Commit | Line | Data |
---|---|---|
eb6d42ea SR |
1 | ftrace - Function Tracer |
2 | ======================== | |
3 | ||
4 | Copyright 2008 Red Hat Inc. | |
5 | Author: Steven Rostedt <srostedt@redhat.com> | |
6 | ||
7 | ||
8 | Introduction | |
9 | ------------ | |
10 | ||
11 | Ftrace is an internal tracer designed to help out developers and | |
12 | designers of systems to find what is going on inside the kernel. | |
13 | It can be used for debugging or analyzing latencies and performance | |
14 | issues that take place outside of user-space. | |
15 | ||
16 | Although ftrace is the function tracer, it also includes an | |
17 | infrastructure that allows for other types of tracing. Some of the | |
18 | tracers that are currently in ftrace is a tracer to trace | |
19 | context switches, the time it takes for a high priority task to | |
20 | run after it was woken up, the time interrupts are disabled, and | |
21 | more. | |
22 | ||
23 | ||
24 | The File System | |
25 | --------------- | |
26 | ||
27 | Ftrace uses the debugfs file system to hold the control files as well | |
28 | as the files to display output. | |
29 | ||
30 | To mount the debugfs system: | |
31 | ||
32 | # mkdir /debug | |
33 | # mount -t debugfs nodev /debug | |
34 | ||
35 | ||
36 | That's it! (assuming that you have ftrace configured into your kernel) | |
37 | ||
38 | After mounting the debugfs, you can see a directory called | |
39 | "tracing". This directory contains the control and output files | |
40 | of ftrace. Here is a list of some of the key files: | |
41 | ||
42 | ||
43 | Note: all time values are in microseconds. | |
44 | ||
45 | current_tracer : This is used to set or display the current tracer | |
46 | that is configured. | |
47 | ||
48 | available_tracers : This holds the different types of tracers that | |
49 | has been compiled into the kernel. The tracers | |
50 | listed here can be configured by echoing in their | |
51 | name into current_tracer. | |
52 | ||
53 | tracing_enabled : This sets or displays whether the current_tracer | |
54 | is activated and tracing or not. Echo 0 into this | |
55 | file to disable the tracer or 1 (or non-zero) to | |
56 | enable it. | |
57 | ||
58 | trace : This file holds the output of the trace in a human readable | |
59 | format. | |
60 | ||
61 | latency_trace : This file shows the same trace but the information | |
62 | is organized more to display possible latencies | |
63 | in the system. | |
64 | ||
65 | trace_pipe : 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 | |
68 | is retrieved. Unlike the "trace" and "latency_trace" | |
69 | files, this file is a consumer. This means reading | |
70 | from this file causes sequential reads to display | |
71 | more current data. Once data is read from this | |
72 | file, it is consumed, and will not be read | |
73 | again with a sequential read. The "trace" and | |
74 | "latency_trace" files are static, and if the | |
75 | tracer isn't adding more data, they will display | |
76 | the same information every time they are read. | |
77 | ||
78 | iter_ctrl : This file lets the user control the amount of data | |
79 | that is displayed in one of the above output | |
80 | files. | |
81 | ||
82 | trace_max_latency : Some of the tracers record the max latency. | |
83 | For example, the time interrupts are disabled. | |
84 | This time is saved in this file. The max trace | |
85 | will also be stored, and displayed by either | |
86 | "trace" or "latency_trace". A new max trace will | |
87 | only be recorded if the latency is greater than | |
88 | the value in this file. (in microseconds) | |
89 | ||
90 | trace_entries : This sets or displays the number of trace | |
91 | entries each CPU buffer can hold. The tracer buffers | |
92 | are the same size for each CPU, so care must be | |
93 | taken when modifying the trace_entries. The number | |
94 | of actually entries will be the number given | |
95 | times the number of possible CPUS. The buffers | |
96 | are saved as individual pages, and the actual entries | |
97 | will always be rounded up to entries per page. | |
98 | ||
99 | This can only be updated when the current_tracer | |
100 | is set to "none". | |
101 | ||
102 | NOTE: It is planned on changing the allocated buffers | |
103 | from being the number of possible CPUS to | |
104 | the number of online CPUS. | |
105 | ||
106 | tracing_cpumask : This is a mask that lets the user only trace | |
107 | on specified CPUS. The format is a hex string | |
108 | representing the CPUS. | |
109 | ||
110 | set_ftrace_filter : When dynamic ftrace is configured in, the | |
111 | code is dynamically modified to disable calling | |
112 | of the function profiler (mcount). This lets | |
113 | tracing be configured in with practically no overhead | |
114 | in performance. This also has a side effect of | |
115 | enabling or disabling specific functions to be | |
116 | traced. Echoing in names of functions into this | |
117 | file will limit the trace to only those files. | |
118 | ||
119 | set_ftrace_notrace: This has the opposite effect that | |
120 | set_ftrace_filter has. Any function that is added | |
121 | here will not be traced. If a function exists | |
122 | in both set_ftrace_filter and set_ftrace_notrace | |
123 | the function will _not_ bet traced. | |
124 | ||
125 | available_filter_functions : When a function is encountered the first | |
126 | time by the dynamic tracer, it is recorded and | |
127 | later the call is converted into a nop. This file | |
128 | lists the functions that have been recorded | |
129 | by the dynamic tracer and these functions can | |
130 | be used to set the ftrace filter by the above | |
131 | "set_ftrace_filter" file. | |
132 | ||
133 | ||
134 | The Tracers | |
135 | ----------- | |
136 | ||
137 | Here are the list of current tracers that can be configured. | |
138 | ||
139 | ftrace - function tracer that uses mcount to trace all functions. | |
140 | It is possible to filter out which functions that are | |
141 | traced when dynamic ftrace is configured in. | |
142 | ||
143 | sched_switch - traces the context switches between tasks. | |
144 | ||
145 | irqsoff - traces the areas that disable interrupts and saves off | |
146 | the trace with the longest max latency. | |
147 | See tracing_max_latency. When a new max is recorded, | |
148 | it replaces the old trace. It is best to view this | |
149 | trace with the latency_trace file. | |
150 | ||
151 | preemptoff - Similar to irqsoff but traces and records the time | |
152 | preemption is disabled. | |
153 | ||
154 | preemptirqsoff - Similar to irqsoff and preemptoff, but traces and | |
155 | records the largest time irqs and/or preemption is | |
156 | disabled. | |
157 | ||
158 | wakeup - Traces and records the max latency that it takes for | |
159 | the highest priority task to get scheduled after | |
160 | it has been woken up. | |
161 | ||
162 | none - This is not a tracer. To remove all tracers from tracing | |
163 | simply echo "none" into current_tracer. | |
164 | ||
165 | ||
166 | Examples of using the tracer | |
167 | ---------------------------- | |
168 | ||
169 | Here are typical examples of using the tracers with only controlling | |
170 | them with the debugfs interface (without using any user-land utilities). | |
171 | ||
172 | Output format: | |
173 | -------------- | |
174 | ||
175 | Here's an example of the output format of the file "trace" | |
176 | ||
177 | -------- | |
178 | # tracer: ftrace | |
179 | # | |
180 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
