Notes on Analysing Behaviour Using Events and Tracepoints

                        Documentation written by Mel Gorman
                PCL information heavily based on email from Ingo Molnar

1. Introduction
===============

Tracepoints (see Documentation/trace/tracepoints.txt) can be used without
creating custom kernel modules to register probe functions using the event
tracing infrastructure.

Simplistically, tracepoints represent important events that can be
taken in conjunction with other tracepoints to build a "Big Picture" of
what is going on within the system. There are a large number of methods for
gathering and interpreting these events. Lacking any current Best Practises,
this document describes some of the methods that can be used.

This document assumes that debugfs is mounted on /sys/kernel/debug and that
the appropriate tracing options have been configured into the kernel. It is
assumed that the PCL tool tools/perf has been installed and is in your path.

2. Listing Available Events
===========================

2.1 Standard Utilities
----------------------

All possible events are visible from /sys/kernel/debug/tracing/events. Simply
calling

  $ find /sys/kernel/debug/tracing/events -type d

will give a fair indication of the number of events available.

2.2 PCL (Performance Counters for Linux)
-------

Discovery and enumeration of all counters and events, including tracepoints,
are available with the perf tool. Getting a list of available events is a
simple case of:

  $ perf list 2>&1 | grep Tracepoint
  ext4:ext4_free_inode                     [Tracepoint event]
  ext4:ext4_request_inode                  [Tracepoint event]
  ext4:ext4_allocate_inode                 [Tracepoint event]
  ext4:ext4_write_begin                    [Tracepoint event]
  ext4:ext4_ordered_write_end              [Tracepoint event]
  [ .... remaining output snipped .... ]


3. Enabling Events
==================

3.1 System-Wide Event Enabling
------------------------------

See Documentation/trace/events.txt for a proper description on how events
can be enabled system-wide. A short example of enabling all events related
to page allocation would look something like:

  $ for i in `find /sys/kernel/debug/tracing/events -name "enable" | grep mm_`; do echo 1 > $i; done

3.2 System-Wide Event Enabling with SystemTap
---------------------------------------------

In SystemTap, tracepoints are accessible using the kernel.trace() function
call. The following is an example that reports every 5 seconds what processes
were allocating the pages.

  global page_allocs

  probe kernel.trace("mm_page_alloc") {
        page_allocs[execname()]++
  }

  function print_count() {
        printf ("%-25s %-s\n", "#Pages Allocated", "Process Name")
        foreach (proc in page_allocs-)
                printf("%-25d %s\n", page_allocs[proc], proc)
        printf ("\n")
        delete page_allocs
  }

  probe timer.s(5) {
          print_count()
  }

3.3 System-Wide Event Enabling with PCL
---------------------------------------

By specifying the -a switch and analysing sleep, the system-wide events
for a duration of time can be examined.

 $ perf stat -a \
        -e kmem:mm_page_alloc -e kmem:mm_page_free_direct \
        -e kmem:mm_pagevec_free \
        sleep 10
 Performance counter stats for 'sleep 10':

           9630  kmem:mm_page_alloc
           2143  kmem:mm_page_free_direct
           7424  kmem:mm_pagevec_free

   10.002577764  seconds time elapsed

Similarly, one could execute a shell and exit it as desired to get a report
at that point.

3.4 Local Event Enabling
------------------------

Documentation/trace/ftrace.txt describes how to enable events on a per-thread
basis using set_ftrace_pid.

3.5 Local Event Enablement with PCL
-----------------------------------

Events can be activated and tracked for the duration of a process on a local
basis using PCL such as follows.

  $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free_direct \
                 -e kmem:mm_pagevec_free ./hackbench 10
  Time: 0.909

    Performance counter stats for './hackbench 10':

          17803  kmem:mm_page_alloc
          12398  kmem:mm_page_free_direct
           4827  kmem:mm_pagevec_free

    0.973913387  seconds time elapsed

4. Event Filtering
==================

Documentation/trace/ftrace.txt covers in-depth how to filter events in
ftrace.  Obviously using grep and awk of trace_pipe is an option as well
as any script reading trace_pipe.

