1 ========================
2 ftrace - Function Tracer
3 ========================
5 Copyright 2008 Red Hat Inc.
7 :Author: Steven Rostedt <srostedt@redhat.com>
8 :License: The GNU Free Documentation License, Version 1.2
9 (dual licensed under the GPL v2)
10 :Original Reviewers: Elias Oltmanns, Randy Dunlap, Andrew Morton,
11 John Kacur, and David Teigland.
13 - Written for: 2.6.28-rc2
15 - Updated for: 4.13 - Copyright 2017 VMware Inc. Steven Rostedt
16 - Converted to rst format - Changbin Du <changbin.du@intel.com>
21 Ftrace is an internal tracer designed to help out developers and
22 designers of systems to find what is going on inside the kernel.
23 It can be used for debugging or analyzing latencies and
24 performance issues that take place outside of user-space.
26 Although ftrace is typically considered the function tracer, it
27 is really a framework of several assorted tracing utilities.
28 There's latency tracing to examine what occurs between interrupts
29 disabled and enabled, as well as for preemption and from a time
30 a task is woken to the task is actually scheduled in.
32 One of the most common uses of ftrace is the event tracing.
33 Throughout the kernel is hundreds of static event points that
34 can be enabled via the tracefs file system to see what is
35 going on in certain parts of the kernel.
37 See events.txt for more information.
40 Implementation Details
41 ----------------------
43 See :doc:`ftrace-design` for details for arch porters and such.
49 Ftrace uses the tracefs file system to hold the control files as
50 well as the files to display output.
52 When tracefs is configured into the kernel (which selecting any ftrace
53 option will do) the directory /sys/kernel/tracing will be created. To mount
54 this directory, you can add to your /etc/fstab file::
56 tracefs /sys/kernel/tracing tracefs defaults 0 0
58 Or you can mount it at run time with::
60 mount -t tracefs nodev /sys/kernel/tracing
62 For quicker access to that directory you may want to make a soft link to
65 ln -s /sys/kernel/tracing /tracing
69 Before 4.1, all ftrace tracing control files were within the debugfs
70 file system, which is typically located at /sys/kernel/debug/tracing.
71 For backward compatibility, when mounting the debugfs file system,
72 the tracefs file system will be automatically mounted at:
74 /sys/kernel/debug/tracing
76 All files located in the tracefs file system will be located in that
77 debugfs file system directory as well.
81 Any selected ftrace option will also create the tracefs file system.
82 The rest of the document will assume that you are in the ftrace directory
83 (cd /sys/kernel/tracing) and will only concentrate on the files within that
84 directory and not distract from the content with the extended
85 "/sys/kernel/tracing" path name.
87 That's it! (assuming that you have ftrace configured into your kernel)
89 After mounting tracefs you will have access to the control and output files
90 of ftrace. Here is a list of some of the key files:
93 Note: all time values are in microseconds.
97 This is used to set or display the current tracer
102 This holds the different types of tracers that
103 have been compiled into the kernel. The
104 tracers listed here can be configured by
105 echoing their name into current_tracer.
109 This sets or displays whether writing to the trace
110 ring buffer is enabled. Echo 0 into this file to disable
111 the tracer or 1 to enable it. Note, this only disables
112 writing to the ring buffer, the tracing overhead may
115 The kernel function tracing_off() can be used within the
116 kernel to disable writing to the ring buffer, which will
117 set this file to "0". User space can re-enable tracing by
118 echoing "1" into the file.
120 Note, the function and event trigger "traceoff" will also
121 set this file to zero and stop tracing. Which can also
122 be re-enabled by user space using this file.
126 This file holds the output of the trace in a human
127 readable format (described below). Note, tracing is temporarily
128 disabled while this file is being read (opened).
132 The output is the same as the "trace" file but this
133 file is meant to be streamed with live tracing.
134 Reads from this file will block until new data is
135 retrieved. Unlike the "trace" file, this file is a
136 consumer. This means reading from this file causes
137 sequential reads to display more current data. Once
138 data is read from this file, it is consumed, and
139 will not be read again with a sequential read. The
140 "trace" file is static, and if the tracer is not
141 adding more data, it will display the same
142 information every time it is read. This file will not
143 disable tracing while being read.
147 This file lets the user control the amount of data
148 that is displayed in one of the above output
149 files. Options also exist to modify how a tracer
150 or events work (stack traces, timestamps, etc).
154 This is a directory that has a file for every available
155 trace option (also in trace_options). Options may also be set
156 or cleared by writing a "1" or "0" respectively into the
157 corresponding file with the option name.
161 Some of the tracers record the max latency.
162 For example, the maximum time that interrupts are disabled.
163 The maximum time is saved in this file. The max trace will also be
164 stored, and displayed by "trace". A new max trace will only be
165 recorded if the latency is greater than the value in this file
168 By echoing in a time into this file, no latency will be recorded
169 unless it is greater than the time in this file.
173 Some latency tracers will record a trace whenever the
174 latency is greater than the number in this file.
175 Only active when the file contains a number greater than 0.
180 This sets or displays the number of kilobytes each CPU
181 buffer holds. By default, the trace buffers are the same size
182 for each CPU. The displayed number is the size of the
183 CPU buffer and not total size of all buffers. The
184 trace buffers are allocated in pages (blocks of memory
185 that the kernel uses for allocation, usually 4 KB in size).
186 If the last page allocated has room for more bytes
187 than requested, the rest of the page will be used,
188 making the actual allocation bigger than requested or shown.
189 ( Note, the size may not be a multiple of the page size
190 due to buffer management meta-data. )
192 Buffer sizes for individual CPUs may vary
193 (see "per_cpu/cpu0/buffer_size_kb" below), and if they do
194 this file will show "X".
196 buffer_total_size_kb:
198 This displays the total combined size of all the trace buffers.
202 If a process is performing tracing, and the ring buffer should be
203 shrunk "freed" when the process is finished, even if it were to be
204 killed by a signal, this file can be used for that purpose. On close
205 of this file, the ring buffer will be resized to its minimum size.
206 Having a process that is tracing also open this file, when the process
207 exits its file descriptor for this file will be closed, and in doing so,
208 the ring buffer will be "freed".
210 It may also stop tracing if disable_on_free option is set.
214 This is a mask that lets the user only trace on specified CPUs.
215 The format is a hex string representing the CPUs.
219 When dynamic ftrace is configured in (see the
220 section below "dynamic ftrace"), the code is dynamically
221 modified (code text rewrite) to disable calling of the
222 function profiler (mcount). This lets tracing be configured
223 in with practically no overhead in performance. This also
224 has a side effect of enabling or disabling specific functions
225 to be traced. Echoing names of functions into this file
226 will limit the trace to only those functions.
227 This influences the tracers "function" and "function_graph"
228 and thus also function profiling (see "function_profile_enabled").
230 The functions listed in "available_filter_functions" are what
231 can be written into this file.
233 This interface also allows for commands to be used. See the
234 "Filter commands" section for more details.
236 As a speed up, since processing strings can't be quite expensive
237 and requires a check of all functions registered to tracing, instead
238 an index can be written into this file. A number (starting with "1")
239 written will instead select the same corresponding at the line position
240 of the "available_filter_functions" file.
244 This has an effect opposite to that of
245 set_ftrace_filter. Any function that is added here will not
246 be traced. If a function exists in both set_ftrace_filter
247 and set_ftrace_notrace, the function will _not_ be traced.
251 Have the function tracer only trace the threads whose PID are
254 If the "function-fork" option is set, then when a task whose
255 PID is listed in this file forks, the child's PID will
256 automatically be added to this file, and the child will be
257 traced by the function tracer as well. This option will also
258 cause PIDs of tasks that exit to be removed from the file.
262 Have the events only trace a task with a PID listed in this file.
263 Note, sched_switch and sched_wake_up will also trace events
266 To have the PIDs of children of tasks with their PID in this file
267 added on fork, enable the "event-fork" option. That option will also
268 cause the PIDs of tasks to be removed from this file when the task
273 Functions listed in this file will cause the function graph
274 tracer to only trace these functions and the functions that
275 they call. (See the section "dynamic ftrace" for more details).
276 Note, set_ftrace_filter and set_ftrace_notrace still affects
277 what functions are being traced.
281 Similar to set_graph_function, but will disable function graph
282 tracing when the function is hit until it exits the function.
283 This makes it possible to ignore tracing functions that are called
284 by a specific function.
286 available_filter_functions:
288 This lists the functions that ftrace has processed and can trace.
289 These are the function names that you can pass to
290 "set_ftrace_filter", "set_ftrace_notrace",
291 "set_graph_function", or "set_graph_notrace".
292 (See the section "dynamic ftrace" below for more details.)
294 dyn_ftrace_total_info:
296 This file is for debugging purposes. The number of functions that
297 have been converted to nops and are available to be traced.
301 This file is more for debugging ftrace, but can also be useful
302 in seeing if any function has a callback attached to it.
303 Not only does the trace infrastructure use ftrace function
304 trace utility, but other subsystems might too. This file
305 displays all functions that have a callback attached to them
306 as well as the number of callbacks that have been attached.
307 Note, a callback may also call multiple functions which will
308 not be listed in this count.
310 If the callback registered to be traced by a function with
311 the "save regs" attribute (thus even more overhead), a 'R'
312 will be displayed on the same line as the function that
313 is returning registers.
315 If the callback registered to be traced by a function with
316 the "ip modify" attribute (thus the regs->ip can be changed),
317 an 'I' will be displayed on the same line as the function that
320 If the architecture supports it, it will also show what callback
321 is being directly called by the function. If the count is greater
322 than 1 it most likely will be ftrace_ops_list_func().
324 If the callback of the function jumps to a trampoline that is
325 specific to a the callback and not the standard trampoline,
326 its address will be printed as well as the function that the
329 function_profile_enabled:
331 When set it will enable all functions with either the function
332 tracer, or if configured, the function graph tracer. It will
333 keep a histogram of the number of functions that were called
334 and if the function graph tracer was configured, it will also keep
335 track of the time spent in those functions. The histogram
336 content can be displayed in the files:
338 trace_stat/function<cpu> ( function0, function1, etc).
342 A directory that holds different tracing stats.
346 Enable dynamic trace points. See kprobetrace.txt.
350 Dynamic trace points stats. See kprobetrace.txt.
354 Used with the function graph tracer. This is the max depth
355 it will trace into a function. Setting this to a value of
356 one will show only the first kernel function that is called
361 This is for tools that read the raw format files. If an event in
362 the ring buffer references a string, only a pointer to the string
363 is recorded into the buffer and not the string itself. This prevents
364 tools from knowing what that string was. This file displays the string
365 and address for the string allowing tools to map the pointers to what
370 Only the pid of the task is recorded in a trace event unless
371 the event specifically saves the task comm as well. Ftrace
372 makes a cache of pid mappings to comms to try to display
373 comms for events. If a pid for a comm is not listed, then
374 "<...>" is displayed in the output.
376 If the option "record-cmd" is set to "0", then comms of tasks
377 will not be saved during recording. By default, it is enabled.
381 By default, 128 comms are saved (see "saved_cmdlines" above). To
382 increase or decrease the amount of comms that are cached, echo
383 in a the number of comms to cache, into this file.
387 If the option "record-tgid" is set, on each scheduling context switch
388 the Task Group ID of a task is saved in a table mapping the PID of
389 the thread to its TGID. By default, the "record-tgid" option is
394 This displays the "snapshot" buffer and also lets the user
395 take a snapshot of the current running trace.
396 See the "Snapshot" section below for more details.
400 When the stack tracer is activated, this will display the
401 maximum stack size it has encountered.
402 See the "Stack Trace" section below.
406 This displays the stack back trace of the largest stack
407 that was encountered when the stack tracer is activated.
408 See the "Stack Trace" section below.
412 This is similar to "set_ftrace_filter" but it limits what
413 functions the stack tracer will check.
417 Whenever an event is recorded into the ring buffer, a
418 "timestamp" is added. This stamp comes from a specified
419 clock. By default, ftrace uses the "local" clock. This
420 clock is very fast and strictly per cpu, but on some
421 systems it may not be monotonic with respect to other
422 CPUs. In other words, the local clocks may not be in sync
423 with local clocks on other CPUs.
425 Usual clocks for tracing::
428 [local] global counter x86-tsc
430 The clock with the square brackets around it is the one in effect.
433 Default clock, but may not be in sync across CPUs
436 This clock is in sync with all CPUs but may
437 be a bit slower than the local clock.
440 This is not a clock at all, but literally an atomic
441 counter. It counts up one by one, but is in sync
442 with all CPUs. This is useful when you need to
443 know exactly the order events occurred with respect to
444 each other on different CPUs.
447 This uses the jiffies counter and the time stamp
448 is relative to the time since boot up.
451 This makes ftrace use the same clock that perf uses.
452 Eventually perf will be able to read ftrace buffers
453 and this will help out in interleaving the data.
456 Architectures may define their own clocks. For
457 example, x86 uses its own TSC cycle clock here.
460 This uses the powerpc timebase register value.
461 This is in sync across CPUs and can also be used
462 to correlate events across hypervisor/guest if
466 This uses the fast monotonic clock (CLOCK_MONOTONIC)
467 which is monotonic and is subject to NTP rate adjustments.
470 This is the raw monotonic clock (CLOCK_MONOTONIC_RAW)
471 which is monotonic but is not subject to any rate adjustments
472 and ticks at the same rate as the hardware clocksource.
475 This is the boot clock (CLOCK_BOOTTIME) and is based on the
476 fast monotonic clock, but also accounts for time spent in
477 suspend. Since the clock access is designed for use in
478 tracing in the suspend path, some side effects are possible
479 if clock is accessed after the suspend time is accounted before
480 the fast mono clock is updated. In this case, the clock update
481 appears to happen slightly sooner than it normally would have.
482 Also on 32-bit systems, it's possible that the 64-bit boot offset
483 sees a partial update. These effects are rare and post
484 processing should be able to handle them. See comments in the
485 ktime_get_boot_fast_ns() function for more information.
487 To set a clock, simply echo the clock name into this file::
489 # echo global > trace_clock
493 This is a very useful file for synchronizing user space
494 with events happening in the kernel. Writing strings into
495 this file will be written into the ftrace buffer.
497 It is useful in applications to open this file at the start
498 of the application and just reference the file descriptor
501 void trace_write(const char *fmt, ...)
511 n = vsnprintf(buf, 256, fmt, ap);
514 write(trace_fd, buf, n);
519 trace_fd = open("trace_marker", WR_ONLY);
521 Note: Writing into the trace_marker file can also initiate triggers
522 that are written into /sys/kernel/tracing/events/ftrace/print/trigger
523 See "Event triggers" in Documentation/trace/events.rst and an
524 example in Documentation/trace/histogram.rst (Section 3.)
