Laurent Dufour <ldufour@xxxxxxxxxxxxx> writes:
> On 03/05/2022, 17:06:41, Michael Ellerman wrote:
>> Laurent Dufour <ldufour@xxxxxxxxxxxxx> writes:
...
>>> diff --git a/arch/powerpc/kernel/rtas.c b/arch/powerpc/kernel/rtas.c
>>> index 1f42aabbbab3..d7775b8c8853 100644
>>> --- a/arch/powerpc/kernel/rtas.c
>>> +++ b/arch/powerpc/kernel/rtas.c
>>> @@ -49,6 +49,11 @@ void enter_rtas(unsigned long);
>>>
>>> static inline void do_enter_rtas(unsigned long args)
>>> {
>>> + unsigned long msr;
>>> +
>>> + msr = mfmsr();
>>> + BUG_ON(!(msr & MSR_RI));
>>
>> I'm not sure about this.
>>
>> We call RTAS in some low-level places, so if we ever hit this BUG_ON
>> then it might cause us to crash badly, or recursively BUG.
>>
>> A WARN_ON_ONCE() might be safer?
>
> I'm afraid a BUG_ON is required here. Since MSR[RI] is set on RTAS exit so
> if it was not set when calling RTAS, that's a real issue and should
> generate unexpected behaviour.
>
> Do you have places in mind where RTAS could be called with !MSR[RI]?
The main one I can think of is if someone is using
CONFIG_UDBG_RTAS_CONSOLE, then udbg_rtascon_putc() is wired up as
udbg_putc() and that might be called from anywhere, including xmon.
There's also RTAS calls in low-level xics interrupt code, that might get
called during panic/crash.
I don't expect any of those places to be called with MSR[RI] unset, but
I'm worried that if we're already crashing and for some reason MSR[RI]
is unset, then that BUG_ON will just make things worse.
eg. imagine taking a BUG_ON() for every character we try to print as
part of an oops.
Admittedly CONFIG_UDBG_RTAS_CONSOLE is old and probably not used much
anymore, but I'm still a bit paranoid :)
cheers
