1========================================================= 2Notes on Analysing Behaviour Using Events and Tracepoints 3========================================================= 4:Author: Mel Gorman (PCL information heavily based on email from Ingo Molnar) 5 61. Introduction 7=============== 8 9Tracepoints (see Documentation/trace/tracepoints.rst) 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/tracing/events. Simply 30calling:: 31 32 $ find /sys/kernel/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.rst 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/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 72 global page_allocs 73 74 probe kernel.trace("mm_page_alloc") { 75 page_allocs[execname()]++ 76 } 77 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) 82 printf ("\n") 83 delete page_allocs 84 } 85 86 probe timer.s(5) { 87 print_count() 88 } 89 903.3 System-Wide Event Enabling with PCL 91--------------------------------------- 92 93By specifying the -a switch and analysing sleep, the system-wide events 94for a duration of time can be examined. 95:: 96 97 $ perf stat -a \ 98 -e kmem:mm_page_alloc -e kmem:mm_page_free \ 99 -e kmem:mm_page_free_batched \ 100 sleep 10 101 Performance counter stats for 'sleep 10': 102 103 9630 kmem:mm_page_alloc 104 2143 kmem:mm_page_free 105 7424 kmem:mm_page_free_batched 106 107 10.002577764 seconds time elapsed 108 109Similarly, one could execute a shell and exit it as desired to get a report 110at that point. 111 1123.4 Local Event Enabling 113------------------------ 114 115Documentation/trace/ftrace.rst describes how to enable events on a per-thread 116basis using set_ftrace_pid. 117 1183.5 Local Event Enablement with PCL 119----------------------------------- 120 121Events can be activated and tracked for the duration of a process on a local 122basis using PCL such as follows. 123:: 124 125 $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free \ 126 -e kmem:mm_page_free_batched ./hackbench 10 127 Time: 0.909 128 129 Performance counter stats for './hackbench 10': 130 131 17803 kmem:mm_page_alloc 132 12398 kmem:mm_page_free 133 4827 kmem:mm_page_free_batched 134 135 0.973913387 seconds time elapsed 136 1374. Event Filtering 138================== 139 140Documentation/trace/ftrace.rst covers in-depth how to filter events in 141ftrace. Obviously using grep and awk of trace_pipe is an option as well 142as any script reading trace_pipe. 143 1445. Analysing Event Variances with PCL 145===================================== 146 147Any workload can exhibit variances between runs and it can be important 148to know what the standard deviation is. By and large, this is left to the 149performance analyst to do it by hand. In the event that the discrete event 150occurrences are useful to the performance analyst, then perf can be used. 151:: 152 153 $ perf stat --repeat 5 -e kmem:mm_page_alloc -e kmem:mm_page_free 154 -e kmem:mm_page_free_batched ./hackbench 10 155 Time: 0.890 156 Time: 0.895 157 Time: 0.915 158 Time: 1.001 159 Time: 0.899 160 161 Performance counter stats for './hackbench 10' (5 runs): 162 163 16630 kmem:mm_page_alloc ( +- 3.542% ) 164 11486 kmem:mm_page_free ( +- 4.771% ) 165 4730 kmem:mm_page_free_batched ( +- 2.325% ) 166 167 0.982653002 seconds time elapsed ( +- 1.448% ) 168 169In the event that some higher-level event is required that depends on some 170aggregation of discrete events, then a script would need to be developed. 171 172Using --repeat, it is also possible to view how events are fluctuating over 173time on a system-wide basis using -a and sleep. 174:: 175 176 $ perf stat -e kmem:mm_page_alloc -e kmem:mm_page_free \ 177 -e kmem:mm_page_free_batched \ 178 -a --repeat 10 \ 179 sleep 1 180 Performance counter stats for 'sleep 1' (10 runs): 181 182 1066 kmem:mm_page_alloc ( +- 26.148% ) 183 182 kmem:mm_page_free ( +- 5.464% ) 184 890 kmem:mm_page_free_batched ( +- 30.079% ) 185 186 1.002251757 seconds time elapsed ( +- 0.005% ) 187 1886. Higher-Level Analysis with Helper Scripts 189============================================ 190 191When events are enabled the events that are triggering can be read from 192/sys/kernel/tracing/trace_pipe in human-readable format although binary 