1 ftrace - Function Tracer
2 ========================
4 Copyright 2008 Red Hat Inc.
5 Author: Steven Rostedt <srostedt@redhat.com>
6 License: The GNU Free Documentation License, Version 1.2
7 (dual licensed under the GPL v2)
8 Original Reviewers: Elias Oltmanns, Randy Dunlap, Andrew Morton,
9 John Kacur, and David Teigland.
10 Written for: 2.6.28-rc2
12 Updated for: 4.13 - Copyright 2017 VMware Inc. Steven Rostedt
17 Ftrace is an internal tracer designed to help out developers and
18 designers of systems to find what is going on inside the kernel.
19 It can be used for debugging or analyzing latencies and
20 performance issues that take place outside of user-space.
22 Although ftrace is typically considered the function tracer, it
23 is really a frame work of several assorted tracing utilities.
24 There's latency tracing to examine what occurs between interrupts
25 disabled and enabled, as well as for preemption and from a time
26 a task is woken to the task is actually scheduled in.
28 One of the most common uses of ftrace is the event tracing.
29 Through out the kernel is hundreds of static event points that
30 can be enabled via the tracefs file system to see what is
31 going on in certain parts of the kernel.
33 See events.txt for more information.
36 Implementation Details
37 ----------------------
39 See ftrace-design.txt for details for arch porters and such.
45 Ftrace uses the tracefs file system to hold the control files as
46 well as the files to display output.
48 When tracefs is configured into the kernel (which selecting any ftrace
49 option will do) the directory /sys/kernel/tracing will be created. To mount
50 this directory, you can add to your /etc/fstab file:
52 tracefs /sys/kernel/tracing tracefs defaults 0 0
54 Or you can mount it at run time with:
56 mount -t tracefs nodev /sys/kernel/tracing
58 For quicker access to that directory you may want to make a soft link to
61 ln -s /sys/kernel/tracing /tracing
65 Before 4.1, all ftrace tracing control files were within the debugfs
66 file system, which is typically located at /sys/kernel/debug/tracing.
67 For backward compatibility, when mounting the debugfs file system,
68 the tracefs file system will be automatically mounted at:
70 /sys/kernel/debug/tracing
72 All files located in the tracefs file system will be located in that
73 debugfs file system directory as well.
77 Any selected ftrace option will also create the tracefs file system.
78 The rest of the document will assume that you are in the ftrace directory
79 (cd /sys/kernel/tracing) and will only concentrate on the files within that
80 directory and not distract from the content with the extended
81 "/sys/kernel/tracing" path name.
83 That's it! (assuming that you have ftrace configured into your kernel)
85 After mounting tracefs you will have access to the control and output files
86 of ftrace. Here is a list of some of the key files:
89 Note: all time values are in microseconds.
93 This is used to set or display the current tracer
98 This holds the different types of tracers that
99 have been compiled into the kernel. The
100 tracers listed here can be configured by
101 echoing their name into current_tracer.
105 This sets or displays whether writing to the trace
106 ring buffer is enabled. Echo 0 into this file to disable
107 the tracer or 1 to enable it. Note, this only disables
108 writing to the ring buffer, the tracing overhead may
111 The kernel function tracing_off() can be used within the
112 kernel to disable writing to the ring buffer, which will
113 set this file to "0". User space can re-enable tracing by
114 echoing "1" into the file.
116 Note, the function and event trigger "traceoff" will also
117 set this file to zero and stop tracing. Which can also
118 be re-enabled by user space using this file.
122 This file holds the output of the trace in a human
123 readable format (described below). Note, tracing is temporarily
124 disabled while this file is being read (opened).
128 The output is the same as the "trace" file but this
129 file is meant to be streamed with live tracing.
130 Reads from this file will block until new data is
131 retrieved. Unlike the "trace" file, this file is a
132 consumer. This means reading from this file causes
133 sequential reads to display more current data. Once
134 data is read from this file, it is consumed, and
135 will not be read again with a sequential read. The
136 "trace" file is static, and if the tracer is not
137 adding more data, it will display the same
138 information every time it is read. This file will not
139 disable tracing while being read.
143 This file lets the user control the amount of data
144 that is displayed in one of the above output
145 files. Options also exist to modify how a tracer
146 or events work (stack traces, timestamps, etc).
150 This is a directory that has a file for every available
151 trace option (also in trace_options). Options may also be set
152 or cleared by writing a "1" or "0" respectively into the
153 corresponding file with the option name.
157 Some of the tracers record the max latency.
158 For example, the maximum time that interrupts are disabled.
159 The maximum time is saved in this file. The max trace will also be
160 stored, and displayed by "trace". A new max trace will only be
161 recorded if the latency is greater than the value in this file
164 By echoing in a time into this file, no latency will be recorded
165 unless it is greater than the time in this file.
169 Some latency tracers will record a trace whenever the
170 latency is greater than the number in this file.
171 Only active when the file contains a number greater than 0.
176 This sets or displays the number of kilobytes each CPU
177 buffer holds. By default, the trace buffers are the same size
178 for each CPU. The displayed number is the size of the
179 CPU buffer and not total size of all buffers. The
180 trace buffers are allocated in pages (blocks of memory
181 that the kernel uses for allocation, usually 4 KB in size).
182 If the last page allocated has room for more bytes
183 than requested, the rest of the page will be used,
184 making the actual allocation bigger than requested or shown.
185 ( Note, the size may not be a multiple of the page size
186 due to buffer management meta-data. )
188 Buffer sizes for individual CPUs may vary
189 (see "per_cpu/cpu0/buffer_size_kb" below), and if they do
190 this file will show "X".
192 buffer_total_size_kb:
194 This displays the total combined size of all the trace buffers.
198 If a process is performing tracing, and the ring buffer should be
199 shrunk "freed" when the process is finished, even if it were to be
200 killed by a signal, this file can be used for that purpose. On close
201 of this file, the ring buffer will be resized to its minimum size.
202 Having a process that is tracing also open this file, when the process
203 exits its file descriptor for this file will be closed, and in doing so,
204 the ring buffer will be "freed".
206 It may also stop tracing if disable_on_free option is set.
210 This is a mask that lets the user only trace on specified CPUs.
211 The format is a hex string representing the CPUs.
215 When dynamic ftrace is configured in (see the
216 section below "dynamic ftrace"), the code is dynamically
217 modified (code text rewrite) to disable calling of the
218 function profiler (mcount). This lets tracing be configured
219 in with practically no overhead in performance. This also
220 has a side effect of enabling or disabling specific functions
221 to be traced. Echoing names of functions into this file
222 will limit the trace to only those functions.
224 The functions listed in "available_filter_functions" are what
225 can be written into this file.
227 This interface also allows for commands to be used. See the
228 "Filter commands" section for more details.
232 This has an effect opposite to that of
233 set_ftrace_filter. Any function that is added here will not
234 be traced. If a function exists in both set_ftrace_filter
235 and set_ftrace_notrace, the function will _not_ be traced.
239 Have the function tracer only trace the threads whose PID are
242 If the "function-fork" option is set, then when a task whose
243 PID is listed in this file forks, the child's PID will
244 automatically be added to this file, and the child will be
245 traced by the function tracer as well. This option will also
246 cause PIDs of tasks that exit to be removed from the file.
250 Have the events only trace a task with a PID listed in this file.
251 Note, sched_switch and sched_wake_up will also trace events
254 To have the PIDs of children of tasks with their PID in this file
255 added on fork, enable the "event-fork" option. That option will also
256 cause the PIDs of tasks to be removed from this file when the task
261 Functions listed in this file will cause the function graph
262 tracer to only trace these functions and the functions that
263 they call. (See the section "dynamic ftrace" for more details).
267 Similar to set_graph_function, but will disable function graph
268 tracing when the function is hit until it exits the function.
269 This makes it possible to ignore tracing functions that are called
270 by a specific function.
272 available_filter_functions:
274 This lists the functions that ftrace has processed and can trace.
275 These are the function names that you can pass to
276 "set_ftrace_filter" or "set_ftrace_notrace".
277 (See the section "dynamic ftrace" below for more details.)
279 dyn_ftrace_total_info:
281 This file is for debugging purposes. The number of functions that
282 have been converted to nops and are available to be traced.
286 This file is more for debugging ftrace, but can also be useful
287 in seeing if any function has a callback attached to it.
288 Not only does the trace infrastructure use ftrace function
289 trace utility, but other subsystems might too. This file
290 displays all functions that have a callback attached to them
291 as well as the number of callbacks that have been attached.
292 Note, a callback may also call multiple functions which will
293 not be listed in this count.
295 If the callback registered to be traced by a function with
296 the "save regs" attribute (thus even more overhead), a 'R'
297 will be displayed on the same line as the function that
298 is returning registers.
300 If the callback registered to be traced by a function with
301 the "ip modify" attribute (thus the regs->ip can be changed),
302 an 'I' will be displayed on the same line as the function that
305 If the architecture supports it, it will also show what callback
306 is being directly called by the function. If the count is greater
307 than 1 it most likely will be ftrace_ops_list_func().
309 If the callback of the function jumps to a trampoline that is
310 specific to a the callback and not the standard trampoline,
311 its address will be printed as well as the function that the
314 function_profile_enabled:
316 When set it will enable all functions with either the function
317 tracer, or if configured, the function graph tracer. It will
318 keep a histogram of the number of functions that were called
319 and if the function graph tracer was configured, it will also keep
320 track of the time spent in those functions. The histogram
321 content can be displayed in the files:
323 trace_stats/function<cpu> ( function0, function1, etc).
