Greetings, Here comes yet another take at the problem of correlating perf samples, timestamped in kernel (with - de-facto - sched_clock values), with performance-related events (be it debug information from JIT engines or energy sensor data obtained via USB or hwmon) generated in user space.
The first patch adds an additional timestamp field in the perf sample data, which can be requested for any perf event along with normal PERF_SAMPLE_TIME. Events with both values appearing periodically in the perf data allow user code to translate raw monotonic time (obtained via POSIX clock API) to sched_clock domain. Although any perf event can be used, the natural choice would be a sched_switch trace event (for processes with root permissions) or a hrtimer-based PERF_COUNT_SW_CPU_CLOCK. One question I haven't found answer to is: could it be even more generic? As in: would it be possible to request a time value from any of the available time sources? It doesn't make sense, I believe, to have a PERF_SAMPLE_* for each possibility. An extra flags in perf_event_attrs maybe? (we still have 39 spare bits there) The second patch, functionally orthogonal but complementing the first one, replicates the "trace_maker" idea from ftrace in the perf world. Instead of a sysfs file, there is an ioctl command, which simply injects a new type of software event into the buffer. The argument can point at a zero-terminated string of PAGE_SIZE max lenght. If provided, it will be copied to the "raw" part of a sample. Of course the event can sample a monotonic clock as well, if used with the above, so one gets means of both synchronisation and time stamp approximation. One doubt I have here is the ioctl argument. It takes a strong now, like trace_marker does. But maybe a simple integer value would suffice? After all the ioctl can be only generated by the "owner" of the perf stream (unlike in trace_marker case, where "anyone" can write to it), so we could rely on him to have a dictionary of events of some sort. On another note, I proposed this subject for the tracing microconference on Plumbers next month. Hope to have some good discussion there. Maybe even a conclusion? (I wish... ;-) Thanks! Pawel Pawel Moll (2): perf: Add sampling of the raw monotonic clock perf: Marker software event and ioctl include/linux/perf_event.h | 2 ++ include/uapi/linux/perf_event.h | 6 ++++- kernel/events/core.c | 55 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 62 insertions(+), 1 deletion(-) -- 1.9.1 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
