Hi Alexandru,
On Wed, Nov 06, 2019 at 01:02:21PM +0000, Alexandru Elisei wrote:
>
> On 10/28/19 1:05 PM, Christoffer Dall wrote:
> > On CPUs that support S2FWB (Armv8.4+), KVM configures the stage 2 page
> > tables to override the memory attributes of memory accesses, regardless
> > of the stage 1 page table configurations, and also when the stage 1 MMU
> > is turned off. This results in all memory accesses to RAM being
> > cacheable, including during early boot of the guest.
> >
> > On CPUs without this feature, memory accesses were non-cacheable during
> > boot until the guest turned on the stage 1 MMU, and we had to detect
> > when the guest turned on the MMU, such that we could invalidate all cache
> > entries and ensure a consistent view of memory with the MMU turned on.
> > When the guest turned on the caches, we would call stage2_flush_vm()
> > from kvm_toggle_cache().
> >
> > However, stage2_flush_vm() walks all the stage 2 tables, and calls
> > __kvm_flush-dcache_pte, which on a system with S2FWD does ... absolutely
> > nothing.
> >
> > We can avoid that whole song and dance, and simply not set TVM when
> > creating a VM on a system that has S2FWB.
> >
> > Signed-off-by: Christoffer Dall <christoffer.dall@xxxxxxx>
> > Reviewed-by: Mark Rutland <mark.rutland@xxxxxxx>
> > ---
> > I was only able to test this on the model with cache modeling enabled,
> > but even removing TVM from HCR_EL2 without having FWB also worked with
> > that setup, so the testing of this has been light. It seems like it
> > should obviously work, but it would be good if someone with access to
> > appropriate hardware could give this a spin.
> >
> > arch/arm64/include/asm/kvm_arm.h | 3 +--
> > arch/arm64/include/asm/kvm_emulate.h | 12 +++++++++++-
> > 2 files changed, 12 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/kvm_arm.h b/arch/arm64/include/asm/kvm_arm.h
> > index ddf9d762ac62..6e5d839f42b5 100644
> > --- a/arch/arm64/include/asm/kvm_arm.h
> > +++ b/arch/arm64/include/asm/kvm_arm.h
> > @@ -61,7 +61,6 @@
> > * RW: 64bit by default, can be overridden for 32bit VMs
> > * TAC: Trap ACTLR
> > * TSC: Trap SMC
> > - * TVM: Trap VM ops (until M+C set in SCTLR_EL1)
> > * TSW: Trap cache operations by set/way
> > * TWE: Trap WFE
> > * TWI: Trap WFI
> > @@ -74,7 +73,7 @@
> > * SWIO: Turn set/way invalidates into set/way clean+invalidate
> > */
> > #define HCR_GUEST_FLAGS (HCR_TSC | HCR_TSW | HCR_TWE | HCR_TWI | HCR_VM | \
> > - HCR_TVM | HCR_BSU_IS | HCR_FB | HCR_TAC | \
> > + HCR_BSU_IS | HCR_FB | HCR_TAC | \
> > HCR_AMO | HCR_SWIO | HCR_TIDCP | HCR_RW | HCR_TLOR | \
> > HCR_FMO | HCR_IMO)
> > #define HCR_VIRT_EXCP_MASK (HCR_VSE | HCR_VI | HCR_VF)
> > diff --git a/arch/arm64/include/asm/kvm_emulate.h b/arch/arm64/include/asm/kvm_emulate.h
> > index d69c1efc63e7..70509799a2a9 100644
> > --- a/arch/arm64/include/asm/kvm_emulate.h
> > +++ b/arch/arm64/include/asm/kvm_emulate.h
> > @@ -53,8 +53,18 @@ static inline void vcpu_reset_hcr(struct kvm_vcpu *vcpu)
> > /* trap error record accesses */
> > vcpu->arch.hcr_el2 |= HCR_TERR;
> > }
> > - if (cpus_have_const_cap(ARM64_HAS_STAGE2_FWB))
> > +
> > + if (cpus_have_const_cap(ARM64_HAS_STAGE2_FWB)) {
> > vcpu->arch.hcr_el2 |= HCR_FWB;
> > + } else {
> > + /*
> > + * For non-FWB CPUs, we trap VM ops (HCR_EL2.TVM) until M+C
> > + * get set in SCTLR_EL1 such that we can detect when the guest
> > + * MMU gets turned off and do the necessary cache maintenance
> > + * then.
> > + */
> > + vcpu->arch.hcr_el2 &= ~HCR_TVM;
> > + }
> >
> > if (test_bit(KVM_ARM_VCPU_EL1_32BIT, vcpu->arch.features))
> > vcpu->arch.hcr_el2 &= ~HCR_RW;
>
> This patch makes sense to me: when FWB is available, the guest memory is cacheable
> even when the stage 1 MMU is disabled, which means it's now impossible to have a
> situation where the data in memory is newer than the data in the cache.
>
> I tested the patch with the fix suggested by Marc by doing a linux boot and then a
> 'ls -R /', and by running kvm-unit-tests in a loop a couple dozen times. For what
> it's worth:
>
> Tested-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
>
> I do need to point out that I haven't been able to make a guest misbehave when FWB
> is not enabled *and* KVM doesn't do a stage2_flush_vm when the stage 1 MMU is
> enabled. I tried to write two different tests in kvm-unit-tests:
>
> 1. With the MMU never enabled, the test tells the host to read a value from memory
> (so a cache line is allocated), writes another value to the same memory location,
> and then enables the MMU and reads the memory back. I always got the latest value
> that was written while the MMU was off.
>
> 2. One thread tells the host to read the memory location in a loop (to make sure
> that the cache line doesn't get evicted), while the other thread writes a value
> with the MMU off, enables the MMU and reads the memory back. I still got the
> latest value written with the MMU off.
>
> I can share the source code for the tests, if anyone is interested; I'm also open
> to other suggestions.
>
Thanks for the thoroughness here. I also wasn't able to produce an
error on the model, so I think we can conclude that it's at least no
worse than the original code, and since we all agree that this should be
correct, then I think it's fair that Marc has merged the patch.
It can't hurt to post the code you wrote for the test; someone might
pick that up in the future to test something similar.
Thanks,
Christoffer
_______________________________________________
kvmarm mailing list
kvmarm@xxxxxxxxxxxxxxxxxxxxx
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
