summaryrefslogtreecommitdiffstats
path: root/documentation
diff options
context:
space:
mode:
authorScott Rifenbark <scott.m.rifenbark@intel.com>2013-01-16 12:49:45 -0800
committerRichard Purdie <richard.purdie@linuxfoundation.org>2013-01-27 13:56:02 +0000
commit86478ff9486516ed428e470afbe76f3cd857888c (patch)
treeb883e2138166f3a9daf986931779a8f4058169bf /documentation
parent487a76f7c7bfdb47ddec76659794834e84ed04eb (diff)
downloadpoky-86478ff9486516ed428e470afbe76f3cd857888c.tar.gz
profile-manual: Updated indentation and sectioning
Had the sectioning off and it was preventing the manual from making. I also added more indentation as needed throughout most of the manual up to the error. (From yocto-docs rev: 5de3544593a63e96b349babc177970d8605e0098) Signed-off-by: Scott Rifenbark <scott.m.rifenbark@intel.com> Signed-off-by: Richard Purdie <richard.purdie@linuxfoundation.org>
Diffstat (limited to 'documentation')
-rw-r--r--documentation/profile-manual/profile-manual-usage.xml1257
1 files changed, 629 insertions, 628 deletions
diff --git a/documentation/profile-manual/profile-manual-usage.xml b/documentation/profile-manual/profile-manual-usage.xml
index cfe916a795..51c483b700 100644
--- a/documentation/profile-manual/profile-manual-usage.xml
+++ b/documentation/profile-manual/profile-manual-usage.xml
@@ -108,28 +108,28 @@
108 top System profiling tool. 108 top System profiling tool.
109 109
110 See 'perf help COMMAND' for more information on a specific command. 110 See 'perf help COMMAND' for more information on a specific command.
111 </literallayout> 111 </literallayout>
112 </para> 112 </para>
113 113
114 <section id='using-perf-to-do-basic-profiling'> 114 <section id='using-perf-to-do-basic-profiling'>
115 <title>Using perf to do Basic Profiling</title> 115 <title>Using perf to do Basic Profiling</title>
116 116
117 <para> 117 <para>
118 As a simple test case, we'll profile the 'wget' of a fairly large 118 As a simple test case, we'll profile the 'wget' of a fairly large
119 file, which is a minimally interesting case because it has both 119 file, which is a minimally interesting case because it has both
120 file and network I/O aspects, and at least in the case of standard 120 file and network I/O aspects, and at least in the case of standard
121 Yocto images, it's implemented as part of busybox, so the methods 121 Yocto images, it's implemented as part of busybox, so the methods
122 we use to analyze it can be used in a very similar way to the whole 122 we use to analyze it can be used in a very similar way to the whole
123 host of supported busybox applets in Yocto. 123 host of supported busybox applets in Yocto.
124 <literallayout class='monospaced'> 124 <literallayout class='monospaced'>
125 root@crownbay:~# rm linux-2.6.19.2.tar.bz2; \ 125 root@crownbay:~# rm linux-2.6.19.2.tar.bz2; \
126 wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 126 wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
127 </literallayout> 127 </literallayout>
128 The quickest and easiest way to get some basic overall data about 128 The quickest and easiest way to get some basic overall data about
129 what's going on for a particular workload it to profile it using 129 what's going on for a particular workload it to profile it using
130 'perf stat'. 'perf stat' basically profiles using a few default 130 'perf stat'. 'perf stat' basically profiles using a few default
131 counters and displays the summed counts at the end of the run: 131 counters and displays the summed counts at the end of the run:
132 <literallayout class='monospaced'> 132 <literallayout class='monospaced'>
133 root@crownbay:~# perf stat wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 133 root@crownbay:~# perf stat wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
134 Connecting to downloads.yoctoproject.org (140.211.169.59:80) 134 Connecting to downloads.yoctoproject.org (140.211.169.59:80)
135 linux-2.6.19.2.tar.b 100% |***************************************************| 41727k 0:00:00 ETA 135 linux-2.6.19.2.tar.b 100% |***************************************************| 41727k 0:00:00 ETA
@@ -148,19 +148,19 @@
148 19550329 branch-misses # 11.82% of all branches 148 19550329 branch-misses # 11.82% of all branches
149 149
150 59.836627620 seconds time elapsed 150 59.836627620 seconds time elapsed
151 </literallayout> 151 </literallayout>
152 Many times such a simple-minded test doesn't yield much of 152 Many times such a simple-minded test doesn't yield much of
153 interest, but sometimes it does (see Real-world Yocto bug 153 interest, but sometimes it does (see Real-world Yocto bug
154 (slow loop-mounted write speed)). 154 (slow loop-mounted write speed)).
155 </para> 155 </para>
156 156
157 <para> 157 <para>
158 Also, note that 'perf stat' isn't restricted to a fixed set of 158 Also, note that 'perf stat' isn't restricted to a fixed set of
159 counters - basically any event listed in the output of 'perf list' 159 counters - basically any event listed in the output of 'perf list'
160 can be tallied by 'perf stat'. For example, suppose we wanted to 160 can be tallied by 'perf stat'. For example, suppose we wanted to
161 see a summary of all the events related to kernel memory 161 see a summary of all the events related to kernel memory
162 allocation/freeing along with cache hits and misses: 162 allocation/freeing along with cache hits and misses:
163 <literallayout class='monospaced'> 163 <literallayout class='monospaced'>
164 root@crownbay:~# perf stat -e kmem:* -e cache-references -e cache-misses wget http:// downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 164 root@crownbay:~# perf stat -e kmem:* -e cache-references -e cache-misses wget http:// downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
165 Connecting to downloads.yoctoproject.org (140.211.169.59:80) 165 Connecting to downloads.yoctoproject.org (140.211.169.59:80)
166 linux-2.6.19.2.tar.b 100% |***************************************************| 41727k 0:00:00 ETA 166 linux-2.6.19.2.tar.b 100% |***************************************************| 41727k 0:00:00 ETA
@@ -183,28 +183,28 @@
183 2917740 cache-misses # 4.365 % of all cache refs 183 2917740 cache-misses # 4.365 % of all cache refs
184 184
185 44.831023415 seconds time elapsed 185 44.831023415 seconds time elapsed
186 </literallayout> 186 </literallayout>
187 So 'perf stat' gives us a nice easy way to get a quick overview of 187 So 'perf stat' gives us a nice easy way to get a quick overview of
188 what might be happening for a set of events, but normally we'd 188 what might be happening for a set of events, but normally we'd
189 need a little more detail in order to understand what's going on 189 need a little more detail in order to understand what's going on
190 in a way that we can act on in a useful way. 190 in a way that we can act on in a useful way.
191 </para> 191 </para>
192 192
193 <para> 193 <para>
194 To dive down into a next level of detail, we can use 'perf 194 To dive down into a next level of detail, we can use 'perf
195 record'/'perf report' which will collect profiling data and 195 record'/'perf report' which will collect profiling data and
196 present it to use using an interactive text-based UI (or 196 present it to use using an interactive text-based UI (or
197 simply as text if we specify --stdio to 'perf report'). 197 simply as text if we specify --stdio to 'perf report').
198 </para> 198 </para>
199 199
200 <para> 200 <para>
201 As our first attempt at profiling this workload, we'll simply 201 As our first attempt at profiling this workload, we'll simply
202 run 'perf record', handing it the workload we want to profile 202 run 'perf record', handing it the workload we want to profile
203 (everything after 'perf record' and any perf options we hand 203 (everything after 'perf record' and any perf options we hand
204 it - here none - will be executedin a new shell). perf collects 204 it - here none - will be executedin a new shell). perf collects
205 samples until the process exits and records them in a file named 205 samples until the process exits and records them in a file named
206 'perf.data' in the current working directory. 206 'perf.data' in the current working directory.
207 <literallayout class='monospaced'> 207 <literallayout class='monospaced'>
208 root@crownbay:~# perf record wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 208 root@crownbay:~# perf record wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
209 209
210 Connecting to downloads.yoctoproject.org (140.211.169.59:80) 210 Connecting to downloads.yoctoproject.org (140.211.169.59:80)
@@ -215,44 +215,44 @@
215 To see the results in a 'text-based UI' (tui), simply run 215 To see the results in a 'text-based UI' (tui), simply run
216 'perf report', which will read the perf.data file in the current 216 'perf report', which will read the perf.data file in the current
217 working directory and display the results in an interactive UI: 217 working directory and display the results in an interactive UI:
218 <literallayout class='monospaced'> 218 <literallayout class='monospaced'>
219 root@crownbay:~# perf report 219 root@crownbay:~# perf report
220 </literallayout> 220 </literallayout>
221 </para> 221 </para>
222 222
223 <para> 223 <para>
224 <imagedata fileref="figures/perf-wget-flat-stripped.png" width="6in" depth="7in" align="center" scalefit="1" /> 224 <imagedata fileref="figures/perf-wget-flat-stripped.png" width="6in" depth="7in" align="center" scalefit="1" />
225 </para> 225 </para>
226 226
227 <para> 227 <para>
228 The above screenshot displays a 'flat' profile, one entry for 228 The above screenshot displays a 'flat' profile, one entry for
229 each 'bucket' corresponding to the functions that were profiled 229 each 'bucket' corresponding to the functions that were profiled
230 during the profiling run, ordered from the most popular to the 230 during the profiling run, ordered from the most popular to the
231 least (perf has options to sort in various orders and keys as 231 least (perf has options to sort in various orders and keys as
232 well as display entries only above a certain threshold and so 232 well as display entries only above a certain threshold and so
233 on - see the perf documentation for details). Note that this 233 on - see the perf documentation for details). Note that this
234 includes both userspace functions (entries containing a [.]) and 234 includes both userspace functions (entries containing a [.]) and
235 kernel functions accounted to the process (entries containing 235 kernel functions accounted to the process (entries containing
236 a [k]). (perf has command-line modifiers that can be used to 236 a [k]). (perf has command-line modifiers that can be used to
237 restrict the profiling to kernel or userspace, among others). 237 restrict the profiling to kernel or userspace, among others).
