[lkml]   [2009]   [Sep]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    Subject[PATCH 00/18] tracing/ring-buffer: fixes for latency tracer [2.6.32]

    I can port these over to -rt if you want, or you could just
    add them yourself. I know the spinlocks need to be converted to atomic.
    Just let me know if you want me to do it.


    This patch series started out fixing a single bug that Arnaldo ran into
    with the -rt patch series. I've hit it too and wrote a patch to solve it.
    Unfortunately the patch was racey and error prone and would crash easily
    in Arnaldo's setup.

    The problem was with the wakeup latency tracer. When it was developed,
    it controlled what went into the ring buffer. But now that we have
    events, which can write into the ring buffer at anytime unless they are
    disabled. When a max latency is encountered, the buffers are swapped
    and the running ring buffer is reset. But if a commit is happening
    at this time, it will be corrupted.

    Luckily the ring buffer is robust enough to detect this and instead of
    crashing the kernel, it safely disables the ring buffer and outputs
    a back trace. But this does not help the developer that wants the trace.

    Talking with Thomas Gleixner, we both agreed that resetting the ring
    buffer on the fly was dangerous. Instead we should insert a marker and
    use that as the start of the trace.

    I thought this would be an easy change, but after a week of debugging,
    fixing bugs just to create two new ones for every one I fixed, this
    was not as easy as I thought.

    The problem was that you can not write a marker in a buffer that is
    not on the same CPU as the writer. I tried to add flags to make the
    ring buffer reset itself when a new write happens on that CPU but
    that just proved to be even more racey.

    Finally (something I should have realized from day 1, not 4 days later)
    I realized I already had a marker. The latency tracers always record
    the timestamp of when the trace began. I can simply ignore any trace
    event that happened before that timestamp. This ended up working out
    very well. The code is quite simple and solid.

    But this journey through the marker bits was not all in vain. I actually
    came across several outstanding bugs in both the tracer and the ring buffer.
    Nothing major, but enough to be fixed.

    Not just the latency tracers could cause a corruption, but the reset
    of the ring buffers by the switching of plugins could also cause it.
    All resets must synchronize the disabling of the ring buffers with any
    current writers.

    Next I found that the swapping of the buffers with the wakeup tracer
    was causing issues. The events would use the trace_array variable *tr
    to access the buffers. tr->buffer to reserve space on the buffer
    and then tr->buffer to commit it. The problem is that the wakeup
    tracer swaps the tr->buffer with the max_tr.buffer. The commit can
    happen on another buffer than what it started with. This again would
    be detected by the ring buffer and would shut it down. The solution
    here was to change the API to the writing of ftrace to pass in
    the buffer directly and not the trace_array. Thus the events would pass
    in the same buffer for both the reserve and the commit. It's OK
    to update the max_tr on the wakeup tracer.

    The irqsoff trace posed a different problem. It does not swap the
    buffers, but it swaps a per cpu buffer within the buffer. This
    can be done because the irqsoff tracer only records per cpu and does
    not record the events of other cpus. But if a latency is detected
    in the middle of a commit (commits can be preempted by interrupts) then
    we can not swap the buffer. This is a tougher problem and I currently
    do not have solution since it would require a copy of one buffer to the
    other. Perhaps in the future. But instead of just failing, I added a
    print into the buffer to notify the users that this has occurred.
    They will now see:

    # tracer: irqsoff
    # irqsoff latency trace v1.1.5 on 2.6.31-rc5
    # --------------------------------------------------------------------
    # latency: 151 us, #1/1, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
    # -----------------
    # | task: -4328 (uid:0 nice:0 policy:0 rt_prio:0)
    # -----------------
    # => started at: save_args
    # => ended at: __do_softirq
    # _------=> CPU#
    # / _-----=> irqs-off
    # | / _----=> need-resched
    # || / _---=> hardirq/softirq
    # ||| / _--=> preempt-depth
    # |||| /
    # ||||| delay
    # cmd pid ||||| time | caller
    # \ / ||||| \ | /
    bash-4328 3d.s5 167us : update_max_tr_single: Failed to swap buffers due to commit in progress

    Which is better than a wrong trace. It still shows the max latency and the
    start and stop points. Note, this run had lockdep enabled with locking
    events on, which spews a lot of events after a latency has been hit
    (the 167 entry with respect to the 151 max). This should not be too big of
    a deal since it took a while loop of constant resettng of the max latency
    and greping for this failure to hit it.

    Because the irqsoff tracer swaps the internal ring buffer cpus,
    a check needs to be added in the recording to handle this. Because this
    check is only needed for swapping the internal buffers, and the irqsoff
    (and preemptoff) tracer is the only user. I only do this if they
    are configured (keep the overhead down).

    Please pull the latest tip/tracing/core tree, which can be found at:


    Steven Rostedt (18):
    ring-buffer: do not reset while in a commit
    ring-buffer: do not swap buffers during a commit
    ring-buffer: remove unnecessary cpu_relax
    ring-buffer: fix ring_buffer_read crossing pages
    ring-buffer: remove ring_buffer_event_discard
    ring-buffer: do not count discarded events
    ring-buffer: disable all cpu buffers when one finds a problem
    tracing: print out start and stop in latency traces
    tracing: disable update max tracer while reading trace
    tracing: disable buffers and synchronize_sched before resetting
    tracing: remove users of tracing_reset
    tracing: use timestamp to determine start of latency traces
    tracing: make tracing_reset safe for external use
    tracing: pass around ring buffer instead of tracer
    tracing: add trace_array_printk for internal tracers to use
    tracing: report error in trace if we fail to swap latency buffer
    ring-buffer: check for swapped buffers in start of committing
    ring-buffer: only enable ring_buffer_swap_cpu when needed

    include/linux/ftrace_event.h | 15 ++-
    include/linux/ring_buffer.h | 23 +--
    include/trace/ftrace.h | 15 ++-
    kernel/trace/Kconfig | 8 +
    kernel/trace/blktrace.c | 12 +-
    kernel/trace/kmemtrace.c | 4 +-
    kernel/trace/ring_buffer.c | 172 +++++++++++++-------
    kernel/trace/trace.c | 297 ++++++++++++++++++++++++----------
    kernel/trace/trace.h | 26 ++--
    kernel/trace/trace_boot.c | 16 +-
    kernel/trace/trace_events.c | 6 +-
    kernel/trace/trace_functions_graph.c | 14 +-
    kernel/trace/trace_irqsoff.c | 3 +-
    kernel/trace/trace_mmiotrace.c | 10 +-
    kernel/trace/trace_power.c | 22 ++-
    kernel/trace/trace_sched_switch.c | 18 ++-
    kernel/trace/trace_sched_wakeup.c | 7 +-
    kernel/trace/trace_syscalls.c | 18 ++-
    18 files changed, 444 insertions(+), 242 deletions(-)

     \ /
      Last update: 2009-09-05 02:05    [W:0.029 / U:56.528 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site