PageRenderTime 91ms CodeModel.GetById 11ms RepoModel.GetById 0ms app.codeStats 0ms

/Documentation/trace/ftrace.txt

https://bitbucket.org/sammyz/iscream_thunderc-2.6.35-rebase
Plain Text | 1946 lines | 1558 code | 388 blank | 0 comment | 0 complexity | dcf3f528e237b56b0eeeb66bce092168 MD5 | raw file
Possible License(s): GPL-2.0, LGPL-2.0, AGPL-1.0
  1. ftrace - Function Tracer
  2. ========================
  3. Copyright 2008 Red Hat Inc.
  4. Author: Steven Rostedt <srostedt@redhat.com>
  5. License: The GNU Free Documentation License, Version 1.2
  6. (dual licensed under the GPL v2)
  7. Reviewers: Elias Oltmanns, Randy Dunlap, Andrew Morton,
  8. John Kacur, and David Teigland.
  9. Written for: 2.6.28-rc2
  10. Introduction
  11. ------------
  12. Ftrace is an internal tracer designed to help out developers and
  13. designers of systems to find what is going on inside the kernel.
  14. It can be used for debugging or analyzing latencies and
  15. performance issues that take place outside of user-space.
  16. Although ftrace is the function tracer, it also includes an
  17. infrastructure that allows for other types of tracing. Some of
  18. the tracers that are currently in ftrace include a tracer to
  19. trace context switches, the time it takes for a high priority
  20. task to run after it was woken up, the time interrupts are
  21. disabled, and more (ftrace allows for tracer plugins, which
  22. means that the list of tracers can always grow).
  23. Implementation Details
  24. ----------------------
  25. See ftrace-design.txt for details for arch porters and such.
  26. The File System
  27. ---------------
  28. Ftrace uses the debugfs file system to hold the control files as
  29. well as the files to display output.
  30. When debugfs is configured into the kernel (which selecting any ftrace
  31. option will do) the directory /sys/kernel/debug will be created. To mount
  32. this directory, you can add to your /etc/fstab file:
  33. debugfs /sys/kernel/debug debugfs defaults 0 0
  34. Or you can mount it at run time with:
  35. mount -t debugfs nodev /sys/kernel/debug
  36. For quicker access to that directory you may want to make a soft link to
  37. it:
  38. ln -s /sys/kernel/debug /debug
  39. Any selected ftrace option will also create a directory called tracing
  40. within the debugfs. The rest of the document will assume that you are in
  41. the ftrace directory (cd /sys/kernel/debug/tracing) and will only concentrate
  42. on the files within that directory and not distract from the content with
  43. the extended "/sys/kernel/debug/tracing" path name.
  44. That's it! (assuming that you have ftrace configured into your kernel)
  45. After mounting the debugfs, you can see a directory called
  46. "tracing". This directory contains the control and output files
  47. of ftrace. Here is a list of some of the key files:
  48. Note: all time values are in microseconds.
  49. current_tracer:
  50. This is used to set or display the current tracer
  51. that is configured.
  52. available_tracers:
  53. This holds the different types of tracers that
  54. have been compiled into the kernel. The
  55. tracers listed here can be configured by
  56. echoing their name into current_tracer.
  57. tracing_enabled:
  58. This sets or displays whether the current_tracer
  59. is activated and tracing or not. Echo 0 into this
  60. file to disable the tracer or 1 to enable it.
  61. trace:
  62. This file holds the output of the trace in a human
  63. readable format (described below).
  64. trace_pipe:
  65. The output is the same as the "trace" file but this
  66. file is meant to be streamed with live tracing.
  67. Reads from this file will block until new data is
  68. retrieved. Unlike the "trace" file, this file is a
  69. consumer. This means reading from this file causes
  70. sequential reads to display more current data. Once
  71. data is read from this file, it is consumed, and
  72. will not be read again with a sequential read. The
  73. "trace" file is static, and if the tracer is not
  74. adding more data,they will display the same
  75. information every time they are read.
  76. trace_options:
  77. This file lets the user control the amount of data
  78. that is displayed in one of the above output
  79. files.
  80. tracing_max_latency:
  81. Some of the tracers record the max latency.
  82. For example, the time interrupts are disabled.
  83. This time is saved in this file. The max trace
  84. will also be stored, and displayed by "trace".
  85. A new max trace will only be recorded if the
  86. latency is greater than the value in this
  87. file. (in microseconds)
  88. buffer_size_kb:
  89. This sets or displays the number of kilobytes each CPU
  90. buffer can hold. The tracer buffers are the same size
  91. for each CPU. The displayed number is the size of the
  92. CPU buffer and not total size of all buffers. The
  93. trace buffers are allocated in pages (blocks of memory
  94. that the kernel uses for allocation, usually 4 KB in size).
  95. If the last page allocated has room for more bytes
  96. than requested, the rest of the page will be used,
  97. making the actual allocation bigger than requested.
  98. ( Note, the size may not be a multiple of the page size
  99. due to buffer management overhead. )
  100. This can only be updated when the current_tracer
  101. is set to "nop".
  102. tracing_cpumask:
  103. This is a mask that lets the user only trace
  104. on specified CPUS. The format is a hex string
  105. representing the CPUS.
  106. set_ftrace_filter:
  107. When dynamic ftrace is configured in (see the
  108. section below "dynamic ftrace"), the code is dynamically
  109. modified (code text rewrite) to disable calling of the
  110. function profiler (mcount). This lets tracing be configured
  111. in with practically no overhead in performance. This also
  112. has a side effect of enabling or disabling specific functions
  113. to be traced. Echoing names of functions into this file
  114. will limit the trace to only those functions.
  115. This interface also allows for commands to be used. See the
  116. "Filter commands" section for more details.
  117. set_ftrace_notrace:
  118. This has an effect opposite to that of
  119. set_ftrace_filter. Any function that is added here will not
  120. be traced. If a function exists in both set_ftrace_filter
  121. and set_ftrace_notrace, the function will _not_ be traced.
  122. set_ftrace_pid:
  123. Have the function tracer only trace a single thread.
  124. set_graph_function:
  125. Set a "trigger" function where tracing should start
  126. with the function graph tracer (See the section
  127. "dynamic ftrace" for more details).
  128. available_filter_functions:
  129. This lists the functions that ftrace
  130. has processed and can trace. These are the function
  131. names that you can pass to "set_ftrace_filter" or
  132. "set_ftrace_notrace". (See the section "dynamic ftrace"
  133. below for more details.)
  134. The Tracers
  135. -----------
  136. Here is the list of current tracers that may be configured.
  137. "function"
  138. Function call tracer to trace all kernel functions.
  139. "function_graph"
  140. Similar to the function tracer except that the
  141. function tracer probes the functions on their entry
  142. whereas the function graph tracer traces on both entry
  143. and exit of the functions. It then provides the ability
  144. to draw a graph of function calls similar to C code
  145. source.
  146. "sched_switch"
  147. Traces the context switches and wakeups between tasks.
  148. "irqsoff"
  149. Traces the areas that disable interrupts and saves
  150. the trace with the longest max latency.
  151. See tracing_max_latency. When a new max is recorded,
  152. it replaces the old trace. It is best to view this
  153. trace with the latency-format option enabled.
  154. "preemptoff"
  155. Similar to irqsoff but traces and records the amount of
  156. time for which preemption is disabled.
  157. "preemptirqsoff"
  158. Similar to irqsoff and preemptoff, but traces and
  159. records the largest time for which irqs and/or preemption
  160. is disabled.
  161. "wakeup"
  162. Traces and records the max latency that it takes for
  163. the highest priority task to get scheduled after
  164. it has been woken up.
  165. "hw-branch-tracer"
  166. Uses the BTS CPU feature on x86 CPUs to traces all
  167. branches executed.
  168. "nop"
  169. This is the "trace nothing" tracer. To remove all
  170. tracers from tracing simply echo "nop" into
  171. current_tracer.
  172. Examples of using the tracer
  173. ----------------------------
  174. Here are typical examples of using the tracers when controlling
  175. them only with the debugfs interface (without using any
  176. user-land utilities).
  177. Output format:
  178. --------------
  179. Here is an example of the output format of the file "trace"
  180. --------
  181. # tracer: function
  182. #
  183. # TASK-PID CPU# TIMESTAMP FUNCTION
  184. # | | | | |
  185. bash-4251 [01] 10152.583854: path_put <-path_walk
  186. bash-4251 [01] 10152.583855: dput <-path_put
  187. bash-4251 [01] 10152.583855: _atomic_dec_and_lock <-dput
  188. --------
  189. A header is printed with the tracer name that is represented by
  190. the trace. In this case the tracer is "function". Then a header
  191. showing the format. Task name "bash", the task PID "4251", the
  192. CPU that it was running on "01", the timestamp in <secs>.<usecs>
  193. format, the function name that was traced "path_put" and the
  194. parent function that called this function "path_walk". The
  195. timestamp is the time at which the function was entered.
  196. The sched_switch tracer also includes tracing of task wakeups
  197. and context switches.
  198. ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 2916:115:S
  199. ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 10:115:S
  200. ksoftirqd/1-7 [01] 1453.070013: 7:115:R ==> 10:115:R
  201. events/1-10 [01] 1453.070013: 10:115:S ==> 2916:115:R
  202. kondemand/1-2916 [01] 1453.070013: 2916:115:S ==> 7:115:R
  203. ksoftirqd/1-7 [01] 1453.070013: 7:115:S ==> 0:140:R
  204. Wake ups are represented by a "+" and the context switches are
  205. shown as "==>". The format is:
  206. Context switches:
  207. Previous task Next Task
  208. <pid>:<prio>:<state> ==> <pid>:<prio>:<state>
  209. Wake ups:
  210. Current task Task waking up
  211. <pid>:<prio>:<state> + <pid>:<prio>:<state>
  212. The prio is the internal kernel priority, which is the inverse
  213. of the priority that is usually displayed by user-space tools.
  214. Zero represents the highest priority (99). Prio 100 starts the
  215. "nice" priorities with 100 being equal to nice -20 and 139 being
  216. nice 19. The prio "140" is reserved for the idle task which is
  217. the lowest priority thread (pid 0).