528 This is similar to trace_marker above, but is meant for for binary data
529 to be written to it, where a tool can be used to parse the data
534 Add dynamic tracepoints in programs.
539 Uprobe statistics. See uprobetrace.txt
543 This is a way to make multiple trace buffers where different
544 events can be recorded in different buffers.
545 See "Instances" section below.
549 This is the trace event directory. It holds event tracepoints
550 (also known as static tracepoints) that have been compiled
551 into the kernel. It shows what event tracepoints exist
552 and how they are grouped by system. There are "enable"
553 files at various levels that can enable the tracepoints
554 when a "1" is written to them.
556 See events.txt for more information.
560 By echoing in the event into this file, will enable that event.
562 See events.txt for more information.
566 A list of events that can be enabled in tracing.
568 See events.txt for more information.
572 Certain tracers may change the timestamp mode used when
573 logging trace events into the event buffer. Events with
574 different modes can coexist within a buffer but the mode in
575 effect when an event is logged determines which timestamp mode
576 is used for that event. The default timestamp mode is
579 Usual timestamp modes for tracing:
584 The timestamp mode with the square brackets around it is the
587 delta: Default timestamp mode - timestamp is a delta against
588 a per-buffer timestamp.
590 absolute: The timestamp is a full timestamp, not a delta
591 against some other value. As such it takes up more
592 space and is less efficient.
596 Directory for the Hardware Latency Detector.
597 See "Hardware Latency Detector" section below.
601 This is a directory that contains the trace per_cpu information.
603 per_cpu/cpu0/buffer_size_kb:
605 The ftrace buffer is defined per_cpu. That is, there's a separate
606 buffer for each CPU to allow writes to be done atomically,
607 and free from cache bouncing. These buffers may have different
608 size buffers. This file is similar to the buffer_size_kb
609 file, but it only displays or sets the buffer size for the
610 specific CPU. (here cpu0).
614 This is similar to the "trace" file, but it will only display
615 the data specific for the CPU. If written to, it only clears
616 the specific CPU buffer.
618 per_cpu/cpu0/trace_pipe
620 This is similar to the "trace_pipe" file, and is a consuming
621 read, but it will only display (and consume) the data specific
624 per_cpu/cpu0/trace_pipe_raw
626 For tools that can parse the ftrace ring buffer binary format,
627 the trace_pipe_raw file can be used to extract the data
628 from the ring buffer directly. With the use of the splice()
629 system call, the buffer data can be quickly transferred to
630 a file or to the network where a server is collecting the
633 Like trace_pipe, this is a consuming reader, where multiple
634 reads will always produce different data.
636 per_cpu/cpu0/snapshot:
638 This is similar to the main "snapshot" file, but will only
639 snapshot the current CPU (if supported). It only displays
640 the content of the snapshot for a given CPU, and if
641 written to, only clears this CPU buffer.
643 per_cpu/cpu0/snapshot_raw:
645 Similar to the trace_pipe_raw, but will read the binary format
646 from the snapshot buffer for the given CPU.
650 This displays certain stats about the ring buffer:
653 The number of events that are still in the buffer.
656 The number of lost events due to overwriting when
660 Should always be zero.
661 This gets set if so many events happened within a nested
662 event (ring buffer is re-entrant), that it fills the
663 buffer and starts dropping events.
666 Bytes actually read (not overwritten).
669 The oldest timestamp in the buffer
672 The current timestamp
675 Events lost due to overwrite option being off.
678 The number of events read.
683 Here is the list of current tracers that may be configured.
687 Function call tracer to trace all kernel functions.
691 Similar to the function tracer except that the
692 function tracer probes the functions on their entry
693 whereas the function graph tracer traces on both entry
694 and exit of the functions. It then provides the ability
695 to draw a graph of function calls similar to C code
700 The block tracer. The tracer used by the blktrace user
705 The Hardware Latency tracer is used to detect if the hardware
706 produces any latency. See "Hardware Latency Detector" section
711 Traces the areas that disable interrupts and saves
712 the trace with the longest max latency.
713 See tracing_max_latency. When a new max is recorded,
714 it replaces the old trace. It is best to view this
715 trace with the latency-format option enabled, which
716 happens automatically when the tracer is selected.
720 Similar to irqsoff but traces and records the amount of
721 time for which preemption is disabled.
725 Similar to irqsoff and preemptoff, but traces and
726 records the largest time for which irqs and/or preemption
731 Traces and records the max latency that it takes for
732 the highest priority task to get scheduled after
733 it has been woken up.
734 Traces all tasks as an average developer would expect.
738 Traces and records the max latency that it takes for just
739 RT tasks (as the current "wakeup" does). This is useful
740 for those interested in wake up timings of RT tasks.
744 Traces and records the max latency that it takes for
745 a SCHED_DEADLINE task to be woken (as the "wakeup" and
750 A special tracer that is used to trace binary module.
751 It will trace all the calls that a module makes to the
752 hardware. Everything it writes and reads from the I/O
757 This tracer can be configured when tracing likely/unlikely
758 calls within the kernel. It will trace when a likely and
759 unlikely branch is hit and if it was correct in its prediction
764 This is the "trace nothing" tracer. To remove all
765 tracers from tracing simply echo "nop" into
771 For most ftrace commands, failure modes are obvious and communicated
772 using standard return codes.
774 For other more involved commands, extended error information may be
775 available via the tracing/error_log file. For the commands that
776 support it, reading the tracing/error_log file after an error will
777 display more detailed information about what went wrong, if
778 information is available. The tracing/error_log file is a circular
779 error log displaying a small number (currently, 8) of ftrace errors
780 for the last (8) failed commands.
782 The extended error information and usage takes the form shown in
785 # echo xxx > /sys/kernel/debug/tracing/events/sched/sched_wakeup/trigger
786 echo: write error: Invalid argument
788 # cat /sys/kernel/debug/tracing/error_log
789 [ 5348.887237] location: error: Couldn't yyy: zzz
792 [ 7517.023364] location: error: Bad rrr: sss
796 To clear the error log, echo the empty string into it::
798 # echo > /sys/kernel/debug/tracing/error_log
800 Examples of using the tracer
801 ----------------------------
803 Here are typical examples of using the tracers when controlling
804 them only with the tracefs interface (without using any
805 user-land utilities).
810 Here is an example of the output format of the file "trace"::
814 # entries-in-buffer/entries-written: 140080/250280 #P:4
817 # / _----=> need-resched
818 # | / _---=> hardirq/softirq
819 # || / _--=> preempt-depth
821 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
823 bash-1977 [000] .... 17284.993652: sys_close <-system_call_fastpath
824 bash-1977 [000] .... 17284.993653: __close_fd <-sys_close
825 bash-1977 [000] .... 17284.993653: _raw_spin_lock <-__close_fd
826 sshd-1974 [003] .... 17284.993653: __srcu_read_unlock <-fsnotify
827 bash-1977 [000] .... 17284.993654: add_preempt_count <-_raw_spin_lock
828 bash-1977 [000] ...1 17284.993655: _raw_spin_unlock <-__close_fd
829 bash-1977 [000] ...1 17284.993656: sub_preempt_count <-_raw_spin_unlock
830 bash-1977 [000] .... 17284.993657: filp_close <-__close_fd
831 bash-1977 [000] .... 17284.993657: dnotify_flush <-filp_close
832 sshd-1974 [003] .... 17284.993658: sys_select <-system_call_fastpath
835 A header is printed with the tracer name that is represented by
836 the trace. In this case the tracer is "function". Then it shows the
837 number of events in the buffer as well as the total number of entries
838 that were written. The difference is the number of entries that were
839 lost due to the buffer filling up (250280 - 140080 = 110200 events
842 The header explains the content of the events. Task name "bash", the task
843 PID "1977", the CPU that it was running on "000", the latency format
844 (explained below), the timestamp in <secs>.<usecs> format, the
845 function name that was traced "sys_close" and the parent function that
846 called this function "system_call_fastpath". The timestamp is the time
847 at which the function was entered.
852 When the latency-format option is enabled or when one of the latency
853 tracers is set, the trace file gives somewhat more information to see
854 why a latency happened. Here is a typical trace::
858 # irqsoff latency trace v1.1.5 on 3.8.0-test+
859 # --------------------------------------------------------------------
860 # latency: 259 us, #4/4, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
862 # | task: ps-6143 (uid:0 nice:0 policy:0 rt_prio:0)
864 # => started at: __lock_task_sighand
865 # => ended at: _raw_spin_unlock_irqrestore
869 # / _-----=> irqs-off
870 # | / _----=> need-resched
871 # || / _---=> hardirq/softirq
872 # ||| / _--=> preempt-depth
874 # cmd pid ||||| time | caller
876 ps-6143 2d... 0us!: trace_hardirqs_off <-__lock_task_sighand
877 ps-6143 2d..1 259us+: trace_hardirqs_on <-_raw_spin_unlock_irqrestore
878 ps-6143 2d..1 263us+: time_hardirqs_on <-_raw_spin_unlock_irqrestore
879 ps-6143 2d..1 306us : <stack trace>
880 => trace_hardirqs_on_caller
882 => _raw_spin_unlock_irqrestore
889 => system_call_fastpath
892 This shows that the current tracer is "irqsoff" tracing the time
893 for which interrupts were disabled. It gives the trace version (which
894 never changes) and the version of the kernel upon which this was executed on
895 (3.8). Then it displays the max latency in microseconds (259 us). The number
896 of trace entries displayed and the total number (both are four: #4/4).
897 VP, KP, SP, and HP are always zero and are reserved for later use.
898 #P is the number of online CPUs (#P:4).
900 The task is the process that was running when the latency
901 occurred. (ps pid: 6143).
903 The start and stop (the functions in which the interrupts were
904 disabled and enabled respectively) that caused the latencies:
906 - __lock_task_sighand is where the interrupts were disabled.
907 - _raw_spin_unlock_irqrestore is where they were enabled again.
909 The next lines after the header are the trace itself. The header
910 explains which is which.
912 cmd: The name of the process in the trace.
914 pid: The PID of that process.
916 CPU#: The CPU which the process was running on.
918 irqs-off: 'd' interrupts are disabled. '.' otherwise.
919 .. caution:: If the architecture does not support a way to
920 read the irq flags variable, an 'X' will always
924 - 'N' both TIF_NEED_RESCHED and PREEMPT_NEED_RESCHED is set,
925 - 'n' only TIF_NEED_RESCHED is set,
926 - 'p' only PREEMPT_NEED_RESCHED is set,
930 - 'Z' - NMI occurred inside a hardirq
931 - 'z' - NMI is running
932 - 'H' - hard irq occurred inside a softirq.
933 - 'h' - hard irq is running
934 - 's' - soft irq is running
935 - '.' - normal context.
937 preempt-depth: The level of preempt_disabled
939 The above is mostly meaningful for kernel developers.
942 When the latency-format option is enabled, the trace file
943 output includes a timestamp relative to the start of the
944 trace. This differs from the output when latency-format
945 is disabled, which includes an absolute timestamp.
948 This is just to help catch your eye a bit better. And
949 needs to be fixed to be only relative to the same CPU.
950 The marks are determined by the difference between this
951 current trace and the next trace.
953 - '$' - greater than 1 second
954 - '@' - greater than 100 millisecond
955 - '*' - greater than 10 millisecond
956 - '#' - greater than 1000 microsecond
957 - '!' - greater than 100 microsecond
958 - '+' - greater than 10 microsecond
959 - ' ' - less than or equal to 10 microsecond.
961 The rest is the same as the 'trace' file.
963 Note, the latency tracers will usually end with a back trace
964 to easily find where the latency occurred.
969 The trace_options file (or the options directory) is used to control
970 what gets printed in the trace output, or manipulate the tracers.
971 To see what is available, simply cat the file::
1002 To disable one of the options, echo in the option prepended with
1005 echo noprint-parent > trace_options
1007 To enable an option, leave off the "no"::
1009 echo sym-offset > trace_options
1011 Here are the available options:
1014 On function traces, display the calling (parent)
1015 function as well as the function being traced.
1019 bash-4000 [01] 1477.606694: simple_strtoul <-kstrtoul
1022 bash-4000 [01] 1477.606694: simple_strtoul
1026 Display not only the function name, but also the
1027 offset in the function. For example, instead of
1028 seeing just "ktime_get", you will see
1029 "ktime_get+0xb/0x20".
1033 bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0
1036 This will also display the function address as well
1037 as the function name.
1041 bash-4000 [01] 1477.606694: simple_strtoul <c0339346>
1044 This deals with the trace file when the
1045 latency-format option is enabled.
1048 bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \
1049 (+0.000ms): simple_strtoul (kstrtoul)
1052 This will display raw numbers. This option is best for
1053 use with user applications that can translate the raw
1054 numbers better than having it done in the kernel.
1057 Similar to raw, but the numbers will be in a hexadecimal format.
1060 This will print out the formats in raw binary.
1063 When set, reading trace_pipe will not block when polled.
1066 Can disable trace_printk() from writing into the buffer.
1069 It is sometimes confusing when the CPU buffers are full
1070 and one CPU buffer had a lot of events recently, thus
1071 a shorter time frame, were another CPU may have only had
1072 a few events, which lets it have older events. When
1073 the trace is reported, it shows the oldest events first,
1074 and it may look like only one CPU ran (the one with the
1075 oldest events). When the annotate option is set, it will
1076 display when a new CPU buffer started::
1078 <idle>-0 [001] dNs4 21169.031481: wake_up_idle_cpu <-add_timer_on
1079 <idle>-0 [001] dNs4 21169.031482: _raw_spin_unlock_irqrestore <-add_timer_on
1080 <idle>-0 [001] .Ns4 21169.031484: sub_preempt_count <-_raw_spin_unlock_irqrestore
1081 ##### CPU 2 buffer started ####
1082 <idle>-0 [002] .N.1 21169.031484: rcu_idle_exit <-cpu_idle
1083 <idle>-0 [001] .Ns3 21169.031484: _raw_spin_unlock <-clocksource_watchdog
1084 <idle>-0 [001] .Ns3 21169.031485: sub_preempt_count <-_raw_spin_unlock
1087 This option changes the trace. It records a
1088 stacktrace of the current user space thread after
1092 when user stacktrace are enabled, look up which
1093 object the address belongs to, and print a
1094 relative address. This is especially useful when
1095 ASLR is on, otherwise you don't get a chance to
1096 resolve the address to object/file/line after
1097 the app is no longer running
1099 The lookup is performed when you read
1100 trace,trace_pipe. Example::
1102 a.out-1623 [000] 40874.465068: /root/a.out[+0x480] <-/root/a.out[+0
1103 x494] <- /root/a.out[+0x4a8] <- /lib/libc-2.7.so[+0x1e1a6]
1107 When set, trace_printk()s will only show the format
1108 and not their parameters (if trace_bprintk() or
1109 trace_bputs() was used to save the trace_printk()).
1112 Show only the event data. Hides the comm, PID,
1113 timestamp, CPU, and other useful data.
