Re: [RFC PATCH 0/3] Enable kprobe to monitor sdei event handler

[James Morse]

Hi Xiongfeng Wang,

On 12/04/2019 13:04, Xiongfeng Wang wrote:
> When I use kprobe to monitor a sdei event handler,

Don't do this! SDEI is like an NMI, it isn't safe to kprobe it as it can interrupt the
kprobe code, causing it become re-entrant.


> the CPU will hang. It's
> because when I probe the event handler, the instruction will be replaced with 
> brk instruction and brk exception is unmaskable. But 'vbar_el1' contains 
> 'tramp_vectors' in '_sdei_handler' when SDEI events interrupt userspace, so
> we will go to the wrong place if brk exception happens.

This was lucky! Its even more fun if the SDEI event interrupted a guest: the kvm vectors
will give you a hyp-panic.

The __kprobes and NOKPROBE_SYMBOL() litter should stop you doing this.


> I notice that 'ghes_sdei_normal_callback' call several funtions that are not
> marked as 'nokprobe'.

Bother. We should probably blacklist those too, its not safe.


> So I was wondering if we can enable kprobe in '_sdei_handler'.

I don't think this can be done safely.


If you need to monitor your SDEI event handler you can just use printk(). Once nmi_enter()
has been called these are safe as they stash data in a per-cpu buffer. The SDEI handler
will exit via the IRQ vector if it can, which will cause this buffer to be flushed to the
console in a timely manner.


Why do you need to kprobe an NMI handler?



Thanks!

James