1TRACE-CMD-RECORD(1) TRACE-CMD-RECORD(1)
2
3
4
6 trace-cmd-record - record a trace from the Ftrace Linux internal tracer
7
9 trace-cmd record [OPTIONS] [command]
10
12 The trace-cmd(1) record command will set up the Ftrace Linux kernel
13 tracer to record the specified plugins or events that happen while the
14 command executes. If no command is given, then it will record until the
15 user hits Ctrl-C.
16
17 The record command of trace-cmd will set up the Ftrace tracer to start
18 tracing the various events or plugins that are given on the command
19 line. It will then create a number of tracing processes (one per CPU)
20 that will start recording from the kernel ring buffer straight into
21 temporary files. When the command is complete (or Ctrl-C is hit) all
22 the files will be combined into a trace.dat file that can later be read
23 (see trace-cmd-report(1)).
24
26 -p tracer
27 Specify a tracer. Tracers usually do more than just trace an event.
28 Common tracers are: function, function_graph, preemptirqsoff,
29 irqsoff, preemptoff and wakeup. A tracer must be supported by the
30 running kernel. To see a list of available tracers, see
31 trace-cmd-list(1).
32
33 -e event
34 Specify an event to trace. Various static trace points have been
35 added to the Linux kernel. They are grouped by subsystem where you
36 can enable all events of a given subsystem or specify specific
37 events to be enabled. The event is of the format
38 "subsystem:event-name". You can also just specify the subsystem
39 without the :event-name or the event-name without the "subsystem:".
40 Using "-e sched_switch" will enable the "sched_switch" event where
41 as, "-e sched" will enable all events under the "sched" subsystem.
42
43 The 'event' can also contain glob expressions. That is, "*stat*" will
44 select all events (or subsystems) that have the characters "stat" in their
45 names.
46
47 The keyword 'all' can be used to enable all events.
48
49 -a
50 Every event that is being recorded has its output format file saved
51 in the output file to be able to display it later. But if other
52 events are enabled in the trace without trace-cmd’s knowledge, the
53 formats of those events will not be recorded and trace-cmd report
54 will not be able to display them. If this is the case, then specify
55 the -a option and the format for all events in the system will be
56 saved.
57
58 -T
59 Enable a stacktrace on each event. For example:
60
61 <idle>-0 [003] 58549.289091: sched_switch: kworker/0:1:0 [120] R ==> trace-cmd:2603 [120]
62 <idle>-0 [003] 58549.289092: kernel_stack: <stack trace>
63 => schedule (ffffffff814b260e)
64 => cpu_idle (ffffffff8100a38c)
65 => start_secondary (ffffffff814ab828)
66
67 --func-stack
68 Enable a stack trace on all functions. Note this is only applicable
69 for the "function" plugin tracer, and will only take effect if the
70 -l option is used and succeeds in limiting functions. If the
71 function tracer is not filtered, and the stack trace is enabled,
72 you can live lock the machine.
73
74 -f filter
75 Specify a filter for the previous event. This must come after a -e.
76 This will filter what events get recorded based on the content of
77 the event. Filtering is passed to the kernel directly so what
78 filtering is allowed may depend on what version of the kernel you
79 have. Basically, it will let you use C notation to check if an
80 event should be processed or not.
81
82 ==, >=, <=, >, <, &, |, && and ||
83
84 The above are usually safe to use to compare fields.
85
86 --no-filter
87 Do not filter out the trace-cmd threads. By default, the threads
88 are filtered out to not be traced by events. This option will have
89 the trace-cmd threads also be traced.
90
91 -R trigger
92 Specify a trigger for the previous event. This must come after a
93 -e. This will add a given trigger to the given event. To only
94 enable the trigger and not the event itself, then place the event
95 after the -v option.
96
97 See Documentation/trace/events.txt in the Linux kernel source for more
98 information on triggers.
