[PATCH v3] KVM: PPC: Tick accounting should defer vtime accounting 'til after IRQ handling

[Laurent Vivier]

On 27/10/2021 16:21, Nicholas Piggin wrote:
> From: Laurent Vivier <lvivier at redhat.com>
> 
> Commit 112665286d08 ("KVM: PPC: Book3S HV: Context tracking exit guest
> context before enabling irqs") moved guest_exit() into the interrupt
> protected area to avoid wrong context warning (or worse). The problem is
> that tick-based time accounting has not yet been updated at this point
> (because it depends on the timer interrupt firing), so the guest time
> gets incorrectly accounted to system time.
> 
> To fix the problem, follow the x86 fix in commit 160457140187 ("Defer
> vtime accounting 'til after IRQ handling"), and allow host IRQs to run
> before accounting the guest exit time.
> 
> In the case vtime accounting is enabled, this is not required because TB
> is used directly for accounting.
> 
> Before this patch, with CONFIG_TICK_CPU_ACCOUNTING=y in the host and a
> guest running a kernel compile, the 'guest' fields of /proc/stat are
> stuck at zero. With the patch they can be observed increasing roughly as
> expected.
> 
> Fixes: e233d54d4d97 ("KVM: booke: use __kvm_guest_exit")
> Fixes: 112665286d08 ("KVM: PPC: Book3S HV: Context tracking exit guest context before enabling irqs")
> Cc: <stable at vger.kernel.org> # 5.12
> Signed-off-by: Laurent Vivier <lvivier at redhat.com>
> [np: only required for tick accounting, add Book3E fix, tweak changelog]
> Signed-off-by: Nicholas Piggin <npiggin at gmail.com>
> ---
> Since v2:
> - I took over the patch with Laurent's blessing.
> - Changed to avoid processing IRQs if we do have vtime accounting
>    enabled.
> - Changed so in either case the accounting is called with irqs disabled.
> - Added similar Book3E fix.
> - Rebased on upstream, tested, observed bug and confirmed fix.
> 
>   arch/powerpc/kvm/book3s_hv.c | 30 ++++++++++++++++++++++++++++--
>   arch/powerpc/kvm/booke.c     | 16 +++++++++++++++-
>   2 files changed, 43 insertions(+), 3 deletions(-)
> 

Tested-by: Laurent Vivier <lvivier at redhat.com>

Checked with mpstat that time is accounted to %guest while a stress-ng test is running in 
the guest. Checked there is no warning in the host kernellogs.

Thanks,
Laurent