327 A directory that holds different tracing stats.
331 Enable dynamic trace points. See kprobetrace.txt.
335 Dynamic trace points stats. See kprobetrace.txt.
339 Used with the function graph tracer. This is the max depth
340 it will trace into a function. Setting this to a value of
341 one will show only the first kernel function that is called
346 This is for tools that read the raw format files. If an event in
347 the ring buffer references a string, only a pointer to the string
348 is recorded into the buffer and not the string itself. This prevents
349 tools from knowing what that string was. This file displays the string
350 and address for the string allowing tools to map the pointers to what
355 Only the pid of the task is recorded in a trace event unless
356 the event specifically saves the task comm as well. Ftrace
357 makes a cache of pid mappings to comms to try to display
358 comms for events. If a pid for a comm is not listed, then
359 "<...>" is displayed in the output.
361 If the option "record-cmd" is set to "0", then comms of tasks
362 will not be saved during recording. By default, it is enabled.
366 By default, 128 comms are saved (see "saved_cmdlines" above). To
367 increase or decrease the amount of comms that are cached, echo
368 in a the number of comms to cache, into this file.
372 If the option "record-tgid" is set, on each scheduling context switch
373 the Task Group ID of a task is saved in a table mapping the PID of
374 the thread to its TGID. By default, the "record-tgid" option is
379 This displays the "snapshot" buffer and also lets the user
380 take a snapshot of the current running trace.
381 See the "Snapshot" section below for more details.
385 When the stack tracer is activated, this will display the
386 maximum stack size it has encountered.
387 See the "Stack Trace" section below.
391 This displays the stack back trace of the largest stack
392 that was encountered when the stack tracer is activated.
393 See the "Stack Trace" section below.
397 This is similar to "set_ftrace_filter" but it limits what
398 functions the stack tracer will check.
402 Whenever an event is recorded into the ring buffer, a
403 "timestamp" is added. This stamp comes from a specified
404 clock. By default, ftrace uses the "local" clock. This
405 clock is very fast and strictly per cpu, but on some
406 systems it may not be monotonic with respect to other
407 CPUs. In other words, the local clocks may not be in sync
408 with local clocks on other CPUs.
410 Usual clocks for tracing:
413 [local] global counter x86-tsc
415 The clock with the square brackets around it is the one
418 local: Default clock, but may not be in sync across CPUs
420 global: This clock is in sync with all CPUs but may
421 be a bit slower than the local clock.
423 counter: This is not a clock at all, but literally an atomic
424 counter. It counts up one by one, but is in sync
425 with all CPUs. This is useful when you need to
426 know exactly the order events occurred with respect to
427 each other on different CPUs.
429 uptime: This uses the jiffies counter and the time stamp
430 is relative to the time since boot up.
432 perf: This makes ftrace use the same clock that perf uses.
433 Eventually perf will be able to read ftrace buffers
434 and this will help out in interleaving the data.
436 x86-tsc: Architectures may define their own clocks. For
437 example, x86 uses its own TSC cycle clock here.
439 ppc-tb: This uses the powerpc timebase register value.
440 This is in sync across CPUs and can also be used
441 to correlate events across hypervisor/guest if
444 mono: This uses the fast monotonic clock (CLOCK_MONOTONIC)
445 which is monotonic and is subject to NTP rate adjustments.
448 This is the raw monotonic clock (CLOCK_MONOTONIC_RAW)
449 which is montonic but is not subject to any rate adjustments
450 and ticks at the same rate as the hardware clocksource.
452 boot: This is the boot clock (CLOCK_BOOTTIME) and is based on the
453 fast monotonic clock, but also accounts for time spent in
454 suspend. Since the clock access is designed for use in
455 tracing in the suspend path, some side effects are possible
456 if clock is accessed after the suspend time is accounted before
457 the fast mono clock is updated. In this case, the clock update
458 appears to happen slightly sooner than it normally would have.
459 Also on 32-bit systems, it's possible that the 64-bit boot offset
460 sees a partial update. These effects are rare and post
461 processing should be able to handle them. See comments in the
462 ktime_get_boot_fast_ns() function for more information.
464 To set a clock, simply echo the clock name into this file.
466 echo global > trace_clock
470 This is a very useful file for synchronizing user space
471 with events happening in the kernel. Writing strings into
472 this file will be written into the ftrace buffer.
474 It is useful in applications to open this file at the start
475 of the application and just reference the file descriptor
478 void trace_write(const char *fmt, ...)
488 n = vsnprintf(buf, 256, fmt, ap);
491 write(trace_fd, buf, n);
496 trace_fd = open("trace_marker", WR_ONLY);
500 This is similar to trace_marker above, but is meant for for binary data
501 to be written to it, where a tool can be used to parse the data
506 Add dynamic tracepoints in programs.
511 Uprobe statistics. See uprobetrace.txt
515 This is a way to make multiple trace buffers where different
516 events can be recorded in different buffers.
517 See "Instances" section below.
521 This is the trace event directory. It holds event tracepoints
522 (also known as static tracepoints) that have been compiled
523 into the kernel. It shows what event tracepoints exist
524 and how they are grouped by system. There are "enable"
525 files at various levels that can enable the tracepoints
526 when a "1" is written to them.
528 See events.txt for more information.
532 By echoing in the event into this file, will enable that event.
534 See events.txt for more information.
538 A list of events that can be enabled in tracing.
540 See events.txt for more information.
544 Directory for the Hardware Latency Detector.
545 See "Hardware Latency Detector" section below.
549 This is a directory that contains the trace per_cpu information.
551 per_cpu/cpu0/buffer_size_kb:
553 The ftrace buffer is defined per_cpu. That is, there's a separate
554 buffer for each CPU to allow writes to be done atomically,
555 and free from cache bouncing. These buffers may have different
556 size buffers. This file is similar to the buffer_size_kb
557 file, but it only displays or sets the buffer size for the
558 specific CPU. (here cpu0).
562 This is similar to the "trace" file, but it will only display
563 the data specific for the CPU. If written to, it only clears
564 the specific CPU buffer.
566 per_cpu/cpu0/trace_pipe
568 This is similar to the "trace_pipe" file, and is a consuming
569 read, but it will only display (and consume) the data specific
572 per_cpu/cpu0/trace_pipe_raw
574 For tools that can parse the ftrace ring buffer binary format,
575 the trace_pipe_raw file can be used to extract the data
576 from the ring buffer directly. With the use of the splice()
577 system call, the buffer data can be quickly transferred to
578 a file or to the network where a server is collecting the
581 Like trace_pipe, this is a consuming reader, where multiple
582 reads will always produce different data.
584 per_cpu/cpu0/snapshot:
586 This is similar to the main "snapshot" file, but will only
587 snapshot the current CPU (if supported). It only displays
588 the content of the snapshot for a given CPU, and if
589 written to, only clears this CPU buffer.
591 per_cpu/cpu0/snapshot_raw:
593 Similar to the trace_pipe_raw, but will read the binary format
594 from the snapshot buffer for the given CPU.
598 This displays certain stats about the ring buffer:
600 entries: The number of events that are still in the buffer.
602 overrun: The number of lost events due to overwriting when
605 commit overrun: Should always be zero.
606 This gets set if so many events happened within a nested
607 event (ring buffer is re-entrant), that it fills the
608 buffer and starts dropping events.
610 bytes: Bytes actually read (not overwritten).
612 oldest event ts: The oldest timestamp in the buffer
614 now ts: The current timestamp
616 dropped events: Events lost due to overwrite option being off.
618 read events: The number of events read.
623 Here is the list of current tracers that may be configured.
627 Function call tracer to trace all kernel functions.
631 Similar to the function tracer except that the
632 function tracer probes the functions on their entry
633 whereas the function graph tracer traces on both entry
634 and exit of the functions. It then provides the ability
635 to draw a graph of function calls similar to C code
640 The block tracer. The tracer used by the blktrace user
645 The Hardware Latency tracer is used to detect if the hardware
646 produces any latency. See "Hardware Latency Detector" section
651 Traces the areas that disable interrupts and saves
652 the trace with the longest max latency.
653 See tracing_max_latency. When a new max is recorded,
654 it replaces the old trace. It is best to view this
655 trace with the latency-format option enabled, which
656 happens automatically when the tracer is selected.
660 Similar to irqsoff but traces and records the amount of
661 time for which preemption is disabled.
665 Similar to irqsoff and preemptoff, but traces and
666 records the largest time for which irqs and/or preemption
671 Traces and records the max latency that it takes for
672 the highest priority task to get scheduled after
673 it has been woken up.
674 Traces all tasks as an average developer would expect.
678 Traces and records the max latency that it takes for just
679 RT tasks (as the current "wakeup" does). This is useful
680 for those interested in wake up timings of RT tasks.
684 Traces and records the max latency that it takes for
685 a SCHED_DEADLINE task to be woken (as the "wakeup" and
690 A special tracer that is used to trace binary module.
691 It will trace all the calls that a module makes to the
692 hardware. Everything it writes and reads from the I/O
697 This tracer can be configured when tracing likely/unlikely
698 calls within the kernel. It will trace when a likely and
699 unlikely branch is hit and if it was correct in its prediction
704 This is the "trace nothing" tracer. To remove all
705 tracers from tracing simply echo "nop" into
709 Examples of using the tracer
710 ----------------------------
712 Here are typical examples of using the tracers when controlling
713 them only with the tracefs interface (without using any
714 user-land utilities).