5. Analysing Event Variances with PCL
=====================================

Any workload can exhibit variances between runs and it can be important
to know what the standard deviation is. By and large, this is left to the
performance analyst to do it by hand. In the event that the discrete event
occurrences are useful to the performance analyst, then perf can be used.

  $ perf stat --repeat 5 -e kmem:mm_page_alloc -e kmem:mm_page_free_direct
                        -e kmem:mm_pagevec_free ./hackbench 10
  Time: 0.890
  Time: 0.895
  Time: 0.915
  Time: 1.001
  Time: 0.899

   Performance counter stats for './hackbench 10' (5 runs):

          16630  kmem:mm_page_alloc         ( +-   3.542% )
          11486  kmem:mm_page_free_direct   ( +-   4.771% )
           4730  kmem:mm_pagevec_free       ( +-   2.325% )

    0.982653002  seconds time elapsed   ( +-   1.448% )

In the event that some higher-level event is required that depends on some
aggregation of discrete events, then a script would need to be developed.

Using --repeat, it is also possible to view how events are fluctuating over
time on a system-wide basis using -a and sleep.

  $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free_direct \
                -e kmem:mm_pagevec_free \
                -a --repeat 10 \
                sleep 1
  Performance counter stats for 'sleep 1' (10 runs):

           1066  kmem:mm_page_alloc         ( +-  26.148% )
            182  kmem:mm_page_free_direct   ( +-   5.464% )
            890  kmem:mm_pagevec_free       ( +-  30.079% )

    1.002251757  seconds time elapsed   ( +-   0.005% )

6. Higher-Level Analysis with Helper Scripts
============================================

When events are enabled the events that are triggering can be read from
/sys/kernel/debug/tracing/trace_pipe in human-readable format although binary
options exist as well. By post-processing the output, further information can
be gathered on-line as appropriate. Examples of post-processing might include

  o Reading information from /proc for the PID that triggered the event
  o Deriving a higher-level event from a series of lower-level events.
  o Calculating latencies between two events

Documentation/trace/postprocess/trace-pagealloc-postprocess.pl is an example
script that can read trace_pipe from STDIN or a copy of a trace. When used
on-line, it can be interrupted once to generate a report without exiting
and twice to exit.

Simplistically, the script just reads STDIN and counts up events but it
also can do more such as

  o Derive high-level events from many low-level events. If a number of pages
    are freed to the main allocator from the per-CPU lists, it recognises
    that as one per-CPU drain even though there is no specific tracepoint
    for that event
  o It can aggregate based on PID or individual process number
  o In the event memory is getting externally fragmented, it reports
    on whether the fragmentation event was severe or moderate.
  o When receiving an event about a PID, it can record who the parent was so
    that if large numbers of events are coming from very short-lived
    processes, the parent process responsible for creating all the helpers
    can be identified

7. Lower-Level Analysis with PCL
================================

There may also be a requirement to identify what functions within a program
were generating events within the kernel. To begin this sort of analysis, the
data must be recorded. At the time of writing, this required root:

  $ perf record -c 1 \
        -e kmem:mm_page_alloc -e kmem:mm_page_free_direct \
        -e kmem:mm_pagevec_free \
        ./hackbench 10
  Time: 0.894
  [ perf record: Captured and wrote 0.733 MB perf.data (~32010 samples) ]

Note the use of '-c 1' to set the event period to sample. The default sample
period is quite high to minimise overhead but the information collected can be
very coarse as a result.

This record outputted a file called perf.data which can be analysed using
perf report.

