1 Notes on Analysing Behaviour Using Events and Tracepoints 2 3 Documentation written by Mel Gorman 4 PCL information heavily based on email from Ingo Molnar 5 61. Introduction 7=============== 8 9Tracepoints (see Documentation/trace/tracepoints.txt) can be used without 10creating custom kernel modules to register probe functions using the event 11tracing infrastructure. 12 13Simplistically, tracepoints represent important events that can be 14taken in conjunction with other tracepoints to build a "Big Picture" of 15what is going on within the system. There are a large number of methods for 16gathering and interpreting these events. Lacking any current Best Practises, 17this document describes some of the methods that can be used. 18 19This document assumes that debugfs is mounted on /sys/kernel/debug and that 20the appropriate tracing options have been configured into the kernel. It is 21assumed that the PCL tool tools/perf has been installed and is in your path. 22 232. Listing Available Events 24=========================== 25 262.1 Standard Utilities 27---------------------- 28 29All possible events are visible from /sys/kernel/debug/tracing/events. Simply 30calling 31 32 $ find /sys/kernel/debug/tracing/events -type d 33 34will give a fair indication of the number of events available. 35 362.2 PCL (Performance Counters for Linux) 37------- 38 39Discovery and enumeration of all counters and events, including tracepoints, 40are available with the perf tool. Getting a list of available events is a 41simple case of: 42 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 .... ] 50 51 523. Enabling Events 53================== 54 553.1 System-Wide Event Enabling 56------------------------------ 57 58See Documentation/trace/events.txt for a proper description on how events 59can be enabled system-wide. A short example of enabling all events related 60to page allocation would look something like: 61 62 $ for i in `find /sys/kernel/debug/tracing/events -name "enable" | grep mm_`; do echo 1 > $i; done 63 643.2 System-Wide Event Enabling with SystemTap 65--------------------------------------------- 66 67In SystemTap, tracepoints are accessible using the kernel.trace() function 68call. The following is an example that reports every 5 seconds what processes 69were allocating the pages. 70 71 global page_allocs 72 73 probe kernel.trace("mm_page_alloc") { 74 page_allocs[execname()]++ 75 } 76 77 function print_count() { 78 printf ("%-25s %-s\n", "#Pages Allocated", "Process Name") 79 foreach (proc in page_allocs-) 80 printf("%-25d %s\n", page_allocs[proc], proc) 81 printf ("\n") 82 delete page_allocs 83 } 84 85 probe timer.s(5) { 86 print_count() 87 } 88 893.3 System-Wide Event Enabling with PCL 90--------------------------------------- 91 92By specifying the -a switch and analysing sleep, the system-wide events 93for a duration of time can be examined. 94 95 $ perf stat -a \ 96 -e kmem:mm_page_alloc -e kmem:mm_page_free \ 97 -e kmem:mm_page_free_batched \ 98 sleep 10 99 Performance counter stats for 'sleep 10': 100 101 9630 kmem:mm_page_alloc 102 2143 kmem:mm_page_free 103 7424 kmem:mm_page_free_batched 104 105 10.002577764 seconds time elapsed 106 107Similarly, one could execute a shell and exit it as desired to get a report 108at that point. 109 1103.4 Local Event Enabling 111------------------------ 112 113Documentation/trace/ftrace.txt describes how to enable events on a per-thread 114basis using set_ftrace_pid. 115 1163.5 Local Event Enablement with PCL 117----------------------------------- 118 119Events can be activated and tracked for the duration of a process on a local 120basis using PCL such as follows. 