1116 This option changes the trace output. When it is enabled,
1117 the trace displays additional information about the
1118 latency, as described in "Latency trace format".
1121 When any event or tracer is enabled, a hook is enabled
1122 in the sched_switch trace point to fill comm cache
1123 with mapped pids and comms. But this may cause some
1124 overhead, and if you only care about pids, and not the
1125 name of the task, disabling this option can lower the
1126 impact of tracing. See "saved_cmdlines".
1129 When any event or tracer is enabled, a hook is enabled
1130 in the sched_switch trace point to fill the cache of
1131 mapped Thread Group IDs (TGID) mapping to pids. See
1135 This controls what happens when the trace buffer is
1136 full. If "1" (default), the oldest events are
1137 discarded and overwritten. If "0", then the newest
1138 events are discarded.
1139 (see per_cpu/cpu0/stats for overrun and dropped)
1142 When the free_buffer is closed, tracing will
1143 stop (tracing_on set to 0).
1146 Shows the interrupt, preempt count, need resched data.
1147 When disabled, the trace looks like::
1151 # entries-in-buffer/entries-written: 144405/9452052 #P:4
1153 # TASK-PID CPU# TIMESTAMP FUNCTION
1155 <idle>-0 [002] 23636.756054: ttwu_do_activate.constprop.89 <-try_to_wake_up
1156 <idle>-0 [002] 23636.756054: activate_task <-ttwu_do_activate.constprop.89
1157 <idle>-0 [002] 23636.756055: enqueue_task <-activate_task
1161 When set, the trace_marker is writable (only by root).
1162 When disabled, the trace_marker will error with EINVAL
1166 When set, tasks with PIDs listed in set_event_pid will have
1167 the PIDs of their children added to set_event_pid when those
1168 tasks fork. Also, when tasks with PIDs in set_event_pid exit,
1169 their PIDs will be removed from the file.
1172 The latency tracers will enable function tracing
1173 if this option is enabled (default it is). When
1174 it is disabled, the latency tracers do not trace
1175 functions. This keeps the overhead of the tracer down
1176 when performing latency tests.
1179 When set, tasks with PIDs listed in set_ftrace_pid will
1180 have the PIDs of their children added to set_ftrace_pid
1181 when those tasks fork. Also, when tasks with PIDs in
1182 set_ftrace_pid exit, their PIDs will be removed from the
1186 When set, the latency tracers (irqsoff, wakeup, etc) will
1187 use function graph tracing instead of function tracing.
1190 When set, a stack trace is recorded after any trace event
1194 Enable branch tracing with the tracer. This enables branch
1195 tracer along with the currently set tracer. Enabling this
1196 with the "nop" tracer is the same as just enabling the
1199 .. tip:: Some tracers have their own options. They only appear in this
1200 file when the tracer is active. They always appear in the
1204 Here are the per tracer options:
1206 Options for function tracer:
1209 When set, a stack trace is recorded after every
1210 function that is recorded. NOTE! Limit the functions
1211 that are recorded before enabling this, with
1212 "set_ftrace_filter" otherwise the system performance
1213 will be critically degraded. Remember to disable
1214 this option before clearing the function filter.
1216 Options for function_graph tracer:
1218 Since the function_graph tracer has a slightly different output
1219 it has its own options to control what is displayed.
1222 When set, the "overrun" of the graph stack is
1223 displayed after each function traced. The
1224 overrun, is when the stack depth of the calls
1225 is greater than what is reserved for each task.
1226 Each task has a fixed array of functions to
1227 trace in the call graph. If the depth of the
1228 calls exceeds that, the function is not traced.
1229 The overrun is the number of functions missed
1230 due to exceeding this array.
1233 When set, the CPU number of the CPU where the trace
1234 occurred is displayed.
1237 When set, if the function takes longer than
1238 A certain amount, then a delay marker is
1239 displayed. See "delay" above, under the
1243 Unlike other tracers, the process' command line
1244 is not displayed by default, but instead only
1245 when a task is traced in and out during a context
1246 switch. Enabling this options has the command
1247 of each process displayed at every line.
1250 At the end of each function (the return)
1251 the duration of the amount of time in the
1252 function is displayed in microseconds.
1255 When set, the timestamp is displayed at each line.
1258 When disabled, functions that happen inside an
1259 interrupt will not be traced.
1262 When set, the return event will include the function
1263 that it represents. By default this is off, and
1264 only a closing curly bracket "}" is displayed for
1265 the return of a function.
1268 When running function graph tracer, to include
1269 the time a task schedules out in its function.
1270 When enabled, it will account time the task has been
1271 scheduled out as part of the function call.
1274 When running function profiler with function graph tracer,
1275 to include the time to call nested functions. When this is
1276 not set, the time reported for the function will only
1277 include the time the function itself executed for, not the
1278 time for functions that it called.
1280 Options for blk tracer:
1283 Shows a more minimalistic output.
1289 When interrupts are disabled, the CPU can not react to any other
1290 external event (besides NMIs and SMIs). This prevents the timer
1291 interrupt from triggering or the mouse interrupt from letting
1292 the kernel know of a new mouse event. The result is a latency
1293 with the reaction time.
1295 The irqsoff tracer tracks the time for which interrupts are
1296 disabled. When a new maximum latency is hit, the tracer saves
1297 the trace leading up to that latency point so that every time a
1298 new maximum is reached, the old saved trace is discarded and the
1301 To reset the maximum, echo 0 into tracing_max_latency. Here is
1304 # echo 0 > options/function-trace
1305 # echo irqsoff > current_tracer
1306 # echo 1 > tracing_on
1307 # echo 0 > tracing_max_latency
1310 # echo 0 > tracing_on
1314 # irqsoff latency trace v1.1.5 on 3.8.0-test+
1315 # --------------------------------------------------------------------
1316 # latency: 16 us, #4/4, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1318 # | task: swapper/0-0 (uid:0 nice:0 policy:0 rt_prio:0)
1320 # => started at: run_timer_softirq
1321 # => ended at: run_timer_softirq
1325 # / _-----=> irqs-off
1326 # | / _----=> need-resched
1327 # || / _---=> hardirq/softirq
1328 # ||| / _--=> preempt-depth
1330 # cmd pid ||||| time | caller
1332 <idle>-0 0d.s2 0us+: _raw_spin_lock_irq <-run_timer_softirq
1333 <idle>-0 0dNs3 17us : _raw_spin_unlock_irq <-run_timer_softirq
1334 <idle>-0 0dNs3 17us+: trace_hardirqs_on <-run_timer_softirq
1335 <idle>-0 0dNs3 25us : <stack trace>
1336 => _raw_spin_unlock_irq
1337 => run_timer_softirq
1342 => smp_apic_timer_interrupt
1343 => apic_timer_interrupt
1348 => x86_64_start_reservations
1349 => x86_64_start_kernel
1351 Here we see that that we had a latency of 16 microseconds (which is
1352 very good). The _raw_spin_lock_irq in run_timer_softirq disabled
1353 interrupts. The difference between the 16 and the displayed
1354 timestamp 25us occurred because the clock was incremented
1355 between the time of recording the max latency and the time of
1356 recording the function that had that latency.
1358 Note the above example had function-trace not set. If we set
1359 function-trace, we get a much larger output::
1361 with echo 1 > options/function-trace
1365 # irqsoff latency trace v1.1.5 on 3.8.0-test+
1366 # --------------------------------------------------------------------
1367 # latency: 71 us, #168/168, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1369 # | task: bash-2042 (uid:0 nice:0 policy:0 rt_prio:0)
1371 # => started at: ata_scsi_queuecmd
1372 # => ended at: ata_scsi_queuecmd
1376 # / _-----=> irqs-off
1377 # | / _----=> need-resched
1378 # || / _---=> hardirq/softirq
1379 # ||| / _--=> preempt-depth
1381 # cmd pid ||||| time | caller
1383 bash-2042 3d... 0us : _raw_spin_lock_irqsave <-ata_scsi_queuecmd
1384 bash-2042 3d... 0us : add_preempt_count <-_raw_spin_lock_irqsave
1385 bash-2042 3d..1 1us : ata_scsi_find_dev <-ata_scsi_queuecmd
1386 bash-2042 3d..1 1us : __ata_scsi_find_dev <-ata_scsi_find_dev
1387 bash-2042 3d..1 2us : ata_find_dev.part.14 <-__ata_scsi_find_dev
1388 bash-2042 3d..1 2us : ata_qc_new_init <-__ata_scsi_queuecmd
1389 bash-2042 3d..1 3us : ata_sg_init <-__ata_scsi_queuecmd
1390 bash-2042 3d..1 4us : ata_scsi_rw_xlat <-__ata_scsi_queuecmd
1391 bash-2042 3d..1 4us : ata_build_rw_tf <-ata_scsi_rw_xlat
1393 bash-2042 3d..1 67us : delay_tsc <-__delay
1394 bash-2042 3d..1 67us : add_preempt_count <-delay_tsc
1395 bash-2042 3d..2 67us : sub_preempt_count <-delay_tsc
1396 bash-2042 3d..1 67us : add_preempt_count <-delay_tsc
1397 bash-2042 3d..2 68us : sub_preempt_count <-delay_tsc
1398 bash-2042 3d..1 68us+: ata_bmdma_start <-ata_bmdma_qc_issue
1399 bash-2042 3d..1 71us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd
1400 bash-2042 3d..1 71us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd
1401 bash-2042 3d..1 72us+: trace_hardirqs_on <-ata_scsi_queuecmd
1402 bash-2042 3d..1 120us : <stack trace>
1403 => _raw_spin_unlock_irqrestore
1404 => ata_scsi_queuecmd
1405 => scsi_dispatch_cmd
1407 => __blk_run_queue_uncond
1410 => generic_make_request
1413 => __ext3_get_inode_loc
1422 => user_path_at_empty
1427 => system_call_fastpath
1430 Here we traced a 71 microsecond latency. But we also see all the
1431 functions that were called during that time. Note that by
1432 enabling function tracing, we incur an added overhead. This
1433 overhead may extend the latency times. But nevertheless, this
1434 trace has provided some very helpful debugging information.
1436 If we prefer function graph output instead of function, we can set
1437 display-graph option::
1439 with echo 1 > options/display-graph
1443 # irqsoff latency trace v1.1.5 on 4.20.0-rc6+
1444 # --------------------------------------------------------------------
1445 # latency: 3751 us, #274/274, CPU#0 | (M:desktop VP:0, KP:0, SP:0 HP:0 #P:4)
1447 # | task: bash-1507 (uid:0 nice:0 policy:0 rt_prio:0)
1449 # => started at: free_debug_processing
1450 # => ended at: return_to_handler
1454 # / _----=> need-resched
1455 # | / _---=> hardirq/softirq
1456 # || / _--=> preempt-depth
1458 # REL TIME CPU TASK/PID |||| DURATION FUNCTION CALLS
1459 # | | | | |||| | | | | | |
1460 0 us | 0) bash-1507 | d... | 0.000 us | _raw_spin_lock_irqsave();
1461 0 us | 0) bash-1507 | d..1 | 0.378 us | do_raw_spin_trylock();
1462 1 us | 0) bash-1507 | d..2 | | set_track() {
1463 2 us | 0) bash-1507 | d..2 | | save_stack_trace() {
1464 2 us | 0) bash-1507 | d..2 | | __save_stack_trace() {
1465 3 us | 0) bash-1507 | d..2 | | __unwind_start() {
1466 3 us | 0) bash-1507 | d..2 | | get_stack_info() {
1467 3 us | 0) bash-1507 | d..2 | 0.351 us | in_task_stack();
1468 4 us | 0) bash-1507 | d..2 | 1.107 us | }
1470 3750 us | 0) bash-1507 | d..1 | 0.516 us | do_raw_spin_unlock();
1471 3750 us | 0) bash-1507 | d..1 | 0.000 us | _raw_spin_unlock_irqrestore();
1472 3764 us | 0) bash-1507 | d..1 | 0.000 us | tracer_hardirqs_on();
1473 bash-1507 0d..1 3792us : <stack trace>
1474 => free_debug_processing
1483 => search_binary_handler
1484 => __do_execve_file.isra.32
1487 => entry_SYSCALL_64_after_hwframe
1492 When preemption is disabled, we may be able to receive
1493 interrupts but the task cannot be preempted and a higher
1494 priority task must wait for preemption to be enabled again
1495 before it can preempt a lower priority task.
1497 The preemptoff tracer traces the places that disable preemption.
1498 Like the irqsoff tracer, it records the maximum latency for
1499 which preemption was disabled. The control of preemptoff tracer
1500 is much like the irqsoff tracer.
1503 # echo 0 > options/function-trace
1504 # echo preemptoff > current_tracer
1505 # echo 1 > tracing_on
1506 # echo 0 > tracing_max_latency
1509 # echo 0 > tracing_on
1511 # tracer: preemptoff
1513 # preemptoff latency trace v1.1.5 on 3.8.0-test+
1514 # --------------------------------------------------------------------
1515 # latency: 46 us, #4/4, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1517 # | task: sshd-1991 (uid:0 nice:0 policy:0 rt_prio:0)
1519 # => started at: do_IRQ
1520 # => ended at: do_IRQ
1524 # / _-----=> irqs-off
1525 # | / _----=> need-resched
1526 # || / _---=> hardirq/softirq
1527 # ||| / _--=> preempt-depth
1529 # cmd pid ||||| time | caller
1531 sshd-1991 1d.h. 0us+: irq_enter <-do_IRQ
1532 sshd-1991 1d..1 46us : irq_exit <-do_IRQ
1533 sshd-1991 1d..1 47us+: trace_preempt_on <-do_IRQ
1534 sshd-1991 1d..1 52us : <stack trace>
1535 => sub_preempt_count
1541 This has some more changes. Preemption was disabled when an
1542 interrupt came in (notice the 'h'), and was enabled on exit.