719 Here is an example of the output format of the file "trace"
724 # entries-in-buffer/entries-written: 140080/250280 #P:4
727 # / _----=> need-resched
728 # | / _---=> hardirq/softirq
729 # || / _--=> preempt-depth
731 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
733 bash-1977 [000] .... 17284.993652: sys_close <-system_call_fastpath
734 bash-1977 [000] .... 17284.993653: __close_fd <-sys_close
735 bash-1977 [000] .... 17284.993653: _raw_spin_lock <-__close_fd
736 sshd-1974 [003] .... 17284.993653: __srcu_read_unlock <-fsnotify
737 bash-1977 [000] .... 17284.993654: add_preempt_count <-_raw_spin_lock
738 bash-1977 [000] ...1 17284.993655: _raw_spin_unlock <-__close_fd
739 bash-1977 [000] ...1 17284.993656: sub_preempt_count <-_raw_spin_unlock
740 bash-1977 [000] .... 17284.993657: filp_close <-__close_fd
741 bash-1977 [000] .... 17284.993657: dnotify_flush <-filp_close
742 sshd-1974 [003] .... 17284.993658: sys_select <-system_call_fastpath
745 A header is printed with the tracer name that is represented by
746 the trace. In this case the tracer is "function". Then it shows the
747 number of events in the buffer as well as the total number of entries
748 that were written. The difference is the number of entries that were
749 lost due to the buffer filling up (250280 - 140080 = 110200 events
752 The header explains the content of the events. Task name "bash", the task
753 PID "1977", the CPU that it was running on "000", the latency format
754 (explained below), the timestamp in <secs>.<usecs> format, the
755 function name that was traced "sys_close" and the parent function that
756 called this function "system_call_fastpath". The timestamp is the time
757 at which the function was entered.
762 When the latency-format option is enabled or when one of the latency
763 tracers is set, the trace file gives somewhat more information to see
764 why a latency happened. Here is a typical trace.
768 # irqsoff latency trace v1.1.5 on 3.8.0-test+
769 # --------------------------------------------------------------------
770 # latency: 259 us, #4/4, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
772 # | task: ps-6143 (uid:0 nice:0 policy:0 rt_prio:0)
774 # => started at: __lock_task_sighand
775 # => ended at: _raw_spin_unlock_irqrestore
779 # / _-----=> irqs-off
780 # | / _----=> need-resched
781 # || / _---=> hardirq/softirq
782 # ||| / _--=> preempt-depth
784 # cmd pid ||||| time | caller
786 ps-6143 2d... 0us!: trace_hardirqs_off <-__lock_task_sighand
787 ps-6143 2d..1 259us+: trace_hardirqs_on <-_raw_spin_unlock_irqrestore
788 ps-6143 2d..1 263us+: time_hardirqs_on <-_raw_spin_unlock_irqrestore
789 ps-6143 2d..1 306us : <stack trace>
790 => trace_hardirqs_on_caller
792 => _raw_spin_unlock_irqrestore
799 => system_call_fastpath
802 This shows that the current tracer is "irqsoff" tracing the time
803 for which interrupts were disabled. It gives the trace version (which
804 never changes) and the version of the kernel upon which this was executed on
805 (3.8). Then it displays the max latency in microseconds (259 us). The number
806 of trace entries displayed and the total number (both are four: #4/4).
807 VP, KP, SP, and HP are always zero and are reserved for later use.
808 #P is the number of online CPUs (#P:4).
810 The task is the process that was running when the latency
811 occurred. (ps pid: 6143).
813 The start and stop (the functions in which the interrupts were
814 disabled and enabled respectively) that caused the latencies:
816 __lock_task_sighand is where the interrupts were disabled.
817 _raw_spin_unlock_irqrestore is where they were enabled again.
819 The next lines after the header are the trace itself. The header
820 explains which is which.
822 cmd: The name of the process in the trace.
824 pid: The PID of that process.
826 CPU#: The CPU which the process was running on.
828 irqs-off: 'd' interrupts are disabled. '.' otherwise.
829 Note: If the architecture does not support a way to
830 read the irq flags variable, an 'X' will always
834 'N' both TIF_NEED_RESCHED and PREEMPT_NEED_RESCHED is set,
835 'n' only TIF_NEED_RESCHED is set,
836 'p' only PREEMPT_NEED_RESCHED is set,
840 'Z' - NMI occurred inside a hardirq
842 'H' - hard irq occurred inside a softirq.
843 'h' - hard irq is running
844 's' - soft irq is running
845 '.' - normal context.
847 preempt-depth: The level of preempt_disabled
849 The above is mostly meaningful for kernel developers.
851 time: When the latency-format option is enabled, the trace file
852 output includes a timestamp relative to the start of the
853 trace. This differs from the output when latency-format
854 is disabled, which includes an absolute timestamp.
856 delay: This is just to help catch your eye a bit better. And
857 needs to be fixed to be only relative to the same CPU.
858 The marks are determined by the difference between this
859 current trace and the next trace.
860 '$' - greater than 1 second
861 '@' - greater than 100 milisecond
862 '*' - greater than 10 milisecond
863 '#' - greater than 1000 microsecond
864 '!' - greater than 100 microsecond
865 '+' - greater than 10 microsecond
866 ' ' - less than or equal to 10 microsecond.
868 The rest is the same as the 'trace' file.
870 Note, the latency tracers will usually end with a back trace
871 to easily find where the latency occurred.
876 The trace_options file (or the options directory) is used to control
877 what gets printed in the trace output, or manipulate the tracers.
878 To see what is available, simply cat the file:
909 To disable one of the options, echo in the option prepended with
912 echo noprint-parent > trace_options
914 To enable an option, leave off the "no".
916 echo sym-offset > trace_options
918 Here are the available options:
920 print-parent - On function traces, display the calling (parent)
921 function as well as the function being traced.
924 bash-4000 [01] 1477.606694: simple_strtoul <-kstrtoul
927 bash-4000 [01] 1477.606694: simple_strtoul
930 sym-offset - Display not only the function name, but also the
931 offset in the function. For example, instead of
932 seeing just "ktime_get", you will see
933 "ktime_get+0xb/0x20".
936 bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0
938 sym-addr - this will also display the function address as well
939 as the function name.
942 bash-4000 [01] 1477.606694: simple_strtoul <c0339346>
944 verbose - This deals with the trace file when the
945 latency-format option is enabled.
947 bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \
948 (+0.000ms): simple_strtoul (kstrtoul)
950 raw - This will display raw numbers. This option is best for
951 use with user applications that can translate the raw
952 numbers better than having it done in the kernel.
954 hex - Similar to raw, but the numbers will be in a hexadecimal
957 bin - This will print out the formats in raw binary.
959 block - When set, reading trace_pipe will not block when polled.
961 trace_printk - Can disable trace_printk() from writing into the buffer.
963 annotate - It is sometimes confusing when the CPU buffers are full
964 and one CPU buffer had a lot of events recently, thus
965 a shorter time frame, were another CPU may have only had
966 a few events, which lets it have older events. When
967 the trace is reported, it shows the oldest events first,
968 and it may look like only one CPU ran (the one with the
969 oldest events). When the annotate option is set, it will
970 display when a new CPU buffer started:
972 <idle>-0 [001] dNs4 21169.031481: wake_up_idle_cpu <-add_timer_on
973 <idle>-0 [001] dNs4 21169.031482: _raw_spin_unlock_irqrestore <-add_timer_on
974 <idle>-0 [001] .Ns4 21169.031484: sub_preempt_count <-_raw_spin_unlock_irqrestore
975 ##### CPU 2 buffer started ####
976 <idle>-0 [002] .N.1 21169.031484: rcu_idle_exit <-cpu_idle
977 <idle>-0 [001] .Ns3 21169.031484: _raw_spin_unlock <-clocksource_watchdog
978 <idle>-0 [001] .Ns3 21169.031485: sub_preempt_count <-_raw_spin_unlock
980 userstacktrace - This option changes the trace. It records a
981 stacktrace of the current user space thread after
984 sym-userobj - when user stacktrace are enabled, look up which
985 object the address belongs to, and print a
986 relative address. This is especially useful when
987 ASLR is on, otherwise you don't get a chance to
988 resolve the address to object/file/line after
989 the app is no longer running
991 The lookup is performed when you read
992 trace,trace_pipe. Example:
994 a.out-1623 [000] 40874.465068: /root/a.out[+0x480] <-/root/a.out[+0
995 x494] <- /root/a.out[+0x4a8] <- /lib/libc-2.7.so[+0x1e1a6]
998 printk-msg-only - When set, trace_printk()s will only show the format
999 and not their parameters (if trace_bprintk() or
1000 trace_bputs() was used to save the trace_printk()).
1002 context-info - Show only the event data. Hides the comm, PID,
1003 timestamp, CPU, and other useful data.
1005 latency-format - This option changes the trace output. When it is enabled,
1006 the trace displays additional information about the
1007 latency, as described in "Latency trace format".
