Em Fri, Jul 22, 2022 at 04:22:12PM -0400, Steven Rostedt escreveu:
> [ Added the user space perf folks ]
>
> On Fri, 22 Jul 2022 18:45:30 +0000
> Chuck Lever III <chuck.lever@xxxxxxxxxx> wrote:
>
> > >> +TRACE_EVENT(xs_data_ready,
> > >> + TP_PROTO(
> > >> + const struct rpc_xprt *xprt
> > >> + ),
> > >> +
> > >> + TP_ARGS(xprt),
> > >> +
> > >> + TP_STRUCT__entry(
> > >> + __sockaddr(addr, xprt->addrlen)
> > >> + ),
> > >> +
> > >> + TP_fast_assign(
> > >> + __assign_sockaddr(addr, &xprt->addr, xprt->addrlen);
> > >> + ),
> > >> +
> > >> + TP_printk("peer=%pISpc", __get_sockaddr(addr))
> > >
> > > NACK. Please resolve and store the string up front instead of storing
> > > the sockaddr. Most versions of perf can't resolve those kernel-specific
> > > %p printks and just end up barfing on them.
> >
> > Interesting. We added get_sockaddr() to avoid this issue in
> > trace-cmd. Sounds like perf needs to be fixed up too, or
> > maybe this is another case of having an old libtraceevent?
> >
> > Meanwhile, I can revert this back to the old way of handling
> > presentation addresses.
> >
>
> Hmm, I thought that perf now uses the external libtraceevent.
>
> Perhaps it hasn't been updated to the latest release that has the ability
> to parse this.
>
> Maybe just install
>
> git://git.kernel.org/pub/scm/libs/libtrace/libtraceevent.git
To use it one has to use:
make -C tools/perf LIBTRACEEVENT_DYNAMIC=1
Then we get it linked with libtraceevent-devel:
$ ldd ~/bin/perf | grep traceevent
libtraceevent.so.1 => /lib64/libtraceevent.so.1 (0x00007faa50f93000)
$
Perhaps it'd be better to check if libtracevent-devel is installed and
use it, falling back to tools/lib/traceevent/ and then adding a warning
that the in-tree codebase is being used?
- Arnaldo