181 | # | | | | | | |
182 | bash-4251 [01] 10152.583854: path_put <-path_walk | |
183 | bash-4251 [01] 10152.583855: dput <-path_put | |
184 | bash-4251 [01] 10152.583855: _atomic_dec_and_lock <-dput | |
185 | -------- | |
186 | ||
187 | A header is printed with the trace that is represented. In this case | |
188 | the tracer is "ftrace". Then a header showing the format. Task name | |
189 | "bash", the task PID "4251", the CPU that it was running on | |
190 | "01", the timestamp in <secs>.<usecs> format, the function name that was | |
191 | traced "path_put" and the parent function that called this function | |
192 | "path_walk". | |
193 | ||
194 | The sched_switch tracer also includes tracing of task wake ups and | |
195 | context switches. | |
196 | ||
197 | ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 2916:115:S | |
198 | ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 10:115:S | |
199 | ksoftirqd/1-7 [01] 1453.070013: 7:115:R ==> 10:115:R | |
200 | events/1-10 [01] 1453.070013: 10:115:S ==> 2916:115:R | |
201 | kondemand/1-2916 [01] 1453.070013: 2916:115:S ==> 7:115:R | |
202 | ksoftirqd/1-7 [01] 1453.070013: 7:115:S ==> 0:140:R | |
203 | ||
204 | Wake ups are represented by a "+" and the context switches show | |
205 | "==>". The format is: | |
206 | ||
207 | Context switches: | |
208 | ||
209 | Previous task Next Task | |
210 | ||
211 | <pid>:<prio>:<state> ==> <pid>:<prio>:<state> | |
212 | ||
213 | Wake ups: | |
214 | ||
215 | Current task Task waking up | |
216 | ||
217 | <pid>:<prio>:<state> + <pid>:<prio>:<state> | |
218 | ||
219 | The prio is the internal kernel priority, which is inverse to the | |
220 | priority that is usually displayed by user-space tools. Zero represents | |
221 | the highest priority (99). Prio 100 starts the "nice" priorities with | |
222 | 100 being equal to nice -20 and 139 being nice 19. The prio "140" is | |
223 | reserved for the idle task which is the lowest priority thread (pid 0). | |
224 | ||
225 | ||
226 | Latency trace format | |
227 | -------------------- | |
228 | ||
229 | For traces that display latency times, the latency_trace file gives | |
230 | a bit more information to see why a latency happened. Here's a typical | |
231 | trace. | |
232 | ||
233 | # tracer: irqsoff | |
234 | # | |
235 | irqsoff latency trace v1.1.5 on 2.6.26-rc8 | |
236 | -------------------------------------------------------------------- | |
237 | latency: 97 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
238 | ----------------- | |
239 | | task: swapper-0 (uid:0 nice:0 policy:0 rt_prio:0) | |
240 | ----------------- | |
241 | => started at: apic_timer_interrupt | |
242 | => ended at: do_softirq | |
243 | ||
244 | # _------=> CPU# | |
245 | # / _-----=> irqs-off | |
246 | # | / _----=> need-resched | |
247 | # || / _---=> hardirq/softirq | |
248 | # ||| / _--=> preempt-depth | |
249 | # |||| / | |
250 | # ||||| delay | |
251 | # cmd pid ||||| time | caller | |
252 | # \ / ||||| \ | / | |
253 | <idle>-0 0d..1 0us+: trace_hardirqs_off_thunk (apic_timer_interrupt) | |
254 | <idle>-0 0d.s. 97us : __do_softirq (do_softirq) | |
255 | <idle>-0 0d.s1 98us : trace_hardirqs_on (do_softirq) | |
256 | ||
257 | ||
258 | vim:ft=help | |
259 | ||
260 | ||
261 | This shows that the current tracer is "irqsoff" tracing the time | |
262 | interrupts are disabled. It gives the trace version and the kernel | |
263 | this was executed on (2.6.26-rc8). Then it displays the max latency | |
264 | in microsecs (97 us). The number of trace entries displayed | |
265 | by the total number recorded (both are three: #3/3). The type of | |
266 | preemption that was used (PREEMPT). VP, KP, SP, and HP are always zero | |
267 | and reserved for later use. #P is the number of online CPUS (#P:2). | |
268 | ||
269 | The task is the process that was running when the latency happened. | |
270 | (swapper pid: 0). | |
271 | ||
272 | The start and stop that caused the latencies: | |
273 | ||
274 | apic_timer_interrupt is where the interrupts were disabled. | |
275 | do_softirq is where they were enabled again. | |
276 | ||
277 | The next lines after the header are the trace itself. The header | |
278 | explains which is which. | |
279 | ||
280 | cmd: The name of the process in the trace. | |
281 | ||
282 | pid: The PID of that process. | |
283 | ||
284 | CPU#: The CPU that the process was running on. | |
285 | ||
286 | irqs-off: 'd' interrupts are disabled. '.' otherwise. | |
287 | ||
288 | need-resched: 'N' task need_resched is set, '.' otherwise. | |
289 | ||
290 | hardirq/softirq: | |
291 | 'H' - hard irq happened inside a softirq. | |
292 | 'h' - hard irq is running | |
293 | 's' - soft irq is running | |
294 | '.' - normal context. | |
295 | ||
296 | preempt-depth: The level of preempt_disabled | |
297 | ||
298 | The above is mostly meaningful for kernel developers. | |
299 | ||
300 | time: This differs from the trace output where as the trace output | |
301 | contained a absolute timestamp. This timestamp is relative | |
302 | to the start of the first entry in the the trace. | |
303 | ||
304 | delay: This is just to help catch your eye a bit better. And | |
305 | needs to be fixed to be only relative to the same CPU. | |
306 | The marks is determined by the difference between this | |
307 | current trace and the next trace. | |
308 | '!' - greater than preempt_mark_thresh (default 100) | |
309 | '+' - greater than 1 microsecond | |
310 | ' ' - less than or equal to 1 microsecond. | |
311 | ||
312 | The rest is the same as the 'trace' file. | |
313 | ||
314 | ||
315 | iter_ctrl | |
316 | --------- | |
317 | ||
318 | The iter_ctrl file is used to control what gets printed in the trace | |
319 | output. To see what is available, simply cat the file: | |
320 | ||
321 | cat /debug/tracing/iter_ctrl | |
322 | print-parent nosym-offset nosym-addr noverbose noraw nohex nobin \ | |
323 | noblock nostacktrace nosched-tree | |
324 | ||
325 | To disable one of the options, echo in the option appended with "no". | |
326 | ||
327 | echo noprint-parent > /debug/tracing/iter_ctrl | |
328 | ||
329 | To enable an option, leave off the "no". | |
330 | ||
331 | echo sym-offest > /debug/tracing/iter_ctrl | |
332 | ||
333 | Here are the available options: | |
334 | ||
335 | print-parent - On function traces, display the calling function | |
336 | as well as the function being traced. | |
337 | ||
338 | print-parent: | |
339 | bash-4000 [01] 1477.606694: simple_strtoul <-strict_strtoul | |
340 | ||
341 | noprint-parent: | |
342 | bash-4000 [01] 1477.606694: simple_strtoul | |
343 | ||
344 | ||
345 | sym-offset - Display not only the function name, but also the offset | |
346 | in the function. For example, instead of seeing just | |
347 | "ktime_get" you will see "ktime_get+0xb/0x20" | |
348 | ||
349 | sym-offset: | |
350 | bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0 | |
351 | ||
352 | sym-addr - this will also display the function address as well as | |
353 | the function name. | |
354 | ||
355 | sym-addr: | |
356 | bash-4000 [01] 1477.606694: simple_strtoul <c0339346> | |
357 | ||
358 | verbose - This deals with the latency_trace file. | |
359 | ||
360 | bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \ | |
361 | (+0.000ms): simple_strtoul (strict_strtoul) | |
362 | ||
363 | raw - This will display raw numbers. This option is best for use with | |
364 | user applications that can translate the raw numbers better than | |
365 | having it done in the kernel. | |
366 | ||
