On Wed, May 19, 2021 at 09:13:08PM +0200, Joerg Roedel wrote: > Hi Peter, > > thanks for your review. > > On Wed, May 19, 2021 at 07:54:50PM +0200, Peter Zijlstra wrote: > > On Wed, May 19, 2021 at 03:52:48PM +0200, Joerg Roedel wrote: > > > --- a/arch/x86/kernel/sev.c > > > +++ b/arch/x86/kernel/sev.c > > > @@ -1343,9 +1343,10 @@ DEFINE_IDTENTRY_VC_SAFE_STACK(exc_vmm_communication) > > > return; > > > } > > > > > > + instrumentation_begin(); > > > + > > > irq_state = irqentry_nmi_enter(regs); > > > lockdep_assert_irqs_disabled(); > > > - instrumentation_begin(); > > > > > > /* > > > * This is invoked through an interrupt gate, so IRQs are disabled. The > > > > That's just plain wrong. No instrumentation is allowed before you enter > > the exception context. > > Okay. > > > > + irqentry_nmi_exit(regs, irq_state); > > > + > > > > And this is wrong too; because at this point the handler doesn't run in > > _any_ context anymore, certainly not one you can call regular C code > > from. > > The #VC handler is at this point not running on the IST stack anymore, > but on the stack it came from or on the task stack. So my believe was > that at this point it inherits the context it came from (just like the > page-fault handler). But I also don't fully understand the context > tracking, so is my assumption wrong? Being on the right stack is only part of the issue; you also need to make sure your runtime environment is set up. Regular kernel C expects a whole lot of things to be present; esp. so with all the debug options on. The irqentry_*_enter() family of functions very carefully sets up this environment and the irqentry_*_exit() undoes it again. Before and after you really cannot run normal code. Just an example, RCU might not be watching, it might think the CPU is in userspace and advance the GP while you're relying on it not doing so. Similarly lockdep is in some undefined state and any lock used can trigger random 'funny' things. Just because this is 'C', doesn't immediately mean you can go call any random function. Up until recently most of this was in ASM. There's a reason for the noinstr annotations.
