Re: [PATCH v2 05/13] perf: Force architectures to opt-in to guest callbacks

From: Sean Christopherson
Date: Tue Sep 21 2021 - 17:30:04 EST


On Tue, Sep 21, 2021, Paolo Bonzini wrote:
> On 28/08/21 21:47, Peter Zijlstra wrote:
> > > +config HAVE_GUEST_PERF_EVENTS
> > > + bool
> > depends on HAVE_KVM
>
> It won't really do anything, since Kconfig does not detects conflicts
> between select' and 'depends on' clauses.

It does throw a WARN, though the build doesn't fail.

WARNING: unmet direct dependencies detected for HAVE_GUEST_PERF_EVENTS
Depends on [n]: HAVE_KVM [=n] && HAVE_PERF_EVENTS [=y]
Selected by [y]:
- ARM64 [=y]

WARNING: unmet direct dependencies detected for HAVE_GUEST_PERF_EVENTS
Depends on [n]: HAVE_KVM [=n] && HAVE_PERF_EVENTS [=y]
Selected by [y]:
- ARM64 [=y]

WARNING: unmet direct dependencies detected for HAVE_GUEST_PERF_EVENTS
Depends on [n]: HAVE_KVM [=n] && HAVE_PERF_EVENTS [=y]
Selected by [y]:
- ARM64 [=y]

> Rather, should the symbol be selected by KVM, instead of ARM64 and X86?

By KVM, you mean KVM in arm64 and x86, correct? Because HAVE_GUEST_PERF_EVENTS
should not be selected for s390, PPC, or MIPS.

Oh, and Xen also uses the callbacks on x86, which means the HAVE_KVM part is
arguabably wrong, even though it's guaranteed to be true for the XEN_PV case.
I'll drop that dependency and send out a separate series to clean up the arm64
side of HAVE_KVM.

The reason I didn't bury HAVE_GUEST_PERF_EVENTS under KVM (and XEN_PV) is that
there are number of references to the callbacks throught perf and I didn't want
to create #ifdef hell.

But I think I figured out a not-awful solution. If there are wrappers+stubs for
the guest callback users, then the new Kconfig can be selected on-demand instead
of unconditionally by arm64 and x86. That has the added bonus of eliminating
the relevant code paths for !KVM (and !XEN_PV on x86), with or without static_call.
It also obviates the needs for __KVM_WANT_GUEST_PERF_EVENTS or whatever I called
that thing.

It more or less requires defining the static calls in generic perf, but I think
that actually ends up being good thing as it consolidates more code without
introducing more #ifdefs. The diffstats for the static_call() conversions are
also quite nice.

include/linux/perf_event.h | 28 ++++++----------------------
kernel/events/core.c | 15 +++++++++++++++
2 files changed, 21 insertions(+), 22 deletions(-)

I'll try to get a new version out today or tomorrow.