Re: [PATCH 2/2] x86/MCE: Add command line option to extend MCE Records pool

[Yazen Ghannam]

On 2/11/2024 6:14 AM, Borislav Petkov wrote:
> On Sat, Feb 10, 2024 at 03:15:26PM -0600, Naik, Avadhut wrote:
> > IIUC, you wouldn't want to extend the pool through late_initcall().
> > Instead, you would want for memory to be allocated (on the heap) and
> > size of the pool to be set at the very beginning i.e. when the pool
> > is created (~2 seconds, according to dmesg timestamps).
> >
> > Please correct me if I have understood wrong.
>
> Nah, you got it right. I went, looked and realized that we have to do
> this early dance because we have no allocator yet. And we can't move
> this gen_pool allocation to later, when we *do* have an allocator
> because MCA is up and logging already.
>
> But your extending approach doesn't fly in all cases either:
>
> gen_pool_add->gen_pool_add_virt->gen_pool_add_owner
>
> it grabs the pool->lock spinlock and adds to &pool->chunks while, at the
> exact same time, gen_pool_alloc(), in *NMI* context iterates over that
> same &pool->chunks in the case we're logging an MCE at exact that same
> time when you're extending the buffer.
>
> And Tony already said that in the thread you're quoting:
>
> https://lore.kernel.org/linux-edac/SJ1PR11MB60832922E4D036138FF390FAFCD7A@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
>
> So no, that doesn't work either.

I'm confused why it won't work.

X86 has ARCH_HAVE_NMI_SAFE_CMPXCHG. I expect atomics/caches will still
work in interrupt or #MC context. If not, then we'd have a fatal error
that causes a hardware reset or a kernel panic before we get to logging,
I think.

Or is the issue when running on the same CPU? In this case, either
&pool->chunks was updated before taking the #MC, so the extra memory
is there and can be used. Or it wasn't updated, so the extra memory is
not available during the #MC which is the same behavior as now.

I need to look more at the genpool code, but I thought I'd ask too.

Thanks,
Yazen