Re: [patch V4 part 1 29/36] x86/mce: Send #MC singal from task work

[Andy Lutomirski <luto@amacapital.net>]

> On May 14, 2020, at 9:03 AM, Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote:
> 
> ----- On May 14, 2020, at 10:17 AM, Borislav Petkov bp@alien8.de wrote:
> 
>> + Tony.
>> 
>>> On Tue, May 05, 2020 at 03:16:31PM +0200, Thomas Gleixner wrote:
>>> From: Peter Zijlstra <peterz@infradead.org>
>>> 
>>> Convert #MC over to using task_work_add(); it will run the same code
>>> slightly later, on the return to user path of the same exception.
>>> 
>>> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
>>> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
>>> Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
>>> ---
>>> arch/x86/kernel/cpu/mce/core.c |   56 ++++++++++++++++++++++-------------------
>>> include/linux/sched.h          |    6 ++++
>>> 2 files changed, 37 insertions(+), 25 deletions(-)
>> 
>> I like this:
>> 
>> Reviewed-by: Borislav Petkov <bp@suse.de>
> 
> What I am not fully grasping here is whether this patch preserves the instruction
> pointer (and possibly other relevant information for siginfo_t) triggering the
> exception in a scenario where we have:
> 
> - #MC triggered, queuing task work,
> - unrelated signal happens to be delivered to task,
> - exit to usermode loop handles do_signal first,
> - then it runs task work.

If anyone wants to ponder this, I suspect that we have lots of delightful bugs in our handling of cr2, trapnr, and error_code in signals. We should move them to the sigcontext, at least in kernel, and fix up ucontext when we deliver the signal.  The current code can’t possibly be correct.