367 | hex - similar to raw, but the numbers will be in a hexadecimal format. | |
368 | ||
369 | bin - This will print out the formats in raw binary. | |
370 | ||
371 | block - TBD (needs update) | |
372 | ||
373 | stacktrace - This is one of the options that changes the trace itself. | |
374 | When a trace is recorded, so is the stack of functions. | |
375 | This allows for back traces of trace sites. | |
376 | ||
377 | sched-tree - TBD (any users??) | |
378 | ||
379 | ||
380 | sched_switch | |
381 | ------------ | |
382 | ||
383 | This tracer simply records schedule switches. Here's an example | |
384 | on how to implement it. | |
385 | ||
386 | # echo sched_switch > /debug/tracing/current_tracer | |
387 | # echo 1 > /debug/tracing/tracing_enabled | |
388 | # sleep 1 | |
389 | # echo 0 > /debug/tracing/tracing_enabled | |
390 | # cat /debug/tracing/trace | |
391 | ||
392 | # tracer: sched_switch | |
393 | # | |
394 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
395 | # | | | | | | |
396 | bash-3997 [01] 240.132281: 3997:120:R + 4055:120:R | |
397 | bash-3997 [01] 240.132284: 3997:120:R ==> 4055:120:R | |
398 | sleep-4055 [01] 240.132371: 4055:120:S ==> 3997:120:R | |
399 | bash-3997 [01] 240.132454: 3997:120:R + 4055:120:S | |
400 | bash-3997 [01] 240.132457: 3997:120:R ==> 4055:120:R | |
401 | sleep-4055 [01] 240.132460: 4055:120:D ==> 3997:120:R | |
402 | bash-3997 [01] 240.132463: 3997:120:R + 4055:120:D | |
403 | bash-3997 [01] 240.132465: 3997:120:R ==> 4055:120:R | |
404 | <idle>-0 [00] 240.132589: 0:140:R + 4:115:S | |
405 | <idle>-0 [00] 240.132591: 0:140:R ==> 4:115:R | |
406 | ksoftirqd/0-4 [00] 240.132595: 4:115:S ==> 0:140:R | |
407 | <idle>-0 [00] 240.132598: 0:140:R + 4:115:S | |
408 | <idle>-0 [00] 240.132599: 0:140:R ==> 4:115:R | |
409 | ksoftirqd/0-4 [00] 240.132603: 4:115:S ==> 0:140:R | |
410 | sleep-4055 [01] 240.133058: 4055:120:S ==> 3997:120:R | |
411 | [...] | |
412 | ||
413 | ||
414 | As we have discussed previously about this format, the header shows | |
415 | the name of the trace and points to the options. The "FUNCTION" | |
416 | is a misnomer since here it represents the wake ups and context | |
417 | switches. | |
418 | ||
419 | The sched_switch only lists the wake ups (represented with '+') | |
420 | and context switches ('==>') with the previous task or current | |
421 | first followed by the next task or task waking up. The format for both | |
422 | of these is PID:KERNEL-PRIO:TASK-STATE. Remember that the KERNEL-PRIO | |
423 | is the inverse of the actual priority with zero (0) being the highest | |
424 | priority and the nice values starting at 100 (nice -20). Below is | |
425 | a quick chart to map the kernel priority to user land priorities. | |
426 | ||
427 | Kernel priority: 0 to 99 ==> user RT priority 99 to 0 | |
428 | Kernel priority: 100 to 139 ==> user nice -20 to 19 | |
429 | Kernel priority: 140 ==> idle task priority | |
430 | ||
431 | The task states are: | |
432 | ||
433 | R - running : wants to run, may not actually be running | |
434 | S - sleep : process is waiting to be woken up (handles signals) | |
435 | D - deep sleep : process must be woken up (ignores signals) | |
436 | T - stopped : process suspended | |
437 | t - traced : process is being traced (with something like gdb) | |
438 | Z - zombie : process waiting to be cleaned up | |
439 | X - unknown | |
440 | ||
441 | ||
442 | ftrace_enabled | |
443 | -------------- | |
444 | ||
445 | The following tracers give different output depending on whether | |
446 | or not the sysctl ftrace_enabled is set. To set ftrace_enabled, | |
447 | one can either use the sysctl function or set it via the proc | |
448 | file system interface. | |
449 | ||
450 | sysctl kernel.ftrace_enabled=1 | |
451 | ||
452 | or | |
453 | ||
454 | echo 1 > /proc/sys/kernel/ftrace_enabled | |
455 | ||
456 | To disable ftrace_enabled simply replace the '1' with '0' in | |
457 | the above commands. | |
458 | ||
459 | When ftrace_enabled is set the tracers will also record the functions | |
460 | that are within the trace. The descriptions of the tracers | |
461 | will also show an example with ftrace enabled. | |
462 | ||
463 | ||
464 | irqsoff | |
465 | ------- | |
466 | ||
467 | When interrupts are disabled, the CPU can not react to any other | |
468 | external event (besides NMIs and SMIs). This prevents the timer | |
469 | interrupt from triggering or the mouse interrupt from letting the | |
470 | kernel know of a new mouse event. The result is a latency with the | |
471 | reaction time. | |
472 | ||
473 | The irqsoff tracer tracks the time interrupts are disabled and when | |
474 | they are re-enabled. When a new maximum latency is hit, it saves off | |
475 | the trace so that it may be retrieved at a later time. Every time a | |
476 | new maximum in reached, the old saved trace is discarded and the new | |
477 | trace is saved. | |
478 | ||
479 | To reset the maximum, echo 0 into tracing_max_latency. Here's an | |
480 | example: | |
481 | ||
482 | # echo irqsoff > /debug/tracing/current_tracer | |
483 | # echo 0 > /debug/tracing/tracing_max_latency | |
484 | # echo 1 > /debug/tracing/tracing_enabled | |
485 | # ls -ltr | |
486 | [...] | |
487 | # echo 0 > /debug/tracing/tracing_enabled | |
488 | # cat /debug/tracing/latency_trace | |
489 | # tracer: irqsoff | |
490 | # | |
491 | irqsoff latency trace v1.1.5 on 2.6.26-rc8 | |
492 | -------------------------------------------------------------------- | |
493 | latency: 6 us, #3/3, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
494 | ----------------- | |
495 | | task: bash-4269 (uid:0 nice:0 policy:0 rt_prio:0) | |
496 | ----------------- | |
497 | => started at: copy_page_range | |
498 | => ended at: copy_page_range | |
499 | ||
500 | # _------=> CPU# | |
501 | # / _-----=> irqs-off | |
502 | # | / _----=> need-resched | |
503 | # || / _---=> hardirq/softirq | |
504 | # ||| / _--=> preempt-depth | |
505 | # |||| / | |
506 | # ||||| delay | |
507 | # cmd pid ||||| time | caller | |
508 | # \ / ||||| \ | / | |
509 | bash-4269 1...1 0us+: _spin_lock (copy_page_range) | |
510 | bash-4269 1...1 7us : _spin_unlock (copy_page_range) | |
511 | bash-4269 1...2 7us : trace_preempt_on (copy_page_range) | |
512 | ||
513 | ||
514 | vim:ft=help | |
515 | ||
516 | Here we see that that we had a latency of 6 microsecs (which is | |
517 | very good). The spin_lock in copy_page_range disabled interrupts. | |
518 | The difference between the 6 and the displayed timestamp 7us is | |
519 | because the clock must have incremented between the time of recording | |
520 | the max latency and recording the function that had that latency. | |
521 | ||
522 | Note the above had ftrace_enabled not set. If we set the ftrace_enabled | |
523 | we get a much larger output: | |
524 | ||
525 | # tracer: irqsoff | |
526 | # | |
527 | irqsoff latency trace v1.1.5 on 2.6.26-rc8 | |
528 | -------------------------------------------------------------------- | |
529 | latency: 50 us, #101/101, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
530 | ----------------- | |
531 | | task: ls-4339 (uid:0 nice:0 policy:0 rt_prio:0) | |
532 | ----------------- | |
533 | => started at: __alloc_pages_internal | |
534 | => ended at: __alloc_pages_internal | |
535 | ||
536 | # _------=> CPU# | |
537 | # / _-----=> irqs-off | |
538 | # | / _----=> need-resched | |
539 | # || / _---=> hardirq/softirq | |
540 | # ||| / _--=> preempt-depth | |
541 | # |||| / | |
542 | # ||||| delay | |
543 | # cmd pid ||||| time | caller | |
544 | # \ / ||||| \ | / | |