1009 record-cmd - When any event or tracer is enabled, a hook is enabled
1010 in the sched_switch trace point to fill comm cache
1011 with mapped pids and comms. But this may cause some
1012 overhead, and if you only care about pids, and not the
1013 name of the task, disabling this option can lower the
1014 impact of tracing. See "saved_cmdlines".
1016 record-tgid - When any event or tracer is enabled, a hook is enabled
1017 in the sched_switch trace point to fill the cache of
1018 mapped Thread Group IDs (TGID) mapping to pids. See
1021 overwrite - This controls what happens when the trace buffer is
1022 full. If "1" (default), the oldest events are
1023 discarded and overwritten. If "0", then the newest
1024 events are discarded.
1025 (see per_cpu/cpu0/stats for overrun and dropped)
1027 disable_on_free - When the free_buffer is closed, tracing will
1028 stop (tracing_on set to 0).
1030 irq-info - Shows the interrupt, preempt count, need resched data.
1031 When disabled, the trace looks like:
1035 # entries-in-buffer/entries-written: 144405/9452052 #P:4
1037 # TASK-PID CPU# TIMESTAMP FUNCTION
1039 <idle>-0 [002] 23636.756054: ttwu_do_activate.constprop.89 <-try_to_wake_up
1040 <idle>-0 [002] 23636.756054: activate_task <-ttwu_do_activate.constprop.89
1041 <idle>-0 [002] 23636.756055: enqueue_task <-activate_task
1044 markers - When set, the trace_marker is writable (only by root).
1045 When disabled, the trace_marker will error with EINVAL
1048 event-fork - When set, tasks with PIDs listed in set_event_pid will have
1049 the PIDs of their children added to set_event_pid when those
1050 tasks fork. Also, when tasks with PIDs in set_event_pid exit,
1051 their PIDs will be removed from the file.
1053 function-trace - The latency tracers will enable function tracing
1054 if this option is enabled (default it is). When
1055 it is disabled, the latency tracers do not trace
1056 functions. This keeps the overhead of the tracer down
1057 when performing latency tests.
1059 function-fork - When set, tasks with PIDs listed in set_ftrace_pid will
1060 have the PIDs of their children added to set_ftrace_pid
1061 when those tasks fork. Also, when tasks with PIDs in
1062 set_ftrace_pid exit, their PIDs will be removed from the
1065 display-graph - When set, the latency tracers (irqsoff, wakeup, etc) will
1066 use function graph tracing instead of function tracing.
1068 stacktrace - When set, a stack trace is recorded after any trace event
1071 branch - Enable branch tracing with the tracer. This enables branch
1072 tracer along with the currently set tracer. Enabling this
1073 with the "nop" tracer is the same as just enabling the
1076 Note: Some tracers have their own options. They only appear in this
1077 file when the tracer is active. They always appear in the
1081 Here are the per tracer options:
1083 Options for function tracer:
1085 func_stack_trace - When set, a stack trace is recorded after every
1086 function that is recorded. NOTE! Limit the functions
1087 that are recorded before enabling this, with
1088 "set_ftrace_filter" otherwise the system performance
1089 will be critically degraded. Remember to disable
1090 this option before clearing the function filter.
1092 Options for function_graph tracer:
1094 Since the function_graph tracer has a slightly different output
1095 it has its own options to control what is displayed.
1097 funcgraph-overrun - When set, the "overrun" of the graph stack is
1098 displayed after each function traced. The
1099 overrun, is when the stack depth of the calls
1100 is greater than what is reserved for each task.
1101 Each task has a fixed array of functions to
1102 trace in the call graph. If the depth of the
1103 calls exceeds that, the function is not traced.
1104 The overrun is the number of functions missed
1105 due to exceeding this array.
1107 funcgraph-cpu - When set, the CPU number of the CPU where the trace
1108 occurred is displayed.
1110 funcgraph-overhead - When set, if the function takes longer than
1111 A certain amount, then a delay marker is
1112 displayed. See "delay" above, under the
1115 funcgraph-proc - Unlike other tracers, the process' command line
1116 is not displayed by default, but instead only
1117 when a task is traced in and out during a context
1118 switch. Enabling this options has the command
1119 of each process displayed at every line.
1121 funcgraph-duration - At the end of each function (the return)
1122 the duration of the amount of time in the
1123 function is displayed in microseconds.
1125 funcgraph-abstime - When set, the timestamp is displayed at each
1128 funcgraph-irqs - When disabled, functions that happen inside an
1129 interrupt will not be traced.
1131 funcgraph-tail - When set, the return event will include the function
1132 that it represents. By default this is off, and
1133 only a closing curly bracket "}" is displayed for
1134 the return of a function.
1136 sleep-time - When running function graph tracer, to include
1137 the time a task schedules out in its function.
1138 When enabled, it will account time the task has been
1139 scheduled out as part of the function call.
1141 graph-time - When running function profiler with function graph tracer,
1142 to include the time to call nested functions. When this is
1143 not set, the time reported for the function will only
1144 include the time the function itself executed for, not the
1145 time for functions that it called.
1147 Options for blk tracer:
1149 blk_classic - Shows a more minimalistic output.
1155 When interrupts are disabled, the CPU can not react to any other
1156 external event (besides NMIs and SMIs). This prevents the timer
1157 interrupt from triggering or the mouse interrupt from letting
1158 the kernel know of a new mouse event. The result is a latency
1159 with the reaction time.
1161 The irqsoff tracer tracks the time for which interrupts are
1162 disabled. When a new maximum latency is hit, the tracer saves
1163 the trace leading up to that latency point so that every time a
1164 new maximum is reached, the old saved trace is discarded and the
1167 To reset the maximum, echo 0 into tracing_max_latency. Here is
1170 # echo 0 > options/function-trace
1171 # echo irqsoff > current_tracer
1172 # echo 1 > tracing_on
1173 # echo 0 > tracing_max_latency
1176 # echo 0 > tracing_on
1180 # irqsoff latency trace v1.1.5 on 3.8.0-test+
1181 # --------------------------------------------------------------------
1182 # latency: 16 us, #4/4, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1184 # | task: swapper/0-0 (uid:0 nice:0 policy:0 rt_prio:0)
1186 # => started at: run_timer_softirq
1187 # => ended at: run_timer_softirq
1191 # / _-----=> irqs-off
1192 # | / _----=> need-resched
1193 # || / _---=> hardirq/softirq
1194 # ||| / _--=> preempt-depth
1196 # cmd pid ||||| time | caller
1198 <idle>-0 0d.s2 0us+: _raw_spin_lock_irq <-run_timer_softirq
1199 <idle>-0 0dNs3 17us : _raw_spin_unlock_irq <-run_timer_softirq
1200 <idle>-0 0dNs3 17us+: trace_hardirqs_on <-run_timer_softirq
1201 <idle>-0 0dNs3 25us : <stack trace>
1202 => _raw_spin_unlock_irq
1203 => run_timer_softirq
1208 => smp_apic_timer_interrupt
1209 => apic_timer_interrupt
1214 => x86_64_start_reservations
1215 => x86_64_start_kernel
1217 Here we see that that we had a latency of 16 microseconds (which is
1218 very good). The _raw_spin_lock_irq in run_timer_softirq disabled
1219 interrupts. The difference between the 16 and the displayed
1220 timestamp 25us occurred because the clock was incremented
1221 between the time of recording the max latency and the time of
1222 recording the function that had that latency.
1224 Note the above example had function-trace not set. If we set
1225 function-trace, we get a much larger output:
1227 with echo 1 > options/function-trace
1231 # irqsoff latency trace v1.1.5 on 3.8.0-test+
1232 # --------------------------------------------------------------------
1233 # latency: 71 us, #168/168, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1235 # | task: bash-2042 (uid:0 nice:0 policy:0 rt_prio:0)
1237 # => started at: ata_scsi_queuecmd
1238 # => ended at: ata_scsi_queuecmd
1242 # / _-----=> irqs-off
1243 # | / _----=> need-resched
1244 # || / _---=> hardirq/softirq
1245 # ||| / _--=> preempt-depth
1247 # cmd pid ||||| time | caller
1249 bash-2042 3d... 0us : _raw_spin_lock_irqsave <-ata_scsi_queuecmd
1250 bash-2042 3d... 0us : add_preempt_count <-_raw_spin_lock_irqsave
1251 bash-2042 3d..1 1us : ata_scsi_find_dev <-ata_scsi_queuecmd
1252 bash-2042 3d..1 1us : __ata_scsi_find_dev <-ata_scsi_find_dev
1253 bash-2042 3d..1 2us : ata_find_dev.part.14 <-__ata_scsi_find_dev
1254 bash-2042 3d..1 2us : ata_qc_new_init <-__ata_scsi_queuecmd
1255 bash-2042 3d..1 3us : ata_sg_init <-__ata_scsi_queuecmd
1256 bash-2042 3d..1 4us : ata_scsi_rw_xlat <-__ata_scsi_queuecmd
1257 bash-2042 3d..1 4us : ata_build_rw_tf <-ata_scsi_rw_xlat
1259 bash-2042 3d..1 67us : delay_tsc <-__delay
1260 bash-2042 3d..1 67us : add_preempt_count <-delay_tsc
1261 bash-2042 3d..2 67us : sub_preempt_count <-delay_tsc
1262 bash-2042 3d..1 67us : add_preempt_count <-delay_tsc
1263 bash-2042 3d..2 68us : sub_preempt_count <-delay_tsc
1264 bash-2042 3d..1 68us+: ata_bmdma_start <-ata_bmdma_qc_issue
1265 bash-2042 3d..1 71us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd
1266 bash-2042 3d..1 71us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd
1267 bash-2042 3d..1 72us+: trace_hardirqs_on <-ata_scsi_queuecmd
1268 bash-2042 3d..1 120us : <stack trace>
1269 => _raw_spin_unlock_irqrestore
1270 => ata_scsi_queuecmd
1271 => scsi_dispatch_cmd
1273 => __blk_run_queue_uncond
1276 => generic_make_request
1279 => __ext3_get_inode_loc
1288 => user_path_at_empty
1293 => system_call_fastpath
1296 Here we traced a 71 microsecond latency. But we also see all the
1297 functions that were called during that time. Note that by
1298 enabling function tracing, we incur an added overhead. This
1299 overhead may extend the latency times. But nevertheless, this
1300 trace has provided some very helpful debugging information.
