> On Jun 18, 2020, at 12:02 PM, Peter Zijlstra <pet...@infradead.org> wrote:
>
> On Thu, Jun 18, 2020 at 11:36:53AM -0700, Andy Lutomirski wrote:
>
>> I wasn't imagining going far down the rabbit hole at all -- I think
>> that, at most, we should cover the path for when the fault wasn't a
>> BUG/WARN in the first place. I admit that, for #UD in particular,
>> this isn't a big deal, but if it were a different vector, this could
>> matter.
>
> Right, so there's 3 cases for ud2:
>
> - WARN; ud2, bug_entry, recovers
> - BUG; ud2, bug_entry, dies
> - UBSAN; ud2, !bug_entry, dies
4. The #UD matches an extable entry. I don’t know whether this ever happens for
real.
The failure is still a bit farfetched: we’d need an extable to hit in an
inconsistent state where we blow up due to a lack of entry handling.
>
> Nothing else should be generating ud2 instructions, any other #UD goes
> into handle_invalid_op() -> do_error_trap() -> ... -> die().
>
> [ while there, we should probably restructure do_trap() to have
> cond_local_irq_enable() _after_ do_trap_no_signal(). ]
>
> We could probably change is_valid_bugaddr() to not use
> probe_kernel_address(), because if it couldn't read the instruction,
> we'd not be getting #UD in the first place.
>
> If we've gotten rid of probe_kernel_address() we can noinstr/inline that
> and then we can only call into report_bug() IFF ud2.
>
> Does that make things 'better' ? This can only go realy bad if there's a
> 1 byte instruction that triggers #UD, but I think that was ruled out.
>
>
> ---
> diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
> index c26751e303f1..275a621f1aff 100644
> --- a/arch/x86/kernel/traps.c
> +++ b/arch/x86/kernel/traps.c
> @@ -91,10 +91,7 @@ int is_valid_bugaddr(unsigned long addr)
> if (addr < TASK_SIZE_MAX)
> return 0;
>
> - if (probe_kernel_address((unsigned short *)addr, ud))
> - return 0;
> -
> - return ud == INSN_UD0 || ud == INSN_UD2;
> + return *(unsigned short *)addr == INSN_UD2;
> }
I’m okay with this, at least until we get PKRS or kernel XO memory. But
probe_kernel_addr() would be wrong then, too. We need probe_kernel_text().
But I think you might need some IRQ fiddling. With your patch, a WARN with IRQs
on will execute the printk code with IRQs off without lockstep handling, and an
appropriately configured debugging kernel may get a recursive splat. Or if irq
tracing somehow notices that IRQs got turned off, the warning recovery might
return back to an IF=1 context with IRQs traced as off.
So maybe also do an untraced cond_local_irq_enable()? After all, if we’re
trying to report a bug from IRQs on, it should be okay to have IRQs on while
reporting it. It might even work better than having IRQs off.
>
> static nokprobe_inline int
> @@ -220,15 +217,17 @@ static noinstr bool handle_bug(struct pt_regs *regs)
> {
> bool handled = false;
>
> - /*
> - * All lies, just get the WARN/BUG out.
> - */
> - instrumentation_begin();
> - if (report_bug(regs->ip, regs) == BUG_TRAP_TYPE_WARN) {
> - regs->ip += LEN_UD2;
> - handled = true;
> + if (is_valid_bugaddr(regs->ip)) {
> + /*
> + * All lies, just get the WARN/BUG out.
> + */
> + instrumentation_begin();
> + if (report_bug(regs->ip, regs) == BUG_TRAP_TYPE_WARN) {
> + regs->ip += LEN_UD2;
> + handled = true;
> + }
> + instrumentation_end();
> }
> - instrumentation_end();
>
> return handled;
> }