99
100 -v
101 This will cause all events specified after it on the command line
102 to not be traced. This is useful for selecting a subsystem to be
103 traced but to leave out various events. For Example: "-e sched -v
104 -e "*stat\*"" will enable all events in the sched subsystem except
105 those that have "stat" in their names.
106
107 Note: the *-v* option was taken from the way grep(1) inverts the following
108 matches.
109
110 -F
111 This will filter only the executable that is given on the command
112 line. If no command is given, then it will filter itself (pretty
113 pointless). Using -F will let you trace only events that are caused
114 by the given command.
115
116 -P pid
117 Similar to -F but lets you specify a process ID to trace.
118
119 -c
120 Used with either -F (or -P if kernel supports it) to trace the
121 process' children too.
122
123 -C clock
124 Set the trace clock to "clock".
125
126 Use trace-cmd(1) list -C to see what clocks are available.
127
128 -o output-file
129 By default, trace-cmd report will create a trace.dat file. You can
130 specify a different file to write to with the -o option.
131
132 -l function-name
133 This will limit the function and function_graph tracers to only
134 trace the given function name. More than one -l may be specified on
135 the command line to trace more than one function. The limited use
136 of glob expressions are also allowed. These are match* to only
137 filter functions that start with match. *match to only filter
138 functions that end with match. *match\* to only filter on
139 functions that contain match.
140
141 -g function-name
142 This option is for the function_graph plugin. It will graph the
143 given function. That is, it will only trace the function and all
144 functions that it calls. You can have more than one -g on the
145 command line.
146
147 -n function-name
148 This has the opposite effect of -l. The function given with the -n
149 option will not be traced. This takes precedence, that is, if you
150 include the same function for both -n and -l, it will not be
151 traced.
152
153 -d
154 Some tracer plugins enable the function tracer by default. Like the
155 latency tracers. This option prevents the function tracer from
156 being enabled at start up.
157
158 -D
159 The option -d will try to use the function-trace option to disable
160 the function tracer (if available), otherwise it defaults to the
161 proc file: /proc/sys/kernel/ftrace_enabled, but will not touch it
162 if the function-trace option is available. The -D option will
163 disable both the ftrace_enabled proc file as well as the
164 function-trace option if it exists.
165
166 Note, this disable function tracing for all users, which includes users
167 outside of ftrace tracers (stack_tracer, perf, etc).
168
169 -O option
170 Ftrace has various options that can be enabled or disabled. This
171 allows you to set them. Appending the text no to an option disables
172 it. For example: "-O nograph-time" will disable the "graph-time"
173 Ftrace option.
174
175 -s interval
176 The processes that trace-cmd creates to record from the ring buffer
177 need to wake up to do the recording. Setting the interval to zero
178 will cause the processes to wakeup every time new data is written
179 into the buffer. But since Ftrace is recording kernel activity, the
180 act of this processes going back to sleep may cause new events into
181 the ring buffer which will wake the process back up. This will
182 needlessly add extra data into the ring buffer.
183
184 The 'interval' metric is microseconds. The default is set to 1000 (1 ms).
185 This is the time each recording process will sleep before waking up to
186 record any new data that was written to the ring buffer.
187
188 -r priority
189 The priority to run the capture threads at. In a busy system the
190 trace capturing threads may be staved and events can be lost. This
191 increases the priority of those threads to the real time (FIFO)
192 priority. But use this option with care, it can also change the
193 behaviour of the system being traced.
194
195 -b size
196 This sets the ring buffer size to size kilobytes. Because the
197 Ftrace ring buffer is per CPU, this size is the size of each per
198 CPU ring buffer inside the kernel. Using "-b 10000" on a machine
199 with 4 CPUs will make Ftrace have a total buffer size of 40 Megs.
200
201 -B buffer-name
202 If the kernel supports multiple buffers, this will add a buffer
203 with the given name. If the buffer name already exists, that buffer
204 is just reset and will not be deleted at the end of record
205 execution. If the buffer is created, it will be removed at the end
206 of execution (unless the -k is set, or start command was used).
