Re: Possible nohz-full/RCU issue in arm64 KVM

[Mark Rutland <mark.rutland@arm.com>]

On Fri, Dec 17, 2021 at 04:54:22PM +0100, Paolo Bonzini wrote:
> On 12/17/21 15:38, Mark Rutland wrote:
> > For example kvm_guest_enter_irqoff() calls guest_enter_irq_off() which calls
> > vtime_account_guest_enter(), but kvm_guest_exit_irqoff() doesn't call
> > guest_exit_irq_off() and the call to vtime_account_guest_exit() is open-coded
> > elsewhere. Also, guest_enter_irq_off() conditionally calls
> > rcu_virt_note_context_switch(), but I can't immediately spot anything on the
> > exit side that corresponded with that, which looks suspicious.
> 
> rcu_note_context_switch() is a point-in-time notification; it's not strictly
> necessary, but it may improve performance a bit by avoiding unnecessary IPIs
> from the RCU subsystem.
> 
> There's no benefit from doing it when you're back from the guest, because at
> that point the CPU is just running normal kernel code.

I see.

My main issue here was just that it's really difficult to see how the
entry/exit logic is balanced, and I reckon we can solve that by splitting
guest_{enter,exit}_irqoff() into helper functions to handle the vtime
accounting separately from the context tracking, so that arch code can do
something like:

  guest_timing_enter_irqoff();
  
  guest_eqs_enter_irqoff();
  < actually run vCPU here >
  guest_eqs_exit_irqoff();
  
  < handle pending IRQs here >
  
  guest_timing_exit_irqoff();

... which I hope should work for RISC-V too.

I've had a go, and I've pushed out a WIP to:

  https://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git/log/?h=arm64/kvm/rcu

I also see we'll need to add some lockdep/irq-tracing management to arm64, and
it probably makes sense to fold that into common helpers, so I'll have a play
with that tomorrow.

Thanks,
Mark.