1543 But we also see that interrupts have been disabled when entering
1544 the preempt off section and leaving it (the 'd'). We do not know if
1545 interrupts were enabled in the mean time or shortly after this
1549 # tracer: preemptoff
1551 # preemptoff latency trace v1.1.5 on 3.8.0-test+
1552 # --------------------------------------------------------------------
1553 # latency: 83 us, #241/241, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1555 # | task: bash-1994 (uid:0 nice:0 policy:0 rt_prio:0)
1557 # => started at: wake_up_new_task
1558 # => ended at: task_rq_unlock
1562 # / _-----=> irqs-off
1563 # | / _----=> need-resched
1564 # || / _---=> hardirq/softirq
1565 # ||| / _--=> preempt-depth
1567 # cmd pid ||||| time | caller
1569 bash-1994 1d..1 0us : _raw_spin_lock_irqsave <-wake_up_new_task
1570 bash-1994 1d..1 0us : select_task_rq_fair <-select_task_rq
1571 bash-1994 1d..1 1us : __rcu_read_lock <-select_task_rq_fair
1572 bash-1994 1d..1 1us : source_load <-select_task_rq_fair
1573 bash-1994 1d..1 1us : source_load <-select_task_rq_fair
1575 bash-1994 1d..1 12us : irq_enter <-smp_apic_timer_interrupt
1576 bash-1994 1d..1 12us : rcu_irq_enter <-irq_enter
1577 bash-1994 1d..1 13us : add_preempt_count <-irq_enter
1578 bash-1994 1d.h1 13us : exit_idle <-smp_apic_timer_interrupt
1579 bash-1994 1d.h1 13us : hrtimer_interrupt <-smp_apic_timer_interrupt
1580 bash-1994 1d.h1 13us : _raw_spin_lock <-hrtimer_interrupt
1581 bash-1994 1d.h1 14us : add_preempt_count <-_raw_spin_lock
1582 bash-1994 1d.h2 14us : ktime_get_update_offsets <-hrtimer_interrupt
1584 bash-1994 1d.h1 35us : lapic_next_event <-clockevents_program_event
1585 bash-1994 1d.h1 35us : irq_exit <-smp_apic_timer_interrupt
1586 bash-1994 1d.h1 36us : sub_preempt_count <-irq_exit
1587 bash-1994 1d..2 36us : do_softirq <-irq_exit
1588 bash-1994 1d..2 36us : __do_softirq <-call_softirq
1589 bash-1994 1d..2 36us : __local_bh_disable <-__do_softirq
1590 bash-1994 1d.s2 37us : add_preempt_count <-_raw_spin_lock_irq
1591 bash-1994 1d.s3 38us : _raw_spin_unlock <-run_timer_softirq
1592 bash-1994 1d.s3 39us : sub_preempt_count <-_raw_spin_unlock
1593 bash-1994 1d.s2 39us : call_timer_fn <-run_timer_softirq
1595 bash-1994 1dNs2 81us : cpu_needs_another_gp <-rcu_process_callbacks
1596 bash-1994 1dNs2 82us : __local_bh_enable <-__do_softirq
1597 bash-1994 1dNs2 82us : sub_preempt_count <-__local_bh_enable
1598 bash-1994 1dN.2 82us : idle_cpu <-irq_exit
1599 bash-1994 1dN.2 83us : rcu_irq_exit <-irq_exit
1600 bash-1994 1dN.2 83us : sub_preempt_count <-irq_exit
1601 bash-1994 1.N.1 84us : _raw_spin_unlock_irqrestore <-task_rq_unlock
1602 bash-1994 1.N.1 84us+: trace_preempt_on <-task_rq_unlock
1603 bash-1994 1.N.1 104us : <stack trace>
1604 => sub_preempt_count
1605 => _raw_spin_unlock_irqrestore
1613 The above is an example of the preemptoff trace with
1614 function-trace set. Here we see that interrupts were not disabled
1615 the entire time. The irq_enter code lets us know that we entered
1616 an interrupt 'h'. Before that, the functions being traced still
1617 show that it is not in an interrupt, but we can see from the
1618 functions themselves that this is not the case.
1623 Knowing the locations that have interrupts disabled or
1624 preemption disabled for the longest times is helpful. But
1625 sometimes we would like to know when either preemption and/or
1626 interrupts are disabled.
1628 Consider the following code::
1630 local_irq_disable();
1631 call_function_with_irqs_off();
1633 call_function_with_irqs_and_preemption_off();
1635 call_function_with_preemption_off();
1638 The irqsoff tracer will record the total length of
1639 call_function_with_irqs_off() and
1640 call_function_with_irqs_and_preemption_off().
1642 The preemptoff tracer will record the total length of
1643 call_function_with_irqs_and_preemption_off() and
1644 call_function_with_preemption_off().
1646 But neither will trace the time that interrupts and/or
1647 preemption is disabled. This total time is the time that we can
1648 not schedule. To record this time, use the preemptirqsoff
1651 Again, using this trace is much like the irqsoff and preemptoff
1655 # echo 0 > options/function-trace
1656 # echo preemptirqsoff > current_tracer
1657 # echo 1 > tracing_on
1658 # echo 0 > tracing_max_latency
1661 # echo 0 > tracing_on
1663 # tracer: preemptirqsoff
1665 # preemptirqsoff latency trace v1.1.5 on 3.8.0-test+
1666 # --------------------------------------------------------------------
1667 # latency: 100 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1669 # | task: ls-2230 (uid:0 nice:0 policy:0 rt_prio:0)
1671 # => started at: ata_scsi_queuecmd
1672 # => ended at: ata_scsi_queuecmd
1676 # / _-----=> irqs-off
1677 # | / _----=> need-resched
1678 # || / _---=> hardirq/softirq
1679 # ||| / _--=> preempt-depth
1681 # cmd pid ||||| time | caller
1683 ls-2230 3d... 0us+: _raw_spin_lock_irqsave <-ata_scsi_queuecmd
1684 ls-2230 3...1 100us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd
1685 ls-2230 3...1 101us+: trace_preempt_on <-ata_scsi_queuecmd
1686 ls-2230 3...1 111us : <stack trace>
1687 => sub_preempt_count
1688 => _raw_spin_unlock_irqrestore
1689 => ata_scsi_queuecmd
1690 => scsi_dispatch_cmd
1692 => __blk_run_queue_uncond
1695 => generic_make_request
1700 => htree_dirblock_to_tree
1701 => ext3_htree_fill_tree
1705 => system_call_fastpath
1708 The trace_hardirqs_off_thunk is called from assembly on x86 when
1709 interrupts are disabled in the assembly code. Without the
1710 function tracing, we do not know if interrupts were enabled
1711 within the preemption points. We do see that it started with
1714 Here is a trace with function-trace set::
1716 # tracer: preemptirqsoff
1718 # preemptirqsoff latency trace v1.1.5 on 3.8.0-test+
1719 # --------------------------------------------------------------------
1720 # latency: 161 us, #339/339, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1722 # | task: ls-2269 (uid:0 nice:0 policy:0 rt_prio:0)
1724 # => started at: schedule
1725 # => ended at: mutex_unlock
1729 # / _-----=> irqs-off
1730 # | / _----=> need-resched
1731 # || / _---=> hardirq/softirq
1732 # ||| / _--=> preempt-depth
1734 # cmd pid ||||| time | caller
1736 kworker/-59 3...1 0us : __schedule <-schedule
1737 kworker/-59 3d..1 0us : rcu_preempt_qs <-rcu_note_context_switch
1738 kworker/-59 3d..1 1us : add_preempt_count <-_raw_spin_lock_irq
1739 kworker/-59 3d..2 1us : deactivate_task <-__schedule
1740 kworker/-59 3d..2 1us : dequeue_task <-deactivate_task
1741 kworker/-59 3d..2 2us : update_rq_clock <-dequeue_task
1742 kworker/-59 3d..2 2us : dequeue_task_fair <-dequeue_task
1743 kworker/-59 3d..2 2us : update_curr <-dequeue_task_fair
1744 kworker/-59 3d..2 2us : update_min_vruntime <-update_curr
1745 kworker/-59 3d..2 3us : cpuacct_charge <-update_curr
1746 kworker/-59 3d..2 3us : __rcu_read_lock <-cpuacct_charge
1747 kworker/-59 3d..2 3us : __rcu_read_unlock <-cpuacct_charge
1748 kworker/-59 3d..2 3us : update_cfs_rq_blocked_load <-dequeue_task_fair
1749 kworker/-59 3d..2 4us : clear_buddies <-dequeue_task_fair
1750 kworker/-59 3d..2 4us : account_entity_dequeue <-dequeue_task_fair
1751 kworker/-59 3d..2 4us : update_min_vruntime <-dequeue_task_fair
1752 kworker/-59 3d..2 4us : update_cfs_shares <-dequeue_task_fair
1753 kworker/-59 3d..2 5us : hrtick_update <-dequeue_task_fair
1754 kworker/-59 3d..2 5us : wq_worker_sleeping <-__schedule
1755 kworker/-59 3d..2 5us : kthread_data <-wq_worker_sleeping
1756 kworker/-59 3d..2 5us : put_prev_task_fair <-__schedule
1757 kworker/-59 3d..2 6us : pick_next_task_fair <-pick_next_task
1758 kworker/-59 3d..2 6us : clear_buddies <-pick_next_task_fair
1759 kworker/-59 3d..2 6us : set_next_entity <-pick_next_task_fair
1760 kworker/-59 3d..2 6us : update_stats_wait_end <-set_next_entity
1761 ls-2269 3d..2 7us : finish_task_switch <-__schedule
1762 ls-2269 3d..2 7us : _raw_spin_unlock_irq <-finish_task_switch
1763 ls-2269 3d..2 8us : do_IRQ <-ret_from_intr
1764 ls-2269 3d..2 8us : irq_enter <-do_IRQ
1765 ls-2269 3d..2 8us : rcu_irq_enter <-irq_enter
1766 ls-2269 3d..2 9us : add_preempt_count <-irq_enter
1767 ls-2269 3d.h2 9us : exit_idle <-do_IRQ
1769 ls-2269 3d.h3 20us : sub_preempt_count <-_raw_spin_unlock
1770 ls-2269 3d.h2 20us : irq_exit <-do_IRQ
1771 ls-2269 3d.h2 21us : sub_preempt_count <-irq_exit
1772 ls-2269 3d..3 21us : do_softirq <-irq_exit
1773 ls-2269 3d..3 21us : __do_softirq <-call_softirq
1774 ls-2269 3d..3 21us+: __local_bh_disable <-__do_softirq
1775 ls-2269 3d.s4 29us : sub_preempt_count <-_local_bh_enable_ip
1776 ls-2269 3d.s5 29us : sub_preempt_count <-_local_bh_enable_ip
1777 ls-2269 3d.s5 31us : do_IRQ <-ret_from_intr
1778 ls-2269 3d.s5 31us : irq_enter <-do_IRQ
1779 ls-2269 3d.s5 31us : rcu_irq_enter <-irq_enter
1781 ls-2269 3d.s5 31us : rcu_irq_enter <-irq_enter
1782 ls-2269 3d.s5 32us : add_preempt_count <-irq_enter
1783 ls-2269 3d.H5 32us : exit_idle <-do_IRQ
1784 ls-2269 3d.H5 32us : handle_irq <-do_IRQ
1785 ls-2269 3d.H5 32us : irq_to_desc <-handle_irq
1786 ls-2269 3d.H5 33us : handle_fasteoi_irq <-handle_irq
1788 ls-2269 3d.s5 158us : _raw_spin_unlock_irqrestore <-rtl8139_poll
1789 ls-2269 3d.s3 158us : net_rps_action_and_irq_enable.isra.65 <-net_rx_action
1790 ls-2269 3d.s3 159us : __local_bh_enable <-__do_softirq
1791 ls-2269 3d.s3 159us : sub_preempt_count <-__local_bh_enable
1792 ls-2269 3d..3 159us : idle_cpu <-irq_exit
1793 ls-2269 3d..3 159us : rcu_irq_exit <-irq_exit
1794 ls-2269 3d..3 160us : sub_preempt_count <-irq_exit
1795 ls-2269 3d... 161us : __mutex_unlock_slowpath <-mutex_unlock
1796 ls-2269 3d... 162us+: trace_hardirqs_on <-mutex_unlock
1797 ls-2269 3d... 186us : <stack trace>
1798 => __mutex_unlock_slowpath
1805 => system_call_fastpath
1807 This is an interesting trace. It started with kworker running and
1808 scheduling out and ls taking over. But as soon as ls released the
1809 rq lock and enabled interrupts (but not preemption) an interrupt
1810 triggered. When the interrupt finished, it started running softirqs.
1811 But while the softirq was running, another interrupt triggered.
1812 When an interrupt is running inside a softirq, the annotation is 'H'.
1818 One common case that people are interested in tracing is the
1819 time it takes for a task that is woken to actually wake up.
1820 Now for non Real-Time tasks, this can be arbitrary. But tracing
1821 it none the less can be interesting.
1823 Without function tracing::
1825 # echo 0 > options/function-trace
1826 # echo wakeup > current_tracer
1827 # echo 1 > tracing_on
1828 # echo 0 > tracing_max_latency
1830 # echo 0 > tracing_on
1834 # wakeup latency trace v1.1.5 on 3.8.0-test+
1835 # --------------------------------------------------------------------
1836 # latency: 15 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1838 # | task: kworker/3:1H-312 (uid:0 nice:-20 policy:0 rt_prio:0)
1842 # / _-----=> irqs-off
1843 # | / _----=> need-resched
1844 # || / _---=> hardirq/softirq
1845 # ||| / _--=> preempt-depth
1847 # cmd pid ||||| time | caller
1849 <idle>-0 3dNs7 0us : 0:120:R + [003] 312:100:R kworker/3:1H
1850 <idle>-0 3dNs7 1us+: ttwu_do_activate.constprop.87 <-try_to_wake_up
1851 <idle>-0 3d..3 15us : __schedule <-schedule
1852 <idle>-0 3d..3 15us : 0:120:R ==> [003] 312:100:R kworker/3:1H
1854 The tracer only traces the highest priority task in the system
1855 to avoid tracing the normal circumstances. Here we see that
1856 the kworker with a nice priority of -20 (not very nice), took
1857 just 15 microseconds from the time it woke up, to the time it
1860 Non Real-Time tasks are not that interesting. A more interesting
1861 trace is to concentrate only on Real-Time tasks.
1866 In a Real-Time environment it is very important to know the
1867 wakeup time it takes for the highest priority task that is woken
1868 up to the time that it executes. This is also known as "schedule
1869 latency". I stress the point that this is about RT tasks. It is
1870 also important to know the scheduling latency of non-RT tasks,
1871 but the average schedule latency is better for non-RT tasks.
1872 Tools like LatencyTop are more appropriate for such
1875 Real-Time environments are interested in the worst case latency.
1876 That is the longest latency it takes for something to happen,
1877 and not the average. We can have a very fast scheduler that may
1878 only have a large latency once in a while, but that would not
1879 work well with Real-Time tasks. The wakeup_rt tracer was designed
1880 to record the worst case wakeups of RT tasks. Non-RT tasks are
1881 not recorded because the tracer only records one worst case and
1882 tracing non-RT tasks that are unpredictable will overwrite the
1883 worst case latency of RT tasks (just run the normal wakeup
1884 tracer for a while to see that effect).
1886 Since this tracer only deals with RT tasks, we will run this
1887 slightly differently than we did with the previous tracers.
1888 Instead of performing an 'ls', we will run 'sleep 1' under
1889 'chrt' which changes the priority of the task.
