Re: [RFC PATCH] KVM: arm64: don't single-step for non-emulated faults

[Mark Rutland]

On Fri, Nov 09, 2018 at 12:24:41PM +0000, Alex BennÃe wrote:
> Mark Rutland <mark.rutland@xxxxxxx> writes:
> > On Thu, Nov 08, 2018 at 02:38:43PM +0000, Peter Maydell wrote:
> >> On 8 November 2018 at 14:28, Alex BennÃe <alex.bennee@xxxxxxxxxx> wrote:
> >> >
> >> > Mark Rutland <mark.rutland@xxxxxxx> writes:
> >> >> One problem is that I couldn't spot when we advance the PC for an MMIO
> >> >> trap. I presume we do that in the kernel, *after* the MMIO trap, but I
> >> >> can't see where that happens.
> >> >
> >> > Nope it gets done before during decode_hsr in mmio.c:
> >> >
> >> >         /*
> >> >          * The MMIO instruction is emulated and should not be re-executed
> >> >          * in the guest.
> >> >          */
> >> >         kvm_skip_instr(vcpu, kvm_vcpu_trap_il_is32bit(vcpu));
> >>
> >> I think that this attempt to do the PC-advance early is
> >> probably an underlying problem that is not helping the
> >> code structure here.
> >>
> >> An enhancement that's been floated previously is that the
> >> MMIO emulation in userspace should be able to report back
> >> to KVM "nope, that access should generate a guest synchronous
> >> external abort (with ESR_EL1.EA = 0/1)".
> >> If we have that, then we definitely need to not advance the
> >> PC until after userspace has done the emulation and told
> >> us whether the memory access succeeded or not...
> >
> > Yup.
> >
> > I think that we absolutely want to do all the CPU state advancement (PC,
> > SS bit, etc) at the point we apply the effects of the instruction. Not
> > before we emulate the instruction, and not higher/lower in the call
> > stack.
> 
> There is certainly an argument to abstract some of the advancement logic
> so we can make debug related decisions in one place. I don't know how
> much churn we would need to get there.

I'm not saying anything about *decisions*. I'm saying that we can make
the state consistent by advancing the singlestep state in the same way
that HW does, at the instant it advances the PC.

i.e. do that in kvm_skip_instr(), as I've done in my local tree.

That mirrors the HW, and we don't need to special-case any handling for
emulated vs non-emulated instructions.

> Currently most of the guest debug decisions are made in one place as we
> head into the guest run. Generally I don't think the emulation code want
> to know or care about the SS bit or what debug is currently happening
> although I guess the presence of the SS bit could be used to decide on
> exactly what exit type you are going to do - back to guest or out to
> user space. Currently kvm_arm_handle_step_debug on cares about host
> directed debug but we could expand it to raise the appropriate guest
> exception if required.

So long as we consistently advance the singlestep state when we emulate
an instruction, we shouldn't need kvm_arm_handle_step_debug() at all. If
we emulate an instruction, we'll return to the guest with PSTATE.SS
clear, and the HW will generate the exception for us.

This is *vastly* simpler to reason about.

I have local patches which I intend to post shortly.

> > We have a big problem in that guest-directed singlestep and
> > host-directed singlestep don't compose, and given that host-directed
> > singlestep doesn't work reliably today I'd be tempted to rip that out
> > until we've fixed guest-directed singlestep.
> 
> Getting host and guest debug to run at the same time is tricky given we
> end up subverting guest state when the host debug is in control. It did
> make my head spin when I worked on it originally which led to the
> acceptance that guest debug couldn't be expected to work transparently
> while host directed debug was in effect. If we can support it without
> adding complexity then that would be great but it's a pretty niche use
> case.

At the very least we need to define whether the kernel or userspace is
responsible for this. If we say it's userspace's responsibility to
virtualize this when it takes control of guest debug, but QEMU doesn't
do so, that's fine by me.

> I'd be loathed to rip out the current single step support as it does
> actually work pretty well - it's just corner cases with emulated MMIO
> and first step that are failing. Last I looked these were cases x86
> didn't even get right and no one has called to remove it's single step
> support AFAIK.
> 
> > We should have a story for how host-directed debug is handled
> > transparently from the PoV of a guest using guest-directed debug.
> 
> What's the use case for this apart from having a cleaner abstraction?

Providing the architecturally mandated behaviour to the guest.

> Do users really spend time running multiple gdbs at different levels
> in the stack?

It's not just about GDB. Things like kprobes, live patching, etc in a
guest may use singlestep, and this may be *critical* to the correct
operation of a given guest.

People *will* want to debug such features under a hypervisor. I know
that I personally want to be able to do so.

In general, a debugger shouldn't silently corrupt guest state or
behaviour, as KVM_GUESTDBG_SINGLESTEP behaviour effectively does today.

Thanks,
Mark.