121 122 $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free \ 123 -e kmem:mm_page_free_batched ./hackbench 10 124 Time: 0.909 125 126 Performance counter stats for './hackbench 10': 127 128 17803 kmem:mm_page_alloc 129 12398 kmem:mm_page_free 130 4827 kmem:mm_page_free_batched 131 132 0.973913387 seconds time elapsed 133 1344. Event Filtering 135================== 136 137Documentation/trace/ftrace.txt covers in-depth how to filter events in 138ftrace. Obviously using grep and awk of trace_pipe is an option as well 139as any script reading trace_pipe. 140 1415. Analysing Event Variances with PCL 142===================================== 143 144Any workload can exhibit variances between runs and it can be important 145to know what the standard deviation is. By and large, this is left to the 146performance analyst to do it by hand. In the event that the discrete event 147occurrences are useful to the performance analyst, then perf can be used. 148 149 $ perf stat --repeat 5 -e kmem:mm_page_alloc -e kmem:mm_page_free 150 -e kmem:mm_page_free_batched ./hackbench 10 151 Time: 0.890 152 Time: 0.895 153 Time: 0.915 154 Time: 1.001 155 Time: 0.899 156 157 Performance counter stats for './hackbench 10' (5 runs): 158 159 16630 kmem:mm_page_alloc ( +- 3.542% ) 160 11486 kmem:mm_page_free ( +- 4.771% ) 161 4730 kmem:mm_page_free_batched ( +- 2.325% ) 162 163 0.982653002 seconds time elapsed ( +- 1.448% ) 164 165In the event that some higher-level event is required that depends on some 166aggregation of discrete events, then a script would need to be developed. 167 168Using --repeat, it is also possible to view how events are fluctuating over 169time on a system-wide basis using -a and sleep. 170 171 $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free \ 172 -e kmem:mm_page_free_batched \ 173 -a --repeat 10 \ 174 sleep 1 175 Performance counter stats for 'sleep 1' (10 runs): 176 177 1066 kmem:mm_page_alloc ( +- 26.148% ) 178 182 kmem:mm_page_free ( +- 5.464% ) 179 890 kmem:mm_page_free_batched ( +- 30.079% ) 180 181 1.002251757 seconds time elapsed ( +- 0.005% ) 182 1836. Higher-Level Analysis with Helper Scripts 184============================================ 185 186When events are enabled the events that are triggering can be read from 187/sys/kernel/debug/tracing/trace_pipe in human-readable format although binary 188options exist as well. By post-processing the output, further information can 189be gathered on-line as appropriate. Examples of post-processing might include 190 191 o Reading information from /proc for the PID that triggered the event 192 o Deriving a higher-level event from a series of lower-level events. 193 o Calculating latencies between two events 194 195Documentation/trace/postprocess/trace-pagealloc-postprocess.pl is an example 196script that can read trace_pipe from STDIN or a copy of a trace. When used 197on-line, it can be interrupted once to generate a report without exiting 198and twice to exit. 199 200Simplistically, the script just reads STDIN and counts up events but it 201also can do more such as 202 203 o Derive high-level events from many low-level events. If a number of pages 204 are freed to the main allocator from the per-CPU lists, it recognises 205 that as one per-CPU drain even though there is no specific tracepoint 206 for that event 207 o It can aggregate based on PID or individual process number 208 o In the event memory is getting externally fragmented, it reports 209 on whether the fragmentation event was severe or moderate. 210 o When receiving an event about a PID, it can record who the parent was so 211 that if large numbers of events are coming from very short-lived 212 processes, the parent process responsible for creating all the helpers 213 can be identified 214 2157. Lower-Level Analysis with PCL 216================================ 217 218There may also be a requirement to identify what functions within a program 219were generating events within the kernel. To begin this sort of analysis, the 220data must be recorded. At the time of writing, this required root: 221 222 $ perf record -c 1 \ 223 -e kmem:mm_page_alloc -e kmem:mm_page_free \ 224 -e kmem:mm_page_free_batched \ 225 ./hackbench 10 226 Time: 0.894 227 [ perf record: Captured and wrote 0.733 MB perf.data (~32010 samples) ] 228 229Note the use of '-c 1' to set the event period to sample. The default sample 230period is quite high to minimise overhead but the information collected can be 231very coarse as a result. 