238 </para> 238 </para>
239 239
240 <para> 240 <para>
241 Notice also that the above report shows an entry for 'busybox', 241 Notice also that the above report shows an entry for 'busybox',
242 which is the executable that implements 'wget' in Yocto, but that 242 which is the executable that implements 'wget' in Yocto, but that
243 instead of a useful function name in that entry, it displays 243 instead of a useful function name in that entry, it displays
244 an not-so-friendly hex value instead. The steps below will show 244 an not-so-friendly hex value instead. The steps below will show
245 how to fix that problem. 245 how to fix that problem.
246 </para> 246 </para>
247 247
248 <para> 248 <para>
249 Before we do that, however, let's try running a different profile, 249 Before we do that, however, let's try running a different profile,
250 one which shows something a little more interesting. The only 250 one which shows something a little more interesting. The only
251 difference between the new profile and the previous one is that 251 difference between the new profile and the previous one is that
252 we'll add the -g option, which will record not just the address 252 we'll add the -g option, which will record not just the address
253 of a sampled function, but the entire callchain to the sampled 253 of a sampled function, but the entire callchain to the sampled
254 function as well: 254 function as well:
255 <literallayout class='monospaced'> 255 <literallayout class='monospaced'>
256 root@crownbay:~# perf record -g wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 256 root@crownbay:~# perf record -g wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
257 Connecting to downloads.yoctoproject.org (140.211.169.59:80) 257 Connecting to downloads.yoctoproject.org (140.211.169.59:80)
258 linux-2.6.19.2.tar.b 100% |************************************************| 41727k 0:00:00 ETA 258 linux-2.6.19.2.tar.b 100% |************************************************| 41727k 0:00:00 ETA
@@ -261,217 +261,217 @@
261 261
262 262
263 root@crownbay:~# perf report 263 root@crownbay:~# perf report
264 </literallayout> 264 </literallayout>
265 </para> 265 </para>
266 266
267 <para> 267 <para>
268 <imagedata fileref="figures/perf-wget-g-copy-to-user-expanded-stripped.png" width="6in" depth="7in" align="center" scalefit="1" /> 268 <imagedata fileref="figures/perf-wget-g-copy-to-user-expanded-stripped.png" width="6in" depth="7in" align="center" scalefit="1" />
269 </para> 269 </para>
270 270
271 <para> 271 <para>
272 Using the callgraph view, we can actually see not only which 272 Using the callgraph view, we can actually see not only which
273 functions took the most time, but we can also see a summary of 273 functions took the most time, but we can also see a summary of
274 how those functions were called and learn something about how the 274 how those functions were called and learn something about how the
275 program interacts with the kernel in the process. 275 program interacts with the kernel in the process.
276 </para> 276 </para>
277 277
278 <para> 278 <para>
279 Notice that each entry in the above screenshot now contains a '+' 279 Notice that each entry in the above screenshot now contains a '+'
280 on the left-hand side. This means that we can expand the entry and 280 on the left-hand side. This means that we can expand the entry and
281 drill down into the callchains that feed into that entry. 281 drill down into the callchains that feed into that entry.
282 Pressing 'enter' on any one of them will expand the callchain 282 Pressing 'enter' on any one of them will expand the callchain
283 (you can also press 'E' to expand them all at the same time or 'C' 283 (you can also press 'E' to expand them all at the same time or 'C'
284 to collapse them all). 284 to collapse them all).
285 </para> 285 </para>
286 286
287 <para> 287 <para>
288 In the screenshot above, we've toggled the __copy_to_user_ll() 288 In the screenshot above, we've toggled the __copy_to_user_ll()
289 entry and several subnodes all the way down. This lets us see 289 entry and several subnodes all the way down. This lets us see
290 which callchains contributed to the profiled __copy_to_user_ll() 290 which callchains contributed to the profiled __copy_to_user_ll()
291 function which contributed 1.77% to the total profile. 291 function which contributed 1.77% to the total profile.
292 </para> 292 </para>
293 293
294 <para> 294 <para>
295 As a bit of background explanation for these callchains, think 295 As a bit of background explanation for these callchains, think
296 about what happens at a high level when you run wget to get a file 296 about what happens at a high level when you run wget to get a file
297 out on the network. Basically what happens is that the data comes 297 out on the network. Basically what happens is that the data comes
298 into the kernel via the network connection (socket) and is passed 298 into the kernel via the network connection (socket) and is passed
299 to the userspace program 'wget' (which is actually a part of 299 to the userspace program 'wget' (which is actually a part of
300 busybox, but that's not important for now), which takes the buffers 300 busybox, but that's not important for now), which takes the buffers
301 the kernel passes to it and writes it to a disk file to save it. 301 the kernel passes to it and writes it to a disk file to save it.
302 </para> 302 </para>
303 303
304 <para> 304 <para>
305 The part of this process that we're looking at in the above call 305 The part of this process that we're looking at in the above call
306 stacks is the part where the kernel passes the data it's read from 306 stacks is the part where the kernel passes the data it's read from
307 the socket down to wget i.e. a copy-to-user. 307 the socket down to wget i.e. a copy-to-user.
308 </para> 308 </para>
309 309
310 <para> 310 <para>
311 Notice also that here there's also a case where the a hex value 311 Notice also that here there's also a case where the a hex value
312 is displayed in the callstack, here in the expanded 312 is displayed in the callstack, here in the expanded
313 sys_clock_gettime() function. Later we'll see it resolve to a 313 sys_clock_gettime() function. Later we'll see it resolve to a
314 userspace function call in busybox. 314 userspace function call in busybox.
315 </para> 315 </para>
316 316
317 <para> 317 <para>
318 <imagedata fileref="figures/perf-wget-g-copy-from-user-expanded-stripped.png" width="6in" depth="7in" align="center" scalefit="1" /> 318 <imagedata fileref="figures/perf-wget-g-copy-from-user-expanded-stripped.png" width="6in" depth="7in" align="center" scalefit="1" />
319 </para> 319 </para>
320 320
321 <para> 321 <para>
322 The above screenshot shows the other half of the journey for the 322 The above screenshot shows the other half of the journey for the
323 data - from the wget program's userspace buffers to disk. To get 323 data - from the wget program's userspace buffers to disk. To get
324 the buffers to disk, the wget program issues a write(2), which 324 the buffers to disk, the wget program issues a write(2), which
325 does a copy-from-user to the kernel, which then takes care via 325 does a copy-from-user to the kernel, which then takes care via
326 some circuitous path (probably also present somewhere in the 326 some circuitous path (probably also present somewhere in the
327 profile data), to get it safely to disk. 327 profile data), to get it safely to disk.
328 </para> 328 </para>
329 329
330 <para> 330 <para>
331 Now that we've seen the basic layout of the profile data and the 331 Now that we've seen the basic layout of the profile data and the
332 basics of how to extract useful information out of it, let's get 332 basics of how to extract useful information out of it, let's get
333 back to the task at hand and see if we can get some basic idea 333 back to the task at hand and see if we can get some basic idea
334 about where the time is spent in the program we're profiling, 334 about where the time is spent in the program we're profiling,
335 wget. Remember that wget is actually implemented as an applet 335 wget. Remember that wget is actually implemented as an applet
336 in busybox, so while the process name is 'wget', the executable 336 in busybox, so while the process name is 'wget', the executable
337 we're actually interested in is busybox. So let's expand the 337 we're actually interested in is busybox. So let's expand the
338 first entry containing busybox: 338 first entry containing busybox:
339 </para> 339 </para>
340 340
341 <para> 341 <para>
342 <imagedata fileref="figures/perf-wget-busybox-expanded-stripped.png" width="6in" depth="7in" align="center" scalefit="1" /> 342 <imagedata fileref="figures/perf-wget-busybox-expanded-stripped.png" width="6in" depth="7in" align="center" scalefit="1" />
343 </para> 343 </para>
344 344
345 <para> 345 <para>
346 Again, before we expanded we saw that the function was labeled 346 Again, before we expanded we saw that the function was labeled
347 with a hex value instead of a symbol as with most of the kernel 347 with a hex value instead of a symbol as with most of the kernel
348 entries. Expanding the busybox entry doesn't make it any better. 348 entries. Expanding the busybox entry doesn't make it any better.
349 </para> 349 </para>
350 350
351 <para> 351 <para>
352 The problem is that perf can't find the symbol information for the 352 The problem is that perf can't find the symbol information for the
353 busybox binary, which is actually stripped out by the Yocto build 353 busybox binary, which is actually stripped out by the Yocto build
354 system. 354 system.
