Re: [RFC PATCH] x86/mce/inject: Add sanity check in inject_mce()

[Ingo Molnar <mingo@kernel.org>]

* Borislav Petkov <bp@alien8.de> wrote:

> On Thu, May 29, 2025 at 11:32:56AM +0800, Xiongfeng Wang wrote:
> > syzbot reported the following kernel panic:
> > 
> > [  306.335489][ T3298] mce: CPUs not responding to MCE broadcast (may include false positives): 1-3
> > [  306.336332][ T3298] Kernel panic - not syncing: Timeout: Not all CPUs entered broadcast exception handler
> > [  306.337786][ T3298] Kernel Offset: 0x17400000 from 0xffffffff81000000 (relocation range: 0xffffffff80000000-0xffffffffbfffffff)
> > 
> > The syzkaller test didn't set 'MCJ_IRQ_BROADCAST | MCJ_NMI_BROADCAST' in
> 
> That's because you're not supposed to inject MCEs with syzkaller 
> doing rando poking at the interface but someone with a brain should 
> use it.

Uhm, avoiding a hard kernel crash in case of a tooling or human error:

  [  306.336332][ T3298] Kernel panic - not syncing: Timeout: Not all CPUs entered broadcast exception handler

... is in fact a proper interface robustness fix, especially if it's so 
incredibly simple as in this case. (The comment might need a bit of 
work but I digress.)

Injection can disrupt the kernel, but it shouldn't be able to trivially 
crash it in this case.

Improving MCE injection robustness might also help syzbot uncover 
*other* kernel bugs, because it can continue testing via the injection 
framework. But your proposed solution:

> Please point your syzkaller somewhere else.

... is how real kernel bugs might be hidden ...

Thanks,

	Ingo