207
208 After a buffer name is stated, all events added after that will be
209 associated with that buffer. If no buffer is specified, or an event
210 is specified before a buffer name, it will be associated with the
211 main (toplevel) buffer.
212
213 trace-cmd record -e sched -B block -e block -B time -e timer sleep 1
214
215 The above is will enable all sched events in the main buffer. It will
216 then create a 'block' buffer instance and enable all block events within
217 that buffer. A 'time' buffer instance is created and all timer events
218 will be enabled for that event.
219
220 -m size
221 The max size in kilobytes that a per cpu buffer should be. Note,
222 due to rounding to page size, the number may not be totally
223 correct. Also, this is performed by switching between two buffers
224 that are half the given size thus the output may not be of the
225 given size even if much more was written.
226
227 Use this to prevent running out of diskspace for long runs.
228
229 -M cpumask
230 Set the cpumask for to trace. It only affects the last buffer
231 instance given. If supplied before any buffer instance, then it
232 affects the main buffer. The value supplied must be a hex number.
233
234 trace-cmd record -p function -M c -B events13 -e all -M 5
235
236 If the -M is left out, then the mask stays the same. To enable all
237 CPUs, pass in a value of '-1'.
238
239 -k
240 By default, when trace-cmd is finished tracing, it will reset the
241 buffers and disable all the tracing that it enabled. This option
242 keeps trace-cmd from disabling the tracer and reseting the buffer.
243 This option is useful for debugging trace-cmd.
244
245 Note: usually trace-cmd will set the "tracing_on" file back to what it
246 was before it was called. This option will leave that file set to zero.
247
248 -i
249 By default, if an event is listed that trace-cmd does not find, it
250 will exit with an error. This option will just ignore events that
251 are listed on the command line but are not found on the system.
252
253 -N host:port
254 If another machine is running "trace-cmd listen", this option is
255 used to have the data sent to that machine with UDP packets.
256 Instead of writing to an output file, the data is sent off to a
257 remote box. This is ideal for embedded machines with little
258 storage, or having a single machine that will keep all the data in
259 a single repository.
260
261 Note: This option is not supported with latency tracer plugins:
262 wakeup, wakeup_rt, irqsoff, preemptoff and preemptirqsoff
263
264 -t
265 This option is used with -N, when there’s a need to send the live
266 data with TCP packets instead of UDP. Although TCP is not nearly as
267 fast as sending the UDP packets, but it may be needed if the
268 network is not that reliable, the amount of data is not that
269 intensive, and a guarantee is needed that all traced information is
270 transfered successfully.
271
272 -q | --quiet
273 For use with recording an application. Suppresses normal output
274 (except for errors) to allow only the application’s output to be
275 displayed.
276
277 --date
278 With the --date option, "trace-cmd" will write timestamps into the
279 trace buffer after it has finished recording. It will then map the
280 timestamp to gettimeofday which will allow wall time output from
281 the timestamps reading the created trace.dat file.
282
283 --max-graph-depth depth
284 Set the maximum depth the function_graph tracer will trace into a
285 function. A value of one will only show where userspace enters the
286 kernel but not any functions called in the kernel. The default is
287 zero, which means no limit.
288
289 --module module
290 Filter a module’s name in function tracing. It is equivalent to
291 adding :mod:module after all other functions being filtered. If no
292 other function filter is listed, then all modules functions will be
293 filtered in the filter.
294
295 '--module snd' is equivalent to '-l :mod:snd'
296
297 '--module snd -l "*jack*"' is equivalent to '-l "*jack*:mod:snd"'
298
299 '--module snd -n "*"' is equivalent to '-n :mod:snd'
300
301 --profile
302 With the --profile option, "trace-cmd" will enable tracing that can
303 be used with trace-cmd-report(1) --profile option. If a tracer -p
304 is not set, and function graph depth is supported by the kernel,
305 then the function_graph tracer will be enabled with a depth of one
306 (only show where userspace enters into the kernel). It will also
307 enable various tracepoints with stack tracing such that the report
308 can show where tasks have been blocked for the longest time.
