Re: [PATCH v4 1/2] rcu: Eliminate rcu_state.nocb_is_setup

From: Leizhen (ThunderTown)
Date: Tue Aug 23 2022 - 11:09:52 EST




On 2022/8/23 0:34, Paul E. McKenney wrote:
> On Wed, Aug 17, 2022 at 09:42:52AM +0800, Zhen Lei wrote:
>> 'rcu_state.nocb_is_setup' is initialized to true only if 'rcu_nocb_mask'
>> successfully allocates memory. So it can be replaced by
>> 'cpumask_available(rcu_nocb_mask)'. More importantly, the latter is more
>> intuitive, and it has been used in several places.
>>
>> Signed-off-by: Zhen Lei <thunder.leizhen@xxxxxxxxxx>
>
> One of the implementations of cpumask_available() does indeed check
> for NULL. But here is the other one:
>
> static inline bool cpumask_available(cpumask_var_t mask)
> {
> return true;
> }

Thanks for the heads-up.

>
> So I have to ask... In a kernel built with CONFIG_CPUMASK_OFFSTACK=n,
> will this change really work?

Yes, I run the test cases on arm64, which does not turn on option
CONFIG_CPUMASK_OFFSTACK by default.

I've listed a combination of build options in the cover-letter. In
this case, cpumask_empty(rcu_nocb_mask) is true.
----
CONFIG_NO_HZ_FULL=n, CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=n, cmdline without rcu_nocbs
[ 0.000000] rcu: Offload RCU callbacks from CPUs: (none).
----

>
> Another important question is "Do all of the existing uses of
> cpumask_available() really work?" Yes, I do believe that they do

The only functional change caused by this patch is:
For clarity, CONFIG_RCU_NOCB_CPU=y and CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=n are
omitted from the following conditions.

When
CONFIG_NO_HZ_FULL=n and boot cmdline without 'rcu_nocbs='.
or
CONFIG_NO_HZ_FULL=y and boot cmdline without 'nohz_full='.

The rdp->nocb_gp_kthread and rdp->nocb_cb_kthread threads are still created.

But we have provided EXPORT_SYMBOL_GPL functions rcu_nocb_cpu_deoffload() and
rcu_nocb_cpu_offload(), which can dynamically modify 'rcu_nocb_mask'. So it
seems appropriate to prepare these threads in advance. Of course, it looks like
only 'rcutorture' currently uses these two functions now.

Otherwise, we can do some optimization:
If none of the CPUs in a 'nocb_gp' group is marked in rcu_nocb_mask, this
grouping does not need to create corresponding threads "rcuog/%d" and "rcuo%c/%d".

And in rcu_init_nohz():
- if (!rcu_state.nocb_is_setup)
+ if (!cpumask_available(rcu_nocb_mask) || cpumask_empty(rcu_nocb_mask))
return;



> work, but it would be good to get another set of eyes on that code.
> "All software developers are blind!" ;-)
>
> Thanx, Paul
>
>> ---
>> kernel/rcu/tree.h | 1 -
>> kernel/rcu/tree_nocb.h | 8 +++-----
>> 2 files changed, 3 insertions(+), 6 deletions(-)
>>
>> diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
>> index d4a97e40ea9c3e2..06f659c63d2d192 100644
>> --- a/kernel/rcu/tree.h
>> +++ b/kernel/rcu/tree.h
>> @@ -375,7 +375,6 @@ struct rcu_state {
>> arch_spinlock_t ofl_lock ____cacheline_internodealigned_in_smp;
>> /* Synchronize offline with */
>> /* GP pre-initialization. */
>> - int nocb_is_setup; /* nocb is setup from boot */
>> };
>>
>> /* Values for rcu_state structure's gp_flags field. */
>> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
>> index 0a5f0ef41484518..ff763e7dc53551f 100644
>> --- a/kernel/rcu/tree_nocb.h
>> +++ b/kernel/rcu/tree_nocb.h
>> @@ -69,7 +69,6 @@ static int __init rcu_nocb_setup(char *str)
>> cpumask_setall(rcu_nocb_mask);
>> }
>> }
>> - rcu_state.nocb_is_setup = true;
>> return 1;
>> }
>> __setup("rcu_nocbs", rcu_nocb_setup);
>> @@ -1215,7 +1214,7 @@ void __init rcu_init_nohz(void)
>> struct rcu_data *rdp;
>>
>> #if defined(CONFIG_RCU_NOCB_CPU_DEFAULT_ALL)
>> - if (!rcu_state.nocb_is_setup) {
>> + if (!cpumask_available(rcu_nocb_mask)) {
>> need_rcu_nocb_mask = true;
>> offload_all = true;
>> }
>> @@ -1235,10 +1234,9 @@ void __init rcu_init_nohz(void)
>> return;
>> }
>> }
>> - rcu_state.nocb_is_setup = true;
>> }
>>
>> - if (!rcu_state.nocb_is_setup)
>> + if (!cpumask_available(rcu_nocb_mask))
>> return;
>>
>> #if defined(CONFIG_NO_HZ_FULL)
>> @@ -1299,7 +1297,7 @@ static void rcu_spawn_cpu_nocb_kthread(int cpu)
>> struct task_struct *t;
>> struct sched_param sp;
>>
>> - if (!rcu_scheduler_fully_active || !rcu_state.nocb_is_setup)
>> + if (!rcu_scheduler_fully_active || !cpumask_available(rcu_nocb_mask))
>> return;
>>
>> /* If there already is an rcuo kthread, then nothing to do. */
>> --
>> 2.25.1
>>
> .
>

--
Regards,
Zhen Lei