On Fri, Jul 12, 2019 at 12:55 PM Y Song <ys114321@xxxxxxxxx> wrote:
>
> On Fri, Jul 12, 2019 at 11:24 AM Andrii Nakryiko
> <andrii.nakryiko@xxxxxxxxx> wrote:
> >
> > On Fri, Jul 12, 2019 at 10:46 AM Ilya Leoshkevich <iii@xxxxxxxxxxxxx> wrote:
> > >
> > > Many s390 setups (most notably, KVM guests) do not have access to
> > > hardware performance events.
> > >
> > > Therefore, use the software event instead.
> > >
> > > Signed-off-by: Ilya Leoshkevich <iii@xxxxxxxxxxxxx>
> > > Acked-by: Vasily Gorbik <gor@xxxxxxxxxxxxx>
> > > ---
> > > tools/testing/selftests/bpf/prog_tests/send_signal.c | 9 +++++++++
> > > 1 file changed, 9 insertions(+)
> > >
> > > diff --git a/tools/testing/selftests/bpf/prog_tests/send_signal.c b/tools/testing/selftests/bpf/prog_tests/send_signal.c
> > > index 67cea1686305..4a45ea0b8448 100644
> > > --- a/tools/testing/selftests/bpf/prog_tests/send_signal.c
> > > +++ b/tools/testing/selftests/bpf/prog_tests/send_signal.c
> > > @@ -176,10 +176,19 @@ static int test_send_signal_tracepoint(void)
> > > static int test_send_signal_nmi(void)
> > > {
> > > struct perf_event_attr attr = {
> > > +#if defined(__s390__)
> > > + /* Many s390 setups (most notably, KVM guests) do not have
> > > + * access to hardware performance events.
> > > + */
> > > + .sample_period = 1,
> > > + .type = PERF_TYPE_SOFTWARE,
> > > + .config = PERF_COUNT_SW_CPU_CLOCK,
> > > +#else
> >
> > Is there any harm in switching all archs to software event? I'd rather
> > avoid all those special arch cases, which will be really hard to test
> > for people without direct access to them.
>
> I still like to do hardware cpu_cycles in order to test nmi.
> In a physical box.
> $ perf list
> List of pre-defined events (to be used in -e):
>
> branch-instructions OR branches [Hardware event]
> branch-misses [Hardware event]
> bus-cycles [Hardware event]
> cache-misses [Hardware event]
> cache-references [Hardware event]
> cpu-cycles OR cycles [Hardware event]
> instructions [Hardware event]
> ref-cycles [Hardware event]
>
> alignment-faults [Software event]
> bpf-output [Software event]
> context-switches OR cs [Software event]
> cpu-clock [Software event]
> cpu-migrations OR migrations [Software event]
> dummy [Software event]
> emulation-faults [Software event]
> major-faults [Software event]
> minor-faults [Software event]
> page-faults OR faults [Software event]
> task-clock [Software event]
>
> L1-dcache-load-misses [Hardware cache event]
> ...
>
> In a VM
> $ perf list
> List of pre-defined events (to be used in -e):
>
> alignment-faults [Software event]
> bpf-output [Software event]
> context-switches OR cs [Software event]
> cpu-clock [Software event]
> cpu-migrations OR migrations [Software event]
> dummy [Software event]
> emulation-faults [Software event]
> major-faults [Software event]
> minor-faults [Software event]
> page-faults OR faults [Software event]
> task-clock [Software event]
>
> msr/smi/ [Kernel PMU
> event]
> msr/tsc/ [Kernel PMU event]
> .....
>
> Is it possible that we detect at runtime whether the hardware
> cpu_cycles available or not?
> If available, let us do hardware one. Otherwise, skip or do the
> software one? The software one does not really do nmi so it will take
> the same code path in kernel as tracepoint.
Yeah, that's what I was worried about.
Ilya, could you please take a look how hard would it be to do this HW
vs SW perf event support?
>
> >
> > > .sample_freq = 50,
> > > .freq = 1,
> > > .type = PERF_TYPE_HARDWARE,
> > > .config = PERF_COUNT_HW_CPU_CYCLES,
> > > +#endif
> > > };
> > >
> > > return test_send_signal_common(&attr, BPF_PROG_TYPE_PERF_EVENT, "perf_event");
> > > --
> > > 2.21.0
> > >