355 </para> 355 </para>
356 356
357 <para> 357 <para>
358 One way around that is to put the following in your local.conf 358 One way around that is to put the following in your local.conf
359 when you build the image: 359 when you build the image:
360 <literallayout class='monospaced'> 360 <literallayout class='monospaced'>
361 INHIBIT_PACKAGE_STRIP = "1" 361 INHIBIT_PACKAGE_STRIP = "1"
362 </literallayout> 362 </literallayout>
363 However, we already have an image with the binaries stripped, 363 However, we already have an image with the binaries stripped,
364 so what can we do to get perf to resolve the symbols? Basically 364 so what can we do to get perf to resolve the symbols? Basically
365 we need to install the debuginfo for the busybox package. 365 we need to install the debuginfo for the busybox package.
366 </para> 366 </para>
367 367
368 <para> 368 <para>
369 To generate the debug info for the packages in the image, we can 369 To generate the debug info for the packages in the image, we can
370 to add dbg-pkgs to EXTRA_IMAGE_FEATURES in local.conf. For example: 370 to add dbg-pkgs to EXTRA_IMAGE_FEATURES in local.conf. For example:
371 <literallayout class='monospaced'> 371 <literallayout class='monospaced'>
372 EXTRA_IMAGE_FEATURES = "debug-tweaks tools-profile dbg-pkgs" 372 EXTRA_IMAGE_FEATURES = "debug-tweaks tools-profile dbg-pkgs"
373 </literallayout> 373 </literallayout>
374 Additionally, in order to generate the type of debuginfo that 374 Additionally, in order to generate the type of debuginfo that
375 perf understands, we also need to add the following to local.conf: 375 perf understands, we also need to add the following to local.conf:
376 <literallayout class='monospaced'> 376 <literallayout class='monospaced'>
377 PACKAGE_DEBUG_SPLIT_STYLE = 'debug-file-directory' 377 PACKAGE_DEBUG_SPLIT_STYLE = 'debug-file-directory'
378 </literallayout> 378 </literallayout>
379 Once we've done that, we can install the debuginfo for busybox. 379 Once we've done that, we can install the debuginfo for busybox.
380 The debug packages once built can be found in 380 The debug packages once built can be found in
381 build/tmp/deploy/rpm/* on the host system. Find the 381 build/tmp/deploy/rpm/* on the host system. Find the
382 busybox-dbg-...rpm file and copy it to the target. For example: 382 busybox-dbg-...rpm file and copy it to the target. For example:
383 <literallayout class='monospaced'> 383 <literallayout class='monospaced'>
384 [trz@empanada core2]$ scp /home/trz/yocto/crownbay-tracing-dbg/build/tmp/deploy/rpm/core2/busybox-dbg-1.20.2-r2.core2.rpm root@192.168.1.31: 384 [trz@empanada core2]$ scp /home/trz/yocto/crownbay-tracing-dbg/build/tmp/deploy/rpm/core2/busybox-dbg-1.20.2-r2.core2.rpm root@192.168.1.31:
385 root@192.168.1.31's password: 385 root@192.168.1.31's password:
386 busybox-dbg-1.20.2-r2.core2.rpm 100% 1826KB 1.8MB/s 00:01 386 busybox-dbg-1.20.2-r2.core2.rpm 100% 1826KB 1.8MB/s 00:01
387 </literallayout> 387 </literallayout>
388 Now install the debug rpm on the target: 388 Now install the debug rpm on the target:
389 <literallayout class='monospaced'> 389 <literallayout class='monospaced'>
390 root@crownbay:~# rpm -i busybox-dbg-1.20.2-r2.core2.rpm 390 root@crownbay:~# rpm -i busybox-dbg-1.20.2-r2.core2.rpm
391 </literallayout> 391 </literallayout>
392 Now that the debuginfo is installed, we see that the busybox 392 Now that the debuginfo is installed, we see that the busybox
393 entries now display their functions symbolically: 393 entries now display their functions symbolically:
394 </para> 394 </para>
395 395
396 <para> 396 <para>
397 <imagedata fileref="figures/perf-wget-busybox-debuginfo.png" width="6in" depth="7in" align="center" scalefit="1" /> 397 <imagedata fileref="figures/perf-wget-busybox-debuginfo.png" width="6in" depth="7in" align="center" scalefit="1" />
398 </para> 398 </para>
399 399
400 <para> 400 <para>
401 If we expand one of the entries and press 'enter' on a leaf node, 401 If we expand one of the entries and press 'enter' on a leaf node,
402 we're presented with a menu of actions we can take to get more 402 we're presented with a menu of actions we can take to get more
403 information related to that entry: 403 information related to that entry:
404 </para> 404 </para>
405 405
406 <para> 406 <para>
407 <imagedata fileref="figures/perf-wget-busybox-dso-zoom-menu.png" width="6in" depth="7in" align="center" scalefit="1" /> 407 <imagedata fileref="figures/perf-wget-busybox-dso-zoom-menu.png" width="6in" depth="7in" align="center" scalefit="1" />
408 </para> 408 </para>
409 409
410 <para> 410 <para>
411 One of these actions allows us to show a view that displays a 411 One of these actions allows us to show a view that displays a
412 busybox-centric view of the profiled functions (in this case we've 412 busybox-centric view of the profiled functions (in this case we've
413 also expanded all the nodes using the 'E' key): 413 also expanded all the nodes using the 'E' key):
414 </para> 414 </para>
415 415
416 <para> 416 <para>
417 <imagedata fileref="figures/perf-wget-busybox-dso-zoom.png" width="6in" depth="7in" align="center" scalefit="1" /> 417 <imagedata fileref="figures/perf-wget-busybox-dso-zoom.png" width="6in" depth="7in" align="center" scalefit="1" />
418 </para> 418 </para>
419 419
420 <para> 420 <para>
421 Finally, we can see that now that the busybox debuginfo is 421 Finally, we can see that now that the busybox debuginfo is
422 installed, the previously unresolved symbol in the 422 installed, the previously unresolved symbol in the
423 sys_clock_gettime() entry mentioned previously is now resolved, 423 sys_clock_gettime() entry mentioned previously is now resolved,
424 and shows that the sys_clock_gettime system call that was the 424 and shows that the sys_clock_gettime system call that was the
425 source of 6.75% of the copy-to-user overhead was initiated by 425 source of 6.75% of the copy-to-user overhead was initiated by
426 the handle_input() busybox function: 426 the handle_input() busybox function:
427 </para> 427 </para>
428 428
429 <para> 429 <para>
430 <imagedata fileref="figures/perf-wget-g-copy-to-user-expanded-debuginfo.png" width="6in" depth="7in" align="center" scalefit="1" /> 430 <imagedata fileref="figures/perf-wget-g-copy-to-user-expanded-debuginfo.png" width="6in" depth="7in" align="center" scalefit="1" />
431 </para> 431 </para>
432 432
433 <para> 433 <para>
434 At the lowest level of detail, we can dive down to the assembly 434 At the lowest level of detail, we can dive down to the assembly
435 level and see which instructions caused the most overhead in a 435 level and see which instructions caused the most overhead in a
436 function. Pressing 'enter' on the 'udhcpc_main' function, we're 436 function. Pressing 'enter' on the 'udhcpc_main' function, we're
437 again presented with a menu: 437 again presented with a menu:
438 </para> 438 </para>
439 439
440 <para> 440 <para>
441 <imagedata fileref="figures/perf-wget-busybox-annotate-menu.png" width="6in" depth="7in" align="center" scalefit="1" /> 441 <imagedata fileref="figures/perf-wget-busybox-annotate-menu.png" width="6in" depth="7in" align="center" scalefit="1" />
442 </para> 442 </para>
443 443
444 <para> 444 <para>
445 Selecting 'Annotate udhcpc_main', we get a detailed listing of 445 Selecting 'Annotate udhcpc_main', we get a detailed listing of
446 percentages by instruction for the udhcpc_main function. From the 446 percentages by instruction for the udhcpc_main function. From the
447 display, we can see that over 50% of the time spent in this 447 display, we can see that over 50% of the time spent in this
448 function is taken up by a couple tests and the move of a 448 function is taken up by a couple tests and the move of a
449 constant (1) to a register: 449 constant (1) to a register:
450 </para> 450 </para>
451 451
452 <para> 452 <para>
453 <imagedata fileref="figures/perf-wget-busybox-annotate-udhcpc.png" width="6in" depth="7in" align="center" scalefit="1" /> 453 <imagedata fileref="figures/perf-wget-busybox-annotate-udhcpc.png" width="6in" depth="7in" align="center" scalefit="1" />
454 </para> 454 </para>
455 455
456 <para> 456 <para>
457 As a segue into tracing, let's try another profile using a 457 As a segue into tracing, let's try another profile using a
458 different counter, something other than the default 'cycles'. 458 different counter, something other than the default 'cycles'.
459 </para> 459 </para>
460 460
461 <para> 461 <para>
462 The tracing and profiling infrastructure in Linux has become 462 The tracing and profiling infrastructure in Linux has become
463 unified in a way that allows us to use the same tool with a 463 unified in a way that allows us to use the same tool with a
464 completely different set of counters, not just the standard 464 completely different set of counters, not just the standard
465 hardware counters that traditionally tools have had to restrict 465 hardware counters that traditionally tools have had to restrict
466 themselves to (of course the traditional tools can also make use 466 themselves to (of course the traditional tools can also make use
467 of the expanded possibilities now available to them, and in some 467 of the expanded possibilities now available to them, and in some
468 cases have, as mentioned previously). 468 cases have, as mentioned previously).