1892 # echo 0 > options/function-trace
1893 # echo wakeup_rt > current_tracer
1894 # echo 1 > tracing_on
1895 # echo 0 > tracing_max_latency
1897 # echo 0 > tracing_on
1903 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+
1904 # --------------------------------------------------------------------
1905 # latency: 5 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1907 # | task: sleep-2389 (uid:0 nice:0 policy:1 rt_prio:5)
1911 # / _-----=> irqs-off
1912 # | / _----=> need-resched
1913 # || / _---=> hardirq/softirq
1914 # ||| / _--=> preempt-depth
1916 # cmd pid ||||| time | caller
1918 <idle>-0 3d.h4 0us : 0:120:R + [003] 2389: 94:R sleep
1919 <idle>-0 3d.h4 1us+: ttwu_do_activate.constprop.87 <-try_to_wake_up
1920 <idle>-0 3d..3 5us : __schedule <-schedule
1921 <idle>-0 3d..3 5us : 0:120:R ==> [003] 2389: 94:R sleep
1924 Running this on an idle system, we see that it only took 5 microseconds
1925 to perform the task switch. Note, since the trace point in the schedule
1926 is before the actual "switch", we stop the tracing when the recorded task
1927 is about to schedule in. This may change if we add a new marker at the
1928 end of the scheduler.
1930 Notice that the recorded task is 'sleep' with the PID of 2389
1931 and it has an rt_prio of 5. This priority is user-space priority
1932 and not the internal kernel priority. The policy is 1 for
1933 SCHED_FIFO and 2 for SCHED_RR.
1935 Note, that the trace data shows the internal priority (99 - rtprio).
1938 <idle>-0 3d..3 5us : 0:120:R ==> [003] 2389: 94:R sleep
1940 The 0:120:R means idle was running with a nice priority of 0 (120 - 120)
1941 and in the running state 'R'. The sleep task was scheduled in with
1942 2389: 94:R. That is the priority is the kernel rtprio (99 - 5 = 94)
1943 and it too is in the running state.
1945 Doing the same with chrt -r 5 and function-trace set.
1948 echo 1 > options/function-trace
1952 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+
1953 # --------------------------------------------------------------------
1954 # latency: 29 us, #85/85, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1956 # | task: sleep-2448 (uid:0 nice:0 policy:1 rt_prio:5)
1960 # / _-----=> irqs-off
1961 # | / _----=> need-resched
1962 # || / _---=> hardirq/softirq
1963 # ||| / _--=> preempt-depth
1965 # cmd pid ||||| time | caller
1967 <idle>-0 3d.h4 1us+: 0:120:R + [003] 2448: 94:R sleep
1968 <idle>-0 3d.h4 2us : ttwu_do_activate.constprop.87 <-try_to_wake_up
1969 <idle>-0 3d.h3 3us : check_preempt_curr <-ttwu_do_wakeup
1970 <idle>-0 3d.h3 3us : resched_curr <-check_preempt_curr
1971 <idle>-0 3dNh3 4us : task_woken_rt <-ttwu_do_wakeup
1972 <idle>-0 3dNh3 4us : _raw_spin_unlock <-try_to_wake_up
1973 <idle>-0 3dNh3 4us : sub_preempt_count <-_raw_spin_unlock
1974 <idle>-0 3dNh2 5us : ttwu_stat <-try_to_wake_up
1975 <idle>-0 3dNh2 5us : _raw_spin_unlock_irqrestore <-try_to_wake_up
1976 <idle>-0 3dNh2 6us : sub_preempt_count <-_raw_spin_unlock_irqrestore
1977 <idle>-0 3dNh1 6us : _raw_spin_lock <-__run_hrtimer
1978 <idle>-0 3dNh1 6us : add_preempt_count <-_raw_spin_lock
1979 <idle>-0 3dNh2 7us : _raw_spin_unlock <-hrtimer_interrupt
1980 <idle>-0 3dNh2 7us : sub_preempt_count <-_raw_spin_unlock
1981 <idle>-0 3dNh1 7us : tick_program_event <-hrtimer_interrupt
1982 <idle>-0 3dNh1 7us : clockevents_program_event <-tick_program_event
1983 <idle>-0 3dNh1 8us : ktime_get <-clockevents_program_event
1984 <idle>-0 3dNh1 8us : lapic_next_event <-clockevents_program_event
1985 <idle>-0 3dNh1 8us : irq_exit <-smp_apic_timer_interrupt
1986 <idle>-0 3dNh1 9us : sub_preempt_count <-irq_exit
1987 <idle>-0 3dN.2 9us : idle_cpu <-irq_exit
1988 <idle>-0 3dN.2 9us : rcu_irq_exit <-irq_exit
1989 <idle>-0 3dN.2 10us : rcu_eqs_enter_common.isra.45 <-rcu_irq_exit
1990 <idle>-0 3dN.2 10us : sub_preempt_count <-irq_exit
1991 <idle>-0 3.N.1 11us : rcu_idle_exit <-cpu_idle
1992 <idle>-0 3dN.1 11us : rcu_eqs_exit_common.isra.43 <-rcu_idle_exit
1993 <idle>-0 3.N.1 11us : tick_nohz_idle_exit <-cpu_idle
1994 <idle>-0 3dN.1 12us : menu_hrtimer_cancel <-tick_nohz_idle_exit
1995 <idle>-0 3dN.1 12us : ktime_get <-tick_nohz_idle_exit
1996 <idle>-0 3dN.1 12us : tick_do_update_jiffies64 <-tick_nohz_idle_exit
1997 <idle>-0 3dN.1 13us : cpu_load_update_nohz <-tick_nohz_idle_exit
1998 <idle>-0 3dN.1 13us : _raw_spin_lock <-cpu_load_update_nohz
1999 <idle>-0 3dN.1 13us : add_preempt_count <-_raw_spin_lock
2000 <idle>-0 3dN.2 13us : __cpu_load_update <-cpu_load_update_nohz
2001 <idle>-0 3dN.2 14us : sched_avg_update <-__cpu_load_update
2002 <idle>-0 3dN.2 14us : _raw_spin_unlock <-cpu_load_update_nohz
2003 <idle>-0 3dN.2 14us : sub_preempt_count <-_raw_spin_unlock
2004 <idle>-0 3dN.1 15us : calc_load_nohz_stop <-tick_nohz_idle_exit
2005 <idle>-0 3dN.1 15us : touch_softlockup_watchdog <-tick_nohz_idle_exit
2006 <idle>-0 3dN.1 15us : hrtimer_cancel <-tick_nohz_idle_exit
2007 <idle>-0 3dN.1 15us : hrtimer_try_to_cancel <-hrtimer_cancel
2008 <idle>-0 3dN.1 16us : lock_hrtimer_base.isra.18 <-hrtimer_try_to_cancel
2009 <idle>-0 3dN.1 16us : _raw_spin_lock_irqsave <-lock_hrtimer_base.isra.18
2010 <idle>-0 3dN.1 16us : add_preempt_count <-_raw_spin_lock_irqsave
2011 <idle>-0 3dN.2 17us : __remove_hrtimer <-remove_hrtimer.part.16
2012 <idle>-0 3dN.2 17us : hrtimer_force_reprogram <-__remove_hrtimer
2013 <idle>-0 3dN.2 17us : tick_program_event <-hrtimer_force_reprogram
2014 <idle>-0 3dN.2 18us : clockevents_program_event <-tick_program_event
2015 <idle>-0 3dN.2 18us : ktime_get <-clockevents_program_event
2016 <idle>-0 3dN.2 18us : lapic_next_event <-clockevents_program_event
2017 <idle>-0 3dN.2 19us : _raw_spin_unlock_irqrestore <-hrtimer_try_to_cancel
2018 <idle>-0 3dN.2 19us : sub_preempt_count <-_raw_spin_unlock_irqrestore
2019 <idle>-0 3dN.1 19us : hrtimer_forward <-tick_nohz_idle_exit
2020 <idle>-0 3dN.1 20us : ktime_add_safe <-hrtimer_forward
2021 <idle>-0 3dN.1 20us : ktime_add_safe <-hrtimer_forward
2022 <idle>-0 3dN.1 20us : hrtimer_start_range_ns <-hrtimer_start_expires.constprop.11
2023 <idle>-0 3dN.1 20us : __hrtimer_start_range_ns <-hrtimer_start_range_ns
2024 <idle>-0 3dN.1 21us : lock_hrtimer_base.isra.18 <-__hrtimer_start_range_ns
2025 <idle>-0 3dN.1 21us : _raw_spin_lock_irqsave <-lock_hrtimer_base.isra.18
2026 <idle>-0 3dN.1 21us : add_preempt_count <-_raw_spin_lock_irqsave
2027 <idle>-0 3dN.2 22us : ktime_add_safe <-__hrtimer_start_range_ns
2028 <idle>-0 3dN.2 22us : enqueue_hrtimer <-__hrtimer_start_range_ns
2029 <idle>-0 3dN.2 22us : tick_program_event <-__hrtimer_start_range_ns
2030 <idle>-0 3dN.2 23us : clockevents_program_event <-tick_program_event
2031 <idle>-0 3dN.2 23us : ktime_get <-clockevents_program_event
2032 <idle>-0 3dN.2 23us : lapic_next_event <-clockevents_program_event
2033 <idle>-0 3dN.2 24us : _raw_spin_unlock_irqrestore <-__hrtimer_start_range_ns
2034 <idle>-0 3dN.2 24us : sub_preempt_count <-_raw_spin_unlock_irqrestore
2035 <idle>-0 3dN.1 24us : account_idle_ticks <-tick_nohz_idle_exit
2036 <idle>-0 3dN.1 24us : account_idle_time <-account_idle_ticks
2037 <idle>-0 3.N.1 25us : sub_preempt_count <-cpu_idle
2038 <idle>-0 3.N.. 25us : schedule <-cpu_idle
2039 <idle>-0 3.N.. 25us : __schedule <-preempt_schedule
2040 <idle>-0 3.N.. 26us : add_preempt_count <-__schedule
2041 <idle>-0 3.N.1 26us : rcu_note_context_switch <-__schedule
2042 <idle>-0 3.N.1 26us : rcu_sched_qs <-rcu_note_context_switch
2043 <idle>-0 3dN.1 27us : rcu_preempt_qs <-rcu_note_context_switch
2044 <idle>-0 3.N.1 27us : _raw_spin_lock_irq <-__schedule
2045 <idle>-0 3dN.1 27us : add_preempt_count <-_raw_spin_lock_irq
2046 <idle>-0 3dN.2 28us : put_prev_task_idle <-__schedule
2047 <idle>-0 3dN.2 28us : pick_next_task_stop <-pick_next_task
2048 <idle>-0 3dN.2 28us : pick_next_task_rt <-pick_next_task
2049 <idle>-0 3dN.2 29us : dequeue_pushable_task <-pick_next_task_rt
2050 <idle>-0 3d..3 29us : __schedule <-preempt_schedule
2051 <idle>-0 3d..3 30us : 0:120:R ==> [003] 2448: 94:R sleep
2053 This isn't that big of a trace, even with function tracing enabled,
2054 so I included the entire trace.
2056 The interrupt went off while when the system was idle. Somewhere
2057 before task_woken_rt() was called, the NEED_RESCHED flag was set,
2058 this is indicated by the first occurrence of the 'N' flag.
2060 Latency tracing and events
2061 --------------------------
2062 As function tracing can induce a much larger latency, but without
2063 seeing what happens within the latency it is hard to know what
2064 caused it. There is a middle ground, and that is with enabling
2068 # echo 0 > options/function-trace
2069 # echo wakeup_rt > current_tracer
2070 # echo 1 > events/enable
2071 # echo 1 > tracing_on
2072 # echo 0 > tracing_max_latency
2074 # echo 0 > tracing_on
2078 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+
2079 # --------------------------------------------------------------------
2080 # latency: 6 us, #12/12, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
2082 # | task: sleep-5882 (uid:0 nice:0 policy:1 rt_prio:5)
2086 # / _-----=> irqs-off
2087 # | / _----=> need-resched
2088 # || / _---=> hardirq/softirq
2089 # ||| / _--=> preempt-depth
2091 # cmd pid ||||| time | caller
2093 <idle>-0 2d.h4 0us : 0:120:R + [002] 5882: 94:R sleep
2094 <idle>-0 2d.h4 0us : ttwu_do_activate.constprop.87 <-try_to_wake_up
2095 <idle>-0 2d.h4 1us : sched_wakeup: comm=sleep pid=5882 prio=94 success=1 target_cpu=002
2096 <idle>-0 2dNh2 1us : hrtimer_expire_exit: hrtimer=ffff88007796feb8
2097 <idle>-0 2.N.2 2us : power_end: cpu_id=2
2098 <idle>-0 2.N.2 3us : cpu_idle: state=4294967295 cpu_id=2
2099 <idle>-0 2dN.3 4us : hrtimer_cancel: hrtimer=ffff88007d50d5e0
2100 <idle>-0 2dN.3 4us : hrtimer_start: hrtimer=ffff88007d50d5e0 function=tick_sched_timer expires=34311211000000 softexpires=34311211000000
2101 <idle>-0 2.N.2 5us : rcu_utilization: Start context switch
2102 <idle>-0 2.N.2 5us : rcu_utilization: End context switch
2103 <idle>-0 2d..3 6us : __schedule <-schedule
2104 <idle>-0 2d..3 6us : 0:120:R ==> [002] 5882: 94:R sleep
2107 Hardware Latency Detector
2108 -------------------------
2110 The hardware latency detector is executed by enabling the "hwlat" tracer.
2112 NOTE, this tracer will affect the performance of the system as it will
2113 periodically make a CPU constantly busy with interrupts disabled.
2116 # echo hwlat > current_tracer
2122 # / _----=> need-resched
2123 # | / _---=> hardirq/softirq
2124 # || / _--=> preempt-depth
2126 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2128 <...>-3638 [001] d... 19452.055471: #1 inner/outer(us): 12/14 ts:1499801089.066141940
2129 <...>-3638 [003] d... 19454.071354: #2 inner/outer(us): 11/9 ts:1499801091.082164365
2130 <...>-3638 [002] dn.. 19461.126852: #3 inner/outer(us): 12/9 ts:1499801098.138150062
2131 <...>-3638 [001] d... 19488.340960: #4 inner/outer(us): 8/12 ts:1499801125.354139633
2132 <...>-3638 [003] d... 19494.388553: #5 inner/outer(us): 8/12 ts:1499801131.402150961
2133 <...>-3638 [003] d... 19501.283419: #6 inner/outer(us): 0/12 ts:1499801138.297435289 nmi-total:4 nmi-count:1
2136 The above output is somewhat the same in the header. All events will have
2137 interrupts disabled 'd'. Under the FUNCTION title there is:
2140 This is the count of events recorded that were greater than the
2141 tracing_threshold (See below).
2143 inner/outer(us): 12/14
2145 This shows two numbers as "inner latency" and "outer latency". The test
2146 runs in a loop checking a timestamp twice. The latency detected within
2147 the two timestamps is the "inner latency" and the latency detected
2148 after the previous timestamp and the next timestamp in the loop is
2149 the "outer latency".
