[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <alpine.LFD.1.10.0809241014250.3265@nehalem.linux-foundation.org>
Date: Wed, 24 Sep 2008 10:25:57 -0700 (PDT)
From: Linus Torvalds <torvalds@...ux-foundation.org>
To: Steven Rostedt <rostedt@...dmis.org>
cc: Peter Zijlstra <peterz@...radead.org>,
Martin Bligh <mbligh@...gle.com>, linux-kernel@...r.kernel.org,
Ingo Molnar <mingo@...e.hu>,
Thomas Gleixner <tglx@...utronix.de>,
Andrew Morton <akpm@...ux-foundation.org>,
prasad@...ux.vnet.ibm.com,
Mathieu Desnoyers <compudj@...stal.dyndns.org>,
"Frank Ch. Eigler" <fche@...hat.com>,
David Wilder <dwilder@...ibm.com>, hch@....de,
Tom Zanussi <zanussi@...cast.net>,
Steven Rostedt <srostedt@...hat.com>
Subject: Re: [RFC PATCH 1/3] Unified trace buffer
On Wed, 24 Sep 2008, Steven Rostedt wrote:
>
> The event counter is just the timestamp (quick patch, simple to fix). The
> term "counter" was bad. It should have been timestamp, which one would
> want a 64bit timestamp.
One definitely would _not_ want the full 64-bit timestamp.
There are two cases:
- lots and lots of events
just do a 32 bit "timestamp delta" to the previous packet (where the
first packet in the queue would be a delta from a value in the queue
header - and we can obviously make that value be the TSC so that the
first delta is always zero, but it may also make sense to make it a
real delta, and the queue header would contain some good
synchronization point value).
- occasional events
Oops. The delta wouldn't fit. So create a new "timestamp update" packet
with a 64-bit thing when doing the reservation. There's obviously no
cost issue (since this would only happen for things where there was a
multi-second delay - or at least an appreciable fraction of a delay -
between events)
This definitely is worth doing. If we have small trace objects (and many
things really do have pretty small traces), using just a 32-bit TSC not
only saves 4 bytes per trace event, but it makes it quite reasonable to
keep the trace data 4-byte-aligned rather than requiring 8-byte alignment.
Of course, if the traces end up being horribly bloated, none of that will
matter. But I really would hope that you we keep the header itself to just
8 bytes (and being 2 4-byte entities), so that small payloads are
reasonable. And that looks doable, if you have a 16-bit "type" and a
16-bit "size" field.
One thing I'd like to do is to also architecturally reserve a few of the
types for internal queue management stuff. Things like "padding" objects
(or a "end-of-ringbuffer" object), and the TSC overflow object, and a
"time sync" object (or heartbeat). So maybe the type would have the high
bit set as a "reserved for internal ringbuffer use" or whatever.
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Powered by blists - more mailing lists