On Tue, 13 Feb 2024 09:32:38 +0000,
Oliver Upton <oliver.upton@xxxxxxxxx> wrote:
>
> LPI translation and injection has been shown to have a significant
> impact on the performance of VM workloads, so it probably makes sense to
> add some signals in this area.
>
> Introduce the concept of a KVM tracepoint that associates with a VM
> stat and use it for the LPI translation cache tracepoints. It isn't too
> uncommon for a kernel hacker to attach to tracepoints, while at the same
> time userspace may open a 'binary stats' FD to peek at the corresponding
> VM stats.
>
> Signed-off-by: Oliver Upton <oliver.upton@xxxxxxxxx>
> ---
> arch/arm64/include/asm/kvm_host.h | 3 ++
> arch/arm64/kvm/guest.c | 5 ++-
> arch/arm64/kvm/vgic/trace.h | 66 +++++++++++++++++++++++++++++++
> arch/arm64/kvm/vgic/vgic-its.c | 14 ++++++-
> include/linux/kvm_host.h | 4 ++
> 5 files changed, 89 insertions(+), 3 deletions(-)
>
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 21c57b812569..6f88b76373a5 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -966,6 +966,9 @@ static inline bool __vcpu_write_sys_reg_to_cpu(u64 val, int reg)
>
> struct kvm_vm_stat {
> struct kvm_vm_stat_generic generic;
> + u64 vgic_its_trans_cache_hit;
> + u64 vgic_its_trans_cache_miss;
> + u64 vgic_its_trans_cache_victim;
> };
>
> struct kvm_vcpu_stat {
> diff --git a/arch/arm64/kvm/guest.c b/arch/arm64/kvm/guest.c
> index aaf1d4939739..354d67251fc2 100644
> --- a/arch/arm64/kvm/guest.c
> +++ b/arch/arm64/kvm/guest.c
> @@ -30,7 +30,10 @@
> #include "trace.h"
>
> const struct _kvm_stats_desc kvm_vm_stats_desc[] = {
> - KVM_GENERIC_VM_STATS()
> + KVM_GENERIC_VM_STATS(),
> + STATS_DESC_COUNTER(VM, vgic_its_trans_cache_hit),
> + STATS_DESC_COUNTER(VM, vgic_its_trans_cache_miss),
> + STATS_DESC_COUNTER(VM, vgic_its_trans_cache_victim)
> };
We've talked about this offline, but I thought I'd make my position
public on these.
I've very concerned that exposing these statistic, however useful they
may be at a given point in time, will eventually become all wrong
and/or misleading.
Case in point, our discussion about this very series, where we landed
on a potential reimplementation of the translation cache as a per-ITS
xarray. If this comes to fruition, these stats will probably be
totally useless (hit monotonically increasing at the rate of 1 per
interrupt, miss being 0 or 1, victim being stuck to 0).
I'm *not* saying that such stats are totally useless. Just that they
are, by definition, tied to a kernel-side implementation choice, which
will evolve. The trouble is that they create an ABI that we need to
support forever.
I wish we could work on something that would allow such statistics to
be *extracted* (as opposed to published). Something like this as a
discussion subject for a future KVM Forum or LPC uConf would get my
full backing (and I'm pretty sure the networking folks have solved
that problem a long time ago).
Thanks,
M.
--
Without deviation from the norm, progress is not possible.
