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

[Harry Yoo <harry.yoo@oracle.com>]

On Tue, Nov 11, 2025 at 11:04:09AM +0800, Qi Zheng wrote:
> 
> On 11/11/25 12:47 AM, Shakeel Butt wrote:
> > On Mon, Nov 10, 2025 at 02:43:21PM +0900, Harry Yoo wrote:
> > > On Mon, Nov 10, 2025 at 12:30:06PM +0800, Qi Zheng wrote:
> > > > > Maybe we could make it safe against re-entrant IRQ handlers by using
> > > > > read-modify-write operations?
> > > > 
> > > > Isn't it because of the RMW operation that we need to use IRQ to
> > > > guarantee atomicity? Or have I misunderstood something?
> > > 
> > > I meant using atomic operations instead of disabling IRQs, like, by
> > > using this_cpu_add() or cmpxchg() instead.
> > 
> > We already have mod_node_page_state() which is safe from IRQs and is
> > optimized to not disable IRQs for archs with HAVE_CMPXCHG_LOCAL which
> > includes x86 and arm64.
> 
> However, in the !CONFIG_HAVE_CMPXCHG_LOCAL case, mod_node_page_state()
> still calls local_irq_save(). Is this feasible in the PREEMPT_RT kernel?

Hmm I was going to say it's necessary, but AFAICT we don't allocate
or free memory in hardirq context on PREEMPT_RT (that's the policy)
and so I'd say it's not necessary to disable IRQs.

Sounds like we still want to disable IRQs only on !PREEMPT_RT on
such architectures?

Not sure how seriously do PREEMPT_RT folks care about architectures
without HAVE_CMPXCHG_LOCAL. (riscv and loongarch have ARCH_SUPPORTS_RT
but doesn't have HAVE_CMPXCHG_LOCAL).

If they do care, this can be done as a separate patch series because
we already call local_irq_{save,restore}() in many places in mm/vmstat.c
if the architecture doesn't not have HAVE_CMPXCHG_LOCAL.

> > Let me send the patch to cleanup the memcg code which uses
> > __mod_node_page_state.

-- 
Cheers,
Harry / Hyeonggon