545 | ls-4339 0...1 0us+: get_page_from_freelist (__alloc_pages_internal) | |
546 | ls-4339 0d..1 3us : rmqueue_bulk (get_page_from_freelist) | |
547 | ls-4339 0d..1 3us : _spin_lock (rmqueue_bulk) | |
548 | ls-4339 0d..1 4us : add_preempt_count (_spin_lock) | |
549 | ls-4339 0d..2 4us : __rmqueue (rmqueue_bulk) | |
550 | ls-4339 0d..2 5us : __rmqueue_smallest (__rmqueue) | |
551 | ls-4339 0d..2 5us : __mod_zone_page_state (__rmqueue_smallest) | |
552 | ls-4339 0d..2 6us : __rmqueue (rmqueue_bulk) | |
553 | ls-4339 0d..2 6us : __rmqueue_smallest (__rmqueue) | |
554 | ls-4339 0d..2 7us : __mod_zone_page_state (__rmqueue_smallest) | |
555 | ls-4339 0d..2 7us : __rmqueue (rmqueue_bulk) | |
556 | ls-4339 0d..2 8us : __rmqueue_smallest (__rmqueue) | |
557 | [...] | |
558 | ls-4339 0d..2 46us : __rmqueue_smallest (__rmqueue) | |
559 | ls-4339 0d..2 47us : __mod_zone_page_state (__rmqueue_smallest) | |
560 | ls-4339 0d..2 47us : __rmqueue (rmqueue_bulk) | |
561 | ls-4339 0d..2 48us : __rmqueue_smallest (__rmqueue) | |
562 | ls-4339 0d..2 48us : __mod_zone_page_state (__rmqueue_smallest) | |
563 | ls-4339 0d..2 49us : _spin_unlock (rmqueue_bulk) | |
564 | ls-4339 0d..2 49us : sub_preempt_count (_spin_unlock) | |
565 | ls-4339 0d..1 50us : get_page_from_freelist (__alloc_pages_internal) | |
566 | ls-4339 0d..2 51us : trace_hardirqs_on (__alloc_pages_internal) | |
567 | ||
568 | ||
569 | vim:ft=help | |
570 | ||
571 | ||
572 | Here we traced a 50 microsecond latency. But we also see all the | |
573 | functions that were called during that time. Note that enabling | |
574 | function tracing we endure an added overhead. This overhead may | |
575 | extend the latency times. But never the less, this trace has provided | |
576 | some very helpful debugging. | |
577 | ||
578 | ||
579 | preemptoff | |
580 | ---------- | |
581 | ||
582 | When preemption is disabled we may be able to receive interrupts but | |
583 | the task can not be preempted and a higher priority task must wait | |
584 | for preemption to be enabled again before it can preempt a lower | |
585 | priority task. | |
586 | ||
587 | The preemptoff tracer traces the places that disables preemption. | |
588 | Like the irqsoff, it records the maximum latency that preemption | |
589 | was disabled. The control of preemptoff is much like the irqsoff. | |
590 | ||
591 | # echo preemptoff > /debug/tracing/current_tracer | |
592 | # echo 0 > /debug/tracing/tracing_max_latency | |
593 | # echo 1 > /debug/tracing/tracing_enabled | |
594 | # ls -ltr | |
595 | [...] | |
596 | # echo 0 > /debug/tracing/tracing_enabled | |
597 | # cat /debug/tracing/latency_trace | |
598 | # tracer: preemptoff | |
599 | # | |
600 | preemptoff latency trace v1.1.5 on 2.6.26-rc8 | |
601 | -------------------------------------------------------------------- | |
602 | latency: 29 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
603 | ----------------- | |
604 | | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0) | |
605 | ----------------- | |
606 | => started at: do_IRQ | |
607 | => ended at: __do_softirq | |
608 | ||
609 | # _------=> CPU# | |
610 | # / _-----=> irqs-off | |
611 | # | / _----=> need-resched | |
612 | # || / _---=> hardirq/softirq | |
613 | # ||| / _--=> preempt-depth | |
614 | # |||| / | |
615 | # ||||| delay | |
616 | # cmd pid ||||| time | caller | |
617 | # \ / ||||| \ | / | |
618 | sshd-4261 0d.h. 0us+: irq_enter (do_IRQ) | |
619 | sshd-4261 0d.s. 29us : _local_bh_enable (__do_softirq) | |
620 | sshd-4261 0d.s1 30us : trace_preempt_on (__do_softirq) | |
621 | ||
622 | ||
623 | vim:ft=help | |
624 | ||
625 | This has some more changes. Preemption was disabled when an interrupt | |
626 | came in (notice the 'h'), and was enabled while doing a softirq. | |
627 | (notice the 's'). But we also see that interrupts have been disabled | |
628 | when entering the preempt off section and leaving it (the 'd'). | |
629 | We do not know if interrupts were enabled in the mean time. | |
630 | ||
631 | # tracer: preemptoff | |
632 | # | |
633 | preemptoff latency trace v1.1.5 on 2.6.26-rc8 | |
634 | -------------------------------------------------------------------- | |
635 | latency: 63 us, #87/87, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
636 | ----------------- | |
637 | | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0) | |
638 | ----------------- | |
639 | => started at: remove_wait_queue | |
640 | => ended at: __do_softirq | |
641 | ||
642 | # _------=> CPU# | |
643 | # / _-----=> irqs-off | |
644 | # | / _----=> need-resched | |
645 | # || / _---=> hardirq/softirq | |
646 | # ||| / _--=> preempt-depth | |
647 | # |||| / | |
648 | # ||||| delay | |
649 | # cmd pid ||||| time | caller | |
650 | # \ / ||||| \ | / | |
651 | sshd-4261 0d..1 0us : _spin_lock_irqsave (remove_wait_queue) | |
652 | sshd-4261 0d..1 1us : _spin_unlock_irqrestore (remove_wait_queue) | |
653 | sshd-4261 0d..1 2us : do_IRQ (common_interrupt) | |
654 | sshd-4261 0d..1 2us : irq_enter (do_IRQ) | |
655 | sshd-4261 0d..1 2us : idle_cpu (irq_enter) | |
656 | sshd-4261 0d..1 3us : add_preempt_count (irq_enter) | |
657 | sshd-4261 0d.h1 3us : idle_cpu (irq_enter) | |
658 | sshd-4261 0d.h. 4us : handle_fasteoi_irq (do_IRQ) | |
659 | [...] | |
660 | sshd-4261 0d.h. 12us : add_preempt_count (_spin_lock) | |
661 | sshd-4261 0d.h1 12us : ack_ioapic_quirk_irq (handle_fasteoi_irq) | |
662 | sshd-4261 0d.h1 13us : move_native_irq (ack_ioapic_quirk_irq) | |
663 | sshd-4261 0d.h1 13us : _spin_unlock (handle_fasteoi_irq) | |
664 | sshd-4261 0d.h1 14us : sub_preempt_count (_spin_unlock) | |
665 | sshd-4261 0d.h1 14us : irq_exit (do_IRQ) | |
666 | sshd-4261 0d.h1 15us : sub_preempt_count (irq_exit) | |
667 | sshd-4261 0d..2 15us : do_softirq (irq_exit) | |
668 | sshd-4261 0d... 15us : __do_softirq (do_softirq) | |
669 | sshd-4261 0d... 16us : __local_bh_disable (__do_softirq) | |
670 | sshd-4261 0d... 16us+: add_preempt_count (__local_bh_disable) | |
671 | sshd-4261 0d.s4 20us : add_preempt_count (__local_bh_disable) | |
672 | sshd-4261 0d.s4 21us : sub_preempt_count (local_bh_enable) | |
673 | sshd-4261 0d.s5 21us : sub_preempt_count (local_bh_enable) | |
674 | [...] | |
675 | sshd-4261 0d.s6 41us : add_preempt_count (__local_bh_disable) | |
676 | sshd-4261 0d.s6 42us : sub_preempt_count (local_bh_enable) | |
677 | sshd-4261 0d.s7 42us : sub_preempt_count (local_bh_enable) | |
678 | sshd-4261 0d.s5 43us : add_preempt_count (__local_bh_disable) | |
679 | sshd-4261 0d.s5 43us : sub_preempt_count (local_bh_enable_ip) | |
680 | sshd-4261 0d.s6 44us : sub_preempt_count (local_bh_enable_ip) | |
681 | sshd-4261 0d.s5 44us : add_preempt_count (__local_bh_disable) | |
682 | sshd-4261 0d.s5 45us : sub_preempt_count (local_bh_enable) | |
683 | [...] | |
684 | sshd-4261 0d.s. 63us : _local_bh_enable (__do_softirq) | |
685 | sshd-4261 0d.s1 64us : trace_preempt_on (__do_softirq) | |
686 | ||
687 | ||
688 | The above is an example of the preemptoff trace with ftrace_enabled | |
689 | set. Here we see that interrupts were disabled the entire time. | |
690 | The irq_enter code lets us know that we entered an interrupt 'h'. | |
691 | Before that, the functions being traced still show that it is not | |
692 | in an interrupt, but we can see by the functions themselves that | |
693 | this is not the case. | |
694 | ||
695 | Notice that the __do_softirq when called doesn't have a preempt_count. | |
696 | It may seem that we missed a preempt enabled. What really happened | |
697 | is that the preempt count is held on the threads stack and we | |
