[RFC,0/4] KVM: arm64: emulation for CTR_EL0

Message

Sebastian Ott March 18, 2024, 11:16 a.m. UTC

Hej folks,

I'm looking into supporting migration between 2 Ampere Altra (Max)
machines (using Neoverse-N1). They are almost identical regarding
their feature id register state except for CTR_EL0.DIC which is set
on one machine but not the other.

For a complete picture it's worth noting that the machine with
CTR_EL0.DIC = 1 also suffers from erratum 1542419 meaning for
userspace we trap the access and fake DIC=0. A KVM guest still sees
the original host value for that register (I assume that this is
intentional so that a guest OS can apply its own workaround to the
erratum).

Anyway, CTR_EL0 is currently marked as invariant and migrating a VM
between those 2 machines using qemu fails.

With the patches below guest access to CTR_EL0 is emulated and
CTR_EL0.DIC can be disabled using KVM_SET_ONE_REG. I'm sending this
as an RFC since I likely missed something obvious and there's still
stuff that needs improving - e.g. I didn't look into FGT. The last
patch adds a tool to dump the KVM register state together with the
writable masks - not sure if that's helpful to others but I've used
this to compare the register state between different machines
(abusing kvm selftests for this is probably not the best idea - I
just wanted to use the neat helpers they provide).

Thanks,
Sebastian

Sebastian Ott (4):
  KVM: arm64: add emulation for CTR_EL0 register
  KVM: arm64: ensure guest access to CTR_EL0 is trapped
  KVM: arm64: show writable masks for feature registers
  KVM: selftests: aarch64: add tool to dump registers

 arch/arm64/include/asm/kvm_emulate.h          |  7 +-
 arch/arm64/kvm/sys_regs.c                     | 63 ++++++++++------
 tools/testing/selftests/kvm/Makefile          |  1 +
 .../testing/selftests/kvm/aarch64/dump_regs.c | 72 +++++++++++++++++++
 4 files changed, 114 insertions(+), 29 deletions(-)
 create mode 100644 tools/testing/selftests/kvm/aarch64/dump_regs.c

Comments

Marc Zyngier March 18, 2024, 3:24 p.m. UTC | #1

On Mon, 18 Mar 2024 11:16:32 +0000,
Sebastian Ott <[e-mail]> wrote:
> 
> Hej folks,
> 
> I'm looking into supporting migration between 2 Ampere Altra (Max)
> machines (using Neoverse-N1). They are almost identical regarding
> their feature id register state except for CTR_EL0.DIC which is set
> on one machine but not the other.
> 
> For a complete picture it's worth noting that the machine with
> CTR_EL0.DIC = 1 also suffers from erratum 1542419 meaning for
> userspace we trap the access and fake DIC=0. A KVM guest still sees
> the original host value for that register (I assume that this is
> intentional so that a guest OS can apply its own workaround to the
> erratum).

Indeed.

The intention is that the EL1 guest will hide DIC for EL0, while EL3
will trap IC instructions from EL0 and replace them with a TLBI.
That's of course assuming that the machine has received an updated
firmware, something that cannot be probed AFAICT.

	M.