1 =========================================================
2 Notes on Analysing Behaviour Using Events and Tracepoints
3 =========================================================
4 :Author: Mel Gorman (PCL information heavily based on email from Ingo Molnar)
9 Tracepoints (see Documentation/trace/tracepoints.rst) can be used without
10 creating custom kernel modules to register probe functions using the event
11 tracing infrastructure.
13 Simplistically, tracepoints represent important events that can be
14 taken in conjunction with other tracepoints to build a "Big Picture" of
15 what is going on within the system. There are a large number of methods for
16 gathering and interpreting these events. Lacking any current Best Practises,
17 this document describes some of the methods that can be used.
19 This document assumes that debugfs is mounted on /sys/kernel/debug and that
20 the appropriate tracing options have been configured into the kernel. It is
21 assumed that the PCL tool tools/perf has been installed and is in your path.
23 2. Listing Available Events
24 ===========================
26 2.1 Standard Utilities
27 ----------------------
29 All possible events are visible from /sys/kernel/tracing/events. Simply
32 $ find /sys/kernel/tracing/events -type d
34 will give a fair indication of the number of events available.
36 2.2 PCL (Performance Counters for Linux)
37 ----------------------------------------
39 Discovery and enumeration of all counters and events, including tracepoints,
40 are available with the perf tool. Getting a list of available events is a
43 $ perf list 2>&1 | grep Tracepoint
44 ext4:ext4_free_inode [Tracepoint event]
45 ext4:ext4_request_inode [Tracepoint event]
46 ext4:ext4_allocate_inode [Tracepoint event]
47 ext4:ext4_write_begin [Tracepoint event]
48 ext4:ext4_ordered_write_end [Tracepoint event]
49 [ .... remaining output snipped .... ]
55 3.1 System-Wide Event Enabling
56 ------------------------------
58 See Documentation/trace/events.rst for a proper description on how events
59 can be enabled system-wide. A short example of enabling all events related
60 to page allocation would look something like::
62 $ for i in `find /sys/kernel/tracing/events -name "enable" | grep mm_`; do echo 1 > $i; done
64 3.2 System-Wide Event Enabling with SystemTap
65 ---------------------------------------------
67 In SystemTap, tracepoints are accessible using the kernel.trace() function
68 call. The following is an example that reports every 5 seconds what processes
69 were allocating the pages.
74 probe kernel.trace("mm_page_alloc") {
75 page_allocs[execname()]++
78 function print_count() {
79 printf ("%-25s %-s\n", "#Pages Allocated", "Process Name")
80 foreach (proc in page_allocs-)
81 printf("%-25d %s\n", page_allocs[proc], proc)
90 3.3 System-Wide Event Enabling with PCL
91 ---------------------------------------
93 By specifying the -a switch and analysing sleep, the system-wide events
94 for a duration of time can be examined.
98 -e kmem:mm_page_alloc -e kmem:mm_page_free \
99 -e kmem:mm_page_free_batched \
101 Performance counter stats for 'sleep 10':
103 9630 kmem:mm_page_alloc
104 2143 kmem:mm_page_free
105 7424 kmem:mm_page_free_batched
107 10.002577764 seconds time elapsed
109 Similarly, one could execute a shell and exit it as desired to get a report
112 3.4 Local Event Enabling
113 ------------------------
115 Documentation/trace/ftrace.rst describes how to enable events on a per-thread
116 basis using set_ftrace_pid.
118 3.5 Local Event Enablement with PCL
119 -----------------------------------
121 Events can be activated and tracked for the duration of a process on a local
122 basis using PCL such as follows.
125 $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free \
126 -e kmem:mm_page_free_batched ./hackbench 10
129 Performance counter stats for './hackbench 10':
131 17803 kmem:mm_page_alloc
132 12398 kmem:mm_page_free
133 4827 kmem:mm_page_free_batched
135 0.973913387 seconds time elapsed
140 Documentation/trace/ftrace.rst covers in-depth how to filter events in
141 ftrace. Obviously using grep and awk of trace_pipe is an option as well
142 as any script reading trace_pipe.
144 5. Analysing Event Variances with PCL
145 =====================================
147 Any workload can exhibit variances between runs and it can be important
148 to know what the standard deviation is. By and large, this is left to the
149 performance analyst to do it by hand. In the event that the discrete event
150 occurrences are useful to the performance analyst, then perf can be used.
153 $ perf stat --repeat 5 -e kmem:mm_page_alloc -e kmem:mm_page_free
154 -e kmem:mm_page_free_batched ./hackbench 10
161 Performance counter stats for './hackbench 10' (5 runs):
163 16630 kmem:mm_page_alloc ( +- 3.542% )
164 11486 kmem:mm_page_free ( +- 4.771% )
165 4730 kmem:mm_page_free_batched ( +- 2.325% )
167 0.982653002 seconds time elapsed ( +- 1.448% )
169 In the event that some higher-level event is required that depends on some
170 aggregation of discrete events, then a script would need to be developed.
172 Using --repeat, it is also possible to view how events are fluctuating over
173 time on a system-wide basis using -a and sleep.
