Re: [Xen-devel] [PATCH 7/7] vm-event/arm: implement support for control-register write vm-events

[Corneliu ZUZU <czuzu@xxxxxxxxxxxxxxx>]

On 6/23/2016 2:11 PM, Julien Grall wrote:
> Hello,
>
> On 23/06/16 06:49, Corneliu ZUZU wrote:
> > On 6/23/2016 8:31 AM, Corneliu ZUZU wrote:
> > > On 6/22/2016 10:41 PM, Julien Grall wrote:
> > > > On 22/06/2016 20:37, Corneliu ZUZU wrote:
> > > > > I've also realized that it's a bit complicated to avoid writing HCR
> > > > > from
> > > > > 2 places.
> > > > > That's because:
> > > > > - p2m_restore_state is part of the process of switching to another 
> > > > > vCPU
> > > > > and the HCR write _must be committed_ here because other components
> > > > > depend on that, like address-translation functions
> > > > > - I want vm_event_vcpu_enter to be called _after_ the switch to the
> > > > > vCPU
> > > > > is completed
> > > > > - I want HCR_TVM to be set in vm_event_vcpu_enter because setting
> > > > > necessary traps _for cr vm-events_ to work should be done there
> > > > > (setting
> > > > > HCR_TVM bit makes sense to be there and the purpose is to centralize
> > > > > operations such as this for code comprehensibility; also, on the X86
> > > > > counterpart a similar operation is done for trapping CR3, so it
> > > > > would be
> > > > > nice to keep the symmetry)
> > > > >
> > > > > Would it be such a stretch to have HCR written in 2 places? (the 
> > > > > second
> > > > > time happens rarely anyway: it's unlikely(..) to have to do the
> > > > > write in
> > > > > vm_event_vcpu_enter)
> > > > Not really. It was mostly to avoid setting/clearing HCR bits in
> > > > different place in the code. It makes more difficult to know what is
> > > > the final result of the register.
> > > >
> > > > Anyway, let's skip it for now, if it is too difficult.
> > > >
> > > > Regards,
> > > Then perhaps something like the following would be suitable:
> > >
> > > 1. store hcr in arch_domain (register_t hcr)
> > >
> > > 2. add a function in asm-arm/processor.h (or where else?) which only
> > > does:
> > >     static inline void update_hcr(struct domain *d)
> > >     {
> > >         WRITE_SYSREG(d->arch.hcr, HCR_EL2);
> > >         isb();
> > >     }
> > >
> > > 3.  modify p2m_restore_state to do:
> > >     n->domain->arch.hcr &= ~HCR_VM;
> > >     update_hcr(n->domain);
> > >     p2m_load_VTTBR(n->domain);
> > >
> > >     n->domain->arch.hcr |= HCR_VM;
> > >
> > >     if ( is_32bit_domain(n->domain) )
> > >         n->domain->arch.hcr &= ~HCR_RW;
> > >     else
> > >         n->domain->arch.hcr |= HCR_RW;
> This is not safe at all, p2m_restore_state is vCPU specific at you 
> modify domain information.
> However, if we store the hcr per domain, overriding every context 
> switch is pointless as the domain will always be 32-bit/64-bit.
Oh right, the RW bit needs not be set/unset anymore with this change.

> > >     update_hcr(n->domain);
> > >
> > >     WRITE_SYSREG(n->arch.sctlr, SCTLR_EL1);
> > >     isb();
> > >
> > > 4. and vcpu_enter_adjust_traps to
> > >
> > >     if ( unlikely(0 != v->domain->arch.monitor.write_ctrlreg_enabled) )
> > >     {
> > >          if ( likely(v->domain->arch.hcr & HCR_TVM) )
> > >              return;
> > >          v->domain->arch.hcr |= HCR_TVM;
> > >     }
> > >     else
> > >     {
> > >          if ( likely(!(v->domain->arch.hcr & HCR_TVM)) )
> > >              return;
> > >          v->domain->arch.hcr &= ~HCR_TVM;
> > >     }
> This does not need to be done in vcpu_enter_adjust_traps everytime. 
> You can set the bit in arch.hcr in DOMCTL_MONITOR_EVENT_WRITE_CTRLREG.
I wanted to keep X86-ARM symmetry and it seemed more intuitive to have 
these kind of adjustments with the vcpu_enter code motion. But now that 
I think about it, given the fact that we have the guarantee that after 
monitor_domctl and before reentering the vCPU p2m_restore_state gets 
called (due to domain_pause/domain_unpause) - thus actually committing 
the hcr update at the proper time - technically monitor_domctl _is_ the 
optimal place to set arch.hcr in. In conclusion, I'm thinking of 
discarding the entire idea of introducing vm_event_vcpu_enter, it seems 
to me now that this would also render a simpler code.
> > >     update_hcr(v->domain);
> > >
> > > That way at least it's easier to follow where update_hcr is called.
> I don't see much reason to store the value in the domain and have 
> multiple update_hcr. If we store the value, then we should only call 
> update_hcr once when returning to the guest.
Yep, that will happen with the above-mentioned discarding of 
vm_event_vcpu_enter idea.
> > And also set the initial value of HCR at the moment of creation, i.e. in
> > arch_domain_create as
> >
> > d->arch.hcr = READ_SYSREG(HCR_EL2)
> We control the value of HCR_EL2, so it would be better to assign the 
> list of flags here.
Right, that will happen too.

> Regards,
Thanks for the useful insights,
Corneliu.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel