Re: [PATCH v2 2/5] preempt/dynamic: Introduce preempt mode accessors

[Valentin Schneider]

On 16/11/21 14:29, Christophe Leroy wrote:
> Le 10/11/2021 à 21:24, Valentin Schneider a écrit :
>> CONFIG_PREEMPT{_NONE, _VOLUNTARY} designate either:
>> o The build-time preemption model when !PREEMPT_DYNAMIC
>> o The default boot-time preemption model when PREEMPT_DYNAMIC
>>
>> IOW, using those on PREEMPT_DYNAMIC kernels is meaningless - the actual
>> model could have been set to something else by the "preempt=foo" cmdline
>> parameter.
>>
>> Introduce a set of helpers to determine the actual preemption mode used by
>> the live kernel.
>>
>> Suggested-by: Marco Elver <elver@xxxxxxxxxx>
>> Signed-off-by: Valentin Schneider <valentin.schneider@xxxxxxx>
>> ---
>>   include/linux/sched.h | 16 ++++++++++++++++
>>   kernel/sched/core.c   | 11 +++++++++++
>>   2 files changed, 27 insertions(+)
>>
>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>> index 5f8db54226af..0640d5622496 100644
>> --- a/include/linux/sched.h
>> +++ b/include/linux/sched.h
>> @@ -2073,6 +2073,22 @@ static inline void cond_resched_rcu(void)
>>   #endif
>>   }
>>
>> +#ifdef CONFIG_PREEMPT_DYNAMIC
>> +
>> +extern bool is_preempt_none(void);
>> +extern bool is_preempt_voluntary(void);
>> +extern bool is_preempt_full(void);
>
> Those are trivial tests supposed to be used in fast pathes. They should
> be static inlines in order to minimise the overhead.
>
>> +
>> +#else
>> +
>> +#define is_preempt_none() IS_ENABLED(CONFIG_PREEMPT_NONE)
>> +#define is_preempt_voluntary() IS_ENABLED(CONFIG_PREEMPT_VOLUNTARY)
>> +#define is_preempt_full() IS_ENABLED(CONFIG_PREEMPT)
>
> Would be better to use static inlines here as well instead of macros.
>

I realize I stripped all ppc folks from the cclist after dropping the ppc
snippet, but you guys might still be interested - my bad. That's done in
v3:

https://lore.kernel.org/lkml/20211112185203.280040-1-valentin.schneider@xxxxxxx/

>> +
>> +#endif
>> +
>> +#define is_preempt_rt() IS_ENABLED(CONFIG_PREEMPT_RT)
>> +
>>   /*
>>    * Does a critical section need to be broken due to another
>>    * task waiting?: (technically does not depend on CONFIG_PREEMPTION,
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 97047aa7b6c2..9db7f77e53c3 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -6638,6 +6638,17 @@ static void __init preempt_dynamic_init(void)
>>      }
>>   }
>>
>> +#define PREEMPT_MODE_ACCESSOR(mode) \
>> +	bool is_preempt_##mode(void)						 \
>> +	{									 \
>> +		WARN_ON_ONCE(preempt_dynamic_mode == preempt_dynamic_undefined); \
>
> Not sure using WARN_ON is a good idea here, as it may be called very
> early, see comment on powerpc patch.

Bah, I was gonna say that you *don't* want users of is_preempt_*() to be
called before the "final" preemption model is set up (such users would need
to make use of static_calls), but I realize there's a debug interface to
flip the preemption model at will... Say an initcall sees
is_preempt_voluntary() and sets things up accordingly, and then the debug
knob switches to preempt_full. I don't think there's much we can really do
here though :/

>
>> +		return preempt_dynamic_mode == preempt_dynamic_##mode;		 \
>> +	}
>
> I'm not sure that's worth a macro. You only have 3 accessors, 2 lines of
> code each. Just define all 3 in plain text.
>
> CONFIG_PREEMPT_DYNAMIC is based on using strategies like static_calls in
> order to minimise the overhead. For those accessors you should use the
> same kind of approach and use things like jump_labels in order to not
> redo the test at each time and minimise overhead as much as possible.
>

That's a valid point, though the few paths that need patching up and don't
make use of static calls already (AFAICT the ppc irq path I was touching in
v2 needs to make use of irqentry_exit_cond_resched()) really seem like
slow-paths.

>> +
>> +PREEMPT_MODE_ACCESSOR(none)
>> +PREEMPT_MODE_ACCESSOR(voluntary)
>> +PREEMPT_MODE_ACCESSOR(full)
>> +
>>   #else /* !CONFIG_PREEMPT_DYNAMIC */
>>
>>   static inline void preempt_dynamic_init(void) { }
>>