jisheng

Re: [PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Jisheng Zhang]

On Wed, 10 Aug 2016 16:49:57 +0800 Jisheng Zhang wrote:

> cpuidle_ops is initialized once by arm_cpuidle_read_ops() during
> initialization, and thereafter is mostly read in arm_cpuidle_suspend()
> 
> The fact that it is mostly read and not written to makes it candidates
> for __read_mostly declarations.
> 
> Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
> ---
>  arch/arm/kernel/cpuidle.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
> index 7dccc96..762e0929 100644
> --- a/arch/arm/kernel/cpuidle.c
> +++ b/arch/arm/kernel/cpuidle.c
> @@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
>  static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
>  	__used __section(__cpuidle_method_of_table_end);
>  
> -static struct cpuidle_ops cpuidle_ops[NR_CPUS];
> +static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;


oops, I just found that "arm: apply more __ro_after_init" from Kees
marked cpuidle_ops as __ro_after_init, that also meet my need, so
let's drop my patch.

Thanks,
Jisheng

Re: [PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Arnd Bergmann]

On Wednesday, August 10, 2016 5:19:26 PM CEST Jisheng Zhang wrote:
> Dear Arnd,
> 
> On Wed, 10 Aug 2016 10:57:57 +0200 Arnd Bergmann wrote:
> 
> > On Wednesday, August 10, 2016 4:49:57 PM CEST Jisheng Zhang wrote:
> > > diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
> > > index 7dccc96..762e0929 100644
> > > --- a/arch/arm/kernel/cpuidle.c
> > > +++ b/arch/arm/kernel/cpuidle.c
> > > @@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
> > >  static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
> > >         __used __section(__cpuidle_method_of_table_end);
> > >  
> > > -static struct cpuidle_ops cpuidle_ops[NR_CPUS];
> > > +static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;  
> > 
> > Should this perhaps be percpu data instead?
> > 
> 
> Per my understanding, percpu is used for those vars with normal read/write
> frequency, while the cpuidle_ops is read mostly, so IMHO, __read_mostly
> is suitable, what do you think?

You are right, __read_mostly is better than the normal .data section here,
but percpu is also better than .data because it saves a little memory
on machines that have few present CPUs than CONFIG_NR_CPUS.

So both have their advantages, we just need to pick a preference.

Actually __ro_after_init would be even better than __read_mostly here
I think, as this is only updated in an __init function. I guess
using that would have the added security advantage of preventing
an attacker from writing to the function pointers when they
find a way to overflow an access in the percpu data section.

	Arnd

Re: [PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Arnd Bergmann]

On Wednesday, August 10, 2016 5:54:03 PM CEST Jisheng Zhang wrote:
> On Wed, 10 Aug 2016 16:49:57 +0800 Jisheng Zhang wrote:
> 
> > cpuidle_ops is initialized once by arm_cpuidle_read_ops() during
> > initialization, and thereafter is mostly read in arm_cpuidle_suspend()
> > 
> > The fact that it is mostly read and not written to makes it candidates
> > for __read_mostly declarations.
> > 
> > Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
> > ---
> >  arch/arm/kernel/cpuidle.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
> > index 7dccc96..762e0929 100644
> > --- a/arch/arm/kernel/cpuidle.c
> > +++ b/arch/arm/kernel/cpuidle.c
> > @@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
> >  static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
> >       __used __section(__cpuidle_method_of_table_end);
> >  
> > -static struct cpuidle_ops cpuidle_ops[NR_CPUS];
> > +static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;
> 
> 
> oops, I just found that "arm: apply more __ro_after_init" from Kees
> marked cpuidle_ops as __ro_after_init, that also meet my need, so
> let's drop my patch.
> 
> 

Ok, just saw this reply and that matches up with what I just
wrote as well.

Thanks,

	Arnd

Re: [PATCH] ARM: cpuidle: declare cpuidle_ops __read_mostly

[Jisheng Zhang]

Dear Arnd,

On Wed, 10 Aug 2016 12:47:21 +0200 Arnd Bergmann wrote:

> On Wednesday, August 10, 2016 5:19:26 PM CEST Jisheng Zhang wrote:
> > Dear Arnd,
> > 
> > On Wed, 10 Aug 2016 10:57:57 +0200 Arnd Bergmann wrote:
> >   
> > > On Wednesday, August 10, 2016 4:49:57 PM CEST Jisheng Zhang wrote:  
> > > > diff --git a/arch/arm/kernel/cpuidle.c b/arch/arm/kernel/cpuidle.c
> > > > index 7dccc96..762e0929 100644
> > > > --- a/arch/arm/kernel/cpuidle.c
> > > > +++ b/arch/arm/kernel/cpuidle.c
> > > > @@ -19,7 +19,7 @@ extern struct of_cpuidle_method __cpuidle_method_of_table[];
> > > >  static const struct of_cpuidle_method __cpuidle_method_of_table_sentinel
> > > >         __used __section(__cpuidle_method_of_table_end);
> > > >  
> > > > -static struct cpuidle_ops cpuidle_ops[NR_CPUS];
> > > > +static struct cpuidle_ops cpuidle_ops[NR_CPUS] __read_mostly;    
> > > 
> > > Should this perhaps be percpu data instead?
> > >   
> > 
> > Per my understanding, percpu is used for those vars with normal read/write
> > frequency, while the cpuidle_ops is read mostly, so IMHO, __read_mostly
> > is suitable, what do you think?  
> 
> You are right, __read_mostly is better than the normal .data section here,
> but percpu is also better than .data because it saves a little memory
> on machines that have few present CPUs than CONFIG_NR_CPUS.
> 
> So both have their advantages, we just need to pick a preference.
> 
> Actually __ro_after_init would be even better than __read_mostly here
> I think, as this is only updated in an __init function. I guess
> using that would have the added security advantage of preventing
> an attacker from writing to the function pointers when they
> find a way to overflow an access in the percpu data section.
> 

Got it, thanks for the detailed explanations. 

And I think the answer to questions:

http://lists.infradead.org/pipermail/linux-arm-kernel/2016-August/448057.html

and

http://lists.infradead.org/pipermail/linux-arm-kernel/2016-August/448059.html

are both "yes"

Thanks,
Jisheng