2151 ts:1499801089.066141940
2153 The absolute timestamp that the event happened.
2155 nmi-total:4 nmi-count:1
2157 On architectures that support it, if an NMI comes in during the
2158 test, the time spent in NMI is reported in "nmi-total" (in
2161 All architectures that have NMIs will show the "nmi-count" if an
2162 NMI comes in during the test.
2167 This gets automatically set to "10" to represent 10
2168 microseconds. This is the threshold of latency that
2169 needs to be detected before the trace will be recorded.
2171 Note, when hwlat tracer is finished (another tracer is
2172 written into "current_tracer"), the original value for
2173 tracing_threshold is placed back into this file.
2175 hwlat_detector/width
2176 The length of time the test runs with interrupts disabled.
2178 hwlat_detector/window
2179 The length of time of the window which the test
2180 runs. That is, the test will run for "width"
2181 microseconds per "window" microseconds
2184 When the test is started. A kernel thread is created that
2185 runs the test. This thread will alternate between CPUs
2186 listed in the tracing_cpumask between each period
2187 (one "window"). To limit the test to specific CPUs
2188 set the mask in this file to only the CPUs that the test
2194 This tracer is the function tracer. Enabling the function tracer
2195 can be done from the debug file system. Make sure the
2196 ftrace_enabled is set; otherwise this tracer is a nop.
2197 See the "ftrace_enabled" section below.
2200 # sysctl kernel.ftrace_enabled=1
2201 # echo function > current_tracer
2202 # echo 1 > tracing_on
2204 # echo 0 > tracing_on
2208 # entries-in-buffer/entries-written: 24799/24799 #P:4
2211 # / _----=> need-resched
2212 # | / _---=> hardirq/softirq
2213 # || / _--=> preempt-depth
2215 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2217 bash-1994 [002] .... 3082.063030: mutex_unlock <-rb_simple_write
2218 bash-1994 [002] .... 3082.063031: __mutex_unlock_slowpath <-mutex_unlock
2219 bash-1994 [002] .... 3082.063031: __fsnotify_parent <-fsnotify_modify
2220 bash-1994 [002] .... 3082.063032: fsnotify <-fsnotify_modify
2221 bash-1994 [002] .... 3082.063032: __srcu_read_lock <-fsnotify
2222 bash-1994 [002] .... 3082.063032: add_preempt_count <-__srcu_read_lock
2223 bash-1994 [002] ...1 3082.063032: sub_preempt_count <-__srcu_read_lock
2224 bash-1994 [002] .... 3082.063033: __srcu_read_unlock <-fsnotify
2228 Note: function tracer uses ring buffers to store the above
2229 entries. The newest data may overwrite the oldest data.
2230 Sometimes using echo to stop the trace is not sufficient because
2231 the tracing could have overwritten the data that you wanted to
2232 record. For this reason, it is sometimes better to disable
2233 tracing directly from a program. This allows you to stop the
2234 tracing at the point that you hit the part that you are
2235 interested in. To disable the tracing directly from a C program,
2236 something like following code snippet can be used::
2240 int main(int argc, char *argv[]) {
2242 trace_fd = open(tracing_file("tracing_on"), O_WRONLY);
2244 if (condition_hit()) {
2245 write(trace_fd, "0", 1);
2251 Single thread tracing
2252 ---------------------
2254 By writing into set_ftrace_pid you can trace a
2255 single thread. For example::
2257 # cat set_ftrace_pid
2259 # echo 3111 > set_ftrace_pid
2260 # cat set_ftrace_pid
2262 # echo function > current_tracer
2266 # TASK-PID CPU# TIMESTAMP FUNCTION
2268 yum-updatesd-3111 [003] 1637.254676: finish_task_switch <-thread_return
2269 yum-updatesd-3111 [003] 1637.254681: hrtimer_cancel <-schedule_hrtimeout_range
2270 yum-updatesd-3111 [003] 1637.254682: hrtimer_try_to_cancel <-hrtimer_cancel
2271 yum-updatesd-3111 [003] 1637.254683: lock_hrtimer_base <-hrtimer_try_to_cancel
2272 yum-updatesd-3111 [003] 1637.254685: fget_light <-do_sys_poll
2273 yum-updatesd-3111 [003] 1637.254686: pipe_poll <-do_sys_poll
2274 # echo > set_ftrace_pid
2278 # TASK-PID CPU# TIMESTAMP FUNCTION
2280 ##### CPU 3 buffer started ####
2281 yum-updatesd-3111 [003] 1701.957688: free_poll_entry <-poll_freewait
2282 yum-updatesd-3111 [003] 1701.957689: remove_wait_queue <-free_poll_entry
2283 yum-updatesd-3111 [003] 1701.957691: fput <-free_poll_entry
2284 yum-updatesd-3111 [003] 1701.957692: audit_syscall_exit <-sysret_audit
2285 yum-updatesd-3111 [003] 1701.957693: path_put <-audit_syscall_exit
2287 If you want to trace a function when executing, you could use
2288 something like this simple program.
2293 #include <sys/types.h>
2294 #include <sys/stat.h>
2300 #define STR(x) _STR(x)
2301 #define MAX_PATH 256
2303 const char *find_tracefs(void)
2305 static char tracefs[MAX_PATH+1];
2306 static int tracefs_found;
2313 if ((fp = fopen("/proc/mounts","r")) == NULL) {
2314 perror("/proc/mounts");
2318 while (fscanf(fp, "%*s %"
2320 "s %99s %*s %*d %*d\n",
2321 tracefs, type) == 2) {
2322 if (strcmp(type, "tracefs") == 0)
2327 if (strcmp(type, "tracefs") != 0) {
2328 fprintf(stderr, "tracefs not mounted");
2332 strcat(tracefs, "/tracing/");
2338 const char *tracing_file(const char *file_name)
2340 static char trace_file[MAX_PATH+1];
2341 snprintf(trace_file, MAX_PATH, "%s/%s", find_tracefs(), file_name);
2345 int main (int argc, char **argv)
2355 ffd = open(tracing_file("current_tracer"), O_WRONLY);
2358 write(ffd, "nop", 3);
2360 fd = open(tracing_file("set_ftrace_pid"), O_WRONLY);
2361 s = sprintf(line, "%d\n", getpid());
2364 write(ffd, "function", 8);
2369 execvp(argv[1], argv+1);
2375 Or this simple script!
2380 tracefs=`sed -ne 's/^tracefs \(.*\) tracefs.*/\1/p' /proc/mounts`
2381 echo nop > $tracefs/tracing/current_tracer
2382 echo 0 > $tracefs/tracing/tracing_on
2383 echo $$ > $tracefs/tracing/set_ftrace_pid
2384 echo function > $tracefs/tracing/current_tracer
2385 echo 1 > $tracefs/tracing/tracing_on
2389 function graph tracer
2390 ---------------------------
2392 This tracer is similar to the function tracer except that it
2393 probes a function on its entry and its exit. This is done by
2394 using a dynamically allocated stack of return addresses in each
2395 task_struct. On function entry the tracer overwrites the return
2396 address of each function traced to set a custom probe. Thus the
2397 original return address is stored on the stack of return address
2400 Probing on both ends of a function leads to special features
2403 - measure of a function's time execution
2404 - having a reliable call stack to draw function calls graph
2406 This tracer is useful in several situations:
2408 - you want to find the reason of a strange kernel behavior and
2409 need to see what happens in detail on any areas (or specific
2412 - you are experiencing weird latencies but it's difficult to
2415 - you want to find quickly which path is taken by a specific
2418 - you just want to peek inside a working kernel and want to see
2423 # tracer: function_graph
2425 # CPU DURATION FUNCTION CALLS
2429 0) | do_sys_open() {
2431 0) | kmem_cache_alloc() {
2432 0) 1.382 us | __might_sleep();
2434 0) | strncpy_from_user() {
2435 0) | might_fault() {
2436 0) 1.389 us | __might_sleep();
2441 0) 0.668 us | _spin_lock();
2442 0) 0.570 us | expand_files();
2443 0) 0.586 us | _spin_unlock();
2446 There are several columns that can be dynamically
2447 enabled/disabled. You can use every combination of options you
2448 want, depending on your needs.
2450 - The cpu number on which the function executed is default
2451 enabled. It is sometimes better to only trace one cpu (see
2452 tracing_cpu_mask file) or you might sometimes see unordered
2453 function calls while cpu tracing switch.
2455 - hide: echo nofuncgraph-cpu > trace_options
2456 - show: echo funcgraph-cpu > trace_options
2458 - The duration (function's time of execution) is displayed on
2459 the closing bracket line of a function or on the same line
2460 than the current function in case of a leaf one. It is default
2463 - hide: echo nofuncgraph-duration > trace_options
2464 - show: echo funcgraph-duration > trace_options
2466 - The overhead field precedes the duration field in case of
2467 reached duration thresholds.
2469 - hide: echo nofuncgraph-overhead > trace_options
2470 - show: echo funcgraph-overhead > trace_options
2471 - depends on: funcgraph-duration
2475 3) # 1837.709 us | } /* __switch_to */
2476 3) | finish_task_switch() {
2477 3) 0.313 us | _raw_spin_unlock_irq();
2479 3) # 1889.063 us | } /* __schedule */
2480 3) ! 140.417 us | } /* __schedule */
2481 3) # 2034.948 us | } /* schedule */
2482 3) * 33998.59 us | } /* schedule_preempt_disabled */
2486 1) 0.260 us | msecs_to_jiffies();
2487 1) 0.313 us | __rcu_read_unlock();
2490 1) 0.313 us | rcu_bh_qs();
2491 1) 0.313 us | __local_bh_enable();
2493 1) 0.365 us | idle_cpu();
2494 1) | rcu_irq_exit() {
2495 1) 0.417 us | rcu_eqs_enter_common.isra.47();
2499 1) @ 119760.2 us | }
2505 2) 0.417 us | scheduler_ipi();
2515 + means that the function exceeded 10 usecs.
2516 ! means that the function exceeded 100 usecs.
2517 # means that the function exceeded 1000 usecs.
2518 * means that the function exceeded 10 msecs.
2519 @ means that the function exceeded 100 msecs.
2520 $ means that the function exceeded 1 sec.
2523 - The task/pid field displays the thread cmdline and pid which
2524 executed the function. It is default disabled.
2526 - hide: echo nofuncgraph-proc > trace_options
2527 - show: echo funcgraph-proc > trace_options
2531 # tracer: function_graph
2533 # CPU TASK/PID DURATION FUNCTION CALLS
2535 0) sh-4802 | | d_free() {
2536 0) sh-4802 | | call_rcu() {
2537 0) sh-4802 | | __call_rcu() {
2538 0) sh-4802 | 0.616 us | rcu_process_gp_end();
2539 0) sh-4802 | 0.586 us | check_for_new_grace_period();
2540 0) sh-4802 | 2.899 us | }
2541 0) sh-4802 | 4.040 us | }
2542 0) sh-4802 | 5.151 us | }
2543 0) sh-4802 | + 49.370 us | }
2546 - The absolute time field is an absolute timestamp given by the
2547 system clock since it started. A snapshot of this time is
2548 given on each entry/exit of functions
2550 - hide: echo nofuncgraph-abstime > trace_options
2551 - show: echo funcgraph-abstime > trace_options
2556 # TIME CPU DURATION FUNCTION CALLS
2558 360.774522 | 1) 0.541 us | }
2559 360.774522 | 1) 4.663 us | }
2560 360.774523 | 1) 0.541 us | __wake_up_bit();
2561 360.774524 | 1) 6.796 us | }
2562 360.774524 | 1) 7.952 us | }
2563 360.774525 | 1) 9.063 us | }
2564 360.774525 | 1) 0.615 us | journal_mark_dirty();
2565 360.774527 | 1) 0.578 us | __brelse();
2566 360.774528 | 1) | reiserfs_prepare_for_journal() {
2567 360.774528 | 1) | unlock_buffer() {
2568 360.774529 | 1) | wake_up_bit() {
2569 360.774529 | 1) | bit_waitqueue() {
2570 360.774530 | 1) 0.594 us | __phys_addr();
2573 The function name is always displayed after the closing bracket
2574 for a function if the start of that function is not in the
2577 Display of the function name after the closing bracket may be
2578 enabled for functions whose start is in the trace buffer,
2579 allowing easier searching with grep for function durations.
2580 It is default disabled.
2582 - hide: echo nofuncgraph-tail > trace_options
2583 - show: echo funcgraph-tail > trace_options
2585 Example with nofuncgraph-tail (default)::
2588 0) | kmem_cache_free() {
2589 0) 0.518 us | __phys_addr();
2593 Example with funcgraph-tail::
2596 0) | kmem_cache_free() {
2597 0) 0.518 us | __phys_addr();
2598 0) 1.757 us | } /* kmem_cache_free() */
2599 0) 2.861 us | } /* putname() */
2601 You can put some comments on specific functions by using
2602 trace_printk() For example, if you want to put a comment inside
2603 the __might_sleep() function, you just have to include
2604 <linux/ftrace.h> and call trace_printk() inside __might_sleep()::
2606 trace_printk("I'm a comment!\n")
2610 1) | __might_sleep() {
2611 1) | /* I'm a comment! */
2615 You might find other useful features for this tracer in the
2616 following "dynamic ftrace" section such as tracing only specific
2622 If CONFIG_DYNAMIC_FTRACE is set, the system will run with
2623 virtually no overhead when function tracing is disabled. The way
2624 this works is the mcount function call (placed at the start of
2625 every kernel function, produced by the -pg switch in gcc),
2626 starts of pointing to a simple return. (Enabling FTRACE will
2627 include the -pg switch in the compiling of the kernel.)
2629 At compile time every C file object is run through the
2630 recordmcount program (located in the scripts directory). This
2631 program will parse the ELF headers in the C object to find all
2632 the locations in the .text section that call mcount. Starting
2633 with gcc version 4.6, the -mfentry has been added for x86, which
2634 calls "__fentry__" instead of "mcount". Which is called before
2635 the creation of the stack frame.
2637 Note, not all sections are traced. They may be prevented by either
2638 a notrace, or blocked another way and all inline functions are not
2639 traced. Check the "available_filter_functions" file to see what functions
2642 A section called "__mcount_loc" is created that holds
2643 references to all the mcount/fentry call sites in the .text section.
2644 The recordmcount program re-links this section back into the
2645 original object. The final linking stage of the kernel will add all these
2646 references into a single table.
2648 On boot up, before SMP is initialized, the dynamic ftrace code
2649 scans this table and updates all the locations into nops. It
2650 also records the locations, which are added to the
2651 available_filter_functions list. Modules are processed as they
2652 are loaded and before they are executed. When a module is
2653 unloaded, it also removes its functions from the ftrace function
2654 list. This is automatic in the module unload code, and the
2655 module author does not need to worry about it.
