Messages in this thread | | | Date | Sun, 20 Mar 2011 21:05:28 +0100 (CET) | From | Thomas Gleixner <> | Subject | Re: [PATCH 1/5] perf events: add support for realtime clock attribute |
| |
On Sun, 20 Mar 2011, David Ahern wrote:
> > > On 03/20/11 13:08, Thomas Gleixner wrote: > > On Sun, 20 Mar 2011, David Ahern wrote: > >> + > >> + if (sample_type & PERF_SAMPLE_REALTIME) > >> + data->realtime = ktime_to_ns(ktime_get_real()); > > > > That's going to live lock on any sample which happens to hit in a > > xtime_lock write locked section. I knew why I suggested adding > > tracepoints to the timekeeping code :) > > Hmmm.... Can you give an example of when perf samples could be generated > with xtime write locked?
NMI triggered samples can hit at any time.
Any tracepoint which we might add to the time keeping code inside the xtime locked regions. And I don't want to see a restriction there.
> What are the rules for when ktime_get_real can be invoked?
Outside of xtime write locked regions.
Thanks,
tglx
| |