Re: [PATCH v1 04/26] mm: vmscan: refactor move_folios_to_lru()

[Harry Yoo]

On Fri, Nov 07, 2025 at 10:32:52PM -0800, Shakeel Butt wrote:
> On Fri, Nov 07, 2025 at 10:20:57PM +0900, Harry Yoo wrote:
> > 
> > Although it's mentioned in the locking documentation, I'm afraid that
> > local_lock is not the right interface to use here. Preemption will be
> > disabled anyway (on both PREEMPT_RT and !PREEMPT_RT) when the stats are
> > updated (in __mod_node_page_state()).
> > 
> > Here we just want to disable IRQ only on !PREEMPT_RT (to update
> > the stats safely).
> 
> I don't think there is a need to disable IRQs. There are three stats
> update functions called in that hunk.
> 
> 1) __mod_lruvec_state
> 2) __count_vm_events
> 3) count_memcg_events
> 
> count_memcg_events() can be called with IRQs. __count_vm_events can be
> replaced with count_vm_events.

Right.

> For __mod_lruvec_state, the
> __mod_node_page_state() inside needs preemption disabled.

The function __mod_node_page_state() disables preemption.
And there's a comment in __mod_zone_page_state():

> /*                                                                      
>  * Accurate vmstat updates require a RMW. On !PREEMPT_RT kernels,           
>  * atomicity is provided by IRQs being disabled -- either explicitly    
>  * or via local_lock_irq. On PREEMPT_RT, local_lock_irq only disables   
>  * CPU migrations and preemption potentially corrupts a counter so          
>  * disable preemption.                                                  
>  */                                                                     
> preempt_disable_nested();

We're relying on IRQs being disabled on !PREEMPT_RT.

Maybe we could make it safe against re-entrant IRQ handlers by using
read-modify-write operations?

-- 
Cheers,
Harry / Hyeonggon

> Easy way would be to just disable/enable preemption instead of IRQs.
> Otherwise go a bit more fine-grained approach i.e. replace
> __count_vm_events with count_vm_events and just disable preemption
> across __mod_node_page_state().