Re: [PATCH v3 28/46] kmsan: entry: handle register passing from uninstrumented code

[Thomas Gleixner <tglx@linutronix.de>]

On Fri, May 06 2022 at 19:41, Alexander Potapenko wrote:
> On Fri, May 6, 2022 at 6:14 PM Thomas Gleixner <tglx@linutronix.de> wrote:
>> sysvec_apic_timer_interrupt() invokes irqentry_enter() _before_
>> set_irq_regs() and irqentry_enter() unpoisons @reg.
>>
>> Confused...
>
> As far as I can tell in this case sysvect_apic_timer_interrupt() is
> called by the following code in arch/x86/kernel/idt.c:
>
>   INTG(LOCAL_TIMER_VECTOR,                asm_sysvec_apic_timer_interrupt),
>
> , which does not use IDTENTRY_SYSVEC framework and thus does not call
> irqentry_enter().

  asm_sysvec_apic_timer_interrupt != sysvec_apic_timer_interrupt

arch/x86/kernel/apic/apic.c:
DEFINE_IDTENTRY_SYSVEC(sysvec_apic_timer_interrupt)
{
        ....

#define DEFINE_IDTENTRY_SYSVEC(func)					\
static void __##func(struct pt_regs *regs);				\
									\
__visible noinstr void func(struct pt_regs *regs)			\
{									\
	irqentry_state_t state = irqentry_enter(regs);			\
        ....
	__##func (regs);						\
        ....
}                                                                       \
		                                                        \
static noinline void __##func(struct pt_regs *regs)

So it goes through that code path _before_ the actual implementation
which does set_irq_regs() is reached.

The callchain is:

  asm_sysvec_apic_timer_interrupt               <- ASM entry in gate
     sysvec_apic_timer_interrupt(regs)          <- noinstr C entry point
        irqentry_enter(regs)                    <- unpoisons @reg
        __sysvec_apic_timer_interrupt(regs)     <- the actual handler
           set_irq_regs(regs)                   <- stores regs
           local_apic_timer_interrupt()
             ...
             tick_handler()                     <- One of the 4 variants
                regs = get_irq_regs();          <- retrieves regs
                update_process_times(user_tick = user_mode(regs))
                   account_process_tick(user_tick)
                      irqtime_account_process_tick(user_tick)
line 382:                } else if { user_tick }   <- KMSAN complains

I'm even more confused now.

> I guess handling those will require wrapping every interrupt gate into
> a function that performs register unpoisoning?

No, guessing does not help here.

The gates point to the ASM entry point, which then invokes the C entry
point. All C entry points use a DEFINE_IDTENTRY variant.

Some of the DEFINE_IDTENTRY_* C entry points are not doing anything in
the macro, but the C function either invokes irqentry_enter() or
irqentry_nmi_enter() open coded _before_ invoking any instrumentable
function. So the unpoisoning of @regs in these functions should tell
KMSAN that @regs or something derived from @regs are not some random
uninitialized values.

There should be no difference between unpoisoning @regs in
irqentry_enter() or in set_irq_regs(), right?

If so, then the problem is definitely _not_ the idt entry code.

> By the way, if it helps, I think we don't necessarily have to call
> kmsan_unpoison_memory() from within the
> instrumentation_begin()/instrumentation_end() region?
> We could move the call to the beginning of irqentry_enter(), removing
> unnecessary duplication.

We could, but then you need to mark unpoison_memory() noinstr too and you
have to add the unpoison into the syscall code. No win and irrelevant to
the problem at hand.

Thanks,

        tglx