469 </para> 469 </para>
470 470
471 <para> 471 <para>
472 We can get a list of the available events that can be used to 472 We can get a list of the available events that can be used to
473 profile a workload via 'perf list': 473 profile a workload via 'perf list':
474 <literallayout class='monospaced'> 474 <literallayout class='monospaced'>
475 root@crownbay:~# perf list 475 root@crownbay:~# perf list
476 476
477 List of pre-defined events (to be used in -e): 477 List of pre-defined events (to be used in -e):
@@ -602,23 +602,23 @@
602 syscalls:sys_exit_unshare [Tracepoint event] 602 syscalls:sys_exit_unshare [Tracepoint event]
603 raw_syscalls:sys_enter [Tracepoint event] 603 raw_syscalls:sys_enter [Tracepoint event]
604 raw_syscalls:sys_exit [Tracepoint event] 604 raw_syscalls:sys_exit [Tracepoint event]
605 </literallayout> 605 </literallayout>
606 </para> 606 </para>
607 607
608 <note> 608 <note>
609 Tying It Together: These are exactly the same set of events defined 609 Tying It Together: These are exactly the same set of events defined
610 by the trace event subsystem and exposed by 610 by the trace event subsystem and exposed by
611 ftrace/tracecmd/kernelshark as files in 611 ftrace/tracecmd/kernelshark as files in
612 /sys/kernel/debug/tracing/events, by SystemTap as 612 /sys/kernel/debug/tracing/events, by SystemTap as
613 kernel.trace("tracepoint_name") and (partially) accessed by LTTng. 613 kernel.trace("tracepoint_name") and (partially) accessed by LTTng.
614 </note> 614 </note>
615 615
616 <para> 616 <para>
617 Only a subset of these would be of interest to us when looking at 617 Only a subset of these would be of interest to us when looking at
618 this workload, so let's choose the most likely subsystems 618 this workload, so let's choose the most likely subsystems
619 (identified by the string before the colon in the Tracepoint events) 619 (identified by the string before the colon in the Tracepoint events)
620 and do a 'perf stat' run using only those wildcarded subsystems: 620 and do a 'perf stat' run using only those wildcarded subsystems:
621 <literallayout class='monospaced'> 621 <literallayout class='monospaced'>
622 root@crownbay:~# perf stat -e skb:* -e net:* -e napi:* -e sched:* -e workqueue:* -e irq:* -e syscalls:* wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 622 root@crownbay:~# perf stat -e skb:* -e net:* -e napi:* -e sched:* -e workqueue:* -e irq:* -e syscalls:* wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
623 Performance counter stats for 'wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2': 623 Performance counter stats for 'wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2':
624 624
@@ -675,26 +675,26 @@
675 . 675 .
676 676
677 58.029710972 seconds time elapsed 677 58.029710972 seconds time elapsed
678 </literallayout> 678 </literallayout>
679 Let's pick one of these tracepoints and tell perf to do a profile 679 Let's pick one of these tracepoints and tell perf to do a profile
680 using it as the sampling event: 680 using it as the sampling event:
681 <literallayout class='monospaced'> 681 <literallayout class='monospaced'>
682 root@crownbay:~# perf record -g -e sched:sched_wakeup wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 682 root@crownbay:~# perf record -g -e sched:sched_wakeup wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
683 </literallayout> 683 </literallayout>
684 </para> 684 </para>
685 685
686 <para> 686 <para>
687 <imagedata fileref="figures/sched-wakeup-profile.png" width="6in" depth="7in" align="center" scalefit="1" /> 687 <imagedata fileref="figures/sched-wakeup-profile.png" width="6in" depth="7in" align="center" scalefit="1" />
688 </para> 688 </para>
689 689
690 <para> 690 <para>
691 The screenshot above shows the results of running a profile using 691 The screenshot above shows the results of running a profile using
692 sched:sched_switch tracepoint, which shows the relative costs of 692 sched:sched_switch tracepoint, which shows the relative costs of
693 various paths to sched_wakeup (note that sched_wakeup is the 693 various paths to sched_wakeup (note that sched_wakeup is the
694 name of the tracepoint - it's actually defined just inside 694 name of the tracepoint - it's actually defined just inside
695 ttwu_do_wakeup(), which accounts for the function name actually 695 ttwu_do_wakeup(), which accounts for the function name actually
696 displayed in the profile: 696 displayed in the profile:
697 <literallayout class='monospaced'> 697 <literallayout class='monospaced'>
698 /* 698 /*
699 * Mark the task runnable and perform wakeup-preemption. 699 * Mark the task runnable and perform wakeup-preemption.
700 */ 700 */
@@ -706,55 +706,55 @@
706 . 706 .
707 . 707 .
708 } 708 }
709 </literallayout> 709 </literallayout>
710 A couple of the more interesting callchains are expanded and 710 A couple of the more interesting callchains are expanded and
711 displayed above, basically some network receive paths that 711 displayed above, basically some network receive paths that
712 presumably end up waking up wget (busybox) when network data is 712 presumably end up waking up wget (busybox) when network data is
713 ready. 713 ready.
714 </para> 714 </para>
715 715
716 <para> 716 <para>
717 Note that because tracepoints are normally used for tracing, 717 Note that because tracepoints are normally used for tracing,
718 the default sampling period for tracepoints is 1 i.e. for 718 the default sampling period for tracepoints is 1 i.e. for
719 tracepoints perf will sample on every event occurrence (this 719 tracepoints perf will sample on every event occurrence (this
720 can be changed using the -c option). This is in contrast to 720 can be changed using the -c option). This is in contrast to
721 hardware counters such as for example the default 'cycles' 721 hardware counters such as for example the default 'cycles'
722 hardware counter used for normal profiling, where sampling 722 hardware counter used for normal profiling, where sampling
723 periods are much higher (in the thousands) because profiling should 723 periods are much higher (in the thousands) because profiling should
724 have as low an overhead as possible and sampling on every cycle w 724 have as low an overhead as possible and sampling on every cycle
725 ould be prohibitively expensive. 725 would be prohibitively expensive.
726 </para> 726 </para>
727 </section> 727 </section>
728 728
729 <section id='using-perf-to-do-basic-tracing'> 729 <section id='using-perf-to-do-basic-tracing'>
730 <title>Using perf to do Basic Tracing</title> 730 <title>Using perf to do Basic Tracing</title>
731 731
732 <para> 732 <para>
733 Profiling is a great tool for solving many problems or for 733 Profiling is a great tool for solving many problems or for
734 getting a high-level view of what's going on with a workload or 734 getting a high-level view of what's going on with a workload or
735 across the system. It is however by definition an approximation, 735 across the system. It is however by definition an approximation,
736 as suggested by the most prominent word associated with it, 736 as suggested by the most prominent word associated with it,
737 'sampling'. On the one hand, it allows a representative picture of 737 'sampling'. On the one hand, it allows a representative picture of
738 what's going on in the system to be cheaply taken, but on the other 738 what's going on in the system to be cheaply taken, but on the other
739 hand, that cheapness limits its utility when that data suggests a 739 hand, that cheapness limits its utility when that data suggests a
740 need to 'dive down' more deeply to discover what's really going 740 need to 'dive down' more deeply to discover what's really going
741 on. In such cases, the only way to see what's really going on is 741 on. In such cases, the only way to see what's really going on is
742 to be able to look at (or summarize more intelligently) the 742 to be able to look at (or summarize more intelligently) the
743 individual steps that go into the higher-level behavior exposed 743 individual steps that go into the higher-level behavior exposed
744 by the coarse-grained profiling data. 744 by the coarse-grained profiling data.
745 </para> 745 </para>
746 746
747 <para> 747 <para>
748 As a concrete example, we can trace all the events we think might 748 As a concrete example, we can trace all the events we think might
749 be applicable to our workload: 749 be applicable to our workload:
750 <literallayout class='monospaced'> 750 <literallayout class='monospaced'>
751 root@crownbay:~# perf record -g -e skb:* -e net:* -e napi:* -e sched:sched_switch -e sched:sched_wakeup -e irq:* 751 root@crownbay:~# perf record -g -e skb:* -e net:* -e napi:* -e sched:sched_switch -e sched:sched_wakeup -e irq:*
752 -e syscalls:sys_enter_read -e syscalls:sys_exit_read -e syscalls:sys_enter_write -e syscalls:sys_exit_write 752 -e syscalls:sys_enter_read -e syscalls:sys_exit_read -e syscalls:sys_enter_write -e syscalls:sys_exit_write
753 wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2 753 wget http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2
754 </literallayout> 754 </literallayout>
755 We can look at the raw trace output using 'perf script' with no 755 We can look at the raw trace output using 'perf script' with no
756 arguments: 756 arguments:
757 <literallayout class='monospaced'> 757 <literallayout class='monospaced'>
758 root@crownbay:~# perf script 758 root@crownbay:~# perf script
759 759
760 perf 1262 [000] 11624.857082: sys_exit_read: 0x0 760 perf 1262 [000] 11624.857082: sys_exit_read: 0x0
@@ -779,90 +779,90 @@
779 wget 1262 [001] 11624.859888: sys_exit_read: 0x400 779 wget 1262 [001] 11624.859888: sys_exit_read: 0x400
780 wget 1262 [001] 11624.859935: sys_enter_read: fd: 0x0003, buf: 0xb7720000, count: 0x0400 780 wget 1262 [001] 11624.859935: sys_enter_read: fd: 0x0003, buf: 0xb7720000, count: 0x0400
781 wget 1262 [001] 11624.859944: sys_exit_read: 0x400 781 wget 1262 [001] 11624.859944: sys_exit_read: 0x400
782 </literallayout> 782 </literallayout>
783 This gives us a detailed timestamped sequence of events that 783 This gives us a detailed timestamped sequence of events that
784 occurred within the workload with respect to those events. 784 occurred within the workload with respect to those events.