1306 When preemption is disabled, we may be able to receive
1307 interrupts but the task cannot be preempted and a higher
1308 priority task must wait for preemption to be enabled again
1309 before it can preempt a lower priority task.
1311 The preemptoff tracer traces the places that disable preemption.
1312 Like the irqsoff tracer, it records the maximum latency for
1313 which preemption was disabled. The control of preemptoff tracer
1314 is much like the irqsoff tracer.
1316 # echo 0 > options/function-trace
1317 # echo preemptoff > current_tracer
1318 # echo 1 > tracing_on
1319 # echo 0 > tracing_max_latency
1322 # echo 0 > tracing_on
1324 # tracer: preemptoff
1326 # preemptoff latency trace v1.1.5 on 3.8.0-test+
1327 # --------------------------------------------------------------------
1328 # latency: 46 us, #4/4, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1330 # | task: sshd-1991 (uid:0 nice:0 policy:0 rt_prio:0)
1332 # => started at: do_IRQ
1333 # => ended at: do_IRQ
1337 # / _-----=> irqs-off
1338 # | / _----=> need-resched
1339 # || / _---=> hardirq/softirq
1340 # ||| / _--=> preempt-depth
1342 # cmd pid ||||| time | caller
1344 sshd-1991 1d.h. 0us+: irq_enter <-do_IRQ
1345 sshd-1991 1d..1 46us : irq_exit <-do_IRQ
1346 sshd-1991 1d..1 47us+: trace_preempt_on <-do_IRQ
1347 sshd-1991 1d..1 52us : <stack trace>
1348 => sub_preempt_count
1354 This has some more changes. Preemption was disabled when an
1355 interrupt came in (notice the 'h'), and was enabled on exit.
1356 But we also see that interrupts have been disabled when entering
1357 the preempt off section and leaving it (the 'd'). We do not know if
1358 interrupts were enabled in the mean time or shortly after this
1361 # tracer: preemptoff
1363 # preemptoff latency trace v1.1.5 on 3.8.0-test+
1364 # --------------------------------------------------------------------
1365 # latency: 83 us, #241/241, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1367 # | task: bash-1994 (uid:0 nice:0 policy:0 rt_prio:0)
1369 # => started at: wake_up_new_task
1370 # => ended at: task_rq_unlock
1374 # / _-----=> irqs-off
1375 # | / _----=> need-resched
1376 # || / _---=> hardirq/softirq
1377 # ||| / _--=> preempt-depth
1379 # cmd pid ||||| time | caller
1381 bash-1994 1d..1 0us : _raw_spin_lock_irqsave <-wake_up_new_task
1382 bash-1994 1d..1 0us : select_task_rq_fair <-select_task_rq
1383 bash-1994 1d..1 1us : __rcu_read_lock <-select_task_rq_fair
1384 bash-1994 1d..1 1us : source_load <-select_task_rq_fair
1385 bash-1994 1d..1 1us : source_load <-select_task_rq_fair
1387 bash-1994 1d..1 12us : irq_enter <-smp_apic_timer_interrupt
1388 bash-1994 1d..1 12us : rcu_irq_enter <-irq_enter
1389 bash-1994 1d..1 13us : add_preempt_count <-irq_enter
1390 bash-1994 1d.h1 13us : exit_idle <-smp_apic_timer_interrupt
1391 bash-1994 1d.h1 13us : hrtimer_interrupt <-smp_apic_timer_interrupt
1392 bash-1994 1d.h1 13us : _raw_spin_lock <-hrtimer_interrupt
1393 bash-1994 1d.h1 14us : add_preempt_count <-_raw_spin_lock
1394 bash-1994 1d.h2 14us : ktime_get_update_offsets <-hrtimer_interrupt
1396 bash-1994 1d.h1 35us : lapic_next_event <-clockevents_program_event
1397 bash-1994 1d.h1 35us : irq_exit <-smp_apic_timer_interrupt
1398 bash-1994 1d.h1 36us : sub_preempt_count <-irq_exit
1399 bash-1994 1d..2 36us : do_softirq <-irq_exit
1400 bash-1994 1d..2 36us : __do_softirq <-call_softirq
1401 bash-1994 1d..2 36us : __local_bh_disable <-__do_softirq
1402 bash-1994 1d.s2 37us : add_preempt_count <-_raw_spin_lock_irq
1403 bash-1994 1d.s3 38us : _raw_spin_unlock <-run_timer_softirq
1404 bash-1994 1d.s3 39us : sub_preempt_count <-_raw_spin_unlock
1405 bash-1994 1d.s2 39us : call_timer_fn <-run_timer_softirq
1407 bash-1994 1dNs2 81us : cpu_needs_another_gp <-rcu_process_callbacks
1408 bash-1994 1dNs2 82us : __local_bh_enable <-__do_softirq
1409 bash-1994 1dNs2 82us : sub_preempt_count <-__local_bh_enable
1410 bash-1994 1dN.2 82us : idle_cpu <-irq_exit
1411 bash-1994 1dN.2 83us : rcu_irq_exit <-irq_exit
1412 bash-1994 1dN.2 83us : sub_preempt_count <-irq_exit
1413 bash-1994 1.N.1 84us : _raw_spin_unlock_irqrestore <-task_rq_unlock
1414 bash-1994 1.N.1 84us+: trace_preempt_on <-task_rq_unlock
1415 bash-1994 1.N.1 104us : <stack trace>
1416 => sub_preempt_count
1417 => _raw_spin_unlock_irqrestore
1425 The above is an example of the preemptoff trace with
1426 function-trace set. Here we see that interrupts were not disabled
1427 the entire time. The irq_enter code lets us know that we entered
1428 an interrupt 'h'. Before that, the functions being traced still
1429 show that it is not in an interrupt, but we can see from the
1430 functions themselves that this is not the case.
1435 Knowing the locations that have interrupts disabled or
1436 preemption disabled for the longest times is helpful. But
1437 sometimes we would like to know when either preemption and/or
1438 interrupts are disabled.
1440 Consider the following code:
1442 local_irq_disable();
1443 call_function_with_irqs_off();
1445 call_function_with_irqs_and_preemption_off();
1447 call_function_with_preemption_off();
1450 The irqsoff tracer will record the total length of
1451 call_function_with_irqs_off() and
1452 call_function_with_irqs_and_preemption_off().
1454 The preemptoff tracer will record the total length of
1455 call_function_with_irqs_and_preemption_off() and
1456 call_function_with_preemption_off().
