Re: [RFC PATCH 1/3] Unified trace buffer

[Ingo Molnar]

* Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> > > Or if we already have the data together:
> > > 
> > > ring_buffer_write(buffer, event_id, length, data);
> > 
> > Don't like the event_id, just stick to plain binary data, and leave 
> > interpretation open to whoemever uses it.
> 
> This is up to debate. I know you don't like this extra event layer, 
> but seriously, all my uses with ring buffers has had some kind of 
> event. But then I'm sure you can argue that if you are using a single 
> type you can can the event.
> 
> I'm open to doing this, but I would like a consensus on this.
> 
> Martin, Thomas, Ingo, Linus?

i'd prefer Peter's simplification and not pass event_id along. Since 
static types are lost anyway (which is the biggest cost and risk of any 
such abstraction), we have to convert between types early on. Whether 
event_id is visible in the API is no big difference.

(It might be cheaper to not pass it along even if everyone ends up using 
it - as it has no semantic meaning anyway.)

pretty much the only generic tracing information is time and payload 
size. ( but even a time key is debatable - there are various resolutions 
needed by different usecases. Some usecases are even fine without any 
timestamps at all - they just want to know the ordering of events and 
that's it.)

i'd like to see some automatic type protection though, as an 
off-by-default debug option: encode the record type on storing and 
double-check it on extraction. So it should be possible to reliably 
store/restore a typed trace buffer and notice corruption early in 
testing.

because there's one thing that is far more important tracer feature than 
sheer performance: robustness.

Automated type checking in debug mode would also mean we could go for 
RLE encoding much more agressively. Most of the risks of a more complex, 
more compressed and pointer-laden data format come from type mismatches 
and the loss of compiler protection against human errors/stupidity. 
(running off the end of the page, misinterpreting a pointer, a record, 
etc.)

	Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/