698 | switched to the softirq stack (4K stacks in effect). The code | |
699 | does not copy the preempt count, but because interrupts are disabled | |
700 | we don't need to worry about it. Having a tracer like this is good | |
701 | to let people know what really happens inside the kernel. | |
702 | ||
703 | ||
704 | preemptirqsoff | |
705 | -------------- | |
706 | ||
707 | Knowing the locations that have interrupts disabled or preemption | |
708 | disabled for the longest times is helpful. But sometimes we would | |
709 | like to know when either preemption and/or interrupts are disabled. | |
710 | ||
711 | The following code: | |
712 | ||
713 | local_irq_disable(); | |
714 | call_function_with_irqs_off(); | |
715 | preempt_disable(); | |
716 | call_function_with_irqs_and_preemption_off(); | |
717 | local_irq_enable(); | |
718 | call_function_with_preemption_off(); | |
719 | preempt_enable(); | |
720 | ||
721 | The irqsoff tracer will record the total length of | |
722 | call_function_with_irqs_off() and | |
723 | call_function_with_irqs_and_preemption_off(). | |
724 | ||
725 | The preemptoff tracer will record the total length of | |
726 | call_function_with_irqs_and_preemption_off() and | |
727 | call_function_with_preemption_off(). | |
728 | ||
729 | But neither will trace the time that interrupts and/or preemption | |
730 | is disabled. This total time is the time that we can not schedule. | |
731 | To record this time, use the preemptirqsoff tracer. | |
732 | ||
733 | Again, using this trace is much like the irqsoff and preemptoff tracers. | |
734 | ||
735 | # echo preemptoff > /debug/tracing/current_tracer | |
736 | # echo 0 > /debug/tracing/tracing_max_latency | |
737 | # echo 1 > /debug/tracing/tracing_enabled | |
738 | # ls -ltr | |
739 | [...] | |
740 | # echo 0 > /debug/tracing/tracing_enabled | |
741 | # cat /debug/tracing/latency_trace | |
742 | # tracer: preemptirqsoff | |
743 | # | |
744 | preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8 | |
745 | -------------------------------------------------------------------- | |
746 | latency: 293 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
747 | ----------------- | |
748 | | task: ls-4860 (uid:0 nice:0 policy:0 rt_prio:0) | |
749 | ----------------- | |
750 | => started at: apic_timer_interrupt | |
751 | => ended at: __do_softirq | |
752 | ||
753 | # _------=> CPU# | |
754 | # / _-----=> irqs-off | |
755 | # | / _----=> need-resched | |
756 | # || / _---=> hardirq/softirq | |
757 | # ||| / _--=> preempt-depth | |
758 | # |||| / | |
759 | # ||||| delay | |
760 | # cmd pid ||||| time | caller | |
761 | # \ / ||||| \ | / | |
762 | ls-4860 0d... 0us!: trace_hardirqs_off_thunk (apic_timer_interrupt) | |
763 | ls-4860 0d.s. 294us : _local_bh_enable (__do_softirq) | |
764 | ls-4860 0d.s1 294us : trace_preempt_on (__do_softirq) | |
765 | ||
766 | ||
767 | vim:ft=help | |
768 | ||
769 | ||
770 | The trace_hardirqs_off_thunk is called from assembly on x86 when | |
771 | interrupts are disabled in the assembly code. Without the function | |
772 | tracing, we don't know if interrupts were enabled within the preemption | |
773 | points. We do see that it started with preemption enabled. | |
774 | ||
775 | Here is a trace with ftrace_enabled set: | |
776 | ||
777 | ||
778 | # tracer: preemptirqsoff | |
779 | # | |
780 | preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8 | |
781 | -------------------------------------------------------------------- | |
782 | latency: 105 us, #183/183, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
783 | ----------------- | |
784 | | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0) | |
785 | ----------------- | |
786 | => started at: write_chan | |
787 | => ended at: __do_softirq | |
788 | ||
789 | # _------=> CPU# | |
790 | # / _-----=> irqs-off | |
791 | # | / _----=> need-resched | |
792 | # || / _---=> hardirq/softirq | |
793 | # ||| / _--=> preempt-depth | |
794 | # |||| / | |
795 | # ||||| delay | |
796 | # cmd pid ||||| time | caller | |
797 | # \ / ||||| \ | / | |
798 | ls-4473 0.N.. 0us : preempt_schedule (write_chan) | |
799 | ls-4473 0dN.1 1us : _spin_lock (schedule) | |
800 | ls-4473 0dN.1 2us : add_preempt_count (_spin_lock) | |
801 | ls-4473 0d..2 2us : put_prev_task_fair (schedule) | |
802 | [...] | |
803 | ls-4473 0d..2 13us : set_normalized_timespec (ktime_get_ts) | |
804 | ls-4473 0d..2 13us : __switch_to (schedule) | |
805 | sshd-4261 0d..2 14us : finish_task_switch (schedule) | |
806 | sshd-4261 0d..2 14us : _spin_unlock_irq (finish_task_switch) | |
807 | sshd-4261 0d..1 15us : add_preempt_count (_spin_lock_irqsave) | |
808 | sshd-4261 0d..2 16us : _spin_unlock_irqrestore (hrtick_set) | |
809 | sshd-4261 0d..2 16us : do_IRQ (common_interrupt) | |
810 | sshd-4261 0d..2 17us : irq_enter (do_IRQ) | |
811 | sshd-4261 0d..2 17us : idle_cpu (irq_enter) | |
812 | sshd-4261 0d..2 18us : add_preempt_count (irq_enter) | |
813 | sshd-4261 0d.h2 18us : idle_cpu (irq_enter) | |
814 | sshd-4261 0d.h. 18us : handle_fasteoi_irq (do_IRQ) | |
815 | sshd-4261 0d.h. 19us : _spin_lock (handle_fasteoi_irq) | |
816 | sshd-4261 0d.h. 19us : add_preempt_count (_spin_lock) | |
817 | sshd-4261 0d.h1 20us : _spin_unlock (handle_fasteoi_irq) | |
818 | sshd-4261 0d.h1 20us : sub_preempt_count (_spin_unlock) | |
819 | [...] | |
820 | sshd-4261 0d.h1 28us : _spin_unlock (handle_fasteoi_irq) | |
821 | sshd-4261 0d.h1 29us : sub_preempt_count (_spin_unlock) | |
822 | sshd-4261 0d.h2 29us : irq_exit (do_IRQ) | |
823 | sshd-4261 0d.h2 29us : sub_preempt_count (irq_exit) | |
824 | sshd-4261 0d..3 30us : do_softirq (irq_exit) | |
825 | sshd-4261 0d... 30us : __do_softirq (do_softirq) | |
826 | sshd-4261 0d... 31us : __local_bh_disable (__do_softirq) | |
827 | sshd-4261 0d... 31us+: add_preempt_count (__local_bh_disable) | |
828 | sshd-4261 0d.s4 34us : add_preempt_count (__local_bh_disable) | |
829 | [...] | |
830 | sshd-4261 0d.s3 43us : sub_preempt_count (local_bh_enable_ip) | |
831 | sshd-4261 0d.s4 44us : sub_preempt_count (local_bh_enable_ip) | |
832 | sshd-4261 0d.s3 44us : smp_apic_timer_interrupt (apic_timer_interrupt) | |
833 | sshd-4261 0d.s3 45us : irq_enter (smp_apic_timer_interrupt) | |
834 | sshd-4261 0d.s3 45us : idle_cpu (irq_enter) | |
835 | sshd-4261 0d.s3 46us : add_preempt_count (irq_enter) | |
836 | sshd-4261 0d.H3 46us : idle_cpu (irq_enter) | |
837 | sshd-4261 0d.H3 47us : hrtimer_interrupt (smp_apic_timer_interrupt) | |
838 | sshd-4261 0d.H3 47us : ktime_get (hrtimer_interrupt) | |
839 | [...] | |
840 | sshd-4261 0d.H3 81us : tick_program_event (hrtimer_interrupt) | |
841 | sshd-4261 0d.H3 82us : ktime_get (tick_program_event) | |
842 | sshd-4261 0d.H3 82us : ktime_get_ts (ktime_get) | |
843 | sshd-4261 0d.H3 83us : getnstimeofday (ktime_get_ts) | |
844 | sshd-4261 0d.H3 83us : set_normalized_timespec (ktime_get_ts) | |
845 | sshd-4261 0d.H3 84us : clockevents_program_event (tick_program_event) | |
846 | sshd-4261 0d.H3 84us : lapic_next_event (clockevents_program_event) | |
847 | sshd-4261 0d.H3 85us : irq_exit (smp_apic_timer_interrupt) | |
848 | sshd-4261 0d.H3 85us : sub_preempt_count (irq_exit) | |
849 | sshd-4261 0d.s4 86us : sub_preempt_count (irq_exit) | |
850 | sshd-4261 0d.s3 86us : add_preempt_count (__local_bh_disable) | |
851 | [...] | |
852 | sshd-4261 0d.s1 98us : sub_preempt_count (net_rx_action) | |
853 | sshd-4261 0d.s. 99us : add_preempt_count (_spin_lock_irq) | |
854 | sshd-4261 0d.s1 99us+: _spin_unlock_irq (run_timer_softirq) | |
