Re: [PATCH] rcu: Remove redundant rcu_read_lock/unlock() in spin_lock critical sections

[Hillf Danton <hdanton@sina.com>]

On Fri, 12 Sep 2025 20:33:31 -0400 Waiman Long wrote:
>On 9/12/25 5:35 PM, Sebastian Andrzej Siewior wrote:
>> On 2025-09-12 17:13:09 [-0400], Waiman Long wrote:
>>> On 9/12/25 2:50 AM, pengdonglin wrote:
>>>> From: pengdonglin <pengdonglin@xiaomi.com>
>>>>
>>>> When CONFIG_PREEMPT_RT is disabled, spin_lock*() operations implicitly
>>>> disable preemption, which provides RCU read-side protection. When
>>>> CONFIG_PREEMPT_RT is enabled, spin_lock*() implementations internally
>>>> manage RCU read-side critical sections.
>>> I have some doubt about your claim that disabling preemption provides RCU
>>> read-side protection. It is true for some flavors but probably not all. I do
>>> know that disabling interrupt will provide RCU read-side protection. So for
>>> spin_lock_irq*() calls, that is valid. I am not sure about spin_lock_bh(),
>>> maybe it applies there too. we need some RCU people to confirm.
>> The claim is valid since Paul merged the three flavours we had. Before
>> that preempt_disable() (and disabling irqs) would match
>> rcu_read_lock_sched(). rcu_read_lock() and rcu_read_lock_bh() were
>> different in terms of grace period and clean up.
>> So _now_ we could remove it if it makes things easier.
>
> Thanks for the clarification.
> 
> In this case, I think the patch description should mention the commit 
> that unify the 3 RCU flavors to make sure that this patch won't be 
> accidentally backport'ed to an older kernel without the necessary 
> prerequisite commit(s).

This change also affects the dereference helpers.

#define rcu_dereference_check(p, c) \
	__rcu_dereference_check((p), __UNIQUE_ID(rcu), \
				(c) || rcu_read_lock_held(), __rcu)