176 $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free \
177 -e kmem:mm_page_free_batched \
180 Performance counter stats for 'sleep 1' (10 runs):
182 1066 kmem:mm_page_alloc ( +- 26.148% )
183 182 kmem:mm_page_free ( +- 5.464% )
184 890 kmem:mm_page_free_batched ( +- 30.079% )
186 1.002251757 seconds time elapsed ( +- 0.005% )
188 6. Higher-Level Analysis with Helper Scripts
189 ============================================
191 When events are enabled the events that are triggering can be read from
192 /sys/kernel/tracing/trace_pipe in human-readable format although binary
193 options exist as well. By post-processing the output, further information can
194 be gathered on-line as appropriate. Examples of post-processing might include
196 - Reading information from /proc for the PID that triggered the event
197 - Deriving a higher-level event from a series of lower-level events.
198 - Calculating latencies between two events
200 Documentation/trace/postprocess/trace-pagealloc-postprocess.pl is an example
201 script that can read trace_pipe from STDIN or a copy of a trace. When used
202 on-line, it can be interrupted once to generate a report without exiting
205 Simplistically, the script just reads STDIN and counts up events but it
206 also can do more such as
208 - Derive high-level events from many low-level events. If a number of pages
209 are freed to the main allocator from the per-CPU lists, it recognises
210 that as one per-CPU drain even though there is no specific tracepoint
212 - It can aggregate based on PID or individual process number
213 - In the event memory is getting externally fragmented, it reports
214 on whether the fragmentation event was severe or moderate.
215 - When receiving an event about a PID, it can record who the parent was so
216 that if large numbers of events are coming from very short-lived
217 processes, the parent process responsible for creating all the helpers
220 7. Lower-Level Analysis with PCL
221 ================================
223 There may also be a requirement to identify what functions within a program
224 were generating events within the kernel. To begin this sort of analysis, the
225 data must be recorded. At the time of writing, this required root:
229 -e kmem:mm_page_alloc -e kmem:mm_page_free \
230 -e kmem:mm_page_free_batched \
233 [ perf record: Captured and wrote 0.733 MB perf.data (~32010 samples) ]
235 Note the use of '-c 1' to set the event period to sample. The default sample
236 period is quite high to minimise overhead but the information collected can be
237 very coarse as a result.
239 This record outputted a file called perf.data which can be analysed using
246 # Overhead Command Shared Object
247 # ........ ......... ................................
249 87.27% hackbench [vdso]
250 6.85% hackbench /lib/i686/cmov/libc-2.9.so
251 2.62% hackbench /lib/ld-2.9.so
253 1.22% hackbench ./hackbench
254 0.48% hackbench [kernel]
255 0.02% perf /lib/i686/cmov/libc-2.9.so
256 0.01% perf /usr/bin/perf
257 0.01% perf /lib/ld-2.9.so
258 0.00% hackbench /lib/i686/cmov/libpthread-2.9.so
260 # (For more details, try: perf report --sort comm,dso,symbol)
263 According to this, the vast majority of events triggered on events
264 within the VDSO. With simple binaries, this will often be the case so let's
265 take a slightly different example. In the course of writing this, it was
266 noticed that X was generating an insane amount of page allocations so let's look
270 $ perf record -c 1 -f \
271 -e kmem:mm_page_alloc -e kmem:mm_page_free \
272 -e kmem:mm_page_free_batched \
275 This was interrupted after a few seconds and
281 # Overhead Command Shared Object
282 # ........ ....... .......................................
285 47.95% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1
286 0.09% Xorg /lib/i686/cmov/libc-2.9.so
289 # (For more details, try: perf report --sort comm,dso,symbol)
292 So, almost half of the events are occurring in a library. To get an idea which
296 $ perf report --sort comm,dso,symbol
299 # Overhead Command Shared Object Symbol
300 # ........ ....... ....................................... ......
302 51.95% Xorg [vdso] [.] 0x000000ffffe424
303 47.93% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 [.] pixmanFillsse2
304 0.09% Xorg /lib/i686/cmov/libc-2.9.so [.] _int_malloc
305 0.01% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 [.] pixman_region32_copy_f
306 0.01% Xorg [kernel] [k] read_hpet
307 0.01% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 [.] get_fast_path
308 0.00% Xorg [kernel] [k] ftrace_trace_userstack
310 To see where within the function pixmanFillsse2 things are going wrong:
313 $ perf annotate pixmanFillsse2
315 0.00 : 34eeb: 0f 18 08 prefetcht0 (%eax)
318 : extern __inline void __attribute__((__gnu_inline__, __always_inline__, _
319 : _mm_store_si128 (__m128i *__P, __m128i __B) : {
321 12.40 : 34eee: 66 0f 7f 80 40 ff ff movdqa %xmm0,-0xc0(%eax)
323 12.40 : 34ef6: 66 0f 7f 80 50 ff ff movdqa %xmm0,-0xb0(%eax)
325 12.39 : 34efe: 66 0f 7f 80 60 ff ff movdqa %xmm0,-0xa0(%eax)
327 12.67 : 34f06: 66 0f 7f 80 70 ff ff movdqa %xmm0,-0x90(%eax)
329 12.58 : 34f0e: 66 0f 7f 40 80 movdqa %xmm0,-0x80(%eax)
330 12.31 : 34f13: 66 0f 7f 40 90 movdqa %xmm0,-0x70(%eax)
331 12.40 : 34f18: 66 0f 7f 40 a0 movdqa %xmm0,-0x60(%eax)
332 12.31 : 34f1d: 66 0f 7f 40 b0 movdqa %xmm0,-0x50(%eax)
334 At a glance, it looks like the time is being spent copying pixmaps to
335 the card. Further investigation would be needed to determine why pixmaps
336 are being copied around so much but a starting point would be to take an
337 ancient build of libpixmap out of the library path where it was totally
338 forgotten about from months ago!