Peter Zijlstra [off-list ref] writes:

On Mon, Jan 07, 2019 at 04:27:27PM +0000, Andrew Murray wrote:

quoted

For drivers that do not support context exclusion let's advertise the
PERF_PMU_CAP_NOEXCLUDE capability. This ensures that perf will
prevent us from handling events where any exclusion flags are set.
Let's also remove the now unnecessary check for exclusion flags.

Signed-off-by: Andrew Murray <redacted>
---
 arch/x86/events/amd/ibs.c          | 13 +------------
 arch/x86/events/amd/power.c        | 10 ++--------
 arch/x86/events/intel/cstate.c     | 12 +++---------
 arch/x86/events/intel/rapl.c       |  9 ++-------
 arch/x86/events/intel/uncore_snb.c |  9 ++-------
 arch/x86/events/msr.c              | 10 ++--------
 6 files changed, 12 insertions(+), 51 deletions(-)

You (correctly) don't add CAP_NO_EXCLUDE to the main x86 pmu code, but
then you also don't check if it handles all the various exclude options
correctly/consistently.

Now; I must admit that that is a bit of a maze, but I think we can at
least add exclude_idle and exclude_hv fails in there, nothing uses those
afaict.

On the various exclude options; they are as follows (IIUC):

  - exclude_guest: we're a HV/host-kernel and we don't want the counter
                   to run when we run a guest context.

  - exclude_host: we're a HV/host-kernel and we don't want the counter
                  to run when we run in host context.

  - exclude_hv: we're a guest and don't want the counter to run in HV
                context.

Now, KVM always implies exclude_hv afaict (for guests)

On Power it mostly does.

There's some host code that can run in real mode (MMU off) and therefore
doesn't do a full context switch out of the guest (including the PMU),
so that's host code that is running while the guest PMCs are still
counting.

cheers