855 | sshd-4261 0d.s. 104us : _local_bh_enable (__do_softirq) | |
856 | sshd-4261 0d.s. 104us : sub_preempt_count (_local_bh_enable) | |
857 | sshd-4261 0d.s. 105us : _local_bh_enable (__do_softirq) | |
858 | sshd-4261 0d.s1 105us : trace_preempt_on (__do_softirq) | |
859 | ||
860 | ||
861 | This is a very interesting trace. It started with the preemption of | |
862 | the ls task. We see that the task had the "need_resched" bit set | |
863 | with the 'N' in the trace. Interrupts are disabled in the spin_lock | |
864 | and the trace started. We see that a schedule took place to run | |
865 | sshd. When the interrupts were enabled we took an interrupt. | |
866 | On return of the interrupt the softirq ran. We took another interrupt | |
867 | while running the softirq as we see with the capital 'H'. | |
868 | ||
869 | ||
870 | wakeup | |
871 | ------ | |
872 | ||
873 | In Real-Time environment it is very important to know the wakeup | |
874 | time it takes for the highest priority task that wakes up to the | |
875 | time it executes. This is also known as "schedule latency". | |
876 | I stress the point that this is about RT tasks. It is also important | |
877 | to know the scheduling latency of non-RT tasks, but the average | |
878 | schedule latency is better for non-RT tasks. Tools like | |
879 | LatencyTop is more appropriate for such measurements. | |
880 | ||
881 | Real-Time environments is interested in the worst case latency. | |
882 | That is the longest latency it takes for something to happen, and | |
883 | not the average. We can have a very fast scheduler that may only | |
884 | have a large latency once in a while, but that would not work well | |
885 | with Real-Time tasks. The wakeup tracer was designed to record | |
886 | the worst case wakeups of RT tasks. Non-RT tasks are not recorded | |
887 | because the tracer only records one worst case and tracing non-RT | |
888 | tasks that are unpredictable will overwrite the worst case latency | |
889 | of RT tasks. | |
890 | ||
891 | Since this tracer only deals with RT tasks, we will run this slightly | |
892 | different than we did with the previous tracers. Instead of performing | |
893 | an 'ls' we will run 'sleep 1' under 'chrt' which changes the | |
894 | priority of the task. | |
895 | ||
896 | # echo wakeup > /debug/tracing/current_tracer | |
897 | # echo 0 > /debug/tracing/tracing_max_latency | |
898 | # echo 1 > /debug/tracing/tracing_enabled | |
899 | # chrt -f 5 sleep 1 | |
900 | # echo 0 > /debug/tracing/tracing_enabled | |
901 | # cat /debug/tracing/latency_trace | |
902 | # tracer: wakeup | |
903 | # | |
904 | wakeup latency trace v1.1.5 on 2.6.26-rc8 | |
905 | -------------------------------------------------------------------- | |
906 | latency: 4 us, #2/2, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
907 | ----------------- | |
908 | | task: sleep-4901 (uid:0 nice:0 policy:1 rt_prio:5) | |
909 | ----------------- | |
910 | ||
911 | # _------=> CPU# | |
912 | # / _-----=> irqs-off | |
913 | # | / _----=> need-resched | |
914 | # || / _---=> hardirq/softirq | |
915 | # ||| / _--=> preempt-depth | |
916 | # |||| / | |
917 | # ||||| delay | |
918 | # cmd pid ||||| time | caller | |
919 | # \ / ||||| \ | / | |
920 | <idle>-0 1d.h4 0us+: try_to_wake_up (wake_up_process) | |
921 | <idle>-0 1d..4 4us : schedule (cpu_idle) | |
922 | ||
923 | ||
924 | vim:ft=help | |
925 | ||
926 | ||
927 | Running this on an idle system we see that it only took 4 microseconds | |
928 | to perform the task switch. Note, since the trace marker in the | |
929 | schedule is before the actual "switch" we stop the tracing when | |
930 | the recorded task is about to schedule in. This may change if | |
931 | we add a new marker at the end of the scheduler. | |
932 | ||
933 | Notice that the recorded task is 'sleep' with the PID of 4901 and it | |
934 | has an rt_prio of 5. This priority is user-space priority and not | |
935 | the internal kernel priority. The policy is 1 for SCHED_FIFO and 2 | |
936 | for SCHED_RR. | |
937 | ||
938 | Doing the same with chrt -r 5 and ftrace_enabled set. | |
939 | ||
940 | # tracer: wakeup | |
941 | # | |
942 | wakeup latency trace v1.1.5 on 2.6.26-rc8 | |
943 | -------------------------------------------------------------------- | |
944 | latency: 50 us, #60/60, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2) | |
945 | ----------------- | |
946 | | task: sleep-4068 (uid:0 nice:0 policy:2 rt_prio:5) | |
947 | ----------------- | |
948 | ||
949 | # _------=> CPU# | |
950 | # / _-----=> irqs-off | |
951 | # | / _----=> need-resched | |
952 | # || / _---=> hardirq/softirq | |
953 | # ||| / _--=> preempt-depth | |
954 | # |||| / | |
955 | # ||||| delay | |
956 | # cmd pid ||||| time | caller | |
957 | # \ / ||||| \ | / | |
958 | ksoftirq-7 1d.H3 0us : try_to_wake_up (wake_up_process) | |
959 | ksoftirq-7 1d.H4 1us : sub_preempt_count (marker_probe_cb) | |
960 | ksoftirq-7 1d.H3 2us : check_preempt_wakeup (try_to_wake_up) | |
961 | ksoftirq-7 1d.H3 3us : update_curr (check_preempt_wakeup) | |
962 | ksoftirq-7 1d.H3 4us : calc_delta_mine (update_curr) | |
963 | ksoftirq-7 1d.H3 5us : __resched_task (check_preempt_wakeup) | |
964 | ksoftirq-7 1d.H3 6us : task_wake_up_rt (try_to_wake_up) | |
965 | ksoftirq-7 1d.H3 7us : _spin_unlock_irqrestore (try_to_wake_up) | |
966 | [...] | |
967 | ksoftirq-7 1d.H2 17us : irq_exit (smp_apic_timer_interrupt) | |
968 | ksoftirq-7 1d.H2 18us : sub_preempt_count (irq_exit) | |
969 | ksoftirq-7 1d.s3 19us : sub_preempt_count (irq_exit) | |
970 | ksoftirq-7 1..s2 20us : rcu_process_callbacks (__do_softirq) | |
971 | [...] | |
972 | ksoftirq-7 1..s2 26us : __rcu_process_callbacks (rcu_process_callbacks) | |
973 | ksoftirq-7 1d.s2 27us : _local_bh_enable (__do_softirq) | |
974 | ksoftirq-7 1d.s2 28us : sub_preempt_count (_local_bh_enable) | |
975 | ksoftirq-7 1.N.3 29us : sub_preempt_count (ksoftirqd) | |
976 | ksoftirq-7 1.N.2 30us : _cond_resched (ksoftirqd) | |
977 | ksoftirq-7 1.N.2 31us : __cond_resched (_cond_resched) | |
978 | ksoftirq-7 1.N.2 32us : add_preempt_count (__cond_resched) | |
979 | ksoftirq-7 1.N.2 33us : schedule (__cond_resched) | |
980 | ksoftirq-7 1.N.2 33us : add_preempt_count (schedule) | |
981 | ksoftirq-7 1.N.3 34us : hrtick_clear (schedule) | |
982 | ksoftirq-7 1dN.3 35us : _spin_lock (schedule) | |
983 | ksoftirq-7 1dN.3 36us : add_preempt_count (_spin_lock) | |
984 | ksoftirq-7 1d..4 37us : put_prev_task_fair (schedule) | |
985 | ksoftirq-7 1d..4 38us : update_curr (put_prev_task_fair) | |
986 | [...] | |
987 | ksoftirq-7 1d..5 47us : _spin_trylock (tracing_record_cmdline) | |
988 | ksoftirq-7 1d..5 48us : add_preempt_count (_spin_trylock) | |
989 | ksoftirq-7 1d..6 49us : _spin_unlock (tracing_record_cmdline) | |
990 | ksoftirq-7 1d..6 49us : sub_preempt_count (_spin_unlock) | |
991 | ksoftirq-7 1d..4 50us : schedule (__cond_resched) | |
992 | ||
993 | The interrupt went off while running ksoftirqd. This task runs at | |
994 | SCHED_OTHER. Why didn't we see the 'N' set early? This may be | |
995 | a harmless bug with x86_32 and 4K stacks. The need_reched() function | |
996 | that tests if we need to reschedule looks on the actual stack. | |
997 | Where as the setting of the NEED_RESCHED bit happens on the | |
998 | task's stack. But because we are in a hard interrupt, the test | |
999 | is with the interrupts stack which has that to be false. We don't | |
1000 | see the 'N' until we switch back to the task's stack. | |
1001 | ||
1002 | ftrace | |
1003 | ------ | |
1004 | ||