  $ perf report
  # Samples: 30922
  #
  # Overhead    Command                     Shared Object
  # ........  .........  ................................
  #
      87.27%  hackbench  [vdso]
       6.85%  hackbench  /lib/i686/cmov/libc-2.9.so
       2.62%  hackbench  /lib/ld-2.9.so
       1.52%       perf  [vdso]
       1.22%  hackbench  ./hackbench
       0.48%  hackbench  [kernel]
       0.02%       perf  /lib/i686/cmov/libc-2.9.so
       0.01%       perf  /usr/bin/perf
       0.01%       perf  /lib/ld-2.9.so
       0.00%  hackbench  /lib/i686/cmov/libpthread-2.9.so
  #
  # (For more details, try: perf report --sort comm,dso,symbol)
  #

According to this, the vast majority of events triggered on events
within the VDSO. With simple binaries, this will often be the case so let's
take a slightly different example. In the course of writing this, it was
noticed that X was generating an insane amount of page allocations so let's look
at it:

  $ perf record -c 1 -f \
                -e kmem:mm_page_alloc -e kmem:mm_page_free_direct \
                -e kmem:mm_pagevec_free \
                -p `pidof X`

This was interrupted after a few seconds and

  $ perf report
  # Samples: 27666
  #
  # Overhead  Command                            Shared Object
  # ........  .......  .......................................
  #
      51.95%     Xorg  [vdso]
      47.95%     Xorg  /opt/gfx-test/lib/libpixman-1.so.0.13.1
       0.09%     Xorg  /lib/i686/cmov/libc-2.9.so
       0.01%     Xorg  [kernel]
  #
  # (For more details, try: perf report --sort comm,dso,symbol)
  #

So, almost half of the events are occurring in a library. To get an idea which
symbol:

  $ perf report --sort comm,dso,symbol
  # Samples: 27666
  #
  # Overhead  Command                            Shared Object  Symbol
  # ........  .......  .......................................  ......
  #
      51.95%     Xorg  [vdso]                                   [.] 0x000000ffffe424
      47.93%     Xorg  /opt/gfx-test/lib/libpixman-1.so.0.13.1  [.] pixmanFillsse2
       0.09%     Xorg  /lib/i686/cmov/libc-2.9.so               [.] _int_malloc
       0.01%     Xorg  /opt/gfx-test/lib/libpixman-1.so.0.13.1  [.] pixman_region32_copy_f
       0.01%     Xorg  [kernel]                                 [k] read_hpet
       0.01%     Xorg  /opt/gfx-test/lib/libpixman-1.so.0.13.1  [.] get_fast_path
       0.00%     Xorg  [kernel]                                 [k] ftrace_trace_userstack

To see where within the function pixmanFillsse2 things are going wrong:

  $ perf annotate pixmanFillsse2
  [ ... ]
    0.00 :         34eeb:       0f 18 08                prefetcht0 (%eax)
         :      }
         :
         :      extern __inline void __attribute__((__gnu_inline__, __always_inline__, _
         :      _mm_store_si128 (__m128i *__P, __m128i __B) :      {
         :        *__P = __B;
   12.40 :         34eee:       66 0f 7f 80 40 ff ff    movdqa %xmm0,-0xc0(%eax)
    0.00 :         34ef5:       ff
   12.40 :         34ef6:       66 0f 7f 80 50 ff ff    movdqa %xmm0,-0xb0(%eax)
    0.00 :         34efd:       ff
   12.39 :         34efe:       66 0f 7f 80 60 ff ff    movdqa %xmm0,-0xa0(%eax)
    0.00 :         34f05:       ff
   12.67 :         34f06:       66 0f 7f 80 70 ff ff    movdqa %xmm0,-0x90(%eax)
    0.00 :         34f0d:       ff
   12.58 :         34f0e:       66 0f 7f 40 80          movdqa %xmm0,-0x80(%eax)
   12.31 :         34f13:       66 0f 7f 40 90          movdqa %xmm0,-0x70(%eax)
   12.40 :         34f18:       66 0f 7f 40 a0          movdqa %xmm0,-0x60(%eax)
   12.31 :         34f1d:       66 0f 7f 40 b0          movdqa %xmm0,-0x50(%eax)

At a glance, it looks like the time is being spent copying pixmaps to
the card.  Further investigation would be needed to determine why pixmaps
are being copied around so much but a starting point would be to take an
ancient build of libpixmap out of the library path where it was totally
forgotten about from months ago!