193options exist as well. By post-processing the output, further information can 194be gathered on-line as appropriate. Examples of post-processing might include 195 196 - 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 199 200Documentation/trace/postprocess/trace-pagealloc-postprocess.pl is an example 201script that can read trace_pipe from STDIN or a copy of a trace. When used 202on-line, it can be interrupted once to generate a report without exiting 203and twice to exit. 204 205Simplistically, the script just reads STDIN and counts up events but it 206also can do more such as 207 208 - 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 211 for that event 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 218 can be identified 219 2207. Lower-Level Analysis with PCL 221================================ 222 223There may also be a requirement to identify what functions within a program 224were generating events within the kernel. To begin this sort of analysis, the 225data must be recorded. At the time of writing, this required root: 226:: 227 228 $ perf record -c 1 \ 229 -e kmem:mm_page_alloc -e kmem:mm_page_free \ 230 -e kmem:mm_page_free_batched \ 231 ./hackbench 10 232 Time: 0.894 233 [ perf record: Captured and wrote 0.733 MB perf.data (~32010 samples) ] 234 235Note the use of '-c 1' to set the event period to sample. The default sample 236period is quite high to minimise overhead but the information collected can be 237very coarse as a result. 238 239This record outputted a file called perf.data which can be analysed using 240perf report. 241:: 242 243 $ perf report 244 # Samples: 30922 245 # 246 # Overhead Command Shared Object 247 # ........ ......... ................................ 248 # 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 252 1.52% perf [vdso] 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 259 # 260 # (For more details, try: perf report --sort comm,dso,symbol) 261 # 262 263According to this, the vast majority of events triggered on events 264within the VDSO. With simple binaries, this will often be the case so let's 265take a slightly different example. In the course of writing this, it was 266noticed that X was generating an insane amount of page allocations so let's look 267at it: 268:: 269 270 $ perf record -c 1 -f \ 271 -e kmem:mm_page_alloc -e kmem:mm_page_free \ 272 -e kmem:mm_page_free_batched \ 273 -p `pidof X` 274 275This was interrupted after a few seconds and 276:: 277 278 $ perf report 279 # Samples: 27666 280 # 281 # Overhead Command Shared Object 282 # ........ ....... ....................................... 283 # 284 51.95% Xorg [vdso] 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 287 0.01% Xorg [kernel] 288 # 289 # (For more details, try: perf report --sort comm,dso,symbol) 290 # 291 292So, almost half of the events are occurring in a library. To get an idea which 293symbol: 294:: 295 296 $ perf report --sort comm,dso,symbol 297 # Samples: 27666 298 # 299 # Overhead Command Shared Object Symbol 300 # ........ ....... ....................................... ...... 301 # 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 309 310To see where within the function pixmanFillsse2 things are going wrong: 311:: 312 313 $ perf annotate pixmanFillsse2 314 [ ... ] 315 0.00 : 34eeb: 0f 18 08 prefetcht0 (%eax) 316 : } 317 : 318 : extern __inline void __attribute__((__gnu_inline__, __always_inline__, _ 319 : _mm_store_si128 (__m128i *__P, __m128i __B) : { 320 : *__P = __B; 321 12.40 : 34eee: 66 0f 7f 80 40 ff ff movdqa %xmm0,-0xc0(%eax) 322 0.00 : 34ef5: ff 323 12.40 : 34ef6: 66 0f 7f 80 50 ff ff movdqa %xmm0,-0xb0(%eax) 324 0.00 : 34efd: ff 325 12.39 : 34efe: 66 0f 7f 80 60 ff ff movdqa %xmm0,-0xa0(%eax) 326 0.00 : 34f05: ff 327 12.67 : 34f06: 66 0f 7f 80 70 ff ff movdqa %xmm0,-0x90(%eax) 328 0.00 : 34f0d: ff 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) 333 334At a glance, it looks like the time is being spent copying pixmaps to 335the card. Further investigation would be needed to determine why pixmaps 336are being copied around so much but a starting point would be to take an 337ancient build of libpixmap out of the library path where it was totally 338forgotten about from months ago! 339