1005 | ftrace is not only the name of the tracing infrastructure, but it | |
1006 | is also a name of one of the tracers. The tracer is the function | |
1007 | tracer. Enabling the function tracer can be done from the | |
1008 | debug file system. Make sure the ftrace_enabled is set otherwise | |
1009 | this tracer is a nop. | |
1010 | ||
1011 | # sysctl kernel.ftrace_enabled=1 | |
1012 | # echo ftrace > /debug/tracing/current_tracer | |
1013 | # echo 1 > /debug/tracing/tracing_enabled | |
1014 | # usleep 1 | |
1015 | # echo 0 > /debug/tracing/tracing_enabled | |
1016 | # cat /debug/tracing/trace | |
1017 | # tracer: ftrace | |
1018 | # | |
1019 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
1020 | # | | | | | | |
1021 | bash-4003 [00] 123.638713: finish_task_switch <-schedule | |
1022 | bash-4003 [00] 123.638714: _spin_unlock_irq <-finish_task_switch | |
1023 | bash-4003 [00] 123.638714: sub_preempt_count <-_spin_unlock_irq | |
1024 | bash-4003 [00] 123.638715: hrtick_set <-schedule | |
1025 | bash-4003 [00] 123.638715: _spin_lock_irqsave <-hrtick_set | |
1026 | bash-4003 [00] 123.638716: add_preempt_count <-_spin_lock_irqsave | |
1027 | bash-4003 [00] 123.638716: _spin_unlock_irqrestore <-hrtick_set | |
1028 | bash-4003 [00] 123.638717: sub_preempt_count <-_spin_unlock_irqrestore | |
1029 | bash-4003 [00] 123.638717: hrtick_clear <-hrtick_set | |
1030 | bash-4003 [00] 123.638718: sub_preempt_count <-schedule | |
1031 | bash-4003 [00] 123.638718: sub_preempt_count <-preempt_schedule | |
1032 | bash-4003 [00] 123.638719: wait_for_completion <-__stop_machine_run | |
1033 | bash-4003 [00] 123.638719: wait_for_common <-wait_for_completion | |
1034 | bash-4003 [00] 123.638720: _spin_lock_irq <-wait_for_common | |
1035 | bash-4003 [00] 123.638720: add_preempt_count <-_spin_lock_irq | |
1036 | [...] | |
1037 | ||
1038 | ||
1039 | Note: It is sometimes better to enable or disable tracing directly from | |
1040 | a program, because the buffer may be overflowed by the echo commands | |
1041 | before you get to the point you want to trace. It is also easier to | |
1042 | stop the tracing at the point that you hit the part that you are | |
1043 | interested in. Since the ftrace buffer is a ring buffer with the | |
1044 | oldest data being overwritten, usually it is sufficient to start the | |
1045 | tracer with an echo command but have you code stop it. Something | |
1046 | like the following is usually appropriate for this. | |
1047 | ||
1048 | int trace_fd; | |
1049 | [...] | |
1050 | int main(int argc, char *argv[]) { | |
1051 | [...] | |
1052 | trace_fd = open("/debug/tracing/tracing_enabled", O_WRONLY); | |
1053 | [...] | |
1054 | if (condition_hit()) { | |
1055 | write(trace_fd, "0", 1); | |
1056 | } | |
1057 | [...] | |
1058 | } | |
1059 | ||
1060 | ||
1061 | dynamic ftrace | |
1062 | -------------- | |
1063 | ||
1064 | If CONFIG_DYNAMIC_FTRACE is set, then the system will run with | |
1065 | virtually no overhead when function tracing is disabled. The way | |
1066 | this works is the mcount function call (placed at the start of | |
1067 | every kernel function, produced by the -pg switch in gcc), starts | |
1068 | of pointing to a simple return. | |
1069 | ||
1070 | When dynamic ftrace is initialized, it calls kstop_machine to make it | |
1071 | act like a uniprocessor so that it can freely modify code without | |
1072 | worrying about other processors executing that same code. At | |
1073 | initialization, the mcount calls are change to call a "record_ip" | |
1074 | function. After this, the first time a kernel function is called, | |
1075 | it has the calling address saved in a hash table. | |
1076 | ||
1077 | Later on the ftraced kernel thread is awoken and will again call | |
1078 | kstop_machine if new functions have been recorded. The ftraced thread | |
1079 | will change all calls to mcount to "nop". Just calling mcount | |
1080 | and having mcount return has shown a 10% overhead. By converting | |
1081 | it to a nop, there is no recordable overhead to the system. | |
1082 | ||
1083 | One special side-effect to the recording of the functions being | |
1084 | traced, is that we can now selectively choose which functions we | |
1085 | want to trace and which ones we want the mcount calls to remain as | |
1086 | nops. | |
1087 | ||
1088 | Two files that contain to the enabling and disabling of recorded | |
1089 | functions are: | |
1090 | ||
1091 | set_ftrace_filter | |
1092 | ||
1093 | and | |
1094 | ||
1095 | set_ftrace_notrace | |
1096 | ||
1097 | A list of available functions that you can add to this files is listed | |
1098 | in: | |
1099 | ||
1100 | available_filter_functions | |
1101 | ||
1102 | # cat /debug/tracing/available_filter_functions | |
1103 | put_prev_task_idle | |
1104 | kmem_cache_create | |
1105 | pick_next_task_rt | |
1106 | get_online_cpus | |
1107 | pick_next_task_fair | |
1108 | mutex_lock | |
1109 | [...] | |
1110 | ||
1111 | If I'm only interested in sys_nanosleep and hrtimer_interrupt: | |
1112 | ||
1113 | # echo sys_nanosleep hrtimer_interrupt \ | |
1114 | > /debug/tracing/set_ftrace_filter | |
1115 | # echo ftrace > /debug/tracing/current_tracer | |
1116 | # echo 1 > /debug/tracing/tracing_enabled | |
1117 | # usleep 1 | |
1118 | # echo 0 > /debug/tracing/tracing_enabled | |
1119 | # cat /debug/tracing/trace | |
1120 | # tracer: ftrace | |
1121 | # | |
1122 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
1123 | # | | | | | | |
1124 | usleep-4134 [00] 1317.070017: hrtimer_interrupt <-smp_apic_timer_interrupt | |
1125 | usleep-4134 [00] 1317.070111: sys_nanosleep <-syscall_call | |
1126 | <idle>-0 [00] 1317.070115: hrtimer_interrupt <-smp_apic_timer_interrupt | |
1127 | ||
1128 | To see what functions are being traced, you can cat the file: | |
1129 | ||
1130 | # cat /debug/tracing/set_ftrace_filter | |
1131 | hrtimer_interrupt | |
1132 | sys_nanosleep | |
1133 | ||
1134 | ||
1135 | Perhaps this isn't enough. The filters also allow simple wild cards. | |
1136 | Only the following is currently available | |
1137 | ||
1138 | <match>* - will match functions that begins with <match> | |
1139 | *<match> - will match functions that end with <match> | |
1140 | *<match>* - will match functions that have <match> in it | |
1141 | ||
1142 | Thats all the wild cards that are allowed. | |
1143 | ||
1144 | <match>*<match> will not work. | |
1145 | ||
1146 | # echo hrtimer_* > /debug/tracing/set_ftrace_filter | |
1147 | ||
1148 | Produces: | |
1149 | ||
1150 | # tracer: ftrace | |
1151 | # | |
1152 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
1153 | # | | | | | | |
1154 | bash-4003 [00] 1480.611794: hrtimer_init <-copy_process | |
1155 | bash-4003 [00] 1480.611941: hrtimer_start <-hrtick_set | |
1156 | bash-4003 [00] 1480.611956: hrtimer_cancel <-hrtick_clear | |
1157 | bash-4003 [00] 1480.611956: hrtimer_try_to_cancel <-hrtimer_cancel | |
1158 | <idle>-0 [00] 1480.612019: hrtimer_get_next_event <-get_next_timer_interrupt | |
1159 | <idle>-0 [00] 1480.612025: hrtimer_get_next_event <-get_next_timer_interrupt | |
1160 | <idle>-0 [00] 1480.612032: hrtimer_get_next_event <-get_next_timer_interrupt | |
1161 | <idle>-0 [00] 1480.612037: hrtimer_get_next_event <-get_next_timer_interrupt | |
1162 | <idle>-0 [00] 1480.612382: hrtimer_get_next_event <-get_next_timer_interrupt | |
1163 | ||
1164 | ||
1165 | Notice that we lost the sys_nanosleep. | |
1166 | ||
1167 | # cat /debug/tracing/set_ftrace_filter | |
1168 | hrtimer_run_queues | |
1169 | hrtimer_run_pending | |
1170 | hrtimer_init | |
