Re: [PATCH 6/8] i386: hvf: Drop hvf_reset_vcpu()

[Roman Bolshakov]

On Thu, Jun 25, 2020 at 03:30:38PM +0200, Paolo Bonzini wrote:
> On 25/06/20 14:36, Roman Bolshakov wrote:
> > 
> > Yes, there's such a place. post-init and post-reset invoke
> > hvf_put_registers() and the latter one calls hvf_put_segments().
> > hvf_put_segments() sets CR4 and CR0 via macvm_set_cr0/macvm_set_cr4
> > using the CR0/CR4 from env. So, the reset is relying on generic QEMU
> > CPUX86State now. LMA in EFER is reset there as well.
> 
> Ok, do you want to send a follow-up or a v2 of this?
> 

I'm still trying to understand what I should do :)

> > I don't know any alternative for PDPTE and VMCS Entry Controls in
> > CPUX86State, that's why I left explicit reset of the VMCS fields in
> > post-reset.
> 
> VMCS entry controls should be handled by macvm_set_cr0 as well, because
> QEMU does not use any except for the LMA bit.  They are initialized zero
> 
>     wvmcs(cpu->hvf_fd, VMCS_ENTRY_CTLS,
>         cap2ctrl(hvf_state->hvf_caps->vmx_cap_entry, 0));
> 
> but in practice the last argument ends up being zero all the time.
> 

macvm_set_cr0() sets/clears LMA in entry controls only in case of
transitions into/out of long mode in enter_long_mode() in
exit_long_mode(), respectively. But macvm_set_cr0() doesn't load
EFER.LMA from CPUX86State into VMCS entry controls during reset and
that's where hvf_put_registers() might not behave properly.

As far as I understand you propose to drop explicit LMA reset in
post-reset and rather impove synchronization between efer and entry
controls in macvm_set_cr0(), right? In that case I don't see a
regression in the series, and if possible I'd prefer a follow up patch
for the issue.

> PDPTEs are not a problem, because they are not used after reset (only if
> CR4.PAE=CR4.PG=1 and EFER.LME=0).
> 

Ok, good, then we're leaving PDPTE initialization as is in post-reset.

Thanks,
Roman