  218. Latency trace format
  219. --------------------
  220. When the latency-format option is enabled, the trace file gives
  221. somewhat more information to see why a latency happened.
  222. Here is a typical trace.
  223. # tracer: irqsoff
  224. #
  225. irqsoff latency trace v1.1.5 on 2.6.26-rc8
  226. --------------------------------------------------------------------
  227. latency: 97 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  228. -----------------
  229. | task: swapper-0 (uid:0 nice:0 policy:0 rt_prio:0)
  230. -----------------
  231. => started at: apic_timer_interrupt
  232. => ended at: do_softirq
  233. # _------=> CPU#
  234. # / _-----=> irqs-off
  235. # | / _----=> need-resched
  236. # || / _---=> hardirq/softirq
  237. # ||| / _--=> preempt-depth
  238. # |||| /
  239. # ||||| delay
  240. # cmd pid ||||| time | caller
  241. # \ / ||||| \ | /
  242. <idle>-0 0d..1 0us+: trace_hardirqs_off_thunk (apic_timer_interrupt)
  243. <idle>-0 0d.s. 97us : __do_softirq (do_softirq)
  244. <idle>-0 0d.s1 98us : trace_hardirqs_on (do_softirq)
  245. This shows that the current tracer is "irqsoff" tracing the time
  246. for which interrupts were disabled. It gives the trace version
  247. and the version of the kernel upon which this was executed on
  248. (2.6.26-rc8). Then it displays the max latency in microsecs (97
  249. us). The number of trace entries displayed and the total number
  250. recorded (both are three: #3/3). The type of preemption that was
  251. used (PREEMPT). VP, KP, SP, and HP are always zero and are
  252. reserved for later use. #P is the number of online CPUS (#P:2).
  253. The task is the process that was running when the latency
  254. occurred. (swapper pid: 0).
  255. The start and stop (the functions in which the interrupts were
  256. disabled and enabled respectively) that caused the latencies:
  257. apic_timer_interrupt is where the interrupts were disabled.
  258. do_softirq is where they were enabled again.
  259. The next lines after the header are the trace itself. The header
  260. explains which is which.
  261. cmd: The name of the process in the trace.
  262. pid: The PID of that process.
  263. CPU#: The CPU which the process was running on.
  264. irqs-off: 'd' interrupts are disabled. '.' otherwise.
  265. Note: If the architecture does not support a way to
  266. read the irq flags variable, an 'X' will always
  267. be printed here.
  268. need-resched: 'N' task need_resched is set, '.' otherwise.
  269. hardirq/softirq:
  270. 'H' - hard irq occurred inside a softirq.
  271. 'h' - hard irq is running
  272. 's' - soft irq is running
  273. '.' - normal context.
  274. preempt-depth: The level of preempt_disabled
  275. The above is mostly meaningful for kernel developers.
  276. time: When the latency-format option is enabled, the trace file
  277. output includes a timestamp relative to the start of the
  278. trace. This differs from the output when latency-format
  279. is disabled, which includes an absolute timestamp.
  280. delay: This is just to help catch your eye a bit better. And
  281. needs to be fixed to be only relative to the same CPU.
  282. The marks are determined by the difference between this
  283. current trace and the next trace.
  284. '!' - greater than preempt_mark_thresh (default 100)
  285. '+' - greater than 1 microsecond
  286. ' ' - less than or equal to 1 microsecond.
  287. The rest is the same as the 'trace' file.
  288. trace_options
  289. -------------
  290. The trace_options file is used to control what gets printed in
  291. the trace output. To see what is available, simply cat the file:
  292. cat trace_options
  293. print-parent nosym-offset nosym-addr noverbose noraw nohex nobin \
  294. noblock nostacktrace nosched-tree nouserstacktrace nosym-userobj
  295. To disable one of the options, echo in the option prepended with
  296. "no".
  297. echo noprint-parent > trace_options
  298. To enable an option, leave off the "no".
  299. echo sym-offset > trace_options
  300. Here are the available options:
  301. print-parent - On function traces, display the calling (parent)
  302. function as well as the function being traced.
  303. print-parent:
  304. bash-4000 [01] 1477.606694: simple_strtoul <-strict_strtoul
  305. noprint-parent:
  306. bash-4000 [01] 1477.606694: simple_strtoul
  307. sym-offset - Display not only the function name, but also the
  308. offset in the function. For example, instead of
  309. seeing just "ktime_get", you will see
  310. "ktime_get+0xb/0x20".
  311. sym-offset:
  312. bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0
  313. sym-addr - this will also display the function address as well
  314. as the function name.
  315. sym-addr:
  316. bash-4000 [01] 1477.606694: simple_strtoul <c0339346>
  317. verbose - This deals with the trace file when the
  318. latency-format option is enabled.
  319. bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \
  320. (+0.000ms): simple_strtoul (strict_strtoul)
  321. raw - This will display raw numbers. This option is best for
  322. use with user applications that can translate the raw
  323. numbers better than having it done in the kernel.
  324. hex - Similar to raw, but the numbers will be in a hexadecimal
  325. format.
  326. bin - This will print out the formats in raw binary.
  327. block - TBD (needs update)
  328. stacktrace - This is one of the options that changes the trace
  329. itself. When a trace is recorded, so is the stack
  330. of functions. This allows for back traces of
  331. trace sites.
  332. userstacktrace - This option changes the trace. It records a
  333. stacktrace of the current userspace thread.
  334. sym-userobj - when user stacktrace are enabled, look up which
  335. object the address belongs to, and print a
  336. relative address. This is especially useful when
  337. ASLR is on, otherwise you don't get a chance to
  338. resolve the address to object/file/line after
  339. the app is no longer running
  340. The lookup is performed when you read
  341. trace,trace_pipe. Example:
  342. a.out-1623 [000] 40874.465068: /root/a.out[+0x480] <-/root/a.out[+0
  343. x494] <- /root/a.out[+0x4a8] <- /lib/libc-2.7.so[+0x1e1a6]
  344. sched-tree - trace all tasks that are on the runqueue, at
  345. every scheduling event. Will add overhead if
  346. there's a lot of tasks running at once.
  347. latency-format - This option changes the trace. When
  348. it is enabled, the trace displays
  349. additional information about the
  350. latencies, as described in "Latency
  351. trace format".
  352. sched_switch
  353. ------------
  354. This tracer simply records schedule switches. Here is an example
  355. of how to use it.
  356. # echo sched_switch > current_tracer
  357. # echo 1 > tracing_enabled
  358. # sleep 1
  359. # echo 0 > tracing_enabled
  360. # cat trace
  361. # tracer: sched_switch
  362. #
  363. # TASK-PID CPU# TIMESTAMP FUNCTION
  364. # | | | | |
  365. bash-3997 [01] 240.132281: 3997:120:R + 4055:120:R
  366. bash-3997 [01] 240.132284: 3997:120:R ==> 4055:120:R
  367. sleep-4055 [01] 240.132371: 4055:120:S ==> 3997:120:R
  368. bash-3997 [01] 240.132454: 3997:120:R + 4055:120:S
  369. bash-3997 [01] 240.132457: 3997:120:R ==> 4055:120:R
  370. sleep-4055 [01] 240.132460: 4055:120:D ==> 3997:120:R
  371. bash-3997 [01] 240.132463: 3997:120:R + 4055:120:D
  372. bash-3997 [01] 240.132465: 3997:120:R ==> 4055:120:R
  373. <idle>-0 [00] 240.132589: 0:140:R + 4:115:S
  374. <idle>-0 [00] 240.132591: 0:140:R ==> 4:115:R
  375. ksoftirqd/0-4 [00] 240.132595: 4:115:S ==> 0:140:R
  376. <idle>-0 [00] 240.132598: 0:140:R + 4:115:S
  377. <idle>-0 [00] 240.132599: 0:140:R ==> 4:115:R
  378. ksoftirqd/0-4 [00] 240.132603: 4:115:S ==> 0:140:R
  379. sleep-4055 [01] 240.133058: 4055:120:S ==> 3997:120:R
  380. [...]
  381. As we have discussed previously about this format, the header
  382. shows the name of the trace and points to the options. The
  383. "FUNCTION" is a misnomer since here it represents the wake ups
  384. and context switches.
  385. The sched_switch file only lists the wake ups (represented with
  386. '+') and context switches ('==>') with the previous task or
  387. current task first followed by the next task or task waking up.
  388. The format for both of these is PID:KERNEL-PRIO:TASK-STATE.
  389. Remember that the KERNEL-PRIO is the inverse of the actual
  390. priority with zero (0) being the highest priority and the nice
  391. values starting at 100 (nice -20). Below is a quick chart to map
  392. the kernel priority to user land priorities.
  393. Kernel Space User Space
  394. ===============================================================
  395. 0(high) to 98(low) user RT priority 99(high) to 1(low)
  396. with SCHED_RR or SCHED_FIFO
  397. ---------------------------------------------------------------
  398. 99 sched_priority is not used in scheduling
  399. decisions(it must be specified as 0)
  400. ---------------------------------------------------------------
  401. 100(high) to 139(low) user nice -20(high) to 19(low)
  402. ---------------------------------------------------------------
  403. 140 idle task priority
  404. ---------------------------------------------------------------
  405. The task states are:
  406. R - running : wants to run, may not actually be running
  407. S - sleep : process is waiting to be woken up (handles signals)
  408. D - disk sleep (uninterruptible sleep) : process must be woken up
  409. (ignores signals)
  410. T - stopped : process suspended
  411. t - traced : process is being traced (with something like gdb)
  412. Z - zombie : process waiting to be cleaned up
  413. X - unknown
  414. ftrace_enabled
  415. --------------
  416. The following tracers (listed below) give different output
  417. depending on whether or not the sysctl ftrace_enabled is set. To
  418. set ftrace_enabled, one can either use the sysctl function or
  419. set it via the proc file system interface.
  420. sysctl kernel.ftrace_enabled=1
  421. or
  422. echo 1 > /proc/sys/kernel/ftrace_enabled
  423. To disable ftrace_enabled simply replace the '1' with '0' in the
  424. above commands.
  425. When ftrace_enabled is set the tracers will also record the
  426. functions that are within the trace. The descriptions of the
  427. tracers will also show an example with ftrace enabled.