1458 But neither will trace the time that interrupts and/or
1459 preemption is disabled. This total time is the time that we can
1460 not schedule. To record this time, use the preemptirqsoff
1463 Again, using this trace is much like the irqsoff and preemptoff
1466 # echo 0 > options/function-trace
1467 # echo preemptirqsoff > current_tracer
1468 # echo 1 > tracing_on
1469 # echo 0 > tracing_max_latency
1472 # echo 0 > tracing_on
1474 # tracer: preemptirqsoff
1476 # preemptirqsoff latency trace v1.1.5 on 3.8.0-test+
1477 # --------------------------------------------------------------------
1478 # latency: 100 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1480 # | task: ls-2230 (uid:0 nice:0 policy:0 rt_prio:0)
1482 # => started at: ata_scsi_queuecmd
1483 # => ended at: ata_scsi_queuecmd
1487 # / _-----=> irqs-off
1488 # | / _----=> need-resched
1489 # || / _---=> hardirq/softirq
1490 # ||| / _--=> preempt-depth
1492 # cmd pid ||||| time | caller
1494 ls-2230 3d... 0us+: _raw_spin_lock_irqsave <-ata_scsi_queuecmd
1495 ls-2230 3...1 100us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd
1496 ls-2230 3...1 101us+: trace_preempt_on <-ata_scsi_queuecmd
1497 ls-2230 3...1 111us : <stack trace>
1498 => sub_preempt_count
1499 => _raw_spin_unlock_irqrestore
1500 => ata_scsi_queuecmd
1501 => scsi_dispatch_cmd
1503 => __blk_run_queue_uncond
1506 => generic_make_request
1511 => htree_dirblock_to_tree
1512 => ext3_htree_fill_tree
1516 => system_call_fastpath
1519 The trace_hardirqs_off_thunk is called from assembly on x86 when
1520 interrupts are disabled in the assembly code. Without the
1521 function tracing, we do not know if interrupts were enabled
1522 within the preemption points. We do see that it started with
1525 Here is a trace with function-trace set:
1527 # tracer: preemptirqsoff
1529 # preemptirqsoff latency trace v1.1.5 on 3.8.0-test+
1530 # --------------------------------------------------------------------
1531 # latency: 161 us, #339/339, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1533 # | task: ls-2269 (uid:0 nice:0 policy:0 rt_prio:0)
1535 # => started at: schedule
1536 # => ended at: mutex_unlock
1540 # / _-----=> irqs-off
1541 # | / _----=> need-resched
1542 # || / _---=> hardirq/softirq
1543 # ||| / _--=> preempt-depth
1545 # cmd pid ||||| time | caller
1547 kworker/-59 3...1 0us : __schedule <-schedule
1548 kworker/-59 3d..1 0us : rcu_preempt_qs <-rcu_note_context_switch
1549 kworker/-59 3d..1 1us : add_preempt_count <-_raw_spin_lock_irq
1550 kworker/-59 3d..2 1us : deactivate_task <-__schedule
1551 kworker/-59 3d..2 1us : dequeue_task <-deactivate_task
1552 kworker/-59 3d..2 2us : update_rq_clock <-dequeue_task
1553 kworker/-59 3d..2 2us : dequeue_task_fair <-dequeue_task
1554 kworker/-59 3d..2 2us : update_curr <-dequeue_task_fair
1555 kworker/-59 3d..2 2us : update_min_vruntime <-update_curr
1556 kworker/-59 3d..2 3us : cpuacct_charge <-update_curr
1557 kworker/-59 3d..2 3us : __rcu_read_lock <-cpuacct_charge
1558 kworker/-59 3d..2 3us : __rcu_read_unlock <-cpuacct_charge
1559 kworker/-59 3d..2 3us : update_cfs_rq_blocked_load <-dequeue_task_fair
1560 kworker/-59 3d..2 4us : clear_buddies <-dequeue_task_fair
1561 kworker/-59 3d..2 4us : account_entity_dequeue <-dequeue_task_fair
1562 kworker/-59 3d..2 4us : update_min_vruntime <-dequeue_task_fair
1563 kworker/-59 3d..2 4us : update_cfs_shares <-dequeue_task_fair
1564 kworker/-59 3d..2 5us : hrtick_update <-dequeue_task_fair
1565 kworker/-59 3d..2 5us : wq_worker_sleeping <-__schedule
1566 kworker/-59 3d..2 5us : kthread_data <-wq_worker_sleeping
1567 kworker/-59 3d..2 5us : put_prev_task_fair <-__schedule
1568 kworker/-59 3d..2 6us : pick_next_task_fair <-pick_next_task
1569 kworker/-59 3d..2 6us : clear_buddies <-pick_next_task_fair
1570 kworker/-59 3d..2 6us : set_next_entity <-pick_next_task_fair
1571 kworker/-59 3d..2 6us : update_stats_wait_end <-set_next_entity
1572 ls-2269 3d..2 7us : finish_task_switch <-__schedule
1573 ls-2269 3d..2 7us : _raw_spin_unlock_irq <-finish_task_switch
1574 ls-2269 3d..2 8us : do_IRQ <-ret_from_intr
1575 ls-2269 3d..2 8us : irq_enter <-do_IRQ
1576 ls-2269 3d..2 8us : rcu_irq_enter <-irq_enter
1577 ls-2269 3d..2 9us : add_preempt_count <-irq_enter
1578 ls-2269 3d.h2 9us : exit_idle <-do_IRQ
1580 ls-2269 3d.h3 20us : sub_preempt_count <-_raw_spin_unlock
1581 ls-2269 3d.h2 20us : irq_exit <-do_IRQ
1582 ls-2269 3d.h2 21us : sub_preempt_count <-irq_exit
1583 ls-2269 3d..3 21us : do_softirq <-irq_exit
1584 ls-2269 3d..3 21us : __do_softirq <-call_softirq
1585 ls-2269 3d..3 21us+: __local_bh_disable <-__do_softirq
1586 ls-2269 3d.s4 29us : sub_preempt_count <-_local_bh_enable_ip
1587 ls-2269 3d.s5 29us : sub_preempt_count <-_local_bh_enable_ip
1588 ls-2269 3d.s5 31us : do_IRQ <-ret_from_intr
1589 ls-2269 3d.s5 31us : irq_enter <-do_IRQ
1590 ls-2269 3d.s5 31us : rcu_irq_enter <-irq_enter
1592 ls-2269 3d.s5 31us : rcu_irq_enter <-irq_enter
1593 ls-2269 3d.s5 32us : add_preempt_count <-irq_enter
1594 ls-2269 3d.H5 32us : exit_idle <-do_IRQ
1595 ls-2269 3d.H5 32us : handle_irq <-do_IRQ
1596 ls-2269 3d.H5 32us : irq_to_desc <-handle_irq
1597 ls-2269 3d.H5 33us : handle_fasteoi_irq <-handle_irq
1599 ls-2269 3d.s5 158us : _raw_spin_unlock_irqrestore <-rtl8139_poll
1600 ls-2269 3d.s3 158us : net_rps_action_and_irq_enable.isra.65 <-net_rx_action
1601 ls-2269 3d.s3 159us : __local_bh_enable <-__do_softirq
1602 ls-2269 3d.s3 159us : sub_preempt_count <-__local_bh_enable
1603 ls-2269 3d..3 159us : idle_cpu <-irq_exit
1604 ls-2269 3d..3 159us : rcu_irq_exit <-irq_exit
1605 ls-2269 3d..3 160us : sub_preempt_count <-irq_exit
1606 ls-2269 3d... 161us : __mutex_unlock_slowpath <-mutex_unlock
1607 ls-2269 3d... 162us+: trace_hardirqs_on <-mutex_unlock
1608 ls-2269 3d... 186us : <stack trace>
1609 => __mutex_unlock_slowpath
1616 => system_call_fastpath
1618 This is an interesting trace. It started with kworker running and
1619 scheduling out and ls taking over. But as soon as ls released the
1620 rq lock and enabled interrupts (but not preemption) an interrupt
1621 triggered. When the interrupt finished, it started running softirqs.
1622 But while the softirq was running, another interrupt triggered.
1623 When an interrupt is running inside a softirq, the annotation is 'H'.
1629 One common case that people are interested in tracing is the
1630 time it takes for a task that is woken to actually wake up.
1631 Now for non Real-Time tasks, this can be arbitrary. But tracing
1632 it none the less can be interesting.
1634 Without function tracing:
1636 # echo 0 > options/function-trace
1637 # echo wakeup > current_tracer
1638 # echo 1 > tracing_on
1639 # echo 0 > tracing_max_latency
1641 # echo 0 > tracing_on
1645 # wakeup latency trace v1.1.5 on 3.8.0-test+
1646 # --------------------------------------------------------------------
1647 # latency: 15 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1649 # | task: kworker/3:1H-312 (uid:0 nice:-20 policy:0 rt_prio:0)
1653 # / _-----=> irqs-off
1654 # | / _----=> need-resched
1655 # || / _---=> hardirq/softirq
1656 # ||| / _--=> preempt-depth
1658 # cmd pid ||||| time | caller
1660 <idle>-0 3dNs7 0us : 0:120:R + [003] 312:100:R kworker/3:1H
1661 <idle>-0 3dNs7 1us+: ttwu_do_activate.constprop.87 <-try_to_wake_up
1662 <idle>-0 3d..3 15us : __schedule <-schedule
1663 <idle>-0 3d..3 15us : 0:120:R ==> [003] 312:100:R kworker/3:1H
1665 The tracer only traces the highest priority task in the system
1666 to avoid tracing the normal circumstances. Here we see that
1667 the kworker with a nice priority of -20 (not very nice), took
1668 just 15 microseconds from the time it woke up, to the time it
1671 Non Real-Time tasks are not that interesting. A more interesting
1672 trace is to concentrate only on Real-Time tasks.
1677 In a Real-Time environment it is very important to know the
1678 wakeup time it takes for the highest priority task that is woken
1679 up to the time that it executes. This is also known as "schedule
1680 latency". I stress the point that this is about RT tasks. It is
1681 also important to know the scheduling latency of non-RT tasks,
1682 but the average schedule latency is better for non-RT tasks.
1683 Tools like LatencyTop are more appropriate for such
1686 Real-Time environments are interested in the worst case latency.
1687 That is the longest latency it takes for something to happen,
1688 and not the average. We can have a very fast scheduler that may
1689 only have a large latency once in a while, but that would not
1690 work well with Real-Time tasks. The wakeup_rt tracer was designed
1691 to record the worst case wakeups of RT tasks. Non-RT tasks are
1692 not recorded because the tracer only records one worst case and
1693 tracing non-RT tasks that are unpredictable will overwrite the
1694 worst case latency of RT tasks (just run the normal wakeup
1695 tracer for a while to see that effect).
1697 Since this tracer only deals with RT tasks, we will run this
1698 slightly differently than we did with the previous tracers.
1699 Instead of performing an 'ls', we will run 'sleep 1' under
1700 'chrt' which changes the priority of the task.
