Re: [PATCH] tools lib traceevent: update KVM plugin

[Paolo Bonzini <pbonzini@redhat.com>]

On 20/10/2015 17:26, Steven Rostedt wrote:
> > > > What happens if you run new perf on an older kernel. Is this new plugin
> > > > going to be screwed up? Plugins should be backward compatible.
> > 
> > If you run new perf on older kernel, the new plugin will print the 
> > "role" field (see kvm_mmu_print_role) slightly incorrectly.  That said, 
> > the existing plugin was _also_ printing the role in a wildly wrong 
> > format, like 2.6.35 vintage; the glevels field was removed by commit 
> > 5b7e0102ae74, "KVM: MMU: Replace role.glevels with role.cr4_pae", in 
> > April 2010.
> 
> Can we add a check if glevels field exists, and do the old format if it
> does? I'm very strict on making sure event-parse works with all
> incarnations of kernels.

Note that this is not a glevels _tracepoint_ field.  The tracepoint
field is "role" and it is a 64-bit integer.  The plugin interprets the
bitfields of that integer, and the position/meaning of the bitfields has
changed.

I guess I could use the mmu_valid_gen field as a proxy (it was
introduced after glevels was removed), and only provide the "very old"
(up to 2.6.34) and "new" (4.2+) formats.  The "old" format (2.6.35-4.1)
was never implemented in the plugin; it may well remain that way.

However, it frankly seems a bit academic.  The parsing _will_ work,
apart from printing a nonsensical role just like it has always done for
the past four years.

Paolo