Hi Zenghui,
On Fri, 24 Apr 2020 12:07:50 +0800
Zenghui Yu <yuzenghui@xxxxxxxxxx> wrote:
> Hi Marc,
>
> On 2020/4/22 20:00, Marc Zyngier wrote:
> > Keeping empty structure as the vcpu state initializer is slightly
> > wasteful: we only want to set pstate, and zero everything else.
> > Just do that.
> >
> > Signed-off-by: Marc Zyngier <maz@xxxxxxxxxx>
> > ---
> > arch/arm64/kvm/reset.c | 20 +++++++++-----------
> > 1 file changed, 9 insertions(+), 11 deletions(-)
> >
> > diff --git a/arch/arm64/kvm/reset.c b/arch/arm64/kvm/reset.c
> > index 241db35a7ef4f..895d7d9ad1866 100644
> > --- a/arch/arm64/kvm/reset.c
> > +++ b/arch/arm64/kvm/reset.c
> > @@ -37,15 +37,11 @@ static u32 kvm_ipa_limit;
> > /*
> > * ARMv8 Reset Values
> > */
> > -static const struct kvm_regs default_regs_reset = {
> > - .regs.pstate = (PSR_MODE_EL1h | PSR_A_BIT | PSR_I_BIT |
> > - PSR_F_BIT | PSR_D_BIT),
> > -};
> > +#define VCPU_RESET_PSTATE_EL1 (PSR_MODE_EL1h | PSR_A_BIT | PSR_I_BIT | \
> > + PSR_F_BIT | PSR_D_BIT)
> > > -static const struct kvm_regs default_regs_reset32 = {
> > - .regs.pstate = (PSR_AA32_MODE_SVC | PSR_AA32_A_BIT |
> > - PSR_AA32_I_BIT | PSR_AA32_F_BIT),
> > -};
> > +#define VCPU_RESET_PSTATE_SVC (PSR_AA32_MODE_SVC | PSR_AA32_A_BIT | \
> > + PSR_AA32_I_BIT | PSR_AA32_F_BIT)
> > > static bool cpu_has_32bit_el1(void)
> > {
> > @@ -261,6 +257,7 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> > const struct kvm_regs *cpu_reset;
> > int ret = -EINVAL;
> > bool loaded;
> > + u32 pstate;
> > > /* Reset PMU outside of the non-preemptible section */
> > kvm_pmu_vcpu_reset(vcpu);
> > @@ -291,16 +288,17 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> > if (test_bit(KVM_ARM_VCPU_EL1_32BIT, vcpu->arch.features)) {
> > if (!cpu_has_32bit_el1())
> > goto out;
> > - cpu_reset = &default_regs_reset32;
> > + pstate = VCPU_RESET_PSTATE_SVC;
> > } else {
> > - cpu_reset = &default_regs_reset;
> > + pstate = VCPU_RESET_PSTATE_EL1;
> > }
> > > break;
> > }
> > > /* Reset core registers */
> > - memcpy(vcpu_gp_regs(vcpu), cpu_reset, sizeof(*cpu_reset));
> > + memset(vcpu_gp_regs(vcpu), 0, sizeof(*cpu_reset));
>
> Be careful that we can *not* use 'sizeof(*cpu_reset)' here anymore. As
> you're going to refactor the layout of the core registers whilst keeping
> the kvm_regs API unchanged. Resetting the whole kvm_regs will go
> corrupting some affected registers and make them temporarily invalid.
> The bad thing will show up after you start moving ELR_EL1 around,
> specifically in patch #20...
Ah, awesome find! Yes, it is pretty obvious now that you point it out.
If I had removed this now useless cpu_reset variable, I'd have spotted
it!
> And the first victim is ... MPIDR_EL1 (the first one in sys_regs array).
> Now you know how this was spotted ;-) I think this should be the root
> cause of what Zengtao had previously reported [*].
It'd be good if Zengtao could confirm that changing this line to
memset(vcpu_gp_regs(vcpu), 0, sizeof(*vcpu_gp_regs(vcpu)));
fixes his problem.
> If these registers are all expected to be reset to architecturally
> UNKNOWN values, I think we can just drop this memset(), though haven't
> check with the ARM ARM carefully.
D1.9.1 ("PE state on reset to AArch64 state"):
"All general-purpose, and SIMD and floating-point registers are
UNKNOWN."
There is a vaguely similar wording for AArch32 (G1.17.1), although it
is only described by omission:
"Immediately after a reset, much of the PE state is UNKNOWN. However,
some of the PE state is defined."
and the GPRs are not part of the list of defined states.
Still, I'm worried to change KVM's behaviour after so long... I'll have
a try with a handful of non-Linux guests and see if anything breaks.
Thanks again,
M.
--
Jazz is not dead. It just smells funny...