309
310 See trace-cmd-profile(1) for more details and examples.
311
312 -H event-hooks
313 Add custom event matching to connect any two events together. When
314 not used with --profile, it will save the parameter and this will
315 be used by trace-cmd report --profile, too. That is:
316
317 trace-cmd record -H hrtimer_expire_entry,hrtimer/hrtimer_expire_exit,hrtimer,sp
318 trace-cmd report --profile
319
320 Will profile hrtimer_expire_entry and hrtimer_expire_ext times.
321
322 See trace-cmd-profile(1) for format.
323
324 -S
325 (for --profile only) Only enable the tracer or events speficied on
326 the command line. With this option, the function_graph tracer is
327 not enabled, nor are any events (like sched_switch), unless they
328 are specifically specified on the command line (i.e. -p function -e
329 sched_switch -e sched_wakeup)
330
331 --ts-offset offset
332 Add an offset for the timestamp in the trace.dat file. This will
333 add a offset option into the trace.dat file such that a trace-cmd
334 report will offset all the timestamps of the events by the given
335 offset. The offset is in raw units. That is, if the event
336 timestamps are in nanoseconds the offset will also be in
337 nanoseconds even if the displayed units are in microseconds.
338
339 --stderr
340 Have output go to stderr instead of stdout, but the output of the
341 command executed will not be changed. This is useful if you want to
342 monitor the output of the command being executed, but not see the
343 output from trace-cmd.
344
346 The basic way to trace all events:
347
348 # trace-cmd record -e all ls > /dev/null
349 # trace-cmd report
350 trace-cmd-13541 [003] 106260.693809: filemap_fault: address=0x128122 offset=0xce
351 trace-cmd-13543 [001] 106260.693809: kmalloc: call_site=81128dd4 ptr=0xffff88003dd83800 bytes_req=768 bytes_alloc=1024 gfp_flags=GFP_KERNEL|GFP_ZERO
352 ls-13545 [002] 106260.693809: kfree: call_site=810a7abb ptr=0x0
353 ls-13545 [002] 106260.693818: sys_exit_write: 0x1
354
355 To use the function tracer with sched switch tracing:
356
357 # trace-cmd record -p function -e sched_switch ls > /dev/null
358 # trace-cmd report
359 ls-13587 [002] 106467.860310: function: hrtick_start_fair <-- pick_next_task_fair
360 ls-13587 [002] 106467.860313: sched_switch: prev_comm=trace-cmd prev_pid=13587 prev_prio=120 prev_state=R ==> next_comm=trace-cmd next_pid=13583 next_prio=120
361 trace-cmd-13585 [001] 106467.860314: function: native_set_pte_at <-- __do_fault
362 trace-cmd-13586 [003] 106467.860314: function: up_read <-- do_page_fault
363 ls-13587 [002] 106467.860317: function: __phys_addr <-- schedule
364 trace-cmd-13585 [001] 106467.860318: function: _raw_spin_unlock <-- __do_fault
365 ls-13587 [002] 106467.860320: function: native_load_sp0 <-- __switch_to
366 trace-cmd-13586 [003] 106467.860322: function: down_read_trylock <-- do_page_fault
367
368 Here is a nice way to find what interrupts have the highest latency:
369
370 # trace-cmd record -p function_graph -e irq_handler_entry -l do_IRQ sleep 10
371 # trace-cmd report
372 <idle>-0 [000] 157412.933969: funcgraph_entry: | do_IRQ() {
373 <idle>-0 [000] 157412.933974: irq_handler_entry: irq=48 name=eth0
