Re: [Qemu-devel] [PATCH] trace: Trace SCSI request lifecycle

[Stefan Hajnoczi]

On Tue, Nov 09, 2010 at 05:47:24PM +0000, Blue Swirl wrote:
> On Mon, Nov 8, 2010 at 9:30 PM, Stefan Hajnoczi <address@hidden> wrote:
> > On Mon, Nov 8, 2010 at 7:18 PM, Blue Swirl <address@hidden> wrote:
> >> On Mon, Nov 8, 2010 at 9:44 AM, Stefan Hajnoczi
> >> <address@hidden> wrote:
> >>> This patch adds trace events for SCSI request allocation, freeing, CDB
> >>> parsing, read, and write.  These trace events can be used to instrument
> >>> the SCSI request lifecycle.
> >>
> >> How about converting also the DPRINTFs in hw/scsi-disk.c to tracepoints?
> >
> > Good idea.  I don't understand the SCSI emulation well enough to trace
> > the rest, so I couldn't do much besides a mechanical conversion.
> 
> I also did only a mechanical conversion of Sparc32 devices earlier. I
> thought about removing or merging some of them, but then I considered
> that more tracepoints may be better to fewer. Is there a downside?
> 
> There's a difference in the viewpoints. DPRINTFs may have been
> designed with debugging internal logic in mind. Alternatively we could
> be interested in only tracing the external interfaces. I don't think
> these views are in conflict since tracepoints are much more precise
> than DPRINTFs.

Having too many trace events makes it harder to choose which ones to
look at.  But when distros build in trace events by default we'll be
thankful for the ones we have available in production.  There needs to
be a balance between too much noise and not enough information.

I'll send a patch to cover more of SCSI soon.

Stefan