1702 # echo 0 > options/function-trace
1703 # echo wakeup_rt > current_tracer
1704 # echo 1 > tracing_on
1705 # echo 0 > tracing_max_latency
1707 # echo 0 > tracing_on
1713 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+
1714 # --------------------------------------------------------------------
1715 # latency: 5 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1717 # | task: sleep-2389 (uid:0 nice:0 policy:1 rt_prio:5)
1721 # / _-----=> irqs-off
1722 # | / _----=> need-resched
1723 # || / _---=> hardirq/softirq
1724 # ||| / _--=> preempt-depth
1726 # cmd pid ||||| time | caller
1728 <idle>-0 3d.h4 0us : 0:120:R + [003] 2389: 94:R sleep
1729 <idle>-0 3d.h4 1us+: ttwu_do_activate.constprop.87 <-try_to_wake_up
1730 <idle>-0 3d..3 5us : __schedule <-schedule
1731 <idle>-0 3d..3 5us : 0:120:R ==> [003] 2389: 94:R sleep
1734 Running this on an idle system, we see that it only took 5 microseconds
1735 to perform the task switch. Note, since the trace point in the schedule
1736 is before the actual "switch", we stop the tracing when the recorded task
1737 is about to schedule in. This may change if we add a new marker at the
1738 end of the scheduler.
1740 Notice that the recorded task is 'sleep' with the PID of 2389
1741 and it has an rt_prio of 5. This priority is user-space priority
1742 and not the internal kernel priority. The policy is 1 for
1743 SCHED_FIFO and 2 for SCHED_RR.
1745 Note, that the trace data shows the internal priority (99 - rtprio).
1747 <idle>-0 3d..3 5us : 0:120:R ==> [003] 2389: 94:R sleep
1749 The 0:120:R means idle was running with a nice priority of 0 (120 - 120)
1750 and in the running state 'R'. The sleep task was scheduled in with
1751 2389: 94:R. That is the priority is the kernel rtprio (99 - 5 = 94)
1752 and it too is in the running state.
1754 Doing the same with chrt -r 5 and function-trace set.
1756 echo 1 > options/function-trace
1760 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+
1761 # --------------------------------------------------------------------
1762 # latency: 29 us, #85/85, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1764 # | task: sleep-2448 (uid:0 nice:0 policy:1 rt_prio:5)
1768 # / _-----=> irqs-off
1769 # | / _----=> need-resched
1770 # || / _---=> hardirq/softirq
1771 # ||| / _--=> preempt-depth
1773 # cmd pid ||||| time | caller
1775 <idle>-0 3d.h4 1us+: 0:120:R + [003] 2448: 94:R sleep
1776 <idle>-0 3d.h4 2us : ttwu_do_activate.constprop.87 <-try_to_wake_up
1777 <idle>-0 3d.h3 3us : check_preempt_curr <-ttwu_do_wakeup
1778 <idle>-0 3d.h3 3us : resched_curr <-check_preempt_curr
1779 <idle>-0 3dNh3 4us : task_woken_rt <-ttwu_do_wakeup
1780 <idle>-0 3dNh3 4us : _raw_spin_unlock <-try_to_wake_up
1781 <idle>-0 3dNh3 4us : sub_preempt_count <-_raw_spin_unlock
1782 <idle>-0 3dNh2 5us : ttwu_stat <-try_to_wake_up
1783 <idle>-0 3dNh2 5us : _raw_spin_unlock_irqrestore <-try_to_wake_up
1784 <idle>-0 3dNh2 6us : sub_preempt_count <-_raw_spin_unlock_irqrestore
1785 <idle>-0 3dNh1 6us : _raw_spin_lock <-__run_hrtimer
1786 <idle>-0 3dNh1 6us : add_preempt_count <-_raw_spin_lock
1787 <idle>-0 3dNh2 7us : _raw_spin_unlock <-hrtimer_interrupt
1788 <idle>-0 3dNh2 7us : sub_preempt_count <-_raw_spin_unlock
1789 <idle>-0 3dNh1 7us : tick_program_event <-hrtimer_interrupt
1790 <idle>-0 3dNh1 7us : clockevents_program_event <-tick_program_event
1791 <idle>-0 3dNh1 8us : ktime_get <-clockevents_program_event
1792 <idle>-0 3dNh1 8us : lapic_next_event <-clockevents_program_event
1793 <idle>-0 3dNh1 8us : irq_exit <-smp_apic_timer_interrupt
1794 <idle>-0 3dNh1 9us : sub_preempt_count <-irq_exit
1795 <idle>-0 3dN.2 9us : idle_cpu <-irq_exit
1796 <idle>-0 3dN.2 9us : rcu_irq_exit <-irq_exit
1797 <idle>-0 3dN.2 10us : rcu_eqs_enter_common.isra.45 <-rcu_irq_exit
1798 <idle>-0 3dN.2 10us : sub_preempt_count <-irq_exit
1799 <idle>-0 3.N.1 11us : rcu_idle_exit <-cpu_idle
1800 <idle>-0 3dN.1 11us : rcu_eqs_exit_common.isra.43 <-rcu_idle_exit
1801 <idle>-0 3.N.1 11us : tick_nohz_idle_exit <-cpu_idle
1802 <idle>-0 3dN.1 12us : menu_hrtimer_cancel <-tick_nohz_idle_exit
1803 <idle>-0 3dN.1 12us : ktime_get <-tick_nohz_idle_exit
1804 <idle>-0 3dN.1 12us : tick_do_update_jiffies64 <-tick_nohz_idle_exit
1805 <idle>-0 3dN.1 13us : cpu_load_update_nohz <-tick_nohz_idle_exit
1806 <idle>-0 3dN.1 13us : _raw_spin_lock <-cpu_load_update_nohz
1807 <idle>-0 3dN.1 13us : add_preempt_count <-_raw_spin_lock
1808 <idle>-0 3dN.2 13us : __cpu_load_update <-cpu_load_update_nohz
1809 <idle>-0 3dN.2 14us : sched_avg_update <-__cpu_load_update
1810 <idle>-0 3dN.2 14us : _raw_spin_unlock <-cpu_load_update_nohz
1811 <idle>-0 3dN.2 14us : sub_preempt_count <-_raw_spin_unlock
1812 <idle>-0 3dN.1 15us : calc_load_nohz_stop <-tick_nohz_idle_exit
1813 <idle>-0 3dN.1 15us : touch_softlockup_watchdog <-tick_nohz_idle_exit
1814 <idle>-0 3dN.1 15us : hrtimer_cancel <-tick_nohz_idle_exit
1815 <idle>-0 3dN.1 15us : hrtimer_try_to_cancel <-hrtimer_cancel
1816 <idle>-0 3dN.1 16us : lock_hrtimer_base.isra.18 <-hrtimer_try_to_cancel
1817 <idle>-0 3dN.1 16us : _raw_spin_lock_irqsave <-lock_hrtimer_base.isra.18
1818 <idle>-0 3dN.1 16us : add_preempt_count <-_raw_spin_lock_irqsave
1819 <idle>-0 3dN.2 17us : __remove_hrtimer <-remove_hrtimer.part.16
1820 <idle>-0 3dN.2 17us : hrtimer_force_reprogram <-__remove_hrtimer
1821 <idle>-0 3dN.2 17us : tick_program_event <-hrtimer_force_reprogram
1822 <idle>-0 3dN.2 18us : clockevents_program_event <-tick_program_event
1823 <idle>-0 3dN.2 18us : ktime_get <-clockevents_program_event
1824 <idle>-0 3dN.2 18us : lapic_next_event <-clockevents_program_event
1825 <idle>-0 3dN.2 19us : _raw_spin_unlock_irqrestore <-hrtimer_try_to_cancel
1826 <idle>-0 3dN.2 19us : sub_preempt_count <-_raw_spin_unlock_irqrestore
1827 <idle>-0 3dN.1 19us : hrtimer_forward <-tick_nohz_idle_exit
1828 <idle>-0 3dN.1 20us : ktime_add_safe <-hrtimer_forward
1829 <idle>-0 3dN.1 20us : ktime_add_safe <-hrtimer_forward
1830 <idle>-0 3dN.1 20us : hrtimer_start_range_ns <-hrtimer_start_expires.constprop.11
1831 <idle>-0 3dN.1 20us : __hrtimer_start_range_ns <-hrtimer_start_range_ns
1832 <idle>-0 3dN.1 21us : lock_hrtimer_base.isra.18 <-__hrtimer_start_range_ns
1833 <idle>-0 3dN.1 21us : _raw_spin_lock_irqsave <-lock_hrtimer_base.isra.18
1834 <idle>-0 3dN.1 21us : add_preempt_count <-_raw_spin_lock_irqsave
1835 <idle>-0 3dN.2 22us : ktime_add_safe <-__hrtimer_start_range_ns
1836 <idle>-0 3dN.2 22us : enqueue_hrtimer <-__hrtimer_start_range_ns
1837 <idle>-0 3dN.2 22us : tick_program_event <-__hrtimer_start_range_ns
1838 <idle>-0 3dN.2 23us : clockevents_program_event <-tick_program_event
1839 <idle>-0 3dN.2 23us : ktime_get <-clockevents_program_event
1840 <idle>-0 3dN.2 23us : lapic_next_event <-clockevents_program_event
1841 <idle>-0 3dN.2 24us : _raw_spin_unlock_irqrestore <-__hrtimer_start_range_ns
1842 <idle>-0 3dN.2 24us : sub_preempt_count <-_raw_spin_unlock_irqrestore
1843 <idle>-0 3dN.1 24us : account_idle_ticks <-tick_nohz_idle_exit
1844 <idle>-0 3dN.1 24us : account_idle_time <-account_idle_ticks
1845 <idle>-0 3.N.1 25us : sub_preempt_count <-cpu_idle
1846 <idle>-0 3.N.. 25us : schedule <-cpu_idle
1847 <idle>-0 3.N.. 25us : __schedule <-preempt_schedule
1848 <idle>-0 3.N.. 26us : add_preempt_count <-__schedule
1849 <idle>-0 3.N.1 26us : rcu_note_context_switch <-__schedule
1850 <idle>-0 3.N.1 26us : rcu_sched_qs <-rcu_note_context_switch
1851 <idle>-0 3dN.1 27us : rcu_preempt_qs <-rcu_note_context_switch
1852 <idle>-0 3.N.1 27us : _raw_spin_lock_irq <-__schedule
1853 <idle>-0 3dN.1 27us : add_preempt_count <-_raw_spin_lock_irq
1854 <idle>-0 3dN.2 28us : put_prev_task_idle <-__schedule
1855 <idle>-0 3dN.2 28us : pick_next_task_stop <-pick_next_task
1856 <idle>-0 3dN.2 28us : pick_next_task_rt <-pick_next_task
1857 <idle>-0 3dN.2 29us : dequeue_pushable_task <-pick_next_task_rt
1858 <idle>-0 3d..3 29us : __schedule <-preempt_schedule
1859 <idle>-0 3d..3 30us : 0:120:R ==> [003] 2448: 94:R sleep
1861 This isn't that big of a trace, even with function tracing enabled,
1862 so I included the entire trace.
