On Fri, Jul 12, 2013 at 11:45:17AM +0200, Ingo Molnar wrote:
> One thing where we could help JITs is to offer a direct channel to any
> perf profiling process: a prctl(SYS_TRACE) which would send a free-form
> string to any profiling task interested in it.
>
> This would be a glorified anonymous write() in essence, without using a
> temporary file.
>
> The advantage would be that the string could be captured as-is and copied
> to the ring-buffer of the profiling task - instead of having to recover it
> later on.
>
> This is a model that I'd generally advocate: a single channel [per
> CPU-ified] for instrumentation/tracing.
'free format text string' is long and cumbersome and requires parsing.
And size is the primary component in speed.
But yes, we could allow injection of something like
struct PERF_RECORD_SYMBOL {
struct perf_event_header header;
u32 pid, tid;
u64 addr;
u64 len;
char symbol[];
};
I still like the idea of actually writing valid ELF DSOs in that that would
also get us the TEXT and allow assembly inspection etc. It might also allow a
JIT to re-map those DSOs and decrease warm-up time -- provided the actual
program didn't change meanwhile.
How to do injection is another thing though; I don't much like prctl(). Then
again, offering a special file like /sys/bus/event_source/sink isn't
particularly pretty either.
Then there is the issue of attaching to an already running JIT; we'd need means
to 'catch' up. The DSOs trivially allow this; the injection not so much.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>