| Message ID | 20210916181510.963449-1-oupton@google.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | KVM: arm64: Add idempotent controls to migrate guest counter | expand |
Hi Oliver, This looks interesting, I am going to try to do my best to review the series. I haven't followed the patches until this version, so I might be missing part of the context when I post my comments. Thanks, Alex On 9/16/21 19:15, Oliver Upton wrote: > Currently, on KVM/arm64, we only allow a VMM to migrate the guest's > virtual counter by-value. Saving and restoring the counter by value is > problematic in the fact that the recorded state is not idempotent. > Furthermore, we obfuscate from userspace the fact that the architecture > actually provides offset-based controls. > > Another issue is that KVM/arm64 doesn't provide userspace with the > controls of the physical counter-timer. This series aims to address both > issues by adding offset-based controls for the virtual and physical > counters. > > Patches 1-2 are refactor changes required to provide offset controls to > userspace and putting in some generic plumbing to use for both physical > and virtual offsets. > > Patch 3 is a minor refactor, creating a helper function to get the > number of timer registers for a particular vCPU. > > Patch 4 exposes a vCPU's virtual offset through the KVM_*_ONE_REG > ioctls. When NV support is added to KVM, CNTVOFF_EL2 will be considered > a guest system register. So, it is safe to expose it now through that > ioctl. > > Patch 5 adds a cpufeature bit to detect 'full' ECV implementations, > providing EL2 with the ability to offset the physical counter-timer. > > Patch 6 exposes a vCPU's physical offset as a vCPU device attribute. > This is deliberate, as the attribute is not architectural; KVM uses this > attribute to track the host<->guest offset. > > Patch 7 is a prepatory change for the sake of physical offset emulation, > as counter-timer traps must be configured separately for each vCPU. > > Patch 8 allows non-ECV hosts to support the physical offset vCPU device > attribute, by trapping and emulating the physical counter registers. > > This series was tested on an Ampere Mt. Jade system (non-ECV, VHE and > nVHE). I did not test this on the FVP, as I need to really figure out > tooling for it on my workstation. > > Applies cleanly to v5.15-rc1 > > v7: http://lore.kernel.org/r/20210816001217.3063400-1-oupton@google.com > > v7 -> v8: > - Only use ECV if !VHE > - Only expose CNTVOFF_EL2 register to userspace with opt-in > - Refer to the direct_ptimer explicitly > > Oliver Upton (8): > KVM: arm64: Refactor update_vtimer_cntvoff() > KVM: arm64: Separate guest/host counter offset values > KVM: arm64: Make a helper function to get nr of timer regs > KVM: arm64: Allow userspace to configure a vCPU's virtual offset > arm64: cpufeature: Enumerate support for FEAT_ECV >= 0x2 > KVM: arm64: Allow userspace to configure a guest's counter-timer > offset > KVM: arm64: Configure timer traps in vcpu_load() for VHE > KVM: arm64: Emulate physical counter offsetting on non-ECV systems > > Documentation/arm64/booting.rst | 7 + > Documentation/virt/kvm/api.rst | 23 +++ > Documentation/virt/kvm/devices/vcpu.rst | 28 ++++ > arch/arm64/include/asm/kvm_host.h | 3 + > arch/arm64/include/asm/sysreg.h | 5 + > arch/arm64/include/uapi/asm/kvm.h | 2 + > arch/arm64/kernel/cpufeature.c | 10 ++ > arch/arm64/kvm/arch_timer.c | 196 +++++++++++++++++++++--- > arch/arm64/kvm/arm.c | 9 +- > arch/arm64/kvm/guest.c | 28 +++- > arch/arm64/kvm/hyp/include/hyp/switch.h | 32 ++++ > arch/arm64/kvm/hyp/nvhe/timer-sr.c | 11 +- > arch/arm64/tools/cpucaps | 1 + > include/clocksource/arm_arch_timer.h | 1 + > include/kvm/arm_arch_timer.h | 14 +- > include/uapi/linux/kvm.h | 1 + > 16 files changed, 337 insertions(+), 34 deletions(-) >
Currently, on KVM/arm64, we only allow a VMM to migrate the guest's virtual counter by-value. Saving and restoring the counter by value is problematic in the fact that the recorded state is not idempotent. Furthermore, we obfuscate from userspace the fact that the architecture actually provides offset-based controls. Another issue is that KVM/arm64 doesn't provide userspace with the controls of the physical counter-timer. This series aims to address both issues by adding offset-based controls for the virtual and physical counters. Patches 1-2 are refactor changes required to provide offset controls to userspace and putting in some generic plumbing to use for both physical and virtual offsets. Patch 3 is a minor refactor, creating a helper function to get the number of timer registers for a particular vCPU. Patch 4 exposes a vCPU's virtual offset through the KVM_*_ONE_REG ioctls. When NV support is added to KVM, CNTVOFF_EL2 will be considered a guest system register. So, it is safe to expose it now through that ioctl. Patch 5 adds a cpufeature bit to detect 'full' ECV implementations, providing EL2 with the ability to offset the physical counter-timer. Patch 6 exposes a vCPU's physical offset as a vCPU device attribute. This is deliberate, as the attribute is not architectural; KVM uses this attribute to track the host<->guest offset. Patch 7 is a prepatory change for the sake of physical offset emulation, as counter-timer traps must be configured separately for each vCPU. Patch 8 allows non-ECV hosts to support the physical offset vCPU device attribute, by trapping and emulating the physical counter registers. This series was tested on an Ampere Mt. Jade system (non-ECV, VHE and nVHE). I did not test this on the FVP, as I need to really figure out tooling for it on my workstation. Applies cleanly to v5.15-rc1 v7: http://lore.kernel.org/r/20210816001217.3063400-1-oupton@google.com v7 -> v8: - Only use ECV if !VHE - Only expose CNTVOFF_EL2 register to userspace with opt-in - Refer to the direct_ptimer explicitly Oliver Upton (8): KVM: arm64: Refactor update_vtimer_cntvoff() KVM: arm64: Separate guest/host counter offset values KVM: arm64: Make a helper function to get nr of timer regs KVM: arm64: Allow userspace to configure a vCPU's virtual offset arm64: cpufeature: Enumerate support for FEAT_ECV >= 0x2 KVM: arm64: Allow userspace to configure a guest's counter-timer offset KVM: arm64: Configure timer traps in vcpu_load() for VHE KVM: arm64: Emulate physical counter offsetting on non-ECV systems Documentation/arm64/booting.rst | 7 + Documentation/virt/kvm/api.rst | 23 +++ Documentation/virt/kvm/devices/vcpu.rst | 28 ++++ arch/arm64/include/asm/kvm_host.h | 3 + arch/arm64/include/asm/sysreg.h | 5 + arch/arm64/include/uapi/asm/kvm.h | 2 + arch/arm64/kernel/cpufeature.c | 10 ++ arch/arm64/kvm/arch_timer.c | 196 +++++++++++++++++++++--- arch/arm64/kvm/arm.c | 9 +- arch/arm64/kvm/guest.c | 28 +++- arch/arm64/kvm/hyp/include/hyp/switch.h | 32 ++++ arch/arm64/kvm/hyp/nvhe/timer-sr.c | 11 +- arch/arm64/tools/cpucaps | 1 + include/clocksource/arm_arch_timer.h | 1 + include/kvm/arm_arch_timer.h | 14 +- include/uapi/linux/kvm.h | 1 + 16 files changed, 337 insertions(+), 34 deletions(-)