On Mon, Feb 29, 2016 at 03:39:35PM +0000, Will Deacon wrote:
> Hi Jan,
>
> I've queued this lot on my perf/updates branch, but I just noticed an
> oddity whilst dealing with some potential conflicts with the kvm tree.
>
> On Thu, Feb 18, 2016 at 05:50:13PM +0100, Jan Glauber wrote:
> > With the long cycle counter bit (LC) disabled the cycle counter is not
> > working on ThunderX SOC (ThunderX only implements Aarch64).
> > Also, according to documentation LC == 0 is deprecated.
> >
> > To keep the code simple the patch does not introduce 64 bit wide counter
> > functions. Instead writing the cycle counter always sets the upper
> > 32 bits so overflow interrupts are generated as before.
> >
> > Original patch from Andrew Pinksi <andrew.pin...@caviumnetworks.com>
> >
> > Signed-off-by: Jan Glauber <jglau...@cavium.com>
> > ---
> > arch/arm64/kernel/perf_event.c | 21 ++++++++++++++++-----
> > 1 file changed, 16 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
> > index 0ed05f6..c68fa98 100644
> > --- a/arch/arm64/kernel/perf_event.c
> > +++ b/arch/arm64/kernel/perf_event.c
> > @@ -405,6 +405,7 @@ static const struct attribute_group
> > *armv8_pmuv3_attr_groups[] = {
> > #define ARMV8_PMCR_D (1 << 3) /* CCNT counts every 64th cpu
> > cycle */
> > #define ARMV8_PMCR_X (1 << 4) /* Export to ETM */
> > #define ARMV8_PMCR_DP (1 << 5) /* Disable CCNT if
> > non-invasive debug*/
> > +#define ARMV8_PMCR_LC (1 << 6) /* Overflow on 64 bit cycle
> > counter */
> > #define ARMV8_PMCR_N_SHIFT 11 /* Number of counters
> > supported */
> > #define ARMV8_PMCR_N_MASK 0x1f
> > #define ARMV8_PMCR_MASK 0x3f /* Mask for writable bits */
>
> You haven't extended this mask to cover the LC bit, so it will be ignored
> by armv8pmu_pmcr_write afaict.
This is weird. I've double checked and I missed this mask. Annoying.
Nevertheless it works for me without the LC bit set.
> How did you test this? I can easily update the mask, but it would be
> good to know that it doesn't end up cause a breakage.
For testing I used:
- perf top and perf record & report
- looked at interrupt numbers in /proc/interrupts
Without the patch _no_ samples at all are recorded and the interrupt does
not occur. With the patch I get samples and see a reasonable number of
interrupts.
Extending the mask so the LC bit is covered would make sense, I'm going
to test this now.
Jan
> Will