  428. irqsoff
  429. -------
  430. When interrupts are disabled, the CPU can not react to any other
  431. external event (besides NMIs and SMIs). This prevents the timer
  432. interrupt from triggering or the mouse interrupt from letting
  433. the kernel know of a new mouse event. The result is a latency
  434. with the reaction time.
  435. The irqsoff tracer tracks the time for which interrupts are
  436. disabled. When a new maximum latency is hit, the tracer saves
  437. the trace leading up to that latency point so that every time a
  438. new maximum is reached, the old saved trace is discarded and the
  439. new trace is saved.
  440. To reset the maximum, echo 0 into tracing_max_latency. Here is
  441. an example:
  442. # echo irqsoff > current_tracer
  443. # echo latency-format > trace_options
  444. # echo 0 > tracing_max_latency
  445. # echo 1 > tracing_enabled
  446. # ls -ltr
  447. [...]
  448. # echo 0 > tracing_enabled
  449. # cat trace
  450. # tracer: irqsoff
  451. #
  452. irqsoff latency trace v1.1.5 on 2.6.26
  453. --------------------------------------------------------------------
  454. latency: 12 us, #3/3, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  455. -----------------
  456. | task: bash-3730 (uid:0 nice:0 policy:0 rt_prio:0)
  457. -----------------
  458. => started at: sys_setpgid
  459. => ended at: sys_setpgid
  460. # _------=> CPU#
  461. # / _-----=> irqs-off
  462. # | / _----=> need-resched
  463. # || / _---=> hardirq/softirq
  464. # ||| / _--=> preempt-depth
  465. # |||| /
  466. # ||||| delay
  467. # cmd pid ||||| time | caller
  468. # \ / ||||| \ | /
  469. bash-3730 1d... 0us : _write_lock_irq (sys_setpgid)
  470. bash-3730 1d..1 1us+: _write_unlock_irq (sys_setpgid)
  471. bash-3730 1d..2 14us : trace_hardirqs_on (sys_setpgid)
  472. Here we see that that we had a latency of 12 microsecs (which is
  473. very good). The _write_lock_irq in sys_setpgid disabled
  474. interrupts. The difference between the 12 and the displayed
  475. timestamp 14us occurred because the clock was incremented
  476. between the time of recording the max latency and the time of
  477. recording the function that had that latency.
  478. Note the above example had ftrace_enabled not set. If we set the
  479. ftrace_enabled, we get a much larger output:
  480. # tracer: irqsoff
  481. #
  482. irqsoff latency trace v1.1.5 on 2.6.26-rc8
  483. --------------------------------------------------------------------
  484. latency: 50 us, #101/101, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  485. -----------------
  486. | task: ls-4339 (uid:0 nice:0 policy:0 rt_prio:0)
  487. -----------------
  488. => started at: __alloc_pages_internal
  489. => ended at: __alloc_pages_internal
  490. # _------=> CPU#
  491. # / _-----=> irqs-off
  492. # | / _----=> need-resched
  493. # || / _---=> hardirq/softirq
  494. # ||| / _--=> preempt-depth
  495. # |||| /
  496. # ||||| delay
  497. # cmd pid ||||| time | caller
  498. # \ / ||||| \ | /
  499. ls-4339 0...1 0us+: get_page_from_freelist (__alloc_pages_internal)
  500. ls-4339 0d..1 3us : rmqueue_bulk (get_page_from_freelist)
  501. ls-4339 0d..1 3us : _spin_lock (rmqueue_bulk)
  502. ls-4339 0d..1 4us : add_preempt_count (_spin_lock)
  503. ls-4339 0d..2 4us : __rmqueue (rmqueue_bulk)
  504. ls-4339 0d..2 5us : __rmqueue_smallest (__rmqueue)
  505. ls-4339 0d..2 5us : __mod_zone_page_state (__rmqueue_smallest)
  506. ls-4339 0d..2 6us : __rmqueue (rmqueue_bulk)
  507. ls-4339 0d..2 6us : __rmqueue_smallest (__rmqueue)
  508. ls-4339 0d..2 7us : __mod_zone_page_state (__rmqueue_smallest)
  509. ls-4339 0d..2 7us : __rmqueue (rmqueue_bulk)
  510. ls-4339 0d..2 8us : __rmqueue_smallest (__rmqueue)
  511. [...]
  512. ls-4339 0d..2 46us : __rmqueue_smallest (__rmqueue)
  513. ls-4339 0d..2 47us : __mod_zone_page_state (__rmqueue_smallest)
  514. ls-4339 0d..2 47us : __rmqueue (rmqueue_bulk)
  515. ls-4339 0d..2 48us : __rmqueue_smallest (__rmqueue)
  516. ls-4339 0d..2 48us : __mod_zone_page_state (__rmqueue_smallest)
  517. ls-4339 0d..2 49us : _spin_unlock (rmqueue_bulk)
  518. ls-4339 0d..2 49us : sub_preempt_count (_spin_unlock)
  519. ls-4339 0d..1 50us : get_page_from_freelist (__alloc_pages_internal)
  520. ls-4339 0d..2 51us : trace_hardirqs_on (__alloc_pages_internal)
  521. Here we traced a 50 microsecond latency. But we also see all the
  522. functions that were called during that time. Note that by
  523. enabling function tracing, we incur an added overhead. This
  524. overhead may extend the latency times. But nevertheless, this
  525. trace has provided some very helpful debugging information.
  526. preemptoff
  527. ----------
  528. When preemption is disabled, we may be able to receive
  529. interrupts but the task cannot be preempted and a higher
  530. priority task must wait for preemption to be enabled again
  531. before it can preempt a lower priority task.
  532. The preemptoff tracer traces the places that disable preemption.
  533. Like the irqsoff tracer, it records the maximum latency for
  534. which preemption was disabled. The control of preemptoff tracer
  535. is much like the irqsoff tracer.
  536. # echo preemptoff > current_tracer
  537. # echo latency-format > trace_options
  538. # echo 0 > tracing_max_latency
  539. # echo 1 > tracing_enabled
  540. # ls -ltr
  541. [...]
  542. # echo 0 > tracing_enabled
  543. # cat trace
  544. # tracer: preemptoff
  545. #
  546. preemptoff latency trace v1.1.5 on 2.6.26-rc8
  547. --------------------------------------------------------------------
  548. latency: 29 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  549. -----------------
  550. | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
  551. -----------------
  552. => started at: do_IRQ
  553. => ended at: __do_softirq
  554. # _------=> CPU#
  555. # / _-----=> irqs-off
  556. # | / _----=> need-resched
  557. # || / _---=> hardirq/softirq
  558. # ||| / _--=> preempt-depth
  559. # |||| /
  560. # ||||| delay
  561. # cmd pid ||||| time | caller
  562. # \ / ||||| \ | /
  563. sshd-4261 0d.h. 0us+: irq_enter (do_IRQ)
  564. sshd-4261 0d.s. 29us : _local_bh_enable (__do_softirq)
  565. sshd-4261 0d.s1 30us : trace_preempt_on (__do_softirq)
  566. This has some more changes. Preemption was disabled when an
  567. interrupt came in (notice the 'h'), and was enabled while doing
  568. a softirq. (notice the 's'). But we also see that interrupts
  569. have been disabled when entering the preempt off section and
  570. leaving it (the 'd'). We do not know if interrupts were enabled
  571. in the mean time.
  572. # tracer: preemptoff
  573. #
  574. preemptoff latency trace v1.1.5 on 2.6.26-rc8
  575. --------------------------------------------------------------------
  576. latency: 63 us, #87/87, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  577. -----------------
  578. | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
  579. -----------------
  580. => started at: remove_wait_queue
  581. => ended at: __do_softirq
  582. # _------=> CPU#
  583. # / _-----=> irqs-off
  584. # | / _----=> need-resched
  585. # || / _---=> hardirq/softirq
  586. # ||| / _--=> preempt-depth
  587. # |||| /
  588. # ||||| delay
  589. # cmd pid ||||| time | caller
  590. # \ / ||||| \ | /
  591. sshd-4261 0d..1 0us : _spin_lock_irqsave (remove_wait_queue)
  592. sshd-4261 0d..1 1us : _spin_unlock_irqrestore (remove_wait_queue)
  593. sshd-4261 0d..1 2us : do_IRQ (common_interrupt)
  594. sshd-4261 0d..1 2us : irq_enter (do_IRQ)
  595. sshd-4261 0d..1 2us : idle_cpu (irq_enter)
  596. sshd-4261 0d..1 3us : add_preempt_count (irq_enter)
  597. sshd-4261 0d.h1 3us : idle_cpu (irq_enter)
  598. sshd-4261 0d.h. 4us : handle_fasteoi_irq (do_IRQ)
  599. [...]
  600. sshd-4261 0d.h. 12us : add_preempt_count (_spin_lock)
  601. sshd-4261 0d.h1 12us : ack_ioapic_quirk_irq (handle_fasteoi_irq)
  602. sshd-4261 0d.h1 13us : move_native_irq (ack_ioapic_quirk_irq)
  603. sshd-4261 0d.h1 13us : _spin_unlock (handle_fasteoi_irq)
  604. sshd-4261 0d.h1 14us : sub_preempt_count (_spin_unlock)
  605. sshd-4261 0d.h1 14us : irq_exit (do_IRQ)
  606. sshd-4261 0d.h1 15us : sub_preempt_count (irq_exit)
  607. sshd-4261 0d..2 15us : do_softirq (irq_exit)
  608. sshd-4261 0d... 15us : __do_softirq (do_softirq)
  609. sshd-4261 0d... 16us : __local_bh_disable (__do_softirq)
  610. sshd-4261 0d... 16us+: add_preempt_count (__local_bh_disable)
  611. sshd-4261 0d.s4 20us : add_preempt_count (__local_bh_disable)
  612. sshd-4261 0d.s4 21us : sub_preempt_count (local_bh_enable)
  613. sshd-4261 0d.s5 21us : sub_preempt_count (local_bh_enable)
  614. [...]
