Re: Unified tracing buffer

[Mathieu Desnoyers <compudj@krystal.dyndns.org>]

* Peter Zijlstra (a.p.zijlstra@chello.nl) wrote:
> On Mon, 2008-09-22 at 09:29 -0700, Martin Bligh wrote:
> > >> In conjunction with the previous email on this thread
> > >> (http://lkml.org/lkml/2008/9/22/160), may I suggest
> > >> the equivalent interfaces in -mm tree (2.6.27-rc5-mm1) to be:
> > >>
> > >> relay_printk(<some struct with default filenames/pathnames>, <string>,
> > >> ....) ;
> > >> relay_dump(<some struct with default filenames/pathnames>, <binary
> > >> data>);
> > >> and
> > >> relay_cleanup_all(<the struct name>); - Single interface that cleans up
> > >> all files/directories/output data created under a logical entity.
> > >
> > > Dude, relayfs is such a bad performing mess that extending it seems like
> > > a bad idea. Better to write something new and delete everything relayfs
> > > related.
> > 
> > There did seem to be pretty universal agreement that we'd rather not
> > use relayfs.
> > 
> > > Also, it seems prudent to separate the ring-buffer implementation from
> > > the event encoding/decoding facilities.
> > 
> > Right - in conversation I had with Mathieu later, he suggested cleaning up
> > relayfs - I fear this will delay us far too long, and get bogged down.
> > If we can get one clean circular buffer implementation, then both
> > relayfs and the tracing could share that common solution,
> 
> Currently only blktrace and kvmtrace use relayfs, and I've heard people
> talk about converting both to use lttng/ftrace infrastructure. At which
> point relayfs is orphaned and ready for removal.
> 

LTTng sits on top of relay for buffer allocation and for the mmap
operation (that's about it, it overrides the rest).

Mathieu


-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68