On Tue, 16 Nov 2021 20:21:52 +0100
Nicolas Saenz Julienne <nsaenzju@xxxxxxxxxx> wrote:
> 'trace-cmd list -C' outputs 'tsc2nsec' on my arm64 machine. TSC is an
> x86 specific counter, so it shouldn't be supported. Fix this by making
> sure TSC exists as a trace clock in trace_tsc2nsec_is_supported().
TSC is just an acronym for Time Stamp Counter. Yes, x86 calls it the TSC,
but really it can be any raw clock.
If arm or any other arch has a raw clock that can write directly to the
ring buffer, and perf has a way to convert that to nsecs, we should still
be able to support it. For instance, I believe that powerpc has a ppc-tb
clock that is similar to the tsc of x86. We probably should allow that too.
Hmm, I wonder if the kernel should have a "raw_clock" if the raw time stamp
is used for any arch, and the tsc2nsec can still be used if raw_clock is
supported.
That said, I agree that we should verify that the raw clock that is
converted is available along with the perf info.
-- Steve
>
> Signed-off-by: Nicolas Saenz Julienne <nsaenzju@xxxxxxxxxx>
> ---
> tracecmd/trace-record.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
> index 1767a6c..78e9566 100644
> --- a/tracecmd/trace-record.c
> +++ b/tracecmd/trace-record.c
> @@ -6122,6 +6122,9 @@ static int get_tsc_nsec(int *shift, int *mult)
>
> bool trace_tsc2nsec_is_supported(void)
> {
> + if (!clock_is_supported(NULL, TSC_CLOCK))
> + return false;
> +
> return get_tsc_nsec(NULL, NULL) == 0;
> }
>
|