  615. sshd-4261 0d.s6 41us : add_preempt_count (__local_bh_disable)
  616. sshd-4261 0d.s6 42us : sub_preempt_count (local_bh_enable)
  617. sshd-4261 0d.s7 42us : sub_preempt_count (local_bh_enable)
  618. sshd-4261 0d.s5 43us : add_preempt_count (__local_bh_disable)
  619. sshd-4261 0d.s5 43us : sub_preempt_count (local_bh_enable_ip)
  620. sshd-4261 0d.s6 44us : sub_preempt_count (local_bh_enable_ip)
  621. sshd-4261 0d.s5 44us : add_preempt_count (__local_bh_disable)
  622. sshd-4261 0d.s5 45us : sub_preempt_count (local_bh_enable)
  623. [...]
  624. sshd-4261 0d.s. 63us : _local_bh_enable (__do_softirq)
  625. sshd-4261 0d.s1 64us : trace_preempt_on (__do_softirq)
  626. The above is an example of the preemptoff trace with
  627. ftrace_enabled set. Here we see that interrupts were disabled
  628. the entire time. The irq_enter code lets us know that we entered
  629. an interrupt 'h'. Before that, the functions being traced still
  630. show that it is not in an interrupt, but we can see from the
  631. functions themselves that this is not the case.
  632. Notice that __do_softirq when called does not have a
  633. preempt_count. It may seem that we missed a preempt enabling.
  634. What really happened is that the preempt count is held on the
  635. thread's stack and we switched to the softirq stack (4K stacks
  636. in effect). The code does not copy the preempt count, but
  637. because interrupts are disabled, we do not need to worry about
  638. it. Having a tracer like this is good for letting people know
  639. what really happens inside the kernel.
  640. preemptirqsoff
  641. --------------
  642. Knowing the locations that have interrupts disabled or
  643. preemption disabled for the longest times is helpful. But
  644. sometimes we would like to know when either preemption and/or
  645. interrupts are disabled.
  646. Consider the following code:
  647. local_irq_disable();
  648. call_function_with_irqs_off();
  649. preempt_disable();
  650. call_function_with_irqs_and_preemption_off();
  651. local_irq_enable();
  652. call_function_with_preemption_off();
  653. preempt_enable();
  654. The irqsoff tracer will record the total length of
  655. call_function_with_irqs_off() and
  656. call_function_with_irqs_and_preemption_off().
  657. The preemptoff tracer will record the total length of
  658. call_function_with_irqs_and_preemption_off() and
  659. call_function_with_preemption_off().
  660. But neither will trace the time that interrupts and/or
  661. preemption is disabled. This total time is the time that we can
  662. not schedule. To record this time, use the preemptirqsoff
  663. tracer.
  664. Again, using this trace is much like the irqsoff and preemptoff
  665. tracers.
  666. # echo preemptirqsoff > current_tracer
  667. # echo latency-format > trace_options
  668. # echo 0 > tracing_max_latency
  669. # echo 1 > tracing_enabled
  670. # ls -ltr
  671. [...]
  672. # echo 0 > tracing_enabled
  673. # cat trace
  674. # tracer: preemptirqsoff
  675. #
  676. preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8
  677. --------------------------------------------------------------------
  678. latency: 293 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  679. -----------------
  680. | task: ls-4860 (uid:0 nice:0 policy:0 rt_prio:0)
  681. -----------------
  682. => started at: apic_timer_interrupt
  683. => ended at: __do_softirq
  684. # _------=> CPU#
  685. # / _-----=> irqs-off
  686. # | / _----=> need-resched
  687. # || / _---=> hardirq/softirq
  688. # ||| / _--=> preempt-depth
  689. # |||| /
  690. # ||||| delay
  691. # cmd pid ||||| time | caller
  692. # \ / ||||| \ | /
  693. ls-4860 0d... 0us!: trace_hardirqs_off_thunk (apic_timer_interrupt)
  694. ls-4860 0d.s. 294us : _local_bh_enable (__do_softirq)
  695. ls-4860 0d.s1 294us : trace_preempt_on (__do_softirq)
  696. The trace_hardirqs_off_thunk is called from assembly on x86 when
  697. interrupts are disabled in the assembly code. Without the
  698. function tracing, we do not know if interrupts were enabled
  699. within the preemption points. We do see that it started with
  700. preemption enabled.
  701. Here is a trace with ftrace_enabled set:
  702. # tracer: preemptirqsoff
  703. #
  704. preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8
  705. --------------------------------------------------------------------
  706. latency: 105 us, #183/183, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  707. -----------------
  708. | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
  709. -----------------
  710. => started at: write_chan
  711. => ended at: __do_softirq
  712. # _------=> CPU#
  713. # / _-----=> irqs-off
  714. # | / _----=> need-resched
  715. # || / _---=> hardirq/softirq
  716. # ||| / _--=> preempt-depth
  717. # |||| /
  718. # ||||| delay
  719. # cmd pid ||||| time | caller
  720. # \ / ||||| \ | /
  721. ls-4473 0.N.. 0us : preempt_schedule (write_chan)
  722. ls-4473 0dN.1 1us : _spin_lock (schedule)
  723. ls-4473 0dN.1 2us : add_preempt_count (_spin_lock)
  724. ls-4473 0d..2 2us : put_prev_task_fair (schedule)
  725. [...]
  726. ls-4473 0d..2 13us : set_normalized_timespec (ktime_get_ts)
  727. ls-4473 0d..2 13us : __switch_to (schedule)
  728. sshd-4261 0d..2 14us : finish_task_switch (schedule)
  729. sshd-4261 0d..2 14us : _spin_unlock_irq (finish_task_switch)
  730. sshd-4261 0d..1 15us : add_preempt_count (_spin_lock_irqsave)
  731. sshd-4261 0d..2 16us : _spin_unlock_irqrestore (hrtick_set)
  732. sshd-4261 0d..2 16us : do_IRQ (common_interrupt)
  733. sshd-4261 0d..2 17us : irq_enter (do_IRQ)
  734. sshd-4261 0d..2 17us : idle_cpu (irq_enter)
  735. sshd-4261 0d..2 18us : add_preempt_count (irq_enter)
  736. sshd-4261 0d.h2 18us : idle_cpu (irq_enter)
  737. sshd-4261 0d.h. 18us : handle_fasteoi_irq (do_IRQ)
  738. sshd-4261 0d.h. 19us : _spin_lock (handle_fasteoi_irq)
  739. sshd-4261 0d.h. 19us : add_preempt_count (_spin_lock)
  740. sshd-4261 0d.h1 20us : _spin_unlock (handle_fasteoi_irq)
  741. sshd-4261 0d.h1 20us : sub_preempt_count (_spin_unlock)
  742. [...]
  743. sshd-4261 0d.h1 28us : _spin_unlock (handle_fasteoi_irq)
  744. sshd-4261 0d.h1 29us : sub_preempt_count (_spin_unlock)
  745. sshd-4261 0d.h2 29us : irq_exit (do_IRQ)
  746. sshd-4261 0d.h2 29us : sub_preempt_count (irq_exit)
  747. sshd-4261 0d..3 30us : do_softirq (irq_exit)
  748. sshd-4261 0d... 30us : __do_softirq (do_softirq)
  749. sshd-4261 0d... 31us : __local_bh_disable (__do_softirq)
  750. sshd-4261 0d... 31us+: add_preempt_count (__local_bh_disable)
  751. sshd-4261 0d.s4 34us : add_preempt_count (__local_bh_disable)
  752. [...]
  753. sshd-4261 0d.s3 43us : sub_preempt_count (local_bh_enable_ip)
  754. sshd-4261 0d.s4 44us : sub_preempt_count (local_bh_enable_ip)
  755. sshd-4261 0d.s3 44us : smp_apic_timer_interrupt (apic_timer_interrupt)
  756. sshd-4261 0d.s3 45us : irq_enter (smp_apic_timer_interrupt)
  757. sshd-4261 0d.s3 45us : idle_cpu (irq_enter)
  758. sshd-4261 0d.s3 46us : add_preempt_count (irq_enter)
  759. sshd-4261 0d.H3 46us : idle_cpu (irq_enter)
  760. sshd-4261 0d.H3 47us : hrtimer_interrupt (smp_apic_timer_interrupt)
  761. sshd-4261 0d.H3 47us : ktime_get (hrtimer_interrupt)
  762. [...]
  763. sshd-4261 0d.H3 81us : tick_program_event (hrtimer_interrupt)
  764. sshd-4261 0d.H3 82us : ktime_get (tick_program_event)
  765. sshd-4261 0d.H3 82us : ktime_get_ts (ktime_get)
  766. sshd-4261 0d.H3 83us : getnstimeofday (ktime_get_ts)
  767. sshd-4261 0d.H3 83us : set_normalized_timespec (ktime_get_ts)
  768. sshd-4261 0d.H3 84us : clockevents_program_event (tick_program_event)
  769. sshd-4261 0d.H3 84us : lapic_next_event (clockevents_program_event)
  770. sshd-4261 0d.H3 85us : irq_exit (smp_apic_timer_interrupt)
  771. sshd-4261 0d.H3 85us : sub_preempt_count (irq_exit)
  772. sshd-4261 0d.s4 86us : sub_preempt_count (irq_exit)
  773. sshd-4261 0d.s3 86us : add_preempt_count (__local_bh_disable)
  774. [...]
  775. sshd-4261 0d.s1 98us : sub_preempt_count (net_rx_action)
  776. sshd-4261 0d.s. 99us : add_preempt_count (_spin_lock_irq)
  777. sshd-4261 0d.s1 99us+: _spin_unlock_irq (run_timer_softirq)
  778. sshd-4261 0d.s. 104us : _local_bh_enable (__do_softirq)
  779. sshd-4261 0d.s. 104us : sub_preempt_count (_local_bh_enable)
  780. sshd-4261 0d.s. 105us : _local_bh_enable (__do_softirq)
  781. sshd-4261 0d.s1 105us : trace_preempt_on (__do_softirq)
  782. This is a very interesting trace. It started with the preemption
  783. of the ls task. We see that the task had the "need_resched" bit
  784. set via the 'N' in the trace. Interrupts were disabled before
  785. the spin_lock at the beginning of the trace. We see that a
  786. schedule took place to run sshd. When the interrupts were
  787. enabled, we took an interrupt. On return from the interrupt
  788. handler, the softirq ran. We took another interrupt while
  789. running the softirq as we see from the capital 'H'.
