Skip to contentSkip to conversation

tracing: Tracing integration - dynamic ftrace for SystemTap

PATCH In review Atsushi Tsuji 3 participants 6 messages Sep 15, 2009 – Sep 16, 2009

lore
Share & tools
lore
#0 lore raw

cover tracing: Tracing integration - dynamic ftrace for SystemTap

Hi,

Currently, there are some tracers using in Linux. Each of them are developed for different users and situations.

Ftrace has own probe mechanism, dynamic ftrace, and various plugins, and it can be used without particular userland tools. It is suitable for kernel developer. The performance counters have own userland tool merged in the kernel tree and it's easy to trace kernel events by using this tool. It is suitable for people who don't have enough knowledge of internal kernel. SystemTap has own scripting language and users can gather kernel internal information suited to their needs by customizing the script. Since it also has features, the log rotation and the initscript, to keep collecting information continually, it is suitable for system administrator and enterprise users.

I think we need to use those tracers as the situation demands, since no tracer exists to meet anyone's needs and to fit any situation. Then, we should consider the tracing mechanisms (kprobe, dynamic ftrace and etc. kernel tracing features) and tracers (user interface) separately, and all tracers should be able to use each tracing mechanisms.

The mechanism of dynamic ftrace is useful for all tracers. However only ftrace can currently use it. I think it can be used for other tracers via register_ and unregister_ftrace_function_probe function. By using it, the overhead of instrumentation can be reduced and gathering information, which cannot be traced because of a large overhead, become available. Also it can enable the performance counters to count events for all kernel functions.

First, I sent patches to integrate ftrace and SystemTap by just exporting ftrace API. Also I fixed minor bugs for __unregister_ftrace_function_probe.

The patches doesn't contain the performance counters integration, since it is still in development. I think following implementations are needed:
- add new type to perf_id (like PERF_TYPE_FTRACE)
- add event ids to struct dyn_ftrace
- add interfaces to reference event ids by user (via debugfs?)
- add probe function to count ftrace events (call do_perf_swcounter_event)

Those patches are for -tip tree.

Thanks,
Atsushi

---
 kernel/trace/ftrace.c |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)
#1 lore raw
Quoted 7 lines · Atsushi Tsuji

The patches doesn't contain the performance counters integration, since it is still in development. I think following implementations are needed:
- add new type to perf_id (like PERF_TYPE_FTRACE)
- add event ids to struct dyn_ftrace
- add interfaces to reference event ids by user (via debugfs?)
- add probe function to count ftrace events (call do_perf_swcounter_event)

That's long since been done,..

#2 lore raw
Quoted 10 lines · Peter Zijlstra
Quoted 7 lines · Atsushi Tsuji

The patches doesn't contain the performance counters integration, since it is still in development. I think following implementations are needed:
- add new type to perf_id (like PERF_TYPE_FTRACE)
- add event ids to struct dyn_ftrace
- add interfaces to reference event ids by user (via debugfs?)
- add probe function to count ftrace events (call do_perf_swcounter_event)

That's long since been done,..

Or are you talking about the mcount thing? Simply expose the mcount hook as a regular tracepoint and perf can already use it.

#3 lore raw
Quoted 10 lines · Peter Zijlstra
Quoted 9 lines · Peter Zijlstra
Quoted 7 lines · Atsushi Tsuji

The patches doesn't contain the performance counters integration, since it is still in development. I think following implementations are needed:
- add new type to perf_id (like PERF_TYPE_FTRACE)
- add event ids to struct dyn_ftrace
- add interfaces to reference event ids by user (via debugfs?)
- add probe function to count ftrace events (call do_perf_swcounter_event)

Quoted 5 lines
Quoted 1 line

That's long since been done,..

Or are you talking about the mcount thing? Simply expose the mcount hook as a regular tracepoint and perf can already use it.

Peter,

I'm not sure what you mean by exposing the mcount hook as a tracepoint. Or you mean make it a special file in debug/tracing/events/ that lets you enable it as a profile trace?

-- Steve

#4 lore raw
Quoted 13 lines · Steven Rostedt
Quoted 12 lines · Peter Zijlstra
Quoted 10 lines · Peter Zijlstra
Quoted 7 lines · Atsushi Tsuji

The patches doesn't contain the performance counters integration, since it is still in development. I think following implementations are needed:
- add new type to perf_id (like PERF_TYPE_FTRACE)
- add event ids to struct dyn_ftrace
- add interfaces to reference event ids by user (via debugfs?)
- add probe function to count ftrace events (call do_perf_swcounter_event)

That's long since been done,..

Quoted 7 lines
Quoted 3 lines

Or are you talking about the mcount thing? Simply expose the mcount hook as a regular tracepoint and perf can already use it.

I'm not sure what you mean by exposing the mcount hook as a tracepoint. Or you mean make it a special file in debug/tracing/events/ that lets you enable it as a profile trace?

Either make it a single tracepoint on which you can apply filters or make it generate a tracepoint per mcount site, I don't really mind ;-)

If kprobes can live in the same space, then surely the function tracer can too ;-)

#5 lore raw
Quoted 20 lines · Peter Zijlstra
Quoted 10 lines · Steven Rostedt
Quoted 9 lines · Peter Zijlstra
Quoted 8 lines · Peter Zijlstra
Quoted 7 lines · Atsushi Tsuji

The patches doesn't contain the performance counters integration, since it is still in development. I think following implementations are needed:
- add new type to perf_id (like PERF_TYPE_FTRACE)
- add event ids to struct dyn_ftrace
- add interfaces to reference event ids by user (via debugfs?)
- add probe function to count ftrace events (call do_perf_swcounter_event)

Quoted 8 lines
Quoted 4 lines
Quoted 1 line

That's long since been done,..

Or are you talking about the mcount thing? Simply expose the mcount hook as a regular tracepoint and perf can already use it.

I'm not sure what you mean by exposing the mcount hook as a tracepoint. Or you mean make it a special file in debug/tracing/events/ that lets you enable it as a profile trace?

Quoted 5 lines

Either make it a single tracepoint on which you can apply filters or make it generate a tracepoint per mcount site, I don't really mind ;-)

If kprobes can live in the same space, then surely the function tracer can too ;-)

Thank you for your comments.

I understood that I need to make the mechanism to dynamically add tracepoints using a mcount hook. I'll try to make it.

Thanks,
Atsushi