On Thu, Jan 09, 2025, Sean Christopherson wrote:
> On Thu, 19 Dec 2024 17:38:58 -0800, Sean Christopherson wrote:
> > Fix a handful of bugs in the binary stats infrastructure, expand support
> > to vCPU-scoped stats, enumerate all KVM stats in selftests, and use the
> > enumerated stats to assert at compile-time that {vm,vcpu}_get_stat() is
> > getting a stat that actually exists.
> >
> > Most of the bugs are benign, and AFAICT, none actually cause problems in
> > the current code base. The worst of the bugs is lack of validation that
> > the requested stat actually exists, which is quite annoying if someone
> > fat fingers a stat name, tries to get a vCPU stat on a VM FD, etc.
> >
> > [...]
>
> Applied 1-7 to kvm-x86 selftests (x86 wants to build tests on the vCPU-scoped
> stats infrastructure).
>
> I'll hold off on the compile-time assertions stuff until there's consensus that
> we want to go that route for all architectures (not at all urgent).
>
> [1/8] KVM: selftests: Fix mostly theoretical leak of VM's binary stats FD
> https://github.com/kvm-x86/linux/commit/b68ec5b6869f
> [2/8] KVM: selftests: Close VM's binary stats FD when releasing VM
> https://github.com/kvm-x86/linux/commit/a59768d6cb64
> [3/8] KVM: selftests: Assert that __vm_get_stat() actually finds a stat
> https://github.com/kvm-x86/linux/commit/52ef723593fe
> [4/8] KVM: selftests: Macrofy vm_get_stat() to auto-generate stat name string
> https://github.com/kvm-x86/linux/commit/7884da344973
> [5/8] KVM: selftests: Add struct and helpers to wrap binary stats cache
> https://github.com/kvm-x86/linux/commit/384544c026f6
> [6/8] KVM: selftests: Get VM's binary stats FD when opening VM
> https://github.com/kvm-x86/linux/commit/6d22ccb1c309
> [7/8] KVM: selftests: Add infrastructure for getting vCPU binary stats
> https://github.com/kvm-x86/linux/commit/60d432517838
Argh, apparently I only tested this series on platforms with high RLIMIT_NOFILE
values. Creating the stats fd for all vCPUs causes kvm_create_max_vcpus to fail
on some of my systems due to doubling the number of fds needed.
One option would be to figure out a clean way to avoid creating the stats fds
for barebones VMs, but x86's recalc_apic_map_test also fails. That test creates
512 (max "supported" by selftests) using vm_create_with_vcpus().
Doubling the rlimit from 1024+100 to 2048+100 for the max test doesn't seem insane,
so my plan is to move the rlimit twiddling into common code, and then account for
the vCPU stats fds.
I'll drop the above commits and post a v2.