2657 When tracing is enabled, the process of modifying the function
2658 tracepoints is dependent on architecture. The old method is to use
2659 kstop_machine to prevent races with the CPUs executing code being
2660 modified (which can cause the CPU to do undesirable things, especially
2661 if the modified code crosses cache (or page) boundaries), and the nops are
2662 patched back to calls. But this time, they do not call mcount
2663 (which is just a function stub). They now call into the ftrace
2666 The new method of modifying the function tracepoints is to place
2667 a breakpoint at the location to be modified, sync all CPUs, modify
2668 the rest of the instruction not covered by the breakpoint. Sync
2669 all CPUs again, and then remove the breakpoint with the finished
2670 version to the ftrace call site.
2672 Some archs do not even need to monkey around with the synchronization,
2673 and can just slap the new code on top of the old without any
2674 problems with other CPUs executing it at the same time.
2676 One special side-effect to the recording of the functions being
2677 traced is that we can now selectively choose which functions we
2678 wish to trace and which ones we want the mcount calls to remain
2681 Two files are used, one for enabling and one for disabling the
2682 tracing of specified functions. They are:
2690 A list of available functions that you can add to these files is
2693 available_filter_functions
2697 # cat available_filter_functions
2706 If I am only interested in sys_nanosleep and hrtimer_interrupt::
2708 # echo sys_nanosleep hrtimer_interrupt > set_ftrace_filter
2709 # echo function > current_tracer
2710 # echo 1 > tracing_on
2712 # echo 0 > tracing_on
2716 # entries-in-buffer/entries-written: 5/5 #P:4
2719 # / _----=> need-resched
2720 # | / _---=> hardirq/softirq
2721 # || / _--=> preempt-depth
2723 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2725 usleep-2665 [001] .... 4186.475355: sys_nanosleep <-system_call_fastpath
2726 <idle>-0 [001] d.h1 4186.475409: hrtimer_interrupt <-smp_apic_timer_interrupt
2727 usleep-2665 [001] d.h1 4186.475426: hrtimer_interrupt <-smp_apic_timer_interrupt
2728 <idle>-0 [003] d.h1 4186.475426: hrtimer_interrupt <-smp_apic_timer_interrupt
2729 <idle>-0 [002] d.h1 4186.475427: hrtimer_interrupt <-smp_apic_timer_interrupt
2731 To see which functions are being traced, you can cat the file:
2734 # cat set_ftrace_filter
2739 Perhaps this is not enough. The filters also allow glob(7) matching.
2742 will match functions that begin with <match>
2744 will match functions that end with <match>
2746 will match functions that have <match> in it
2747 ``<match1>*<match2>``
2748 will match functions that begin with <match1> and end with <match2>
2751 It is better to use quotes to enclose the wild cards,
2752 otherwise the shell may expand the parameters into names
2753 of files in the local directory.
2757 # echo 'hrtimer_*' > set_ftrace_filter
2763 # entries-in-buffer/entries-written: 897/897 #P:4
2766 # / _----=> need-resched
2767 # | / _---=> hardirq/softirq
2768 # || / _--=> preempt-depth
2770 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2772 <idle>-0 [003] dN.1 4228.547803: hrtimer_cancel <-tick_nohz_idle_exit
2773 <idle>-0 [003] dN.1 4228.547804: hrtimer_try_to_cancel <-hrtimer_cancel
2774 <idle>-0 [003] dN.2 4228.547805: hrtimer_force_reprogram <-__remove_hrtimer
2775 <idle>-0 [003] dN.1 4228.547805: hrtimer_forward <-tick_nohz_idle_exit
2776 <idle>-0 [003] dN.1 4228.547805: hrtimer_start_range_ns <-hrtimer_start_expires.constprop.11
2777 <idle>-0 [003] d..1 4228.547858: hrtimer_get_next_event <-get_next_timer_interrupt
2778 <idle>-0 [003] d..1 4228.547859: hrtimer_start <-__tick_nohz_idle_enter
2779 <idle>-0 [003] d..2 4228.547860: hrtimer_force_reprogram <-__rem
2781 Notice that we lost the sys_nanosleep.
2784 # cat set_ftrace_filter
2789 hrtimer_try_to_cancel
2793 hrtimer_force_reprogram
2794 hrtimer_get_next_event
2798 hrtimer_get_remaining
2800 hrtimer_init_sleeper
2803 This is because the '>' and '>>' act just like they do in bash.
2804 To rewrite the filters, use '>'
2805 To append to the filters, use '>>'
2807 To clear out a filter so that all functions will be recorded
2810 # echo > set_ftrace_filter
2811 # cat set_ftrace_filter
2814 Again, now we want to append.
2818 # echo sys_nanosleep > set_ftrace_filter
2819 # cat set_ftrace_filter
2821 # echo 'hrtimer_*' >> set_ftrace_filter
2822 # cat set_ftrace_filter
2827 hrtimer_try_to_cancel
2831 hrtimer_force_reprogram
2832 hrtimer_get_next_event
2837 hrtimer_get_remaining
2839 hrtimer_init_sleeper
2842 The set_ftrace_notrace prevents those functions from being
2846 # echo '*preempt*' '*lock*' > set_ftrace_notrace
2852 # entries-in-buffer/entries-written: 39608/39608 #P:4
2855 # / _----=> need-resched
2856 # | / _---=> hardirq/softirq
2857 # || / _--=> preempt-depth
2859 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2861 bash-1994 [000] .... 4342.324896: file_ra_state_init <-do_dentry_open
2862 bash-1994 [000] .... 4342.324897: open_check_o_direct <-do_last
2863 bash-1994 [000] .... 4342.324897: ima_file_check <-do_last
2864 bash-1994 [000] .... 4342.324898: process_measurement <-ima_file_check
2865 bash-1994 [000] .... 4342.324898: ima_get_action <-process_measurement
2866 bash-1994 [000] .... 4342.324898: ima_match_policy <-ima_get_action
2867 bash-1994 [000] .... 4342.324899: do_truncate <-do_last
2868 bash-1994 [000] .... 4342.324899: should_remove_suid <-do_truncate
2869 bash-1994 [000] .... 4342.324899: notify_change <-do_truncate
2870 bash-1994 [000] .... 4342.324900: current_fs_time <-notify_change
2871 bash-1994 [000] .... 4342.324900: current_kernel_time <-current_fs_time
2872 bash-1994 [000] .... 4342.324900: timespec_trunc <-current_fs_time
2874 We can see that there's no more lock or preempt tracing.
2876 Selecting function filters via index
2877 ------------------------------------
2879 Because processing of strings is expensive (the address of the function
2880 needs to be looked up before comparing to the string being passed in),
2881 an index can be used as well to enable functions. This is useful in the
2882 case of setting thousands of specific functions at a time. By passing
2883 in a list of numbers, no string processing will occur. Instead, the function
2884 at the specific location in the internal array (which corresponds to the
2885 functions in the "available_filter_functions" file), is selected.
2889 # echo 1 > set_ftrace_filter
2891 Will select the first function listed in "available_filter_functions"
2895 # head -1 available_filter_functions
2896 trace_initcall_finish_cb
2898 # cat set_ftrace_filter
2899 trace_initcall_finish_cb
2901 # head -50 available_filter_functions | tail -1
2904 # echo 1 50 > set_ftrace_filter
2905 # cat set_ftrace_filter
2906 trace_initcall_finish_cb
2909 Dynamic ftrace with the function graph tracer
2910 ---------------------------------------------
2912 Although what has been explained above concerns both the
2913 function tracer and the function-graph-tracer, there are some
2914 special features only available in the function-graph tracer.
2916 If you want to trace only one function and all of its children,
2917 you just have to echo its name into set_graph_function::
2919 echo __do_fault > set_graph_function
2921 will produce the following "expanded" trace of the __do_fault()
2925 0) | filemap_fault() {
2926 0) | find_lock_page() {
2927 0) 0.804 us | find_get_page();
2928 0) | __might_sleep() {
2932 0) 0.653 us | _spin_lock();
2933 0) 0.578 us | page_add_file_rmap();
2934 0) 0.525 us | native_set_pte_at();
2935 0) 0.585 us | _spin_unlock();
2936 0) | unlock_page() {
2937 0) 0.541 us | page_waitqueue();
2938 0) 0.639 us | __wake_up_bit();
2942 0) | filemap_fault() {
2943 0) | find_lock_page() {
2944 0) 0.698 us | find_get_page();
2945 0) | __might_sleep() {
2949 0) 0.631 us | _spin_lock();
2950 0) 0.571 us | page_add_file_rmap();
2951 0) 0.526 us | native_set_pte_at();
2952 0) 0.586 us | _spin_unlock();
2953 0) | unlock_page() {
2954 0) 0.533 us | page_waitqueue();
2955 0) 0.638 us | __wake_up_bit();
2959 You can also expand several functions at once::
2961 echo sys_open > set_graph_function
2962 echo sys_close >> set_graph_function
2964 Now if you want to go back to trace all functions you can clear
2965 this special filter via::
2967 echo > set_graph_function
2973 Note, the proc sysctl ftrace_enable is a big on/off switch for the
2974 function tracer. By default it is enabled (when function tracing is
2975 enabled in the kernel). If it is disabled, all function tracing is
2976 disabled. This includes not only the function tracers for ftrace, but
2977 also for any other uses (perf, kprobes, stack tracing, profiling, etc).
2979 Please disable this with care.
2981 This can be disable (and enabled) with::
2983 sysctl kernel.ftrace_enabled=0
2984 sysctl kernel.ftrace_enabled=1
2988 echo 0 > /proc/sys/kernel/ftrace_enabled
2989 echo 1 > /proc/sys/kernel/ftrace_enabled
2995 A few commands are supported by the set_ftrace_filter interface.
2996 Trace commands have the following format::
2998 <function>:<command>:<parameter>
3000 The following commands are supported:
3003 This command enables function filtering per module. The
3004 parameter defines the module. For example, if only the write*
3005 functions in the ext3 module are desired, run:
3007 echo 'write*:mod:ext3' > set_ftrace_filter
3009 This command interacts with the filter in the same way as
3010 filtering based on function names. Thus, adding more functions
3011 in a different module is accomplished by appending (>>) to the
3012 filter file. Remove specific module functions by prepending
3015 echo '!writeback*:mod:ext3' >> set_ftrace_filter
3017 Mod command supports module globbing. Disable tracing for all
3018 functions except a specific module::
3020 echo '!*:mod:!ext3' >> set_ftrace_filter
3022 Disable tracing for all modules, but still trace kernel::
3024 echo '!*:mod:*' >> set_ftrace_filter
3026 Enable filter only for kernel::
3028 echo '*write*:mod:!*' >> set_ftrace_filter
3030 Enable filter for module globbing::
3032 echo '*write*:mod:*snd*' >> set_ftrace_filter
3035 These commands turn tracing on and off when the specified
3036 functions are hit. The parameter determines how many times the
3037 tracing system is turned on and off. If unspecified, there is
3038 no limit. For example, to disable tracing when a schedule bug
3039 is hit the first 5 times, run::
3041 echo '__schedule_bug:traceoff:5' > set_ftrace_filter
3043 To always disable tracing when __schedule_bug is hit::
3045 echo '__schedule_bug:traceoff' > set_ftrace_filter
3047 These commands are cumulative whether or not they are appended
3048 to set_ftrace_filter. To remove a command, prepend it by '!'
3049 and drop the parameter::
3051 echo '!__schedule_bug:traceoff:0' > set_ftrace_filter
3053 The above removes the traceoff command for __schedule_bug
3054 that have a counter. To remove commands without counters::
3056 echo '!__schedule_bug:traceoff' > set_ftrace_filter
3059 Will cause a snapshot to be triggered when the function is hit.
3062 echo 'native_flush_tlb_others:snapshot' > set_ftrace_filter
3064 To only snapshot once:
3067 echo 'native_flush_tlb_others:snapshot:1' > set_ftrace_filter
3069 To remove the above commands::
3071 echo '!native_flush_tlb_others:snapshot' > set_ftrace_filter
3072 echo '!native_flush_tlb_others:snapshot:0' > set_ftrace_filter
3074 - enable_event/disable_event:
3075 These commands can enable or disable a trace event. Note, because
3076 function tracing callbacks are very sensitive, when these commands
3077 are registered, the trace point is activated, but disabled in
3078 a "soft" mode. That is, the tracepoint will be called, but
3079 just will not be traced. The event tracepoint stays in this mode
3080 as long as there's a command that triggers it.
3083 echo 'try_to_wake_up:enable_event:sched:sched_switch:2' > \
3088 <function>:enable_event:<system>:<event>[:count]
3089 <function>:disable_event:<system>:<event>[:count]
3091 To remove the events commands::
3093 echo '!try_to_wake_up:enable_event:sched:sched_switch:0' > \
3095 echo '!schedule:disable_event:sched:sched_switch' > \
3099 When the function is hit, it will dump the contents of the ftrace
3100 ring buffer to the console. This is useful if you need to debug
3101 something, and want to dump the trace when a certain function
3102 is hit. Perhaps it's a function that is called before a triple
3103 fault happens and does not allow you to get a regular dump.
3106 When the function is hit, it will dump the contents of the ftrace
3107 ring buffer for the current CPU to the console. Unlike the "dump"
3108 command, it only prints out the contents of the ring buffer for the
3109 CPU that executed the function that triggered the dump.
3112 When the function is hit, a stack trace is recorded.
3117 The trace_pipe outputs the same content as the trace file, but
3118 the effect on the tracing is different. Every read from
3119 trace_pipe is consumed. This means that subsequent reads will be
3120 different. The trace is live.
3123 # echo function > current_tracer
3124 # cat trace_pipe > /tmp/trace.out &
3126 # echo 1 > tracing_on
3128 # echo 0 > tracing_on
3132 # entries-in-buffer/entries-written: 0/0 #P:4
3135 # / _----=> need-resched
3136 # | / _---=> hardirq/softirq
3137 # || / _--=> preempt-depth
3139 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3143 # cat /tmp/trace.out
3144 bash-1994 [000] .... 5281.568961: mutex_unlock <-rb_simple_write
3145 bash-1994 [000] .... 5281.568963: __mutex_unlock_slowpath <-mutex_unlock
3146 bash-1994 [000] .... 5281.568963: __fsnotify_parent <-fsnotify_modify
3147 bash-1994 [000] .... 5281.568964: fsnotify <-fsnotify_modify
3148 bash-1994 [000] .... 5281.568964: __srcu_read_lock <-fsnotify
3149 bash-1994 [000] .... 5281.568964: add_preempt_count <-__srcu_read_lock
3150 bash-1994 [000] ...1 5281.568965: sub_preempt_count <-__srcu_read_lock
3151 bash-1994 [000] .... 5281.568965: __srcu_read_unlock <-fsnotify
3152 bash-1994 [000] .... 5281.568967: sys_dup2 <-system_call_fastpath
3155 Note, reading the trace_pipe file will block until more input is
3161 Having too much or not enough data can be troublesome in
3162 diagnosing an issue in the kernel. The file buffer_size_kb is
3163 used to modify the size of the internal trace buffers. The
3164 number listed is the number of entries that can be recorded per
3165 CPU. To know the full size, multiply the number of possible CPUs
3166 with the number of entries.