1171 | hrtimer_cancel | |
1172 | hrtimer_try_to_cancel | |
1173 | hrtimer_forward | |
1174 | hrtimer_start | |
1175 | hrtimer_reprogram | |
1176 | hrtimer_force_reprogram | |
1177 | hrtimer_get_next_event | |
1178 | hrtimer_interrupt | |
1179 | hrtimer_nanosleep | |
1180 | hrtimer_wakeup | |
1181 | hrtimer_get_remaining | |
1182 | hrtimer_get_res | |
1183 | hrtimer_init_sleeper | |
1184 | ||
1185 | ||
1186 | This is because the '>' and '>>' act just like they do in bash. | |
1187 | To rewrite the filters, use '>' | |
1188 | To append to the filters, use '>>' | |
1189 | ||
1190 | To clear out a filter so that all functions will be recorded again. | |
1191 | ||
1192 | # echo > /debug/tracing/set_ftrace_filter | |
1193 | # cat /debug/tracing/set_ftrace_filter | |
1194 | # | |
1195 | ||
1196 | Again, now we want to append. | |
1197 | ||
1198 | # echo sys_nanosleep > /debug/tracing/set_ftrace_filter | |
1199 | # cat /debug/tracing/set_ftrace_filter | |
1200 | sys_nanosleep | |
1201 | # echo hrtimer_* >> /debug/tracing/set_ftrace_filter | |
1202 | # cat /debug/tracing/set_ftrace_filter | |
1203 | hrtimer_run_queues | |
1204 | hrtimer_run_pending | |
1205 | hrtimer_init | |
1206 | hrtimer_cancel | |
1207 | hrtimer_try_to_cancel | |
1208 | hrtimer_forward | |
1209 | hrtimer_start | |
1210 | hrtimer_reprogram | |
1211 | hrtimer_force_reprogram | |
1212 | hrtimer_get_next_event | |
1213 | hrtimer_interrupt | |
1214 | sys_nanosleep | |
1215 | hrtimer_nanosleep | |
1216 | hrtimer_wakeup | |
1217 | hrtimer_get_remaining | |
1218 | hrtimer_get_res | |
1219 | hrtimer_init_sleeper | |
1220 | ||
1221 | ||
1222 | The set_ftrace_notrace prevents those functions from being traced. | |
1223 | ||
1224 | # echo '*preempt*' '*lock*' > /debug/tracing/set_ftrace_notrace | |
1225 | ||
1226 | Produces: | |
1227 | ||
1228 | # tracer: ftrace | |
1229 | # | |
1230 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
1231 | # | | | | | | |
1232 | bash-4043 [01] 115.281644: finish_task_switch <-schedule | |
1233 | bash-4043 [01] 115.281645: hrtick_set <-schedule | |
1234 | bash-4043 [01] 115.281645: hrtick_clear <-hrtick_set | |
1235 | bash-4043 [01] 115.281646: wait_for_completion <-__stop_machine_run | |
1236 | bash-4043 [01] 115.281647: wait_for_common <-wait_for_completion | |
1237 | bash-4043 [01] 115.281647: kthread_stop <-stop_machine_run | |
1238 | bash-4043 [01] 115.281648: init_waitqueue_head <-kthread_stop | |
1239 | bash-4043 [01] 115.281648: wake_up_process <-kthread_stop | |
1240 | bash-4043 [01] 115.281649: try_to_wake_up <-wake_up_process | |
1241 | ||
1242 | We can see that there's no more lock or preempt tracing. | |
1243 | ||
1244 | ftraced | |
1245 | ------- | |
1246 | ||
1247 | As mentioned above, when dynamic ftrace is configured in, a kernel | |
1248 | thread wakes up once a second and checks to see if there are mcount | |
1249 | calls that need to be converted into nops. If there is not, then | |
1250 | it simply goes back to sleep. But if there is, it will call | |
1251 | kstop_machine to convert the calls to nops. | |
1252 | ||
1253 | There may be a case that you do not want this added latency. | |
1254 | Perhaps you are doing some audio recording and this activity might | |
1255 | cause skips in the playback. There is an interface to disable | |
1256 | and enable the ftraced kernel thread. | |
1257 | ||
1258 | # echo 0 > /debug/tracing/ftraced_enabled | |
1259 | ||
1260 | This will disable the calling of the kstop_machine to update the | |
1261 | mcount calls to nops. Remember that there's a large overhead | |
1262 | to calling mcount. Without this kernel thread, that overhead will | |
1263 | exist. | |
1264 | ||
1265 | Any write to the ftraced_enabled file will cause the kstop_machine | |
1266 | to run if there are recorded calls to mcount. This means that a | |
1267 | user can manually perform the updates when they want to by simply | |
1268 | echoing a '0' into the ftraced_enabled file. | |
1269 | ||
1270 | The updates are also done at the beginning of enabling a tracer | |
1271 | that uses ftrace function recording. | |
1272 | ||
1273 | ||
1274 | trace_pipe | |
1275 | ---------- | |
1276 | ||
1277 | The trace_pipe outputs the same as trace, but the effect on the | |
1278 | tracing is different. Every read from trace_pipe is consumed. | |
1279 | This means that subsequent reads will be different. The trace | |
1280 | is live. | |
1281 | ||
1282 | # echo ftrace > /debug/tracing/current_tracer | |
1283 | # cat /debug/tracing/trace_pipe > /tmp/trace.out & | |
1284 | [1] 4153 | |
1285 | # echo 1 > /debug/tracing/tracing_enabled | |
1286 | # usleep 1 | |
1287 | # echo 0 > /debug/tracing/tracing_enabled | |
1288 | # cat /debug/tracing/trace | |
1289 | # tracer: ftrace | |
1290 | # | |
1291 | # TASK-PID CPU# TIMESTAMP FUNCTION | |
1292 | # | | | | | | |
1293 | ||
1294 | # | |
1295 | # cat /tmp/trace.out | |
1296 | bash-4043 [00] 41.267106: finish_task_switch <-schedule | |
1297 | bash-4043 [00] 41.267106: hrtick_set <-schedule | |
1298 | bash-4043 [00] 41.267107: hrtick_clear <-hrtick_set | |
1299 | bash-4043 [00] 41.267108: wait_for_completion <-__stop_machine_run | |
1300 | bash-4043 [00] 41.267108: wait_for_common <-wait_for_completion | |
1301 | bash-4043 [00] 41.267109: kthread_stop <-stop_machine_run | |
1302 | bash-4043 [00] 41.267109: init_waitqueue_head <-kthread_stop | |
1303 | bash-4043 [00] 41.267110: wake_up_process <-kthread_stop | |
1304 | bash-4043 [00] 41.267110: try_to_wake_up <-wake_up_process | |
1305 | bash-4043 [00] 41.267111: select_task_rq_rt <-try_to_wake_up | |
1306 | ||
1307 | ||
1308 | Note, reading the trace_pipe will block until more input is added. | |
1309 | By changing the tracer, trace_pipe will issue an EOF. We needed | |
1310 | to set the ftrace tracer _before_ cating the trace_pipe file. | |
1311 | ||
1312 | ||
1313 | trace entries | |
1314 | ------------- | |
1315 | ||
1316 | Having too much or not enough data can be troublesome in diagnosing | |
1317 | some issue in the kernel. The file trace_entries is used to modify | |
1318 | the size of the internal trace buffers. The numbers listed | |
1319 | is the number of entries that can be recorded per CPU. To know | |
1320 | the full size, multiply the number of possible CPUS with the | |
1321 | number of entries. | |
1322 | ||
1323 | # cat /debug/tracing/trace_entries | |
1324 | 65620 | |
1325 | ||
1326 | Note, to modify this you must have tracing fulling disabled. To do that, | |
1327 | echo "none" into the current_tracer. | |
1328 | ||
1329 | # echo none > /debug/tracing/current_tracer | |
1330 | # echo 100000 > /debug/tracing/trace_entries | |
1331 | # cat /debug/tracing/trace_entries | |
1332 | 100045 | |
1333 | ||
1334 | ||
1335 | Notice that we echoed in 100,000 but the size is 100,045. The entries | |
1336 | are held by individual pages. It allocates the number of pages it takes | |
1337 | to fulfill the request. If more entries may fit on the last page | |
1338 | it will add them. | |
1339 | ||
1340 | # echo 1 > /debug/tracing/trace_entries | |
1341 | # cat /debug/tracing/trace_entries | |
1342 | 85 | |
1343 | ||
1344 | This shows us that 85 entries can fit on a single page. | |
1345 | ||
1346 | The number of pages that will be allocated is a percentage of available | |
1347 | memory. Allocating too much will produces an error. | |
1348 | ||
1349 | # echo 1000000000000 > /debug/tracing/trace_entries | |
1350 | -bash: echo: write error: Cannot allocate memory | |
1351 | # cat /debug/tracing/trace_entries | |
1352 | 85 | |
1353 |