[PATCH v2 00/12] KVM: Add idempotent controls for migrating system counter state

[Andrew Jones]

On Fri, Jul 16, 2021 at 09:26:17PM +0000, Oliver Upton wrote:
> KVM's current means of saving/restoring system counters is plagued with
> temporal issues. At least on ARM64 and x86, we migrate the guest's
> system counter by-value through the respective guest system register
> values (cntvct_el0, ia32_tsc). Restoring system counters by-value is
> brittle as the state is not idempotent: the host system counter is still
> oscillating between the attempted save and restore. Furthermore, VMMs
> may wish to transparently live migrate guest VMs, meaning that they
> include the elapsed time due to live migration blackout in the guest
> system counter view. The VMM thread could be preempted for any number of
> reasons (scheduler, L0 hypervisor under nested) between the time that
> it calculates the desired guest counter value and when KVM actually sets
> this counter state.
> 
> Despite the value-based interface that we present to userspace, KVM
> actually has idempotent guest controls by way of system counter offsets.
> We can avoid all of the issues associated with a value-based interface
> by abstracting these offset controls in new ioctls. This series
> introduces new vCPU device attributes to provide userspace access to the
> vCPU's system counter offset.
> 
> Patch 1 adopts Paolo's suggestion, augmenting the KVM_{GET,SET}_CLOCK
> ioctls to provide userspace with a (host_tsc, realtime) instant. This is
> essential for a VMM to perform precise migration of the guest's system
> counters.
> 
> Patches 2-3 add support for x86 by shoehorning the new controls into the
> pre-existing synchronization heuristics.
> 
> Patches 4-5 implement a test for the new additions to
> KVM_{GET,SET}_CLOCK.
> 
> Patches 6-7 implement at test for the tsc offset attribute introduced in
> patch 3.
> 
> Patch 8 adds a device attribute for the arm64 virtual counter-timer
> offset.
> 
> Patch 9 extends the test from patch 7 to cover the arm64 virtual
> counter-timer offset.
> 
> Patch 10 adds a device attribute for the arm64 physical counter-timer
> offset. Currently, this is implemented as a synthetic register, forcing
> the guest to trap to the host and emulating the offset in the fast exit
> path. Later down the line we will have hardware with FEAT_ECV, which
> allows the hypervisor to perform physical counter-timer offsetting in
> hardware (CNTPOFF_EL2).
> 
> Patch 11 extends the test from patch 7 to cover the arm64 physical
> counter-timer offset.
> 
> Patch 12 introduces a benchmark to measure the overhead of emulation in
> patch 10.
> 
> Physical counter benchmark
> --------------------------
> 
> The following data was collected by running 10000 iterations of the
> benchmark test from Patch 6 on an Ampere Mt. Jade reference server, A 2S
> machine with 2 80-core Ampere Altra SoCs. Measurements were collected
> for both VHE and nVHE operation using the `kvm-arm.mode=` command-line
> parameter.
> 
> nVHE
> ----
> 
> +--------------------+--------+---------+
> |       Metric       | Native | Trapped |
> +--------------------+--------+---------+
> | Average            | 54ns   | 148ns   |
> | Standard Deviation | 124ns  | 122ns   |
> | 95th Percentile    | 258ns  | 348ns   |
> +--------------------+--------+---------+
> 
> VHE
> ---
> 
> +--------------------+--------+---------+
> |       Metric       | Native | Trapped |
> +--------------------+--------+---------+
> | Average            | 53ns   | 152ns   |
> | Standard Deviation | 92ns   | 94ns    |
> | 95th Percentile    | 204ns  | 307ns   |
> +--------------------+--------+---------+
> 
> This series applies cleanly to the following commit:
> 
> 1889228d80fe ("KVM: selftests: smm_test: Test SMM enter from L2")
> 
> v1 -> v2:
>   - Reimplemented as vCPU device attributes instead of a distinct ioctl.
>   - Added the (realtime, host_tsc) instant support to
>     KVM_{GET,SET}_CLOCK
>   - Changed the arm64 implementation to broadcast counter offset values
>     to all vCPUs in a guest. This upholds the architectural expectations
>     of a consistent counter-timer across CPUs.
>   - Fixed a bug with traps in VHE mode. We now configure traps on every
>     transition into a guest to handle differing VMs (trapped, emulated).
>

Oops, I see there's a v3 of this series. I'll switch to reviewing that. I
think my comments / r-b's apply to that version as well though.

Thanks,
drew