Re: [Qemu-devel] [PATCH for-2.8 v1 00/60] Modular build of trace event files

["Daniel P. Berrange" <berrange@redhat.com>]

On Fri, Sep 09, 2016 at 01:03:51PM +0200, Lluís Vilanova wrote:
> Daniel P Berrange writes:
> 
> > On Thu, Sep 08, 2016 at 03:23:26PM +0200, Lluís Vilanova wrote:
> >> Daniel P Berrange writes:
> >> 
> >> > I previously split the global trace-events file up into one file
> >> > per-subdirectory to avoid merge conflict hell.
> >> [...]
> >> 
> >> Sorry, I could not find the message where the infrastructure is modified to
> >> provide this. But I think there's a more efficient way to provide modular
> >> auto-generated tracing code without the hierarchical indexing you proposed.
> 
> > [snip]
> 
> >> struct TraceEvent ___trace_events[] = {
> >> {
> >> .name = "eventname",
> >> .sstate = 1,
> >> .dstate = ___trace_eventname_dstate;
> >> }
> >> }
> >> 
> >> TraceEvent *TRACE_EVENTNAME = &___trace_events[...];
> 
> > Life would be simpler if we had the 'bool dstate' as part of the
> > TraceEvent struct, but doing so would essentially be reverting this
> > previous change:
> 
> >   commit 585ec7273e6fdab902b2128bc6c2a8136aafef04
> >   Author: Paolo Bonzini <pbonzini@redhat.com>
> >   Date:   Wed Oct 28 07:06:27 2015 +0100
> 
> >     trace: track enabled events in a separate array
>     
> >     This is more cache friendly on the fast path, where we already have
> >     the event id available.
> 
> > I asked Paolo about this previously and he indicated it was a notable
> > performance improvement, so we can't put dstate back into the TraceEvent
> > struct :-(
> 
> Sorry, there was a typo in my example code that led to this misunderstanding.
> 
> I meant this for the global auto-generated .c:
> 
>   struct TraceEvent {
>       /* ... */
>       bool *dstate;
>   };
> 
>   bool ___TRACE_EVENTNAME_DSTATE;
> 
>   struct TraceEvent ___trace_events[] = {
>       {
>           .name = "eventname",
>           .sstate = 1,
>           .dstate = &___TRACE_EVENTNAME_DSTATE;
>       }
>   }
> 
>   TraceEvent *TRACE_EVENTNAME = &___trace_events[...];
> 
> 
> If you look at the modified macro I pasted for "trace/control-internal.h":
> 
> 
>   #define trace_event_get_state_dynamic_by_id(id) \
>       (unlikely(trace_events_enabled_count) && \
>        (___ ## id ## _DSTATE))
> 
> We're retaining the fast-path performance of the seggregated boolean dstate
> array, while the event descriptor contains a pointer to the per-event dstate
> global boolean variable in case you want to get/set the dstate based on an event
> pointer.

The various  _DSTATE variables are still arbitrarily scattered in
memory, as opposed to in a contiguous cache friendly array, which
is one of the goals of Paolo's original change.

That said, I'm unclear on how much of the performance win from
Paolo's change came from eliminating the struct field de-reference,
vs having the contiguous array. I'm guessing most of the win is
from the former, the latter only being important if we hit multiple
related tracepoints on close succession.

Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|