  790. wakeup
  791. ------
  792. In a Real-Time environment it is very important to know the
  793. wakeup time it takes for the highest priority task that is woken
  794. up to the time that it executes. This is also known as "schedule
  795. latency". I stress the point that this is about RT tasks. It is
  796. also important to know the scheduling latency of non-RT tasks,
  797. but the average schedule latency is better for non-RT tasks.
  798. Tools like LatencyTop are more appropriate for such
  799. measurements.
  800. Real-Time environments are interested in the worst case latency.
  801. That is the longest latency it takes for something to happen,
  802. and not the average. We can have a very fast scheduler that may
  803. only have a large latency once in a while, but that would not
  804. work well with Real-Time tasks. The wakeup tracer was designed
  805. to record the worst case wakeups of RT tasks. Non-RT tasks are
  806. not recorded because the tracer only records one worst case and
  807. tracing non-RT tasks that are unpredictable will overwrite the
  808. worst case latency of RT tasks.
  809. Since this tracer only deals with RT tasks, we will run this
  810. slightly differently than we did with the previous tracers.
  811. Instead of performing an 'ls', we will run 'sleep 1' under
  812. 'chrt' which changes the priority of the task.
  813. # echo wakeup > current_tracer
  814. # echo latency-format > trace_options
  815. # echo 0 > tracing_max_latency
  816. # echo 1 > tracing_enabled
  817. # chrt -f 5 sleep 1
  818. # echo 0 > tracing_enabled
  819. # cat trace
  820. # tracer: wakeup
  821. #
  822. wakeup latency trace v1.1.5 on 2.6.26-rc8
  823. --------------------------------------------------------------------
  824. latency: 4 us, #2/2, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  825. -----------------
  826. | task: sleep-4901 (uid:0 nice:0 policy:1 rt_prio:5)
  827. -----------------
  828. # _------=> CPU#
  829. # / _-----=> irqs-off
  830. # | / _----=> need-resched
  831. # || / _---=> hardirq/softirq
  832. # ||| / _--=> preempt-depth
  833. # |||| /
  834. # ||||| delay
  835. # cmd pid ||||| time | caller
  836. # \ / ||||| \ | /
  837. <idle>-0 1d.h4 0us+: try_to_wake_up (wake_up_process)
  838. <idle>-0 1d..4 4us : schedule (cpu_idle)
  839. Running this on an idle system, we see that it only took 4
  840. microseconds to perform the task switch. Note, since the trace
  841. marker in the schedule is before the actual "switch", we stop
  842. the tracing when the recorded task is about to schedule in. This
  843. may change if we add a new marker at the end of the scheduler.
  844. Notice that the recorded task is 'sleep' with the PID of 4901
  845. and it has an rt_prio of 5. This priority is user-space priority
  846. and not the internal kernel priority. The policy is 1 for
  847. SCHED_FIFO and 2 for SCHED_RR.
  848. Doing the same with chrt -r 5 and ftrace_enabled set.
  849. # tracer: wakeup
  850. #
  851. wakeup latency trace v1.1.5 on 2.6.26-rc8
  852. --------------------------------------------------------------------
  853. latency: 50 us, #60/60, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
  854. -----------------
  855. | task: sleep-4068 (uid:0 nice:0 policy:2 rt_prio:5)
  856. -----------------
  857. # _------=> CPU#
  858. # / _-----=> irqs-off
  859. # | / _----=> need-resched
  860. # || / _---=> hardirq/softirq
  861. # ||| / _--=> preempt-depth
  862. # |||| /
  863. # ||||| delay
  864. # cmd pid ||||| time | caller
  865. # \ / ||||| \ | /
  866. ksoftirq-7 1d.H3 0us : try_to_wake_up (wake_up_process)
  867. ksoftirq-7 1d.H4 1us : sub_preempt_count (marker_probe_cb)
  868. ksoftirq-7 1d.H3 2us : check_preempt_wakeup (try_to_wake_up)
  869. ksoftirq-7 1d.H3 3us : update_curr (check_preempt_wakeup)
  870. ksoftirq-7 1d.H3 4us : calc_delta_mine (update_curr)
  871. ksoftirq-7 1d.H3 5us : __resched_task (check_preempt_wakeup)
  872. ksoftirq-7 1d.H3 6us : task_wake_up_rt (try_to_wake_up)
  873. ksoftirq-7 1d.H3 7us : _spin_unlock_irqrestore (try_to_wake_up)
  874. [...]
  875. ksoftirq-7 1d.H2 17us : irq_exit (smp_apic_timer_interrupt)
  876. ksoftirq-7 1d.H2 18us : sub_preempt_count (irq_exit)
  877. ksoftirq-7 1d.s3 19us : sub_preempt_count (irq_exit)
  878. ksoftirq-7 1..s2 20us : rcu_process_callbacks (__do_softirq)
  879. [...]
  880. ksoftirq-7 1..s2 26us : __rcu_process_callbacks (rcu_process_callbacks)
  881. ksoftirq-7 1d.s2 27us : _local_bh_enable (__do_softirq)
  882. ksoftirq-7 1d.s2 28us : sub_preempt_count (_local_bh_enable)
  883. ksoftirq-7 1.N.3 29us : sub_preempt_count (ksoftirqd)
  884. ksoftirq-7 1.N.2 30us : _cond_resched (ksoftirqd)
  885. ksoftirq-7 1.N.2 31us : __cond_resched (_cond_resched)
  886. ksoftirq-7 1.N.2 32us : add_preempt_count (__cond_resched)
  887. ksoftirq-7 1.N.2 33us : schedule (__cond_resched)
  888. ksoftirq-7 1.N.2 33us : add_preempt_count (schedule)
  889. ksoftirq-7 1.N.3 34us : hrtick_clear (schedule)
  890. ksoftirq-7 1dN.3 35us : _spin_lock (schedule)
  891. ksoftirq-7 1dN.3 36us : add_preempt_count (_spin_lock)
  892. ksoftirq-7 1d..4 37us : put_prev_task_fair (schedule)
  893. ksoftirq-7 1d..4 38us : update_curr (put_prev_task_fair)
  894. [...]
  895. ksoftirq-7 1d..5 47us : _spin_trylock (tracing_record_cmdline)
  896. ksoftirq-7 1d..5 48us : add_preempt_count (_spin_trylock)
  897. ksoftirq-7 1d..6 49us : _spin_unlock (tracing_record_cmdline)
  898. ksoftirq-7 1d..6 49us : sub_preempt_count (_spin_unlock)
  899. ksoftirq-7 1d..4 50us : schedule (__cond_resched)
  900. The interrupt went off while running ksoftirqd. This task runs
  901. at SCHED_OTHER. Why did not we see the 'N' set early? This may
  902. be a harmless bug with x86_32 and 4K stacks. On x86_32 with 4K
  903. stacks configured, the interrupt and softirq run with their own
  904. stack. Some information is held on the top of the task's stack
  905. (need_resched and preempt_count are both stored there). The
  906. setting of the NEED_RESCHED bit is done directly to the task's
  907. stack, but the reading of the NEED_RESCHED is done by looking at
  908. the current stack, which in this case is the stack for the hard
  909. interrupt. This hides the fact that NEED_RESCHED has been set.
  910. We do not see the 'N' until we switch back to the task's
  911. assigned stack.
  912. function
  913. --------
  914. This tracer is the function tracer. Enabling the function tracer
  915. can be done from the debug file system. Make sure the
  916. ftrace_enabled is set; otherwise this tracer is a nop.
  917. # sysctl kernel.ftrace_enabled=1
  918. # echo function > current_tracer
  919. # echo 1 > tracing_enabled
  920. # usleep 1
  921. # echo 0 > tracing_enabled
  922. # cat trace
  923. # tracer: function
  924. #
  925. # TASK-PID CPU# TIMESTAMP FUNCTION
  926. # | | | | |
  927. bash-4003 [00] 123.638713: finish_task_switch <-schedule
  928. bash-4003 [00] 123.638714: _spin_unlock_irq <-finish_task_switch
  929. bash-4003 [00] 123.638714: sub_preempt_count <-_spin_unlock_irq
  930. bash-4003 [00] 123.638715: hrtick_set <-schedule
  931. bash-4003 [00] 123.638715: _spin_lock_irqsave <-hrtick_set
  932. bash-4003 [00] 123.638716: add_preempt_count <-_spin_lock_irqsave
  933. bash-4003 [00] 123.638716: _spin_unlock_irqrestore <-hrtick_set
  934. bash-4003 [00] 123.638717: sub_preempt_count <-_spin_unlock_irqrestore
  935. bash-4003 [00] 123.638717: hrtick_clear <-hrtick_set
  936. bash-4003 [00] 123.638718: sub_preempt_count <-schedule
  937. bash-4003 [00] 123.638718: sub_preempt_count <-preempt_schedule
  938. bash-4003 [00] 123.638719: wait_for_completion <-__stop_machine_run
  939. bash-4003 [00] 123.638719: wait_for_common <-wait_for_completion
  940. bash-4003 [00] 123.638720: _spin_lock_irq <-wait_for_common
  941. bash-4003 [00] 123.638720: add_preempt_count <-_spin_lock_irq
  942. [...]
  943. Note: function tracer uses ring buffers to store the above
  944. entries. The newest data may overwrite the oldest data.
  945. Sometimes using echo to stop the trace is not sufficient because
  946. the tracing could have overwritten the data that you wanted to
  947. record. For this reason, it is sometimes better to disable
  948. tracing directly from a program. This allows you to stop the
  949. tracing at the point that you hit the part that you are
  950. interested in. To disable the tracing directly from a C program,
  951. something like following code snippet can be used:
  952. int trace_fd;
  953. [...]
  954. int main(int argc, char *argv[]) {
  955. [...]
  956. trace_fd = open(tracing_file("tracing_enabled"), O_WRONLY);
  957. [...]
  958. if (condition_hit()) {
  959. write(trace_fd, "0", 1);
  960. }
  961. [...]