374 <idle>-0 [000] 157412.934004: funcgraph_exit: + 36.358 us | }
375 <idle>-0 [000] 157413.895004: funcgraph_entry: | do_IRQ() {
376 <idle>-0 [000] 157413.895011: irq_handler_entry: irq=48 name=eth0
377 <idle>-0 [000] 157413.895026: funcgraph_exit: + 24.014 us | }
378 <idle>-0 [000] 157415.891762: funcgraph_entry: | do_IRQ() {
379 <idle>-0 [000] 157415.891769: irq_handler_entry: irq=48 name=eth0
380 <idle>-0 [000] 157415.891784: funcgraph_exit: + 22.928 us | }
381 <idle>-0 [000] 157415.934869: funcgraph_entry: | do_IRQ() {
382 <idle>-0 [000] 157415.934874: irq_handler_entry: irq=48 name=eth0
383 <idle>-0 [000] 157415.934906: funcgraph_exit: + 37.512 us | }
384 <idle>-0 [000] 157417.888373: funcgraph_entry: | do_IRQ() {
385 <idle>-0 [000] 157417.888381: irq_handler_entry: irq=48 name=eth0
386 <idle>-0 [000] 157417.888398: funcgraph_exit: + 25.943 us | }
387
388 An example of the profile:
389
390 # trace-cmd record --profile sleep 1
391 # trace-cmd report --profile --comm sleep
392 task: sleep-21611
393 Event: sched_switch:R (1) Total: 99442 Avg: 99442 Max: 99442 Min:99442
394 <stack> 1 total:99442 min:99442 max:99442 avg=99442
395 => ftrace_raw_event_sched_switch (0xffffffff8105f812)
396 => __schedule (0xffffffff8150810a)
397 => preempt_schedule (0xffffffff8150842e)
398 => ___preempt_schedule (0xffffffff81273354)
399 => cpu_stop_queue_work (0xffffffff810b03c5)
400 => stop_one_cpu (0xffffffff810b063b)
401 => sched_exec (0xffffffff8106136d)
402 => do_execve_common.isra.27 (0xffffffff81148c89)
403 => do_execve (0xffffffff811490b0)
404 => SyS_execve (0xffffffff811492c4)
405 => return_to_handler (0xffffffff8150e3c8)
406 => stub_execve (0xffffffff8150c699)
407 Event: sched_switch:S (1) Total: 1000506680 Avg: 1000506680 Max: 1000506680 Min:1000506680
408 <stack> 1 total:1000506680 min:1000506680 max:1000506680 avg=1000506680
409 => ftrace_raw_event_sched_switch (0xffffffff8105f812)
410 => __schedule (0xffffffff8150810a)
411 => schedule (0xffffffff815084b8)
412 => do_nanosleep (0xffffffff8150b22c)
413 => hrtimer_nanosleep (0xffffffff8108d647)
414 => SyS_nanosleep (0xffffffff8108d72c)
415 => return_to_handler (0xffffffff8150e3c8)
416 => tracesys_phase2 (0xffffffff8150c304)
417 Event: sched_wakeup:21611 (1) Total: 30326 Avg: 30326 Max: 30326 Min:30326
418 <stack> 1 total:30326 min:30326 max:30326 avg=30326
419 => ftrace_raw_event_sched_wakeup_template (0xffffffff8105f653)
420 => ttwu_do_wakeup (0xffffffff810606eb)
421 => ttwu_do_activate.constprop.124 (0xffffffff810607c8)
422 => try_to_wake_up (0xffffffff8106340a)
423
425 trace-cmd(1), trace-cmd-report(1), trace-cmd-start(1),
426 trace-cmd-stop(1), trace-cmd-extract(1), trace-cmd-reset(1),
427 trace-cmd-split(1), trace-cmd-list(1), trace-cmd-listen(1),
428 trace-cmd-profile(1)
429
431 Written by Steven Rostedt, <rostedt@goodmis.org[1]>
432
434 git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/trace-cmd.git
435
437 Copyright (C) 2010 Red Hat, Inc. Free use of this software is granted
438 under the terms of the GNU Public License (GPL).
439
441 1. rostedt@goodmis.org
442 mailto:rostedt@goodmis.org
443
444
445
446 02/08/2020 TRACE-CMD-RECORD(1)