Re: [PATCH v17 0/3] Improve proc RSS accuracy

[Mathieu Desnoyers]

On 2026-02-27 08:11, Heiko Carstens wrote:
> On Thu, Feb 26, 2026 at 06:12:01PM -0700, Nathan Chancellor wrote:
> > Hi Mathieu,
> >
> > On Thu, Feb 26, 2026 at 02:38:04PM -0500, Mathieu Desnoyers wrote:
> > > I've successfully booted a defconfig s390x next-20260226 kernel in qemu
> > > with 1 and 4 CPUs, and within a nested s390x VM on 2 cpus.
> > >
> > > I guess I'll really need more info about your specific .config and
> > > command line args to help further.
>
> On s390 cpumask_set_cpu(0, mm_cpumask(&init_mm)); in arch_mm_preinit() writes
> out-of-bounds into swap_attrs[] overwriting the terminating NULL.
>
> This seems to happen because the return value of get_rss_stat_items_size() is
> larger than PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE:
>
> PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE: 18688
> get_rss_stat_items_size(): 21504
>
> Here I stopped looking further into this. I guess you will figure out
> immediately what's wrong :)

Indeed!

So in get_rss_stat_items_size() we have:

static inline size_t get_rss_stat_items_size(void)
{
        return percpu_counter_tree_items_size() * NR_MM_COUNTERS;
}

And just above:

#define MM_STRUCT_FLEXIBLE_ARRAY_INIT                                                                   \
{                                                                                                       \
        [0 ... PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE + sizeof(cpumask_t) + MM_CID_STATIC_SIZE - 1] = 0  \
}

Which fails to account for NR_MM_COUNTERS. Does the following fix your issue ?

#define MM_STRUCT_FLEXIBLE_ARRAY_INIT                                                                   \
{                                                                                                       \
        [0 ... (PERCPU_COUNTER_TREE_ITEMS_STATIC_SIZE * NR_MM_COUNTERS) + sizeof(cpumask_t) + MM_CID_STATIC_SIZE - 1] = 0  \
}

It would only cause issues when nr_cpu_ids grows closer to NR_CPUS, which explains
why I could not reproduce it locally.

Thanks,

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com