785 </para> 785 </para>
786 786
787 <para> 787 <para>
788 In many ways, profiling can be viewed as a subset of tracing - 788 In many ways, profiling can be viewed as a subset of tracing -
789 theoretically, if you have a set of trace events that's sufficient 789 theoretically, if you have a set of trace events that's sufficient
790 to capture all the important aspects of a workload, you can derive 790 to capture all the important aspects of a workload, you can derive
791 any of the results or views that a profiling run can. 791 any of the results or views that a profiling run can.
792 </para> 792 </para>
793 793
794 <para> 794 <para>
795 Another aspect of traditional profiling is that while powerful in 795 Another aspect of traditional profiling is that while powerful in
796 many ways, it's limited by the granularity of the underlying data. 796 many ways, it's limited by the granularity of the underlying data.
797 Profiling tools offer various ways of sorting and presenting the 797 Profiling tools offer various ways of sorting and presenting the
798 sample data, which make it much more useful and amenable to user 798 sample data, which make it much more useful and amenable to user
799 experimentation, but in the end it can't be used in an open-ended 799 experimentation, but in the end it can't be used in an open-ended
800 way to extract data that just isn't present as a consequence of 800 way to extract data that just isn't present as a consequence of
801 the fact that conceptually, most of it has been thrown away. 801 the fact that conceptually, most of it has been thrown away.
802 </para> 802 </para>
803 803
804 <para> 804 <para>
805 Full-blown detailed tracing data does however offer the opportunity 805 Full-blown detailed tracing data does however offer the opportunity
806 to manipulate and present the information collected during a 806 to manipulate and present the information collected during a
807 tracing run in an infinite variety of ways. 807 tracing run in an infinite variety of ways.
808 </para> 808 </para>
809 809
810 <para> 810 <para>
811 Another way to look at it is that there are only so many ways that 811 Another way to look at it is that there are only so many ways that
812 the 'primitive' counters can be used on their own to generate 812 the 'primitive' counters can be used on their own to generate
813 interesting output; to get anything more complicated than simple 813 interesting output; to get anything more complicated than simple
814 counts requires some amount of additional logic, which is typically 814 counts requires some amount of additional logic, which is typically
815 very specific to the problem at hand. For example, if we wanted to 815 very specific to the problem at hand. For example, if we wanted to
816 make use of a 'counter' that maps to the value of the time 816 make use of a 'counter' that maps to the value of the time
817 difference between when a process was scheduled to run on a 817 difference between when a process was scheduled to run on a
818 processor and the time it actually ran, we wouldn't expect such 818 processor and the time it actually ran, we wouldn't expect such
819 a counter to exist on its own, but we could derive one called say 819 a counter to exist on its own, but we could derive one called say
820 'wakeup_latency' and use it to extract a useful view of that metric 820 'wakeup_latency' and use it to extract a useful view of that metric
821 from trace data. Likewise, we really can't figure out from standard 821 from trace data. Likewise, we really can't figure out from standard
822 profiling tools how much data every process on the system reads and 822 profiling tools how much data every process on the system reads and
823 writes, along with how many of those reads and writes fail 823 writes, along with how many of those reads and writes fail
824 completely. If we have sufficient trace data, however, we could 824 completely. If we have sufficient trace data, however, we could
825 with the right tools easily extract and present that information, 825 with the right tools easily extract and present that information,
826 but we'd need something other than pre-canned profiling tools to 826 but we'd need something other than pre-canned profiling tools to
827 do that. 827 do that.
828 </para> 828 </para>
829 829
830 <para> 830 <para>
831 Luckily, there is general-purpose way to handle such needs, 831 Luckily, there is general-purpose way to handle such needs,
832 called 'programming languages'. Making programming languages 832 called 'programming languages'. Making programming languages
833 easily available to apply to such problems given the specific 833 easily available to apply to such problems given the specific
834 format of data is called a 'programming language binding' for 834 format of data is called a 'programming language binding' for
835 that data and language. Perf supports two programming language 835 that data and language. Perf supports two programming language
836 bindings, one for Python and one for Perl. 836 bindings, one for Python and one for Perl.
837 </para> 837 </para>
838 838
839 <note> 839 <note>
840 Tying It Together: Language bindings for manipulating and 840 Tying It Together: Language bindings for manipulating and
841 aggregating trace data are of course not a new 841 aggregating trace data are of course not a new
842 idea. One of the first projects to do this was IBM's DProbes 842 idea. One of the first projects to do this was IBM's DProbes
843 dpcc compiler, an ANSI C compiler which targeted a low-level 843 dpcc compiler, an ANSI C compiler which targeted a low-level
844 assembly language running on an in-kernel interpreter on the 844 assembly language running on an in-kernel interpreter on the
845 target system. This is exactly analagous to what Sun's DTrace 845 target system. This is exactly analagous to what Sun's DTrace
846 did, except that DTrace invented its own language for the purpose. 846 did, except that DTrace invented its own language for the purpose.
847 Systemtap, heavily inspired by DTrace, also created its own 847 Systemtap, heavily inspired by DTrace, also created its own
848 one-off language, but rather than running the product on an 848 one-off language, but rather than running the product on an
849 in-kernel interpreter, created an elaborate compiler-based 849 in-kernel interpreter, created an elaborate compiler-based
850 machinery to translate its language into kernel modules written 850 machinery to translate its language into kernel modules written
851 in C. 851 in C.
852 </note> 852 </note>
853 853
854 <para> 854 <para>
855 Now that we have the trace data in perf.data, we can use 855 Now that we have the trace data in perf.data, we can use
856 'perf script -g' to generate a skeleton script with handlers 856 'perf script -g' to generate a skeleton script with handlers
857 for the read/write entry/exit events we recorded: 857 for the read/write entry/exit events we recorded:
858 <literallayout class='monospaced'> 858 <literallayout class='monospaced'>
859 root@crownbay:~# perf script -g python 859 root@crownbay:~# perf script -g python
860 generated Python script: perf-script.py 860 generated Python script: perf-script.py
861 </literallayout> 861 </literallayout>
862 The skeleton script simply creates a python function for each 862 The skeleton script simply creates a python function for each
863 event type in the perf.data file. The body of each function simply 863 event type in the perf.data file. The body of each function simply
864 prints the event name along with its parameters. For example: 864 prints the event name along with its parameters. For example:
865 <literallayout class='monospaced'> 865 <literallayout class='monospaced'>
866 def net__netif_rx(event_name, context, common_cpu, 866 def net__netif_rx(event_name, context, common_cpu,
867 common_secs, common_nsecs, common_pid, common_comm, 867 common_secs, common_nsecs, common_pid, common_comm,
868 skbaddr, len, name): 868 skbaddr, len, name):
@@ -870,10 +870,10 @@
870 common_pid, common_comm) 870 common_pid, common_comm)
871 871
872 print "skbaddr=%u, len=%u, name=%s\n" % (skbaddr, len, name), 872 print "skbaddr=%u, len=%u, name=%s\n" % (skbaddr, len, name),
873 </literallayout> 873 </literallayout>
874 We can run that script directly to print all of the events 874 We can run that script directly to print all of the events
875 contained in the perf.data file: 875 contained in the perf.data file:
876 <literallayout class='monospaced'> 876 <literallayout class='monospaced'>
877 root@crownbay:~# perf script -s perf-script.py 877 root@crownbay:~# perf script -s perf-script.py
878 878
879 in trace_begin 879 in trace_begin
@@ -899,39 +899,39 @@
899 syscalls__sys_exit_read 1 11624.859888770 1262 wget nr=3, ret=1024 899 syscalls__sys_exit_read 1 11624.859888770 1262 wget nr=3, ret=1024
900 syscalls__sys_enter_read 1 11624.859935140 1262 wget nr=3, fd=3, buf=3077701632, count=1024 900 syscalls__sys_enter_read 1 11624.859935140 1262 wget nr=3, fd=3, buf=3077701632, count=1024
901 syscalls__sys_exit_read 1 11624.859944032 1262 wget nr=3, ret=1024 901 syscalls__sys_exit_read 1 11624.859944032 1262 wget nr=3, ret=1024
902 </literallayout> 902 </literallayout>
903 That in itself isn't very useful; after all, we can accomplish 903 That in itself isn't very useful; after all, we can accomplish
904 pretty much the same thing by simply running 'perf script' 904 pretty much the same thing by simply running 'perf script'
905 without arguments in the same directory as the perf.data file. 905 without arguments in the same directory as the perf.data file.
