[lkml]   [2008]   [Jan]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    Subject[RFC PATCH 00/23 -v4] mcount and latency tracing utility -v4
    version 4 of mcount patches:

    changes include:

    Folded several patches together that made up tracer.c

    changed the schedule context switch marker use pointers instead of
    simply listing pid and state.

    Added wakeup latency timings.

    Added John Stultz new RCU implementation of get_monotonic_cycles
    (with special thanks to Mathieu Desnoyers for having the original

    Thanks also go out to Frank Ch. Eigler for helping out on deciding
    how to use that scheduler marker.

    All released version of these patches can be found at:

    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.

    Note: I did investigate using -finstrument-functions but that adds a call
    to both start and end of a function. Using mcount only does the
    beginning of the function. mcount alone adds ~13% overhead. The
    -finstrument-functions added ~19%. Also it caused me to do tricks with
    inline, because it adds the function calls to inline functions as well.

    This patch series implements the code for x86 (32 and 64 bit), but
    other archs can easily be implemented as well (note: ARM and PPC are
    already implemented in -rt)

    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:

    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


    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 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:


    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>

    Latency Tracer Format:

    The format used by the RT patch is a bit more complex. It is designed to
    record a lot of information quickly and dump out a lot too.

    There's two versions of the format. Verbose and non-vebose.


    preemption latency trace v1.1.5 on 2.6.24-rc7-tst
    latency: 89 us, #3/3, CPU#1 | (M:desktop VP:0, KP:0, SP:0 HP:0 #P:4)
    | task: kjournald-600 (uid:0 nice:-5 policy:0 rt_prio:0)
    => started at: _spin_lock_irqsave+0x2a/0x63 <c06310d2>
    => ended at: _spin_unlock_irqrestore+0x32/0x41 <c0631245>

    kjournald 600 1 1 00000000 00000000 [397408f1] 0.003ms (+0.079ms): _spin_lock_irqsave+0x2a/0x63 <c06310d2> (scsi_dispatch_cmd+0x155/0x234 [scsi_mod] <f8867c19>)
    kjournald 600 1 1 00000000 00000001 [39740940] 0.081ms (+0.005ms): _spin_unlock_irqrestore+0x32/0x41 <c0631245> (scsi_dispatch_cmd+0x1be/0x234 [scsi_mod] <f8867c82>)
    kjournald 600 1 1 00000000 00000002 [39740945] 0.087ms (+0.000ms): trace_hardirqs_on_caller+0x74/0x86 <c0508bdc> (_spin_unlock_irqrestore+0x32/0x41 <c0631245>)


    preemption latency trace v1.1.5 on 2.6.24-rc7-tst
    latency: 89 us, #3/3, CPU#2 | (M:desktop VP:0, KP:0, SP:0 HP:0 #P:4)
    | task: kjournald-600 (uid:0 nice:-5 policy:0 rt_prio:0)
    => started at: _spin_lock_irqsave+0x2a/0x63 <c06310d2>
    => ended at: _spin_unlock_irqrestore+0x32/0x41 <c0631245>

    _------=> CPU#
    / _-----=> irqs-off
    | / _----=> need-resched
    || / _---=> hardirq/softirq
    ||| / _--=> preempt-depth
    |||| /
    ||||| delay
    cmd pid ||||| time | caller
    \ / ||||| \ | /
    kjournal-600 1d... 3us+: _spin_lock_irqsave+0x2a/0x63 <c06310d2> (scsi_dispatch_cmd+0x155/0x234 [scsi_mod] <f8867c19>)
    kjournal-600 1d... 81us+: _spin_unlock_irqrestore+0x32/0x41 <c0631245> (scsi_dispatch_cmd+0x1be/0x234 [scsi_mod] <f8867c82>)
    kjournal-600 1d... 87us : trace_hardirqs_on_caller+0x74/0x86 <c0508bdc> (_spin_unlock_irqrestore+0x32/0x41 <c0631245>)

    Debug FS:

    Although enabling and disabling mcount is done through the sysctl:


    The rest of the tracing uses debugfs.


    Here's the available files:

    echo 1 to this enables mcount tracing (if mcount_enabled is set)
    echo 0 to disable the function trace tracing.

    Outputs the function trace in latency_trace format.

    echo 1 to enable function tracing in critical sections timings
    echo 0 to disable

    Outputs the critical section latency

    echo a number (in usecs) into this to record all traces that are
    greater than threshold.

    echo "symonly" to not show the instruction pointers in the trace
    echo "nosymonly" to disable symonly.
    echo "verbose" for verbose output from latency format.
    echo "noverbose" to disable verbose ouput.
    cat iter_ctrl to see the current settings.

    Holds the current max critical latency.
    echo 0 to reset and start tracing.

    Holds the current max wakeup latency (usecs)

    Holds the threshold to print after a latency exceeds it (usecs)

    Turn on or off wakeup latency

    Holds the output of the highest wakeup latency.

    simple output format of the function trace code.


    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


    But this is tracing *every* function call!

    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.

    This will be the last series that has each of the traces as a separate
    buffer. The next series will be more like the RT patch implementation
    of holding a single buffer for all latency traces. The only disadvantage
    of this is that you can only perform one type of latency at a time.
    But this has never been brought up as an issue with the RT patch.

    The user will still be able to switch at runtime which type of latency
    they would like to record.

    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.

    Latency Tracing:

    We can also add trace points to record the time the highest priority task
    needs to wait before running. This too is currently done in the RT patch.

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

    These patches are for the underlining work. We'll see what happens next.

     \ /
      Last update: 2008-01-21 16:27    [W:0.035 / U:68.156 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site