  962. }
  963. Single thread tracing
  964. ---------------------
  965. By writing into set_ftrace_pid you can trace a
  966. single thread. For example:
  967. # cat set_ftrace_pid
  968. no pid
  969. # echo 3111 > set_ftrace_pid
  970. # cat set_ftrace_pid
  971. 3111
  972. # echo function > current_tracer
  973. # cat trace | head
  974. # tracer: function
  975. #
  976. # TASK-PID CPU# TIMESTAMP FUNCTION
  977. # | | | | |
  978. yum-updatesd-3111 [003] 1637.254676: finish_task_switch <-thread_return
  979. yum-updatesd-3111 [003] 1637.254681: hrtimer_cancel <-schedule_hrtimeout_range
  980. yum-updatesd-3111 [003] 1637.254682: hrtimer_try_to_cancel <-hrtimer_cancel
  981. yum-updatesd-3111 [003] 1637.254683: lock_hrtimer_base <-hrtimer_try_to_cancel
  982. yum-updatesd-3111 [003] 1637.254685: fget_light <-do_sys_poll
  983. yum-updatesd-3111 [003] 1637.254686: pipe_poll <-do_sys_poll
  984. # echo -1 > set_ftrace_pid
  985. # cat trace |head
  986. # tracer: function
  987. #
  988. # TASK-PID CPU# TIMESTAMP FUNCTION
  989. # | | | | |
  990. ##### CPU 3 buffer started ####
  991. yum-updatesd-3111 [003] 1701.957688: free_poll_entry <-poll_freewait
  992. yum-updatesd-3111 [003] 1701.957689: remove_wait_queue <-free_poll_entry
  993. yum-updatesd-3111 [003] 1701.957691: fput <-free_poll_entry
  994. yum-updatesd-3111 [003] 1701.957692: audit_syscall_exit <-sysret_audit
  995. yum-updatesd-3111 [003] 1701.957693: path_put <-audit_syscall_exit
  996. If you want to trace a function when executing, you could use
  997. something like this simple program:
  998. #include <stdio.h>
  999. #include <stdlib.h>
  1000. #include <sys/types.h>
  1001. #include <sys/stat.h>
  1002. #include <fcntl.h>
  1003. #include <unistd.h>
  1004. #include <string.h>
  1005. #define _STR(x) #x
  1006. #define STR(x) _STR(x)
  1007. #define MAX_PATH 256
  1008. const char *find_debugfs(void)
  1009. {
  1010. static char debugfs[MAX_PATH+1];
  1011. static int debugfs_found;
  1012. char type[100];
  1013. FILE *fp;
  1014. if (debugfs_found)
  1015. return debugfs;
  1016. if ((fp = fopen("/proc/mounts","r")) == NULL) {
  1017. perror("/proc/mounts");
  1018. return NULL;
  1019. }
  1020. while (fscanf(fp, "%*s %"
  1021. STR(MAX_PATH)
  1022. "s %99s %*s %*d %*d\n",
  1023. debugfs, type) == 2) {
  1024. if (strcmp(type, "debugfs") == 0)
  1025. break;
  1026. }
  1027. fclose(fp);
  1028. if (strcmp(type, "debugfs") != 0) {
  1029. fprintf(stderr, "debugfs not mounted");
  1030. return NULL;
  1031. }
  1032. strcat(debugfs, "/tracing/");
  1033. debugfs_found = 1;
  1034. return debugfs;
  1035. }
  1036. const char *tracing_file(const char *file_name)
  1037. {
  1038. static char trace_file[MAX_PATH+1];
  1039. snprintf(trace_file, MAX_PATH, "%s/%s", find_debugfs(), file_name);
  1040. return trace_file;
  1041. }
  1042. int main (int argc, char **argv)
  1043. {
  1044. if (argc < 1)
  1045. exit(-1);
  1046. if (fork() > 0) {
  1047. int fd, ffd;
  1048. char line[64];
  1049. int s;
  1050. ffd = open(tracing_file("current_tracer"), O_WRONLY);
  1051. if (ffd < 0)
  1052. exit(-1);
  1053. write(ffd, "nop", 3);
  1054. fd = open(tracing_file("set_ftrace_pid"), O_WRONLY);
  1055. s = sprintf(line, "%d\n", getpid());
  1056. write(fd, line, s);
  1057. write(ffd, "function", 8);
  1058. close(fd);
  1059. close(ffd);
  1060. execvp(argv[1], argv+1);
  1061. }
  1062. return 0;
  1063. }
  1064. hw-branch-tracer (x86 only)
  1065. ---------------------------
  1066. This tracer uses the x86 last branch tracing hardware feature to
  1067. collect a branch trace on all cpus with relatively low overhead.
  1068. The tracer uses a fixed-size circular buffer per cpu and only
  1069. traces ring 0 branches. The trace file dumps that buffer in the
  1070. following format:
  1071. # tracer: hw-branch-tracer
  1072. #
  1073. # CPU# TO <- FROM
  1074. 0 scheduler_tick+0xb5/0x1bf <- task_tick_idle+0x5/0x6
  1075. 2 run_posix_cpu_timers+0x2b/0x72a <- run_posix_cpu_timers+0x25/0x72a
  1076. 0 scheduler_tick+0x139/0x1bf <- scheduler_tick+0xed/0x1bf
  1077. 0 scheduler_tick+0x17c/0x1bf <- scheduler_tick+0x148/0x1bf
  1078. 2 run_posix_cpu_timers+0x9e/0x72a <- run_posix_cpu_timers+0x5e/0x72a
  1079. 0 scheduler_tick+0x1b6/0x1bf <- scheduler_tick+0x1aa/0x1bf
  1080. The tracer may be used to dump the trace for the oops'ing cpu on
  1081. a kernel oops into the system log. To enable this,
  1082. ftrace_dump_on_oops must be set. To set ftrace_dump_on_oops, one
  1083. can either use the sysctl function or set it via the proc system
  1084. interface.
  1085. sysctl kernel.ftrace_dump_on_oops=n
  1086. or
  1087. echo n > /proc/sys/kernel/ftrace_dump_on_oops
  1088. If n = 1, ftrace will dump buffers of all CPUs, if n = 2 ftrace will
  1089. only dump the buffer of the CPU that triggered the oops.
  1090. Here's an example of such a dump after a null pointer
  1091. dereference in a kernel module:
  1092. [57848.105921] BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
  1093. [57848.106019] IP: [<ffffffffa0000006>] open+0x6/0x14 [oops]
  1094. [57848.106019] PGD 2354e9067 PUD 2375e7067 PMD 0
  1095. [57848.106019] Oops: 0002 [#1] SMP
  1096. [57848.106019] last sysfs file: /sys/devices/pci0000:00/0000:00:1e.0/0000:20:05.0/local_cpus
  1097. [57848.106019] Dumping ftrace buffer:
  1098. [57848.106019] ---------------------------------
  1099. [...]
  1100. [57848.106019] 0 chrdev_open+0xe6/0x165 <- cdev_put+0x23/0x24
  1101. [57848.106019] 0 chrdev_open+0x117/0x165 <- chrdev_open+0xfa/0x165
  1102. [57848.106019] 0 chrdev_open+0x120/0x165 <- chrdev_open+0x11c/0x165
  1103. [57848.106019] 0 chrdev_open+0x134/0x165 <- chrdev_open+0x12b/0x165
  1104. [57848.106019] 0 open+0x0/0x14 [oops] <- chrdev_open+0x144/0x165
  1105. [57848.106019] 0 page_fault+0x0/0x30 <- open+0x6/0x14 [oops]
  1106. [57848.106019] 0 error_entry+0x0/0x5b <- page_fault+0x4/0x30
  1107. [57848.106019] 0 error_kernelspace+0x0/0x31 <- error_entry+0x59/0x5b
  1108. [57848.106019] 0 error_sti+0x0/0x1 <- error_kernelspace+0x2d/0x31
  1109. [57848.106019] 0 page_fault+0x9/0x30 <- error_sti+0x0/0x1
  1110. [57848.106019] 0 do_page_fault+0x0/0x881 <- page_fault+0x1a/0x30
  1111. [...]
  1112. [57848.106019] 0 do_page_fault+0x66b/0x881 <- is_prefetch+0x1ee/0x1f2
  1113. [57848.106019] 0 do_page_fault+0x6e0/0x881 <- do_page_fault+0x67a/0x881
  1114. [57848.106019] 0 oops_begin+0x0/0x96 <- do_page_fault+0x6e0/0x881
  1115. [57848.106019] 0 trace_hw_branch_oops+0x0/0x2d <- oops_begin+0x9/0x96
  1116. [...]
  1117. [57848.106019] 0 ds_suspend_bts+0x2a/0xe3 <- ds_suspend_bts+0x1a/0xe3
  1118. [57848.106019] ---------------------------------
  1119. [57848.106019] CPU 0
  1120. [57848.106019] Modules linked in: oops
  1121. [57848.106019] Pid: 5542, comm: cat Tainted: G W 2.6.28 #23
  1122. [57848.106019] RIP: 0010:[<ffffffffa0000006>] [<ffffffffa0000006>] open+0x6/0x14 [oops]
  1123. [57848.106019] RSP: 0018:ffff880235457d48 EFLAGS: 00010246
  1124. [...]
  1125. function graph tracer
  1126. ---------------------------
  1127. This tracer is similar to the function tracer except that it
  1128. probes a function on its entry and its exit. This is done by
  1129. using a dynamically allocated stack of return addresses in each
  1130. task_struct. On function entry the tracer overwrites the return
  1131. address of each function traced to set a custom probe. Thus the
  1132. original return address is stored on the stack of return address
  1133. in the task_struct.
  1134. Probing on both ends of a function leads to special features
  1135. such as:
  1136. - measure of a function's time execution
  1137. - having a reliable call stack to draw function calls graph
  1138. This tracer is useful in several situations:
  1139. - you want to find the reason of a strange kernel behavior and
  1140. need to see what happens in detail on any areas (or specific
  1141. ones).
  1142. - you are experiencing weird latencies but it's difficult to
  1143. find its origin.
  1144. - you want to find quickly which path is taken by a specific
  1145. function
  1146. - you just want to peek inside a working kernel and want to see
  1147. what happens there.
