Message-Id: <20080103071609.478486470@goodmis.org>
User-Agent: quilt/0.46-1
Date: Thu, 03 Jan 2008 02:16:09 -0500
From: Steven Rostedt <rostedt@goodmis.org>
To: LKML <linux-kernel@vger.kernel.org>
Cc: Ingo Molnar <mingo@elte.hu>,
       Linus Torvalds <torvalds@linux-foundation.org>,
       Andrew Morton <akpm@linux-foundation.org>,
       Peter Zijlstra <a.p.zijlstra@chello.nl>,
       Christoph Hellwig <hch@infradead.org>,
       Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>,
       Gregory Haskins <ghaskins@novell.com>,
       Arnaldo Carvalho de Melo <acme@ghostprotocols.net>,
       "William L. Irwin" <sparclinux@vger.kernel.org>
Subject: [RFC PATCH 00/11] mcount tracing utility
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 6562
Lines: 175


The following patch series brings to vanilla Linux a bit of the RT kernel
trace facility. This incorporates the "-pg" profiling option of gcc
that will call the "mcount" function for all functions called in
the kernel.

This patch series implements the code for x86 (32 and 64 bit), but
other archs can easily be implemented as well.

Some Background:
----------------

A while back, Ingo Molnar and William Lee Irwin III created a latency tracer
to find problem latency areas in the kernel for the RT patch.  This tracer
became a very integral part of the RT kernel in solving where latency hot
spots were.  One of the features that the latency tracer added was a
function trace.  This function tracer would record all functions that
were called (implemented by the gcc "-pg" option) and would show what was
called when interrupts or preemption was turned off.

This feature is also very helpful in normal debugging. So it's been talked
about taking bits and pieces from the RT latency tracer and bring them
to LKML. But no one had the time to do it.

Arnaldo Carvalho de Melo took a crack at it. He pulled out the mcount
as well as part of the tracing code and made it generic from the point
of the tracing code.  I'm not sure why this stopped. Probably because
Arnaldo is a very busy man, and his efforts had to be utilized elsewhere.

While I still maintain my own Logdev utility:

  http://rostedt.homelinux.com/logdev

I came across a need to do the mcount with logdev too. I was successful
but found that it became very dependent on a lot of code. One thing that
I liked about my logdev utility was that it was very non-intrusive, and has
been easy to port from the Linux 2.0 days. I did not want to burden the
logdev patch with the intrusiveness of mcount (not really that intrusive,
it just needs to add a "notrace" annotation to functions in the kernel
that will cause more conflicts in applying patches for me).

Being close to the holidays, I grabbed Arnaldos old patches and started
massaging them into something that could be useful for logdev, and what
I found out (and talking this over with Arnaldo too) that this can
be much more useful for others as well.

The main thing I changed, was that I made the mcount function itself
generic, and not the dependency on the tracing code.  That is I added

register_mcount_function()
 and
clear_mcount_function()

So when ever mcount is enabled and a function is registered that function
is called for all functions in the kernel that is not labeled with the
"notrace" annotation.

The key thing here is that *any* utility can now hook its own function into
mcount!

The Simple Tracer:
------------------

To show the power of this I also massaged the tracer code that Arnaldo pulled
from the RT patch and made it be a nice example of what can be done
with this.

The function that is registered to mcount has the prototype:

 void func(unsigned long ip, unsigned long parent_ip);

The ip is the address of the function and parent_ip is the address of
the parent function that called it.

The x86_64 version has the assembly call the registered function directly
to save having to do a double function call.

To enable mcount, a sysctl is added:

   /proc/sys/kernel/mcount_enabled

Once mcount is enabled, when a function is registed, it will be called by
all functions. The tracer in this patch series shows how this is done.
It adds a directory in the debugfs, called mctracer. With a ctrl file that
will allow the user have the tracer register its function.  Note, the order
of enabling mcount and registering a function is not important, but both
must be done to initiate the tracing. That is, you can disable tracing
by either disabling mcount or by clearing the registered function.

Only one function may be registered at a time. If another function is
registered, it will simply override what ever was there previously.

Here's a simple example of the tracer output:

CPU 2: hackbench:11867 preempt_schedule+0xc/0x84 <-- avc_has_perm_noaudit+0x45d/0x52c
CPU 1: hackbench:12052 selinux_file_permission+0x10/0x11c <-- security_file_permission+0x16/0x18
CPU 3: hackbench:12017 update_curr+0xe/0x8b <-- put_prev_task_fair+0x24/0x4c
CPU 2: hackbench:11867 avc_audit+0x16/0x9e3 <-- avc_has_perm+0x51/0x63
CPU 0: hackbench:12019 socket_has_perm+0x16/0x7c <-- selinux_socket_sendmsg+0x27/0x3e
CPU 1: hackbench:12052 file_has_perm+0x16/0xbb <-- selinux_file_permission+0x104/0x11c

This is formated like:

 CPU <CPU#>: <task-comm>:<task-pid> <function> <-- <parent-function>


Overhead:
---------

Note that having mcount compiled in seems to show a little overhead.

Here's 3 runs of hackbench 50 without the patches:
Time: 2.137
Time: 2.283
Time: 2.245

 Avg: 2.221

and here's 3 runs with the patches (without tracing on):
Time: 2.738
Time: 2.469
Time: 2.388

  Avg: 2.531

So it is a 13% overhead when enabled (according to hackbench).

But full tracing can cause a bit more problems:

# hackbench 50
Time: 113.350

  113.350!!!!!

But this is tracing *every* function call!


Future:
-------
The way the mcount hook is done here, other utilities can easily add their
own functions. Just care needs to be made not to call anything that is not
marked with notrace, or you will crash the box with recursion. But
even the simple tracer adds a "disabled" feature so in case it happens
to call something that is not marked with notrace, it is a safety net
not to kill the box.

I was originally going to use the relay system to record the data, but
that had a chance of calling functions not marked with notrace. But, if
for example LTTng wanted to use this, it could disable tracing on a CPU
when doing the calls, and this will protect from recusion.

SystemTap:
----------
One thing that Arnaldo and I discussed last year was using systemtap to
add hooks into the kernel to start and stop tracing.  kprobes is too
heavy to do on all funtion calls, but it would be perfect to add to
non hot paths to start the tracer and stop the tracer.

So when debugging the kernel, instead of recompiling with printks
or other markers, you could simply use systemtap to place a trace start
and stop locations and trace the problem areas to see what is happening.


These are just some of the ideas we have with this. And we are sure others
could come up with more.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/