906 </para> 906 </para>
907 907
908 <para> 908 <para>
909 We can however replace the print statements in the generated 909 We can however replace the print statements in the generated
910 function bodies with whatever we want, and thereby make it 910 function bodies with whatever we want, and thereby make it
911 infinitely more useful. 911 infinitely more useful.
912 </para> 912 </para>
913 913
914 <para> 914 <para>
915 As a simple example, let's just replace the print statements in 915 As a simple example, let's just replace the print statements in
916 the function bodies with a simple function that does nothing but 916 the function bodies with a simple function that does nothing but
917 increment a per-event count. When the program is run against a 917 increment a per-event count. When the program is run against a
918 perf.data file, each time a particular event is encountered, 918 perf.data file, each time a particular event is encountered,
919 a tally is incremented for that event. For example: 919 a tally is incremented for that event. For example:
920 <literallayout class='monospaced'> 920 <literallayout class='monospaced'>
921 def net__netif_rx(event_name, context, common_cpu, 921 def net__netif_rx(event_name, context, common_cpu,
922 common_secs, common_nsecs, common_pid, common_comm, 922 common_secs, common_nsecs, common_pid, common_comm,
923 skbaddr, len, name): 923 skbaddr, len, name):
924 inc_counts(event_name) 924 inc_counts(event_name)
925 </literallayout> 925 </literallayout>
926 Each event handler function in the generated code is modified 926 Each event handler function in the generated code is modified
927 to do this. For convenience, we define a common function called 927 to do this. For convenience, we define a common function called
928 inc_counts() that each handler calls; inc_counts simply tallies 928 inc_counts() that each handler calls; inc_counts simply tallies
929 a count for each event using the 'counts' hash, which is a 929 a count for each event using the 'counts' hash, which is a
930 specialized has function that does Perl-like autovivification, a 930 specialized has function that does Perl-like autovivification, a
931 capability that's extremely useful for kinds of multi-level 931 capability that's extremely useful for kinds of multi-level
932 aggregation commonly used in processing traces (see perf's 932 aggregation commonly used in processing traces (see perf's
933 documentation on the Python language binding for details): 933 documentation on the Python language binding for details):
934 <literallayout class='monospaced'> 934 <literallayout class='monospaced'>
935 counts = autodict() 935 counts = autodict()
936 936
937 def inc_counts(event_name): 937 def inc_counts(event_name):
@@ -939,18 +939,18 @@
939 counts[event_name] += 1 939 counts[event_name] += 1
940 except TypeError: 940 except TypeError:
941 counts[event_name] = 1 941 counts[event_name] = 1
942 </literallayout> 942 </literallayout>
943 Finally, at the end of the trace processing run, we want to 943 Finally, at the end of the trace processing run, we want to
944 print the result of all the per-event tallies. For that, we 944 print the result of all the per-event tallies. For that, we
945 use the special 'trace_end()' function: 945 use the special 'trace_end()' function:
946 <literallayout class='monospaced'> 946 <literallayout class='monospaced'>
947 def trace_end(): 947 def trace_end():
948 for event_name, count in counts.iteritems(): 948 for event_name, count in counts.iteritems():
949 print "%-40s %10s\n" % (event_name, count) 949 print "%-40s %10s\n" % (event_name, count)
950 </literallayout> 950 </literallayout>
951 The end result is a summary of all the events recorded in the 951 The end result is a summary of all the events recorded in the
952 trace: 952 trace:
953 <literallayout> 953 <literallayout>
954 skb__skb_copy_datagram_iovec 13148 954 skb__skb_copy_datagram_iovec 13148
955 irq__softirq_entry 4796 955 irq__softirq_entry 4796
956 irq__irq_handler_exit 3805 956 irq__irq_handler_exit 3805
@@ -968,112 +968,112 @@
968 net__netif_rx 2 968 net__netif_rx 2
969 napi__napi_poll 1877 969 napi__napi_poll 1877
970 syscalls__sys_exit_write 8990 970 syscalls__sys_exit_write 8990
971 </literallayout> 971 </literallayout>
972 Note that this is pretty much exactly the same information we get 972 Note that this is pretty much exactly the same information we get
973 from 'perf stat', which goes a little way to support the idea 973 from 'perf stat', which goes a little way to support the idea
974 mentioned previously that given the right kind of trace data, 974 mentioned previously that given the right kind of trace data,
975 higher-level profiling-type summaries can be derived from it. 975 higher-level profiling-type summaries can be derived from it.
976 </para> 976 </para>
977 977
978 <para> 978 <para>
979 Documentation on using the 979 Documentation on using the
980 <ulink url='http://linux.die.net/man/1/perf-script-python'>'perf script' python binding</ulink>. 980 <ulink url='http://linux.die.net/man/1/perf-script-python'>'perf script' python binding</ulink>.
981 </para> 981 </para>
982 </section> 982 </section>
983 983
984 <section id='system-wide-tracing-and-profiling'> 984 <section id='system-wide-tracing-and-profiling'>
985 <title>System-Wide Tracing and Profiling</title> 985 <title>System-Wide Tracing and Profiling</title>
986 986
987 <para> 987 <para>
988 The examples so far have focused on tracing a particular program or 988 The examples so far have focused on tracing a particular program or
989 workload - in other words, every profiling run has specified the 989 workload - in other words, every profiling run has specified the
990 program to profile in the command-line e.g. 'perf record wget ...'. 990 program to profile in the command-line e.g. 'perf record wget ...'.
991 </para> 991 </para>
992 992
993 <para> 993 <para>
994 It's also possible, and more interesting in many cases, to run a 994 It's also possible, and more interesting in many cases, to run a
995 system-wide profile or trace while running the workload in a 995 system-wide profile or trace while running the workload in a
996 separate shell. 996 separate shell.
997 </para> 997 </para>
998 998
999 <para> 999 <para>
1000 To do system-wide profiling or tracing, you typically use 1000 To do system-wide profiling or tracing, you typically use
1001 the -a flag to 'perf record'. 1001 the -a flag to 'perf record'.
1002 </para> 1002 </para>
1003 1003
1004 <para> 1004 <para>
1005 To demonstrate this, open up one window and start the profile 1005 To demonstrate this, open up one window and start the profile
1006 using the -a flag (press Ctrl-C to stop tracing): 1006 using the -a flag (press Ctrl-C to stop tracing):
1007 <literallayout class='monospaced'> 1007 <literallayout class='monospaced'>
1008 root@crownbay:~# perf record -g -a 1008 root@crownbay:~# perf record -g -a
1009 ^C[ perf record: Woken up 6 times to write data ] 1009 ^C[ perf record: Woken up 6 times to write data ]
1010 [ perf record: Captured and wrote 1.400 MB perf.data (~61172 samples) ] 1010 [ perf record: Captured and wrote 1.400 MB perf.data (~61172 samples) ]
1011 </literallayout> 1011 </literallayout>
1012 In another window, run the wget test: 1012 In another window, run the wget test:
1013 <literallayout class='monospaced'> 1013 <literallayout class='monospaced'>
1014 root@crownbay:~# wget <ulink url='http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2'>http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2</ulink> 1014 root@crownbay:~# wget <ulink url='http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2'>http://downloads.yoctoproject.org/mirror/sources/linux-2.6.19.2.tar.bz2</ulink>
1015 Connecting to downloads.yoctoproject.org (140.211.169.59:80) 1015 Connecting to downloads.yoctoproject.org (140.211.169.59:80)
1016 linux-2.6.19.2.tar.b 100% |*******************************| 41727k 0:00:00 ETA 1016 linux-2.6.19.2.tar.b 100% |*******************************| 41727k 0:00:00 ETA
1017 </literallayout> 1017 </literallayout>
1018 Here we see entries not only for our wget load, but for other 1018 Here we see entries not only for our wget load, but for other
1019 processes running on the system as well: 1019 processes running on the system as well:
1020 </para> 1020 </para>
1021 1021
1022 <para> 1022 <para>
1023 <imagedata fileref="figures/perf-systemwide.png" width="6in" depth="7in" align="center" scalefit="1" /> 1023 <imagedata fileref="figures/perf-systemwide.png" width="6in" depth="7in" align="center" scalefit="1" />
1024 </para> 1024 </para>
1025 1025
1026 <para> 1026 <para>
1027 In the snapshot above, we can see callchains that originate in 1027 In the snapshot above, we can see callchains that originate in
1028 libc, and a callchain from Xorg that demonstrates that we're 1028 libc, and a callchain from Xorg that demonstrates that we're
1029 using a proprietary X driver in userspace (notice the presence 1029 using a proprietary X driver in userspace (notice the presence
1030 of 'PVR' and some other unresolvable symbols in the expanded 1030 of 'PVR' and some other unresolvable symbols in the expanded
1031 Xorg callchain). 1031 Xorg callchain).
1032 </para> 1032 </para>
1033 1033
1034 <para> 1034 <para>
1035 Note also that we have both kernel and userspace entries in the 1035 Note also that we have both kernel and userspace entries in the
1036 above snapshot. We can also tell perf to focus on userspace but 1036 above snapshot. We can also tell perf to focus on userspace but
1037 providing a modifier, in this case 'u', to the 'cycles' hardware 1037 providing a modifier, in this case 'u', to the 'cycles' hardware
1038 counter when we record a profile: 1038 counter when we record a profile:
1039 <literallayout class='monospaced'> 1039 <literallayout class='monospaced'>
1040 root@crownbay:~# perf record -g -a -e cycles:u 1040 root@crownbay:~# perf record -g -a -e cycles:u
1041 ^C[ perf record: Woken up 2 times to write data ] 1041 ^C[ perf record: Woken up 2 times to write data ]
1042 [ perf record: Captured and wrote 0.376 MB perf.data (~16443 samples) ] 1042 [ perf record: Captured and wrote 0.376 MB perf.data (~16443 samples) ]
1043 </literallayout> 1043 </literallayout>
1044 </para> 1044 </para>
1045 1045
1046 <para> 1046 <para>
1047 <imagedata fileref="figures/perf-report-cycles-u.png" width="6in" depth="7in" align="center" scalefit="1" /> 1047 <imagedata fileref="figures/perf-report-cycles-u.png" width="6in" depth="7in" align="center" scalefit="1" />
1048 </para> 1048 </para>
1049 1049
1050 <para> 1050 <para>
1051 Notice in the screenshot above, we see only userspace entries ([.]) 1051 Notice in the screenshot above, we see only userspace entries ([.])
