Re: [Qemu-devel] Our use of #include is undisciplined, and what to do about it

[Stefan Hajnoczi]

On Wed, Mar 16, 2016 at 06:27:48PM +0000, Dr. David Alan Gilbert wrote:
> * Stefan Hajnoczi (address@hidden) wrote:
> > On Tue, Mar 15, 2016 at 01:56:47PM +0000, Dr. David Alan Gilbert wrote:
> > > > This would put trace_foo() in generated-tracers-virtio-blk.h and
> > > > trace_bar() in generated-tracers-memory.h.  Source files using tracing
> > > > would need to include headers for relevant sections.
> > > > 
> > > > This way we can narrow the scope of tracing headers and prevent global
> > > > recompilation.
> > > > 
> > > > The fly in the ointment is that trace/control.h defines enum
> > > > TraceEventID, a global numbering of all trace events.  The enum is used
> > > > in trace/contro.h APIs and also in the simpletrace file format.
> > > > 
> > > > If ./trace-event is modified the numbering of trace events could change.
> > > > This would require global recompilation :(.
> > > > 
> > > > So in order to avoid global recompilation we need to eliminate enum
> > > > TraceEventID.  Perhaps it's possible to use TraceEvent* instead of
> > > > TraceEventID.  For the simpletrace backend we would continue to use a
> > > > global ordering but that only affects generated-tracers.c and not header
> > > > files (thankfully!).
> > > 
> > > I think the hardest part is going to be understanding all of the different
> > > tracers and the things they need to know; like that thing about 
> > > simpletrace
> > > needing the single ordering;  I don't know what the other backends need
> > > but I bet they've each got their own surprise.
> > > 
> > > (And yes, I'd love to split them up - I tend to use traceing quite a bit
> > > in migration now and trying to do a bisect over a big migration patch
> > > series takes ages mostly because of the trace.h changes)
> > 
> > Yes, this is the hard part.  The simpletrace.py pretty-printing script
> > takes a trace-events file as input.  I suppose it could alphabetically
> > sort the trace-events filenames from the command-line to produce a
> > global ordering of trace event IDs.  Of course that doesn't work
> > properly if the script is invoked with relative paths from a
> > sub-directory...
> > 
> > Perhaps we need a new script during the QEMU build process that produces
> > a trace-event-ids.csv file.  Then simpletrace.py could take that instead
> > of processing ./trace-events.
> 
> Well that sounds easy - but is it easy to avoid using those fixed IDs
> in any of the trace.h functions that are included everywhere (without
> making them more expensive).

How about this:

Instead of using TraceEventID in tracing APIs, pass the TraceEvent
pointer directly.  The tracetool.py code generator should produce
"extern TraceEvent trace_event_foo;" definitions in headers so that the
users have direct access to the TraceEvents.

Stefan

signature.asc
Description: PGP signature