Re: [PATCH v4] x86/power: Fix 'nosmt' vs. hibernation triple fault during resume

[Sean Christopherson]

On Fri, May 31, 2019 at 02:22:27PM -0700, Andy Lutomirski wrote:
> 
> > On May 31, 2019, at 2:05 PM, Jiri Kosina <jikos@xxxxxxxxxx> wrote:
> > 
> >> On Fri, 31 May 2019, Andy Lutomirski wrote:
> >> 
> >> The Intel SDM Vol 3 34.10 says:
> >> 
> >> If the HLT instruction is restarted, the processor will generate a
> >> memory access to fetch the HLT instruction (if it is
> >> not in the internal cache), and execute a HLT bus transaction. This
> >> behavior results in multiple HLT bus transactions
> >> for the same HLT instruction.
> > 
> > Which basically means that both hibernation and kexec have been broken in 
> > this respect for gazillions of years, and seems like noone noticed. Makes 
> > one wonder what the reason for that might be.
> > 
> > Either SDM is not precise and the refetch actually never happens for real 
> > (or is always in these cases satisfied from I$ perhaps?), or ... ?
> > 
> > So my patch basically puts things back where they have been for ages 
> > (while mwait is obviously much worse, as that gets woken up by the write 
> > to the monitored address, which inevitably does happen during resume), but 
> > seems like SDM is suggesting that we've been in a grey zone wrt RSM at 
> > least for all those ages.
> > 
> > So perhaps we really should ditch resume_play_dead() altogether 
> > eventually, and replace it with sending INIT IPI around instead (and then 
> > waking the CPUs properly via INIT INIT START). I'd still like to do that 
> > for 5.3 though, as that'd be slightly bigger surgery, and conservatively 
> > put things basically back to state they have been up to now for 5.2.
> > 
> 
> 
> Seems reasonable to me.  I would guess that it mostly works because SMI isnât
> all that common and the window where it matters is short.  Or maybe the SDM
> is misleading.

For P6 and later, i.e. all modern CPUs, Intel processors go straight to
halted state and don't fetch/decode the HLT instruction.

P5 actually did a fetch, but from what I can tell that behavior wasn't
carried forward to KNC, unlike other legacy interrupt crud from P5:

[1] https://lkml.kernel.org/r/20190430004504.GH31379@xxxxxxxxxxxxxxx