summaryrefslogtreecommitdiff
path: root/Documentation/trace/tracepoint-analysis.txt
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation/trace/tracepoint-analysis.txt')
-rw-r--r--Documentation/trace/tracepoint-analysis.txt327
1 files changed, 327 insertions, 0 deletions
diff --git a/Documentation/trace/tracepoint-analysis.txt b/Documentation/trace/tracepoint-analysis.txt
new file mode 100644
index 000000000..058cc6c9d
--- /dev/null
+++ b/Documentation/trace/tracepoint-analysis.txt
@@ -0,0 +1,327 @@
+ 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 \
+ -e kmem:mm_page_free_batched \
+ sleep 10
+ Performance counter stats for 'sleep 10':
+
+ 9630 kmem:mm_page_alloc
+ 2143 kmem:mm_page_free
+ 7424 kmem:mm_page_free_batched
+
+ 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 \
+ -e kmem:mm_page_free_batched ./hackbench 10
+ Time: 0.909
+
+ Performance counter stats for './hackbench 10':
+
+ 17803 kmem:mm_page_alloc
+ 12398 kmem:mm_page_free
+ 4827 kmem:mm_page_free_batched
+
+ 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
+ -e kmem:mm_page_free_batched ./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 ( +- 4.771% )
+ 4730 kmem:mm_page_free_batched ( +- 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 \
+ -e kmem:mm_page_free_batched \
+ -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 ( +- 5.464% )
+ 890 kmem:mm_page_free_batched ( +- 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 \
+ -e kmem:mm_page_free_batched \
+ ./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 \
+ -e kmem:mm_page_free_batched \
+ -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!