  1148. # tracer: function_graph
  1149. #
  1150. # CPU DURATION FUNCTION CALLS
  1151. # | | | | | | |
  1152. 0) | sys_open() {
  1153. 0) | do_sys_open() {
  1154. 0) | getname() {
  1155. 0) | kmem_cache_alloc() {
  1156. 0) 1.382 us | __might_sleep();
  1157. 0) 2.478 us | }
  1158. 0) | strncpy_from_user() {
  1159. 0) | might_fault() {
  1160. 0) 1.389 us | __might_sleep();
  1161. 0) 2.553 us | }
  1162. 0) 3.807 us | }
  1163. 0) 7.876 us | }
  1164. 0) | alloc_fd() {
  1165. 0) 0.668 us | _spin_lock();
  1166. 0) 0.570 us | expand_files();
  1167. 0) 0.586 us | _spin_unlock();
  1168. There are several columns that can be dynamically
  1169. enabled/disabled. You can use every combination of options you
  1170. want, depending on your needs.
  1171. - The cpu number on which the function executed is default
  1172. enabled. It is sometimes better to only trace one cpu (see
  1173. tracing_cpu_mask file) or you might sometimes see unordered
  1174. function calls while cpu tracing switch.
  1175. hide: echo nofuncgraph-cpu > trace_options
  1176. show: echo funcgraph-cpu > trace_options
  1177. - The duration (function's time of execution) is displayed on
  1178. the closing bracket line of a function or on the same line
  1179. than the current function in case of a leaf one. It is default
  1180. enabled.
  1181. hide: echo nofuncgraph-duration > trace_options
  1182. show: echo funcgraph-duration > trace_options
  1183. - The overhead field precedes the duration field in case of
  1184. reached duration thresholds.
  1185. hide: echo nofuncgraph-overhead > trace_options
  1186. show: echo funcgraph-overhead > trace_options
  1187. depends on: funcgraph-duration
  1188. ie:
  1189. 0) | up_write() {
  1190. 0) 0.646 us | _spin_lock_irqsave();
  1191. 0) 0.684 us | _spin_unlock_irqrestore();
  1192. 0) 3.123 us | }
  1193. 0) 0.548 us | fput();
  1194. 0) + 58.628 us | }
  1195. [...]
  1196. 0) | putname() {
  1197. 0) | kmem_cache_free() {
  1198. 0) 0.518 us | __phys_addr();
  1199. 0) 1.757 us | }
  1200. 0) 2.861 us | }
  1201. 0) ! 115.305 us | }
  1202. 0) ! 116.402 us | }
  1203. + means that the function exceeded 10 usecs.
  1204. ! means that the function exceeded 100 usecs.
  1205. - The task/pid field displays the thread cmdline and pid which
  1206. executed the function. It is default disabled.
  1207. hide: echo nofuncgraph-proc > trace_options
  1208. show: echo funcgraph-proc > trace_options
  1209. ie:
  1210. # tracer: function_graph
  1211. #
  1212. # CPU TASK/PID DURATION FUNCTION CALLS
  1213. # | | | | | | | | |
  1214. 0) sh-4802 | | d_free() {
  1215. 0) sh-4802 | | call_rcu() {
  1216. 0) sh-4802 | | __call_rcu() {
  1217. 0) sh-4802 | 0.616 us | rcu_process_gp_end();
  1218. 0) sh-4802 | 0.586 us | check_for_new_grace_period();
  1219. 0) sh-4802 | 2.899 us | }
  1220. 0) sh-4802 | 4.040 us | }
  1221. 0) sh-4802 | 5.151 us | }
  1222. 0) sh-4802 | + 49.370 us | }
  1223. - The absolute time field is an absolute timestamp given by the
  1224. system clock since it started. A snapshot of this time is
  1225. given on each entry/exit of functions
  1226. hide: echo nofuncgraph-abstime > trace_options
  1227. show: echo funcgraph-abstime > trace_options
  1228. ie:
  1229. #
  1230. # TIME CPU DURATION FUNCTION CALLS
  1231. # | | | | | | | |
  1232. 360.774522 | 1) 0.541 us | }
  1233. 360.774522 | 1) 4.663 us | }
  1234. 360.774523 | 1) 0.541 us | __wake_up_bit();
  1235. 360.774524 | 1) 6.796 us | }
  1236. 360.774524 | 1) 7.952 us | }
  1237. 360.774525 | 1) 9.063 us | }
  1238. 360.774525 | 1) 0.615 us | journal_mark_dirty();
  1239. 360.774527 | 1) 0.578 us | __brelse();
  1240. 360.774528 | 1) | reiserfs_prepare_for_journal() {
  1241. 360.774528 | 1) | unlock_buffer() {
  1242. 360.774529 | 1) | wake_up_bit() {
  1243. 360.774529 | 1) | bit_waitqueue() {
  1244. 360.774530 | 1) 0.594 us | __phys_addr();
  1245. You can put some comments on specific functions by using
  1246. trace_printk() For example, if you want to put a comment inside
  1247. the __might_sleep() function, you just have to include
  1248. <linux/ftrace.h> and call trace_printk() inside __might_sleep()
  1249. trace_printk("I'm a comment!\n")
  1250. will produce:
  1251. 1) | __might_sleep() {
  1252. 1) | /* I'm a comment! */
  1253. 1) 1.449 us | }
  1254. You might find other useful features for this tracer in the
  1255. following "dynamic ftrace" section such as tracing only specific
  1256. functions or tasks.
  1257. dynamic ftrace
  1258. --------------
  1259. If CONFIG_DYNAMIC_FTRACE is set, the system will run with
  1260. virtually no overhead when function tracing is disabled. The way
  1261. this works is the mcount function call (placed at the start of
  1262. every kernel function, produced by the -pg switch in gcc),
  1263. starts of pointing to a simple return. (Enabling FTRACE will
  1264. include the -pg switch in the compiling of the kernel.)
  1265. At compile time every C file object is run through the
  1266. recordmcount.pl script (located in the scripts directory). This
  1267. script will process the C object using objdump to find all the
  1268. locations in the .text section that call mcount. (Note, only the
  1269. .text section is processed, since processing other sections like
  1270. .init.text may cause races due to those sections being freed).
  1271. A new section called "__mcount_loc" is created that holds
  1272. references to all the mcount call sites in the .text section.
  1273. This section is compiled back into the original object. The
  1274. final linker will add all these references into a single table.
  1275. On boot up, before SMP is initialized, the dynamic ftrace code
  1276. scans this table and updates all the locations into nops. It
  1277. also records the locations, which are added to the
  1278. available_filter_functions list. Modules are processed as they
  1279. are loaded and before they are executed. When a module is
  1280. unloaded, it also removes its functions from the ftrace function
  1281. list. This is automatic in the module unload code, and the
  1282. module author does not need to worry about it.
  1283. When tracing is enabled, kstop_machine is called to prevent
  1284. races with the CPUS executing code being modified (which can
  1285. cause the CPU to do undesirable things), and the nops are
  1286. patched back to calls. But this time, they do not call mcount
  1287. (which is just a function stub). They now call into the ftrace
  1288. infrastructure.
  1289. One special side-effect to the recording of the functions being
  1290. traced is that we can now selectively choose which functions we
  1291. wish to trace and which ones we want the mcount calls to remain
  1292. as nops.
  1293. Two files are used, one for enabling and one for disabling the
  1294. tracing of specified functions. They are:
  1295. set_ftrace_filter
  1296. and
  1297. set_ftrace_notrace
  1298. A list of available functions that you can add to these files is
  1299. listed in:
  1300. available_filter_functions
  1301. # cat available_filter_functions
  1302. put_prev_task_idle
  1303. kmem_cache_create
  1304. pick_next_task_rt
  1305. get_online_cpus
  1306. pick_next_task_fair
  1307. mutex_lock
  1308. [...]
  1309. If I am only interested in sys_nanosleep and hrtimer_interrupt:
  1310. # echo sys_nanosleep hrtimer_interrupt \
  1311. > set_ftrace_filter
  1312. # echo function > current_tracer
  1313. # echo 1 > tracing_enabled
  1314. # usleep 1
  1315. # echo 0 > tracing_enabled
  1316. # cat trace
  1317. # tracer: ftrace
  1318. #
  1319. # TASK-PID CPU# TIMESTAMP FUNCTION
  1320. # | | | | |
  1321. usleep-4134 [00] 1317.070017: hrtimer_interrupt <-smp_apic_timer_interrupt
  1322. usleep-4134 [00] 1317.070111: sys_nanosleep <-syscall_call
  1323. <idle>-0 [00] 1317.070115: hrtimer_interrupt <-smp_apic_timer_interrupt
  1324. To see which functions are being traced, you can cat the file:
  1325. # cat set_ftrace_filter
  1326. hrtimer_interrupt
  1327. sys_nanosleep
  1328. Perhaps this is not enough. The filters also allow simple wild
  1329. cards. Only the following are currently available
  1330. <match>* - will match functions that begin with <match>
  1331. *<match> - will match functions that end with <match>
  1332. *<match>* - will match functions that have <match> in it
  1333. These are the only wild cards which are supported.
  1334. <match>*<match> will not work.
  1335. Note: It is better to use quotes to enclose the wild cards,
  1336. otherwise the shell may expand the parameters into names
  1337. of files in the local directory.