1052 </para> 1052 </para>
1053 1053
1054 <para> 1054 <para>
1055 Finally, we can press 'enter' on a leaf node and select the 'Zoom 1055 Finally, we can press 'enter' on a leaf node and select the 'Zoom
1056 into DSO' menu item to show only entries associated with a 1056 into DSO' menu item to show only entries associated with a
1057 specific DSO. In the screenshot below, we've zoomed into the 1057 specific DSO. In the screenshot below, we've zoomed into the
1058 'libc' DSO which shows all the entries associated with the 1058 'libc' DSO which shows all the entries associated with the
1059 libc-xxx.so DSO. 1059 libc-xxx.so DSO.
1060 </para> 1060 </para>
1061 1061
1062 <para> 1062 <para>
1063 <imagedata fileref="figures/perf-systemwide-libc.png" width="6in" depth="7in" align="center" scalefit="1" /> 1063 <imagedata fileref="figures/perf-systemwide-libc.png" width="6in" depth="7in" align="center" scalefit="1" />
1064 </para> 1064 </para>
1065 1065
1066 <para> 1066 <para>
1067 We can also use the system-wide -a switch to do system-wide 1067 We can also use the system-wide -a switch to do system-wide
1068 tracing. Here we'll trace a couple of scheduler events: 1068 tracing. Here we'll trace a couple of scheduler events:
1069 <literallayout class='monospaced'> 1069 <literallayout class='monospaced'>
1070 root@crownbay:~# perf record -a -e sched:sched_switch -e sched:sched_wakeup 1070 root@crownbay:~# perf record -a -e sched:sched_switch -e sched:sched_wakeup
1071 ^C[ perf record: Woken up 38 times to write data ] 1071 ^C[ perf record: Woken up 38 times to write data ]
1072 [ perf record: Captured and wrote 9.780 MB perf.data (~427299 samples) ] 1072 [ perf record: Captured and wrote 9.780 MB perf.data (~427299 samples) ]
1073 </literallayout> 1073 </literallayout>
1074 We can look at the raw output using 'perf script' with no 1074 We can look at the raw output using 'perf script' with no
1075 arguments: 1075 arguments:
1076 <literallayout class='monospaced'> 1076 <literallayout class='monospaced'>
1077 root@crownbay:~# perf script 1077 root@crownbay:~# perf script
1078 1078
1079 perf 1383 [001] 6171.460045: sched_wakeup: comm=kworker/1:1 pid=21 prio=120 success=1 target_cpu=001 1079 perf 1383 [001] 6171.460045: sched_wakeup: comm=kworker/1:1 pid=21 prio=120 success=1 target_cpu=001
@@ -1086,21 +1086,21 @@
1086 perf 1383 [001] 6171.470058: sched_switch: prev_comm=perf prev_pid=1383 prev_prio=120 prev_state=R+ ==> next_comm=kworker/1:1 next_pid=21 next_prio=120 1086 perf 1383 [001] 6171.470058: sched_switch: prev_comm=perf prev_pid=1383 prev_prio=120 prev_state=R+ ==> next_comm=kworker/1:1 next_pid=21 next_prio=120
1087 kworker/1:1 21 [001] 6171.470082: sched_switch: prev_comm=kworker/1:1 prev_pid=21 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=1383 next_prio=120 1087 kworker/1:1 21 [001] 6171.470082: sched_switch: prev_comm=kworker/1:1 prev_pid=21 prev_prio=120 prev_state=S ==> next_comm=perf next_pid=1383 next_prio=120
1088 perf 1383 [001] 6171.480035: sched_wakeup: comm=kworker/1:1 pid=21 prio=120 success=1 target_cpu=001 1088 perf 1383 [001] 6171.480035: sched_wakeup: comm=kworker/1:1 pid=21 prio=120 success=1 target_cpu=001
1089 </literallayout> 1089 </literallayout>
1090 </para> 1090 </para>
1091
1092 <section id='perf-filtering'>
1093 <title>Filtering</title>
1094 1091
1095 <para> 1092 <section id='perf-filtering'>
1096 Notice that there are a lot of events that don't really have 1093 <title>Filtering</title>
1097 anything to do with what we're interested in, namely events 1094
1098 that schedule 'perf' itself in and out or that wake perf up. 1095 <para>
1099 We can get rid of those by using the '--filter' option - 1096 Notice that there are a lot of events that don't really have
1100 for each event we specify using -e, we can add a --filter 1097 anything to do with what we're interested in, namely events
1101 after that to filter out trace events that contain fields 1098 that schedule 'perf' itself in and out or that wake perf up.
1102 with specific values: 1099 We can get rid of those by using the '--filter' option -
1103 <literallayout class='monospaced'> 1100 for each event we specify using -e, we can add a --filter
1101 after that to filter out trace events that contain fields
1102 with specific values:
1103 <literallayout class='monospaced'>
1104 root@crownbay:~# perf record -a -e sched:sched_switch --filter 'next_comm != perf &amp;&amp; prev_comm != perf' -e sched:sched_wakeup --filter 'comm != perf' 1104 root@crownbay:~# perf record -a -e sched:sched_switch --filter 'next_comm != perf &amp;&amp; prev_comm != perf' -e sched:sched_wakeup --filter 'comm != perf'
1105 ^C[ perf record: Woken up 38 times to write data ] 1105 ^C[ perf record: Woken up 38 times to write data ]
1106 [ perf record: Captured and wrote 9.688 MB perf.data (~423279 samples) ] 1106 [ perf record: Captured and wrote 9.688 MB perf.data (~423279 samples) ]
@@ -1122,59 +1122,59 @@
1122 swapper 0 [000] 7932.326109: sched_wakeup: comm=kworker/0:3 pid=1209 prio=120 success=1 target_cpu=000 1122 swapper 0 [000] 7932.326109: sched_wakeup: comm=kworker/0:3 pid=1209 prio=120 success=1 target_cpu=000
1123 swapper 0 [000] 7932.326171: sched_switch: prev_comm=swapper/0 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=kworker/0:3 next_pid=1209 next_prio=120 1123 swapper 0 [000] 7932.326171: sched_switch: prev_comm=swapper/0 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=kworker/0:3 next_pid=1209 next_prio=120
1124 kworker/0:3 1209 [000] 7932.326214: sched_switch: prev_comm=kworker/0:3 prev_pid=1209 prev_prio=120 prev_state=S ==> next_comm=swapper/0 next_pid=0 next_prio=120 1124 kworker/0:3 1209 [000] 7932.326214: sched_switch: prev_comm=kworker/0:3 prev_pid=1209 prev_prio=120 prev_state=S ==> next_comm=swapper/0 next_pid=0 next_prio=120
1125 </literallayout> 1125 </literallayout>
1126 In this case, we've filtered out all events that have 'perf' 1126 In this case, we've filtered out all events that have 'perf'
1127 in their 'comm' or 'comm_prev' or 'comm_next' fields. Notice 1127 in their 'comm' or 'comm_prev' or 'comm_next' fields. Notice
1128 that there are still events recorded for perf, but notice 1128 that there are still events recorded for perf, but notice
1129 that those events don't have values of 'perf' for the filtered 1129 that those events don't have values of 'perf' for the filtered
1130 fields. To completely filter out anything from perf will 1130 fields. To completely filter out anything from perf will
1131 require a bit more work, but for the purpose of demonstrating 1131 require a bit more work, but for the purpose of demonstrating
1132 how to use filters, it's close enough. 1132 how to use filters, it's close enough.
1133 </para> 1133 </para>
1134 1134
1135 <note> 1135 <note>
1136 Tying It Together: These are exactly the same set of event 1136 Tying It Together: These are exactly the same set of event
1137 filters defined by the trace event subsystem. See the 1137 filters defined by the trace event subsystem. See the
1138 ftrace/tracecmd/kernelshark section for more discussion about 1138 ftrace/tracecmd/kernelshark section for more discussion about
1139 these event filters. 1139 these event filters.
