Re: [PATCH] tracing: Export key trace event symbols

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- Original Message -----
> On Mon, 20 Apr 2015 16:38:11 -0500
> Ron Rechenmacher <ron@fnal.gov> wrote:
> 
> > If symbols are not exported, modules can no longer register additional
> > (module specified) tracepoints like they use to be able to (i.e
> > linux-3.15.x).
> > Somewhere on or about commit de7b2973903c6cc50b31ee5682a69b2219b9919d
> > (Author: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> > Date:   Tue Apr 8 17:26:21 2014 -0400
> > tracepoint: Use struct pointer instead of name hash for reg/unreg
> > tracepoints)
> > modules which attempted to register additional tracing functions would
> > get "Unknown symbol" errors. For example: "... Unknown symbol
> > __tracepoint_sched_switch (err 0)"
> > Symbols can be exported using the kernel's EXPORT_TRACEPOINT_SYMBOL_GPL
> > macro
> > to allow modules to once again register their own tracing functions (for at
> > least some key points in the kernel as provided by this patch).
> > 
> > Signed-off-by: Ron Rechenmacher <ron@fnal.gov>
> > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=96051
> 
> Hi Ron,
> 
> I was talking with Mathieu on IRC and asked him how LTTng gets its
> kernel tracepoints, and he told me he uses
> for_each_kernel_tracepoint(). That will iterate over all tracepoints
> that have been added in the kernel (and is exported GPL).
> 
> You can still use that to get the handle onto any tracepoint you need.
> It's pretty straight forward (I just wrote a simple module to test it
> out), and just compare against the tp->name, to find what you want.
> 
> I still would like to get more usage out of the internal code, but this
> is your work around you wanted. No need to export new symbols. Just a
> little more setup time on module load.

Hi Ron,

Quoting a snippet of your earlier emails:

"ftrace (if one considers ltt-ng) is probably capable of doing what my trace does
except that the timestamp is not TOD (Time Of Day) -- but probably/maybe could
be made to do so???"

FYI, LTTng and Ftrace are two different projects. LTTng features
am out-of-tree kernel and a user-space tracer, whereas Ftrace features
a kernel tracer, readily available from the Linux kernel sources.

The timestamps can now be the monotonic clock for each of perf, ftrace,
and lttng, thanks to the work on nmi-safe clock source done by Thomas
Gleixner (merged in 3.17).

With the nmi-safe clocks, you'll be able to correlate samples
and traces taken by perf, ftrace, and lttng-modules with LTTng
userspace traces (LTTng-UST). This has been one of the goals of the
Common Trace Format (CTF) since its creation. You might want to look
at the perf-to-ctf conversion feature merged in Linux 4.0.

Best regards,

Mathieu



> 
> -- Steve
> 
> Here's my mod....
> 
> 
> #include <linux/module.h>
> #include <linux/ftrace.h>
> #include <linux/tracepoint.h>
> 
> static func(struct tracepoint *tp, void *ignore)
> {
> 	printk("tracepoint: %s\n", tp->name);
> }
> 
> static int __init my_tp_init(void)
> {
> 	for_each_kernel_tracepoint(func, NULL);
> 	return 0;
> }
> 
> static void __exit my_tp_exit(void)
> {
> }
> 
> module_init(my_tp_init);
> module_exit(my_tp_exit);
> 
> MODULE_AUTHOR("My name here");
> MODULE_DESCRIPTION("Me!");
> MODULE_LICENSE("GPL");
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com