  1338. # echo 'hrtimer_*' > set_ftrace_filter
  1339. Produces:
  1340. # tracer: ftrace
  1341. #
  1342. # TASK-PID CPU# TIMESTAMP FUNCTION
  1343. # | | | | |
  1344. bash-4003 [00] 1480.611794: hrtimer_init <-copy_process
  1345. bash-4003 [00] 1480.611941: hrtimer_start <-hrtick_set
  1346. bash-4003 [00] 1480.611956: hrtimer_cancel <-hrtick_clear
  1347. bash-4003 [00] 1480.611956: hrtimer_try_to_cancel <-hrtimer_cancel
  1348. <idle>-0 [00] 1480.612019: hrtimer_get_next_event <-get_next_timer_interrupt
  1349. <idle>-0 [00] 1480.612025: hrtimer_get_next_event <-get_next_timer_interrupt
  1350. <idle>-0 [00] 1480.612032: hrtimer_get_next_event <-get_next_timer_interrupt
  1351. <idle>-0 [00] 1480.612037: hrtimer_get_next_event <-get_next_timer_interrupt
  1352. <idle>-0 [00] 1480.612382: hrtimer_get_next_event <-get_next_timer_interrupt
  1353. Notice that we lost the sys_nanosleep.
  1354. # cat set_ftrace_filter
  1355. hrtimer_run_queues
  1356. hrtimer_run_pending
  1357. hrtimer_init
  1358. hrtimer_cancel
  1359. hrtimer_try_to_cancel
  1360. hrtimer_forward
  1361. hrtimer_start
  1362. hrtimer_reprogram
  1363. hrtimer_force_reprogram
  1364. hrtimer_get_next_event
  1365. hrtimer_interrupt
  1366. hrtimer_nanosleep
  1367. hrtimer_wakeup
  1368. hrtimer_get_remaining
  1369. hrtimer_get_res
  1370. hrtimer_init_sleeper
  1371. This is because the '>' and '>>' act just like they do in bash.
  1372. To rewrite the filters, use '>'
  1373. To append to the filters, use '>>'
  1374. To clear out a filter so that all functions will be recorded
  1375. again:
  1376. # echo > set_ftrace_filter
  1377. # cat set_ftrace_filter
  1378. #
  1379. Again, now we want to append.
  1380. # echo sys_nanosleep > set_ftrace_filter
  1381. # cat set_ftrace_filter
  1382. sys_nanosleep
  1383. # echo 'hrtimer_*' >> set_ftrace_filter
  1384. # cat set_ftrace_filter
  1385. hrtimer_run_queues
  1386. hrtimer_run_pending
  1387. hrtimer_init
  1388. hrtimer_cancel
  1389. hrtimer_try_to_cancel
  1390. hrtimer_forward
  1391. hrtimer_start
  1392. hrtimer_reprogram
  1393. hrtimer_force_reprogram
  1394. hrtimer_get_next_event
  1395. hrtimer_interrupt
  1396. sys_nanosleep
  1397. hrtimer_nanosleep
  1398. hrtimer_wakeup
  1399. hrtimer_get_remaining
  1400. hrtimer_get_res
  1401. hrtimer_init_sleeper
  1402. The set_ftrace_notrace prevents those functions from being
  1403. traced.
  1404. # echo '*preempt*' '*lock*' > set_ftrace_notrace
  1405. Produces:
  1406. # tracer: ftrace
  1407. #
  1408. # TASK-PID CPU# TIMESTAMP FUNCTION
  1409. # | | | | |
  1410. bash-4043 [01] 115.281644: finish_task_switch <-schedule
  1411. bash-4043 [01] 115.281645: hrtick_set <-schedule
  1412. bash-4043 [01] 115.281645: hrtick_clear <-hrtick_set
  1413. bash-4043 [01] 115.281646: wait_for_completion <-__stop_machine_run
  1414. bash-4043 [01] 115.281647: wait_for_common <-wait_for_completion
  1415. bash-4043 [01] 115.281647: kthread_stop <-stop_machine_run
  1416. bash-4043 [01] 115.281648: init_waitqueue_head <-kthread_stop
  1417. bash-4043 [01] 115.281648: wake_up_process <-kthread_stop
  1418. bash-4043 [01] 115.281649: try_to_wake_up <-wake_up_process
  1419. We can see that there's no more lock or preempt tracing.
  1420. Dynamic ftrace with the function graph tracer
  1421. ---------------------------------------------
  1422. Although what has been explained above concerns both the
  1423. function tracer and the function-graph-tracer, there are some
  1424. special features only available in the function-graph tracer.
  1425. If you want to trace only one function and all of its children,
  1426. you just have to echo its name into set_graph_function:
  1427. echo __do_fault > set_graph_function
  1428. will produce the following "expanded" trace of the __do_fault()
  1429. function:
  1430. 0) | __do_fault() {
  1431. 0) | filemap_fault() {
  1432. 0) | find_lock_page() {
  1433. 0) 0.804 us | find_get_page();
  1434. 0) | __might_sleep() {
  1435. 0) 1.329 us | }
  1436. 0) 3.904 us | }
  1437. 0) 4.979 us | }
  1438. 0) 0.653 us | _spin_lock();
  1439. 0) 0.578 us | page_add_file_rmap();
  1440. 0) 0.525 us | native_set_pte_at();
  1441. 0) 0.585 us | _spin_unlock();
  1442. 0) | unlock_page() {
  1443. 0) 0.541 us | page_waitqueue();
  1444. 0) 0.639 us | __wake_up_bit();
  1445. 0) 2.786 us | }
  1446. 0) + 14.237 us | }
  1447. 0) | __do_fault() {
  1448. 0) | filemap_fault() {
  1449. 0) | find_lock_page() {
  1450. 0) 0.698 us | find_get_page();
  1451. 0) | __might_sleep() {
  1452. 0) 1.412 us | }
  1453. 0) 3.950 us | }
  1454. 0) 5.098 us | }
  1455. 0) 0.631 us | _spin_lock();
  1456. 0) 0.571 us | page_add_file_rmap();
  1457. 0) 0.526 us | native_set_pte_at();
  1458. 0) 0.586 us | _spin_unlock();
  1459. 0) | unlock_page() {
  1460. 0) 0.533 us | page_waitqueue();
  1461. 0) 0.638 us | __wake_up_bit();
  1462. 0) 2.793 us | }
  1463. 0) + 14.012 us | }
  1464. You can also expand several functions at once:
  1465. echo sys_open > set_graph_function
  1466. echo sys_close >> set_graph_function
  1467. Now if you want to go back to trace all functions you can clear
  1468. this special filter via:
  1469. echo > set_graph_function
  1470. Filter commands
  1471. ---------------
  1472. A few commands are supported by the set_ftrace_filter interface.
  1473. Trace commands have the following format:
  1474. <function>:<command>:<parameter>
  1475. The following commands are supported:
  1476. - mod
  1477. This command enables function filtering per module. The
  1478. parameter defines the module. For example, if only the write*
  1479. functions in the ext3 module are desired, run:
  1480. echo 'write*:mod:ext3' > set_ftrace_filter
  1481. This command interacts with the filter in the same way as
  1482. filtering based on function names. Thus, adding more functions
  1483. in a different module is accomplished by appending (>>) to the
  1484. filter file. Remove specific module functions by prepending
  1485. '!':
  1486. echo '!writeback*:mod:ext3' >> set_ftrace_filter
  1487. - traceon/traceoff
  1488. These commands turn tracing on and off when the specified
  1489. functions are hit. The parameter determines how many times the
  1490. tracing system is turned on and off. If unspecified, there is
  1491. no limit. For example, to disable tracing when a schedule bug
  1492. is hit the first 5 times, run:
  1493. echo '__schedule_bug:traceoff:5' > set_ftrace_filter
  1494. These commands are cumulative whether or not they are appended
  1495. to set_ftrace_filter. To remove a command, prepend it by '!'
  1496. and drop the parameter:
  1497. echo '!__schedule_bug:traceoff' > set_ftrace_filter
  1498. trace_pipe
  1499. ----------
  1500. The trace_pipe outputs the same content as the trace file, but
  1501. the effect on the tracing is different. Every read from
  1502. trace_pipe is consumed. This means that subsequent reads will be
  1503. different. The trace is live.
  1504. # echo function > current_tracer
  1505. # cat trace_pipe > /tmp/trace.out &
  1506. [1] 4153
  1507. # echo 1 > tracing_enabled
  1508. # usleep 1
  1509. # echo 0 > tracing_enabled
  1510. # cat trace
  1511. # tracer: function
  1512. #
  1513. # TASK-PID CPU# TIMESTAMP FUNCTION
  1514. # | | | | |
  1515. #
  1516. # cat /tmp/trace.out
  1517. bash-4043 [00] 41.267106: finish_task_switch <-schedule
  1518. bash-4043 [00] 41.267106: hrtick_set <-schedule
  1519. bash-4043 [00] 41.267107: hrtick_clear <-hrtick_set
  1520. bash-4043 [00] 41.267108: wait_for_completion <-__stop_machine_run
  1521. bash-4043 [00] 41.267108: wait_for_common <-wait_for_completion
  1522. bash-4043 [00] 41.267109: kthread_stop <-stop_machine_run
  1523. bash-4043 [00] 41.267109: init_waitqueue_head <-kthread_stop
  1524. bash-4043 [00] 41.267110: wake_up_process <-kthread_stop
  1525. bash-4043 [00] 41.267110: try_to_wake_up <-wake_up_process
  1526. bash-4043 [00] 41.267111: select_task_rq_rt <-try_to_wake_up
  1527. Note, reading the trace_pipe file will block until more input is
  1528. added. By changing the tracer, trace_pipe will issue an EOF. We
  1529. needed to set the function tracer _before_ we "cat" the
  1530. trace_pipe file.
  1531. trace entries
  1532. -------------
  1533. Having too much or not enough data can be troublesome in
  1534. diagnosing an issue in the kernel. The file buffer_size_kb is
  1535. used to modify the size of the internal trace buffers. The
  1536. number listed is the number of entries that can be recorded per
  1537. CPU. To know the full size, multiply the number of possible CPUS
  1538. with the number of entries.
  1539. # cat buffer_size_kb
  1540. 1408 (units kilobytes)
  1541. Note, to modify this, you must have tracing completely disabled.
  1542. To do that, echo "nop" into the current_tracer. If the
  1543. current_tracer is not set to "nop", an EINVAL error will be
  1544. returned.
  1545. # echo nop > current_tracer
  1546. # echo 10000 > buffer_size_kb
  1547. # cat buffer_size_kb
  1548. 10000 (units kilobytes)
  1549. The number of pages which will be allocated is limited to a
  1550. percentage of available memory. Allocating too much will produce
  1551. an error.
  1552. # echo 1000000000000 > buffer_size_kb
  1553. -bash: echo: write error: Cannot allocate memory
  1554. # cat buffer_size_kb
  1555. 85
  1556. -----------
  1557. More details can be found in the source code, in the
  1558. kernel/trace/*.c files.