1140 </note> 1140 </note>
1141 1141
1142 <note> 1142 <note>
1143 Tying It Together: These event filters are implemented by a 1143 Tying It Together: These event filters are implemented by a
1144 special-purpose pseudo-interpreter in the kernel and are an 1144 special-purpose pseudo-interpreter in the kernel and are an
1145 integral and indispensable part of the perf design as it 1145 integral and indispensable part of the perf design as it
1146 relates to tracing. kernel-based event filters provide a 1146 relates to tracing. kernel-based event filters provide a
1147 mechanism to precisely throttle the event stream that appears 1147 mechanism to precisely throttle the event stream that appears
1148 in user space, where it makes sense to provide bindings to real 1148 in user space, where it makes sense to provide bindings to real
1149 programming languages for postprocessing the event stream. 1149 programming languages for postprocessing the event stream.
1150 This architecture allows for the intelligent and flexible 1150 This architecture allows for the intelligent and flexible
1151 partitioning of processing between the kernel and user space. 1151 partitioning of processing between the kernel and user space.
1152 Contrast this with other tools such as SystemTap, which does 1152 Contrast this with other tools such as SystemTap, which does
1153 all of its processing in the kernel and as such requires a 1153 all of its processing in the kernel and as such requires a
1154 special project-defined language in order to accommodate that 1154 special project-defined language in order to accommodate that
1155 design, or LTTng, where everything is sent to userspace and 1155 design, or LTTng, where everything is sent to userspace and
1156 as such requires a super-efficient kernel-to-userspace 1156 as such requires a super-efficient kernel-to-userspace
1157 transport mechanism in order to function properly. While 1157 transport mechanism in order to function properly. While
1158 perf certainly can benefit from for instance advances in 1158 perf certainly can benefit from for instance advances in
1159 the design of the transport, it doesn't fundamentally depend 1159 the design of the transport, it doesn't fundamentally depend
1160 on them. Basically, if you find that your perf tracing 1160 on them. Basically, if you find that your perf tracing
1161 application is causing buffer I/O overruns, it probably 1161 application is causing buffer I/O overruns, it probably
1162 means that you aren't taking enough advantage of the 1162 means that you aren't taking enough advantage of the
1163 kernel filtering engine. 1163 kernel filtering engine.
1164 </note> 1164 </note>
1165 </section>
1165 </section> 1166 </section>
1166 </section>
1167 1167
1168 <section id='using-dynamic-tracepoints'> 1168 <section id='using-dynamic-tracepoints'>
1169 <title>Using Dynamic Tracepoints</title> 1169 <title>Using Dynamic Tracepoints</title>
1170 1170
1171 <para> 1171 <para>
1172 perf isn't restricted to the fixed set of static tracepoints 1172 perf isn't restricted to the fixed set of static tracepoints
1173 listed by 'perf list'. Users can also add their own 'dynamic' 1173 listed by 'perf list'. Users can also add their own 'dynamic'
1174 tracepoints anywhere in the kernel. For instance, suppose we 1174 tracepoints anywhere in the kernel. For instance, suppose we
1175 want to define our own tracepoint on do_fork(). We can do that 1175 want to define our own tracepoint on do_fork(). We can do that
1176 using the 'perf probe' perf subcommand: 1176 using the 'perf probe' perf subcommand:
1177 <literallayout class='monospaced'> 1177 <literallayout class='monospaced'>
1178 root@crownbay:~# perf probe do_fork 1178 root@crownbay:~# perf probe do_fork
1179 Added new event: 1179 Added new event:
1180 probe:do_fork (on do_fork) 1180 probe:do_fork (on do_fork)
@@ -1182,13 +1182,13 @@
1182 You can now use it in all perf tools, such as: 1182 You can now use it in all perf tools, such as:
1183 1183
1184 perf record -e probe:do_fork -aR sleep 1 1184 perf record -e probe:do_fork -aR sleep 1
1185 </literallayout> 1185 </literallayout>
1186 Adding a new tracepoint via 'perf probe' results in an event 1186 Adding a new tracepoint via 'perf probe' results in an event
1187 with all the expected files and format in 1187 with all the expected files and format in
1188 /sys/kernel/debug/tracing/events, just the same as for static 1188 /sys/kernel/debug/tracing/events, just the same as for static
1189 tracepoints (as discussed in more detail in the trace events 1189 tracepoints (as discussed in more detail in the trace events
1190 subsystem section: 1190 subsystem section:
1191 <literallayout class='monospaced'> 1191 <literallayout class='monospaced'>
1192 root@crownbay:/sys/kernel/debug/tracing/events/probe/do_fork# ls -al 1192 root@crownbay:/sys/kernel/debug/tracing/events/probe/do_fork# ls -al
1193 drwxr-xr-x 2 root root 0 Oct 28 11:42 . 1193 drwxr-xr-x 2 root root 0 Oct 28 11:42 .
1194 drwxr-xr-x 3 root root 0 Oct 28 11:42 .. 1194 drwxr-xr-x 3 root root 0 Oct 28 11:42 ..
@@ -1210,23 +1210,23 @@
1210 field:unsigned long __probe_ip; offset:12; size:4; signed:0; 1210 field:unsigned long __probe_ip; offset:12; size:4; signed:0;
1211 1211
1212 print fmt: "(%lx)", REC->__probe_ip 1212 print fmt: "(%lx)", REC->__probe_ip
1213 </literallayout> 1213 </literallayout>
1214 We can list all dynamic tracepoints currently in existence: 1214 We can list all dynamic tracepoints currently in existence:
1215 <literallayout class='monospaced'> 1215 <literallayout class='monospaced'>
1216 root@crownbay:~# perf probe -l 1216 root@crownbay:~# perf probe -l
1217 probe:do_fork (on do_fork) 1217 probe:do_fork (on do_fork)
1218 probe:schedule (on schedule) 1218 probe:schedule (on schedule)
1219 </literallayout> 1219 </literallayout>
1220 Let's record system-wide ('sleep 30' is a trick for recording 1220 Let's record system-wide ('sleep 30' is a trick for recording
1221 system-wide but basically do nothing and then wake up after 1221 system-wide but basically do nothing and then wake up after
1222 30 seconds): 1222 30 seconds):
1223 <literallayout class='monospaced'> 1223 <literallayout class='monospaced'>
1224 root@crownbay:~# perf record -g -a -e probe:do_fork sleep 30 1224 root@crownbay:~# perf record -g -a -e probe:do_fork sleep 30
1225 [ perf record: Woken up 1 times to write data ] 1225 [ perf record: Woken up 1 times to write data ]
1226 [ perf record: Captured and wrote 0.087 MB perf.data (~3812 samples) ] 1226 [ perf record: Captured and wrote 0.087 MB perf.data (~3812 samples) ]
1227 </literallayout> 1227 </literallayout>
1228 Using 'perf script' we can see each do_fork event that fired: 1228 Using 'perf script' we can see each do_fork event that fired:
1229 <literallayout class='monospaced'> 1229 <literallayout class='monospaced'>
1230 root@crownbay:~# perf script 1230 root@crownbay:~# perf script
1231 1231
1232 # ======== 1232 # ========
@@ -1265,27 +1265,28 @@
1265 matchbox-deskto 1197 [001] 34237.112138: do_fork: (c1028460) 1265 matchbox-deskto 1197 [001] 34237.112138: do_fork: (c1028460)
1266 matchbox-deskto 1311 [001] 34237.114106: do_fork: (c1028460) 1266 matchbox-deskto 1311 [001] 34237.114106: do_fork: (c1028460)
1267 gaku 1312 [000] 34237.202388: do_fork: (c1028460) 1267 gaku 1312 [000] 34237.202388: do_fork: (c1028460)
1268 </literallayout> 1268 </literallayout>
1269 And using 'perf report' on the same file, we can see the 1269 And using 'perf report' on the same file, we can see the
1270 callgraphs from starting a few programs during those 30 seconds: 1270 callgraphs from starting a few programs during those 30 seconds:
1271 </para> 1271 </para>
1272 1272
1273 <para> 1273 <para>
1274 <imagedata fileref="figures/perf-probe-do_fork-profile.png" width="6in" depth="7in" align="center" scalefit="1" /> 1274 <imagedata fileref="figures/perf-probe-do_fork-profile.png" width="6in" depth="7in" align="center" scalefit="1" />
1275 </para> 1275 </para>
1276 1276
1277 <note> 1277 <note>
1278 Tying It Together: The trace events subsystem accomodate static 1278 Tying It Together: The trace events subsystem accomodate static
1279 and dynamic tracepoints in exactly the same way - there's no 1279 and dynamic tracepoints in exactly the same way - there's no
1280 difference as far as the infrastructure is concerned. See the 1280 difference as far as the infrastructure is concerned. See the
1281 ftrace section for more details on the trace event subsystem. 1281 ftrace section for more details on the trace event subsystem.
1282 </note> 1282 </note>
1283 1283
1284 <note> 1284 <note>
1285 Tying It Together: Dynamic tracepoints are implemented under the 1285 Tying It Together: Dynamic tracepoints are implemented under the
1286 covers by kprobes and uprobes. kprobes and uprobes are also used 1286 covers by kprobes and uprobes. kprobes and uprobes are also used
1287 by and in fact are the main focus of SystemTap. 1287 by and in fact are the main focus of SystemTap.
1288 </note> 1288 </note>
1289 </section>
1289 </section> 1290 </section>
1290 1291
1291 <section id='perf-documentation'> 1292 <section id='perf-documentation'>