232 233This record outputted a file called perf.data which can be analysed using 234perf report. 235 236 $ perf report 237 # Samples: 30922 238 # 239 # Overhead Command Shared Object 240 # ........ ......... ................................ 241 # 242 87.27% hackbench [vdso] 243 6.85% hackbench /lib/i686/cmov/libc-2.9.so 244 2.62% hackbench /lib/ld-2.9.so 245 1.52% perf [vdso] 246 1.22% hackbench ./hackbench 247 0.48% hackbench [kernel] 248 0.02% perf /lib/i686/cmov/libc-2.9.so 249 0.01% perf /usr/bin/perf 250 0.01% perf /lib/ld-2.9.so 251 0.00% hackbench /lib/i686/cmov/libpthread-2.9.so 252 # 253 # (For more details, try: perf report --sort comm,dso,symbol) 254 # 255 256According to this, the vast majority of events triggered on events 257within the VDSO. With simple binaries, this will often be the case so let's 258take a slightly different example. In the course of writing this, it was 259noticed that X was generating an insane amount of page allocations so let's look 260at it: 261 262 $ perf record -c 1 -f \ 263 -e kmem:mm_page_alloc -e kmem:mm_page_free \ 264 -e kmem:mm_page_free_batched \ 265 -p `pidof X` 266 267This was interrupted after a few seconds and 268 269 $ perf report 270 # Samples: 27666 271 # 272 # Overhead Command Shared Object 273 # ........ ....... ....................................... 274 # 275 51.95% Xorg [vdso] 276 47.95% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 277 0.09% Xorg /lib/i686/cmov/libc-2.9.so 278 0.01% Xorg [kernel] 279 # 280 # (For more details, try: perf report --sort comm,dso,symbol) 281 # 282 283So, almost half of the events are occurring in a library. To get an idea which 284symbol: 285 286 $ perf report --sort comm,dso,symbol 287 # Samples: 27666 288 # 289 # Overhead Command Shared Object Symbol 290 # ........ ....... ....................................... ...... 291 # 292 51.95% Xorg [vdso] [.] 0x000000ffffe424 293 47.93% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 [.] pixmanFillsse2 294 0.09% Xorg /lib/i686/cmov/libc-2.9.so [.] _int_malloc 295 0.01% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 [.] pixman_region32_copy_f 296 0.01% Xorg [kernel] [k] read_hpet 297 0.01% Xorg /opt/gfx-test/lib/libpixman-1.so.0.13.1 [.] get_fast_path 298 0.00% Xorg [kernel] [k] ftrace_trace_userstack 299 300To see where within the function pixmanFillsse2 things are going wrong: 301 302 $ perf annotate pixmanFillsse2 303 [ ... ] 304 0.00 : 34eeb: 0f 18 08 prefetcht0 (%eax) 305 : } 306 : 307 : extern __inline void __attribute__((__gnu_inline__, __always_inline__, _ 308 : _mm_store_si128 (__m128i *__P, __m128i __B) : { 309 : *__P = __B; 310 12.40 : 34eee: 66 0f 7f 80 40 ff ff movdqa %xmm0,-0xc0(%eax) 311 0.00 : 34ef5: ff 312 12.40 : 34ef6: 66 0f 7f 80 50 ff ff movdqa %xmm0,-0xb0(%eax) 313 0.00 : 34efd: ff 314 12.39 : 34efe: 66 0f 7f 80 60 ff ff movdqa %xmm0,-0xa0(%eax) 315 0.00 : 34f05: ff 316 12.67 : 34f06: 66 0f 7f 80 70 ff ff movdqa %xmm0,-0x90(%eax) 317 0.00 : 34f0d: ff 318 12.58 : 34f0e: 66 0f 7f 40 80 movdqa %xmm0,-0x80(%eax) 319 12.31 : 34f13: 66 0f 7f 40 90 movdqa %xmm0,-0x70(%eax) 320 12.40 : 34f18: 66 0f 7f 40 a0 movdqa %xmm0,-0x60(%eax) 321 12.31 : 34f1d: 66 0f 7f 40 b0 movdqa %xmm0,-0x50(%eax) 322 323At a glance, it looks like the time is being spent copying pixmaps to 324the card. Further investigation would be needed to determine why pixmaps 325are being copied around so much but a starting point would be to take an 326ancient build of libpixmap out of the library path where it was totally 327forgotten about from months ago! 328