Re: [Patch 0/2] Renaming 'trace' to 'relay' and enhancements to 'relay'

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 8 Aug 2008 09:22:39 +0530 "K.Prasad" <prasad@linux.vnet.ibm.com> wrote:

> On Wed, Aug 06, 2008 at 11:08:10AM -0400, Frank Ch. Eigler wrote:
> > Andrew Morton <akpm@linux-foundation.org> writes:
> > 
> > > [...]
> > >> 	Please find the patches that enhance the 'trace' infrastructure
> > >> (available in the -mm tree) and which introduce two new APIs
> > >> relay_dump() and relay_printk().
> > >> [...]
> > 
> > > I'm a bit perplexed by these trace patches
> > > (http://userweb.kernel.org/~akpm/mmotm/broken-out/trace-code [...]
> > > Is it useful?  Will it be useful?  [...]  I haven't heard much noise
> > > about it and I'm struggling to justify merging it.
> > 
> > Right.
> > 
> > > Also, it's starting to look somewhat similar to ftrace, which also
> > > provides sort of high-bandwidth per-cpu channels into userspace for
> > > tracing purposes.
> > 
> > Perhaps ftrace ought to use this facility for its debugfs-facing bulk
> > data interface rather than an internal one that cannot be used by
> > anyone else.  Perhaps lttng could use it.  Systemtap could.  I believe
> > a grand unification at this level was the idea.
> >
> 
> Hi Andrew,
> 	The 'relay_printk' and 'relay_dump' interfaces were meant to
> provide clutter-free tracing interface for the user who does not want to
> care much beyond getting his trace output to user-space. It greatly
> helps reduce the amount of work that a tracer needs to perform to setup
> and tear-down. For e.g. when the Block I/O tracing code in
> block/blktrace.c was converted to use these interfaces they resulted in
> code-reduction of ~130 lines (http://lkml.org/lkml/2008/5/16/318).
> 
> The default values can work fine for most tunables and are also exposed
> to the user for overriding them.
> 
> These interfaces will be helpful in almost all non-blocking tracers
> (unlike usbmon which displays information in real-time) and uses the
> scalable infrastructure provided by 'relay'.
> 
> As pointed out by Frank earlier, most tracers (including ftrace) can be
> made to use the above-mentioned interfaces resulting in substantial
> savings in terms of LoC and increasing modularity/code re-usability.
> 

Oh, OK, that's a good case.

Was the result of your blktrace conversion compatible with existing
interfaces?

It would be higly persuasive if we were to see at least a prototype
conversion of ftrace to use this new code (hint :))



On a naming note: I am officially utterly bewildered by the number of
subsystems which call themselves "trace" or "footrace" or "tracebar". 
And we have at least one more (ltt) (a footracebar!) heading in our
direction.

y:/usr/src/linux-2.6.27-rc2> find -type f -a -name '*trace*' | wc -l
144

(!)

Is there something we can do to bring order out of chaos here?