3169 # cat buffer_size_kb
3170 1408 (units kilobytes)
3172 Or simply read buffer_total_size_kb
3175 # cat buffer_total_size_kb
3178 To modify the buffer, simple echo in a number (in 1024 byte segments).
3181 # echo 10000 > buffer_size_kb
3182 # cat buffer_size_kb
3183 10000 (units kilobytes)
3185 It will try to allocate as much as possible. If you allocate too
3186 much, it can cause Out-Of-Memory to trigger.
3189 # echo 1000000000000 > buffer_size_kb
3190 -bash: echo: write error: Cannot allocate memory
3191 # cat buffer_size_kb
3194 The per_cpu buffers can be changed individually as well:
3197 # echo 10000 > per_cpu/cpu0/buffer_size_kb
3198 # echo 100 > per_cpu/cpu1/buffer_size_kb
3200 When the per_cpu buffers are not the same, the buffer_size_kb
3201 at the top level will just show an X
3204 # cat buffer_size_kb
3207 This is where the buffer_total_size_kb is useful:
3210 # cat buffer_total_size_kb
3213 Writing to the top level buffer_size_kb will reset all the buffers
3214 to be the same again.
3218 CONFIG_TRACER_SNAPSHOT makes a generic snapshot feature
3219 available to all non latency tracers. (Latency tracers which
3220 record max latency, such as "irqsoff" or "wakeup", can't use
3221 this feature, since those are already using the snapshot
3222 mechanism internally.)
3224 Snapshot preserves a current trace buffer at a particular point
3225 in time without stopping tracing. Ftrace swaps the current
3226 buffer with a spare buffer, and tracing continues in the new
3227 current (=previous spare) buffer.
3229 The following tracefs files in "tracing" are related to this
3234 This is used to take a snapshot and to read the output
3235 of the snapshot. Echo 1 into this file to allocate a
3236 spare buffer and to take a snapshot (swap), then read
3237 the snapshot from this file in the same format as
3238 "trace" (described above in the section "The File
3239 System"). Both reads snapshot and tracing are executable
3240 in parallel. When the spare buffer is allocated, echoing
3241 0 frees it, and echoing else (positive) values clear the
3243 More details are shown in the table below.
3245 +--------------+------------+------------+------------+
3246 |status\\input | 0 | 1 | else |
3247 +==============+============+============+============+
3248 |not allocated |(do nothing)| alloc+swap |(do nothing)|
3249 +--------------+------------+------------+------------+
3250 |allocated | free | swap | clear |
3251 +--------------+------------+------------+------------+
3253 Here is an example of using the snapshot feature.
3256 # echo 1 > events/sched/enable
3261 # entries-in-buffer/entries-written: 71/71 #P:8
3264 # / _----=> need-resched
3265 # | / _---=> hardirq/softirq
3266 # || / _--=> preempt-depth
3268 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3270 <idle>-0 [005] d... 2440.603828: sched_switch: prev_comm=swapper/5 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=snapshot-test-2 next_pid=2242 next_prio=120
3271 sleep-2242 [005] d... 2440.603846: sched_switch: prev_comm=snapshot-test-2 prev_pid=2242 prev_prio=120 prev_state=R ==> next_comm=kworker/5:1 next_pid=60 next_prio=120
3273 <idle>-0 [002] d... 2440.707230: sched_switch: prev_comm=swapper/2 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=snapshot-test-2 next_pid=2229 next_prio=120
3278 # entries-in-buffer/entries-written: 77/77 #P:8
3281 # / _----=> need-resched
3282 # | / _---=> hardirq/softirq
3283 # || / _--=> preempt-depth
3285 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3287 <idle>-0 [007] d... 2440.707395: sched_switch: prev_comm=swapper/7 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=snapshot-test-2 next_pid=2243 next_prio=120
3288 snapshot-test-2-2229 [002] d... 2440.707438: sched_switch: prev_comm=snapshot-test-2 prev_pid=2229 prev_prio=120 prev_state=S ==> next_comm=swapper/2 next_pid=0 next_prio=120
3292 If you try to use this snapshot feature when current tracer is
3293 one of the latency tracers, you will get the following results.
3296 # echo wakeup > current_tracer
3298 bash: echo: write error: Device or resource busy
3300 cat: snapshot: Device or resource busy
3305 In the tracefs tracing directory is a directory called "instances".
3306 This directory can have new directories created inside of it using
3307 mkdir, and removing directories with rmdir. The directory created
3308 with mkdir in this directory will already contain files and other
3309 directories after it is created.
3312 # mkdir instances/foo
3314 buffer_size_kb buffer_total_size_kb events free_buffer per_cpu
3315 set_event snapshot trace trace_clock trace_marker trace_options
3316 trace_pipe tracing_on
3318 As you can see, the new directory looks similar to the tracing directory
3319 itself. In fact, it is very similar, except that the buffer and
3320 events are agnostic from the main director, or from any other
3321 instances that are created.
3323 The files in the new directory work just like the files with the
3324 same name in the tracing directory except the buffer that is used
3325 is a separate and new buffer. The files affect that buffer but do not
3326 affect the main buffer with the exception of trace_options. Currently,
3327 the trace_options affect all instances and the top level buffer
3328 the same, but this may change in future releases. That is, options
3329 may become specific to the instance they reside in.
3331 Notice that none of the function tracer files are there, nor is
3332 current_tracer and available_tracers. This is because the buffers
3333 can currently only have events enabled for them.
3336 # mkdir instances/foo
3337 # mkdir instances/bar
3338 # mkdir instances/zoot
3339 # echo 100000 > buffer_size_kb
3340 # echo 1000 > instances/foo/buffer_size_kb
3341 # echo 5000 > instances/bar/per_cpu/cpu1/buffer_size_kb
3342 # echo function > current_trace
3343 # echo 1 > instances/foo/events/sched/sched_wakeup/enable
3344 # echo 1 > instances/foo/events/sched/sched_wakeup_new/enable
3345 # echo 1 > instances/foo/events/sched/sched_switch/enable
3346 # echo 1 > instances/bar/events/irq/enable
3347 # echo 1 > instances/zoot/events/syscalls/enable
3349 CPU:2 [LOST 11745 EVENTS]
3350 bash-2044 [002] .... 10594.481032: _raw_spin_lock_irqsave <-get_page_from_freelist
3351 bash-2044 [002] d... 10594.481032: add_preempt_count <-_raw_spin_lock_irqsave
3352 bash-2044 [002] d..1 10594.481032: __rmqueue <-get_page_from_freelist
3353 bash-2044 [002] d..1 10594.481033: _raw_spin_unlock <-get_page_from_freelist
3354 bash-2044 [002] d..1 10594.481033: sub_preempt_count <-_raw_spin_unlock
3355 bash-2044 [002] d... 10594.481033: get_pageblock_flags_group <-get_pageblock_migratetype
3356 bash-2044 [002] d... 10594.481034: __mod_zone_page_state <-get_page_from_freelist
3357 bash-2044 [002] d... 10594.481034: zone_statistics <-get_page_from_freelist
3358 bash-2044 [002] d... 10594.481034: __inc_zone_state <-zone_statistics
3359 bash-2044 [002] d... 10594.481034: __inc_zone_state <-zone_statistics
3360 bash-2044 [002] .... 10594.481035: arch_dup_task_struct <-copy_process
3363 # cat instances/foo/trace_pipe
3364 bash-1998 [000] d..4 136.676759: sched_wakeup: comm=kworker/0:1 pid=59 prio=120 success=1 target_cpu=000
3365 bash-1998 [000] dN.4 136.676760: sched_wakeup: comm=bash pid=1998 prio=120 success=1 target_cpu=000
3366 <idle>-0 [003] d.h3 136.676906: sched_wakeup: comm=rcu_preempt pid=9 prio=120 success=1 target_cpu=003
3367 <idle>-0 [003] d..3 136.676909: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=rcu_preempt next_pid=9 next_prio=120
3368 rcu_preempt-9 [003] d..3 136.676916: sched_switch: prev_comm=rcu_preempt prev_pid=9 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120
3369 bash-1998 [000] d..4 136.677014: sched_wakeup: comm=kworker/0:1 pid=59 prio=120 success=1 target_cpu=000
3370 bash-1998 [000] dN.4 136.677016: sched_wakeup: comm=bash pid=1998 prio=120 success=1 target_cpu=000
3371 bash-1998 [000] d..3 136.677018: sched_switch: prev_comm=bash prev_pid=1998 prev_prio=120 prev_state=R+ ==> next_comm=kworker/0:1 next_pid=59 next_prio=120
3372 kworker/0:1-59 [000] d..4 136.677022: sched_wakeup: comm=sshd pid=1995 prio=120 success=1 target_cpu=001
3373 kworker/0:1-59 [000] d..3 136.677025: sched_switch: prev_comm=kworker/0:1 prev_pid=59 prev_prio=120 prev_state=S ==> next_comm=bash next_pid=1998 next_prio=120
3376 # cat instances/bar/trace_pipe
3377 migration/1-14 [001] d.h3 138.732674: softirq_raise: vec=3 [action=NET_RX]
3378 <idle>-0 [001] dNh3 138.732725: softirq_raise: vec=3 [action=NET_RX]
3379 bash-1998 [000] d.h1 138.733101: softirq_raise: vec=1 [action=TIMER]
3380 bash-1998 [000] d.h1 138.733102: softirq_raise: vec=9 [action=RCU]
3381 bash-1998 [000] ..s2 138.733105: softirq_entry: vec=1 [action=TIMER]
3382 bash-1998 [000] ..s2 138.733106: softirq_exit: vec=1 [action=TIMER]
3383 bash-1998 [000] ..s2 138.733106: softirq_entry: vec=9 [action=RCU]
3384 bash-1998 [000] ..s2 138.733109: softirq_exit: vec=9 [action=RCU]
3385 sshd-1995 [001] d.h1 138.733278: irq_handler_entry: irq=21 name=uhci_hcd:usb4
3386 sshd-1995 [001] d.h1 138.733280: irq_handler_exit: irq=21 ret=unhandled
3387 sshd-1995 [001] d.h1 138.733281: irq_handler_entry: irq=21 name=eth0
3388 sshd-1995 [001] d.h1 138.733283: irq_handler_exit: irq=21 ret=handled
3391 # cat instances/zoot/trace
3394 # entries-in-buffer/entries-written: 18996/18996 #P:4
3397 # / _----=> need-resched
3398 # | / _---=> hardirq/softirq
3399 # || / _--=> preempt-depth
3401 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3403 bash-1998 [000] d... 140.733501: sys_write -> 0x2
3404 bash-1998 [000] d... 140.733504: sys_dup2(oldfd: a, newfd: 1)
3405 bash-1998 [000] d... 140.733506: sys_dup2 -> 0x1
3406 bash-1998 [000] d... 140.733508: sys_fcntl(fd: a, cmd: 1, arg: 0)
3407 bash-1998 [000] d... 140.733509: sys_fcntl -> 0x1
3408 bash-1998 [000] d... 140.733510: sys_close(fd: a)
3409 bash-1998 [000] d... 140.733510: sys_close -> 0x0
3410 bash-1998 [000] d... 140.733514: sys_rt_sigprocmask(how: 0, nset: 0, oset: 6e2768, sigsetsize: 8)
3411 bash-1998 [000] d... 140.733515: sys_rt_sigprocmask -> 0x0
3412 bash-1998 [000] d... 140.733516: sys_rt_sigaction(sig: 2, act: 7fff718846f0, oact: 7fff71884650, sigsetsize: 8)
3413 bash-1998 [000] d... 140.733516: sys_rt_sigaction -> 0x0
3415 You can see that the trace of the top most trace buffer shows only
3416 the function tracing. The foo instance displays wakeups and task
3419 To remove the instances, simply delete their directories:
3422 # rmdir instances/foo
3423 # rmdir instances/bar
3424 # rmdir instances/zoot
3426 Note, if a process has a trace file open in one of the instance
3427 directories, the rmdir will fail with EBUSY.
3432 Since the kernel has a fixed sized stack, it is important not to
3433 waste it in functions. A kernel developer must be conscience of
3434 what they allocate on the stack. If they add too much, the system
3435 can be in danger of a stack overflow, and corruption will occur,
3436 usually leading to a system panic.
3438 There are some tools that check this, usually with interrupts
3439 periodically checking usage. But if you can perform a check
3440 at every function call that will become very useful. As ftrace provides
3441 a function tracer, it makes it convenient to check the stack size
3442 at every function call. This is enabled via the stack tracer.
3444 CONFIG_STACK_TRACER enables the ftrace stack tracing functionality.
3445 To enable it, write a '1' into /proc/sys/kernel/stack_tracer_enabled.
3448 # echo 1 > /proc/sys/kernel/stack_tracer_enabled
3450 You can also enable it from the kernel command line to trace
3451 the stack size of the kernel during boot up, by adding "stacktrace"
3452 to the kernel command line parameter.
3454 After running it for a few minutes, the output looks like:
3457 # cat stack_max_size
3461 Depth Size Location (18 entries)
3463 0) 2928 224 update_sd_lb_stats+0xbc/0x4ac
3464 1) 2704 160 find_busiest_group+0x31/0x1f1
3465 2) 2544 256 load_balance+0xd9/0x662
3466 3) 2288 80 idle_balance+0xbb/0x130
3467 4) 2208 128 __schedule+0x26e/0x5b9
3468 5) 2080 16 schedule+0x64/0x66
3469 6) 2064 128 schedule_timeout+0x34/0xe0
3470 7) 1936 112 wait_for_common+0x97/0xf1
3471 8) 1824 16 wait_for_completion+0x1d/0x1f
3472 9) 1808 128 flush_work+0xfe/0x119
3473 10) 1680 16 tty_flush_to_ldisc+0x1e/0x20
3474 11) 1664 48 input_available_p+0x1d/0x5c
3475 12) 1616 48 n_tty_poll+0x6d/0x134
3476 13) 1568 64 tty_poll+0x64/0x7f
3477 14) 1504 880 do_select+0x31e/0x511
3478 15) 624 400 core_sys_select+0x177/0x216
3479 16) 224 96 sys_select+0x91/0xb9
3480 17) 128 128 system_call_fastpath+0x16/0x1b
3482 Note, if -mfentry is being used by gcc, functions get traced before
3483 they set up the stack frame. This means that leaf level functions
3484 are not tested by the stack tracer when -mfentry is used.
3486 Currently, -mfentry is used by gcc 4.6.0 and above on x86 only.
3490 More details can be found in the source code, in the `kernel/trace/*.c` files.
projects / linux.git / blob