1864 The interrupt went off while when the system was idle. Somewhere
1865 before task_woken_rt() was called, the NEED_RESCHED flag was set,
1866 this is indicated by the first occurrence of the 'N' flag.
1868 Latency tracing and events
1869 --------------------------
1870 As function tracing can induce a much larger latency, but without
1871 seeing what happens within the latency it is hard to know what
1872 caused it. There is a middle ground, and that is with enabling
1875 # echo 0 > options/function-trace
1876 # echo wakeup_rt > current_tracer
1877 # echo 1 > events/enable
1878 # echo 1 > tracing_on
1879 # echo 0 > tracing_max_latency
1881 # echo 0 > tracing_on
1885 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+
1886 # --------------------------------------------------------------------
1887 # latency: 6 us, #12/12, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4)
1889 # | task: sleep-5882 (uid:0 nice:0 policy:1 rt_prio:5)
1893 # / _-----=> irqs-off
1894 # | / _----=> need-resched
1895 # || / _---=> hardirq/softirq
1896 # ||| / _--=> preempt-depth
1898 # cmd pid ||||| time | caller
1900 <idle>-0 2d.h4 0us : 0:120:R + [002] 5882: 94:R sleep
1901 <idle>-0 2d.h4 0us : ttwu_do_activate.constprop.87 <-try_to_wake_up
1902 <idle>-0 2d.h4 1us : sched_wakeup: comm=sleep pid=5882 prio=94 success=1 target_cpu=002
1903 <idle>-0 2dNh2 1us : hrtimer_expire_exit: hrtimer=ffff88007796feb8
1904 <idle>-0 2.N.2 2us : power_end: cpu_id=2
1905 <idle>-0 2.N.2 3us : cpu_idle: state=4294967295 cpu_id=2
1906 <idle>-0 2dN.3 4us : hrtimer_cancel: hrtimer=ffff88007d50d5e0
1907 <idle>-0 2dN.3 4us : hrtimer_start: hrtimer=ffff88007d50d5e0 function=tick_sched_timer expires=34311211000000 softexpires=34311211000000
1908 <idle>-0 2.N.2 5us : rcu_utilization: Start context switch
1909 <idle>-0 2.N.2 5us : rcu_utilization: End context switch
1910 <idle>-0 2d..3 6us : __schedule <-schedule
1911 <idle>-0 2d..3 6us : 0:120:R ==> [002] 5882: 94:R sleep
1914 Hardware Latency Detector
1915 -------------------------
1917 The hardware latency detector is executed by enabling the "hwlat" tracer.
1919 NOTE, this tracer will affect the performance of the system as it will
1920 periodically make a CPU constantly busy with interrupts disabled.
1922 # echo hwlat > current_tracer
1928 # / _----=> need-resched
1929 # | / _---=> hardirq/softirq
1930 # || / _--=> preempt-depth
1932 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
1934 <...>-3638 [001] d... 19452.055471: #1 inner/outer(us): 12/14 ts:1499801089.066141940
1935 <...>-3638 [003] d... 19454.071354: #2 inner/outer(us): 11/9 ts:1499801091.082164365
1936 <...>-3638 [002] dn.. 19461.126852: #3 inner/outer(us): 12/9 ts:1499801098.138150062
1937 <...>-3638 [001] d... 19488.340960: #4 inner/outer(us): 8/12 ts:1499801125.354139633
1938 <...>-3638 [003] d... 19494.388553: #5 inner/outer(us): 8/12 ts:1499801131.402150961
1939 <...>-3638 [003] d... 19501.283419: #6 inner/outer(us): 0/12 ts:1499801138.297435289 nmi-total:4 nmi-count:1
1942 The above output is somewhat the same in the header. All events will have
1943 interrupts disabled 'd'. Under the FUNCTION title there is:
1945 #1 - This is the count of events recorded that were greater than the
1946 tracing_threshold (See below).
1948 inner/outer(us): 12/14
1950 This shows two numbers as "inner latency" and "outer latency". The test
1951 runs in a loop checking a timestamp twice. The latency detected within
1952 the two timestamps is the "inner latency" and the latency detected
1953 after the previous timestamp and the next timestamp in the loop is
1954 the "outer latency".
1956 ts:1499801089.066141940
1958 The absolute timestamp that the event happened.
1960 nmi-total:4 nmi-count:1
1962 On architectures that support it, if an NMI comes in during the
1963 test, the time spent in NMI is reported in "nmi-total" (in
1966 All architectures that have NMIs will show the "nmi-count" if an
1967 NMI comes in during the test.
1971 tracing_threshold - This gets automatically set to "10" to represent 10
1972 microseconds. This is the threshold of latency that
1973 needs to be detected before the trace will be recorded.
1975 Note, when hwlat tracer is finished (another tracer is
1976 written into "current_tracer"), the original value for
1977 tracing_threshold is placed back into this file.
1979 hwlat_detector/width - The length of time the test runs with interrupts
1982 hwlat_detector/window - The length of time of the window which the test
1983 runs. That is, the test will run for "width"
1984 microseconds per "window" microseconds
1986 tracing_cpumask - When the test is started. A kernel thread is created that
1987 runs the test. This thread will alternate between CPUs
1988 listed in the tracing_cpumask between each period
1989 (one "window"). To limit the test to specific CPUs
1990 set the mask in this file to only the CPUs that the test
1996 This tracer is the function tracer. Enabling the function tracer
1997 can be done from the debug file system. Make sure the
1998 ftrace_enabled is set; otherwise this tracer is a nop.
1999 See the "ftrace_enabled" section below.
2001 # sysctl kernel.ftrace_enabled=1
2002 # echo function > current_tracer
2003 # echo 1 > tracing_on
2005 # echo 0 > tracing_on
2009 # entries-in-buffer/entries-written: 24799/24799 #P:4
2012 # / _----=> need-resched
2013 # | / _---=> hardirq/softirq
2014 # || / _--=> preempt-depth
2016 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2018 bash-1994 [002] .... 3082.063030: mutex_unlock <-rb_simple_write
2019 bash-1994 [002] .... 3082.063031: __mutex_unlock_slowpath <-mutex_unlock
2020 bash-1994 [002] .... 3082.063031: __fsnotify_parent <-fsnotify_modify
2021 bash-1994 [002] .... 3082.063032: fsnotify <-fsnotify_modify
2022 bash-1994 [002] .... 3082.063032: __srcu_read_lock <-fsnotify
2023 bash-1994 [002] .... 3082.063032: add_preempt_count <-__srcu_read_lock
2024 bash-1994 [002] ...1 3082.063032: sub_preempt_count <-__srcu_read_lock
2025 bash-1994 [002] .... 3082.063033: __srcu_read_unlock <-fsnotify
2029 Note: function tracer uses ring buffers to store the above
2030 entries. The newest data may overwrite the oldest data.
2031 Sometimes using echo to stop the trace is not sufficient because
2032 the tracing could have overwritten the data that you wanted to
2033 record. For this reason, it is sometimes better to disable
2034 tracing directly from a program. This allows you to stop the
2035 tracing at the point that you hit the part that you are
2036 interested in. To disable the tracing directly from a C program,
2037 something like following code snippet can be used:
2041 int main(int argc, char *argv[]) {
2043 trace_fd = open(tracing_file("tracing_on"), O_WRONLY);
2045 if (condition_hit()) {
2046 write(trace_fd, "0", 1);
2052 Single thread tracing
2053 ---------------------
2055 By writing into set_ftrace_pid you can trace a
2056 single thread. For example:
2058 # cat set_ftrace_pid
2060 # echo 3111 > set_ftrace_pid
2061 # cat set_ftrace_pid
2063 # echo function > current_tracer
2067 # TASK-PID CPU# TIMESTAMP FUNCTION
2069 yum-updatesd-3111 [003] 1637.254676: finish_task_switch <-thread_return
2070 yum-updatesd-3111 [003] 1637.254681: hrtimer_cancel <-schedule_hrtimeout_range
2071 yum-updatesd-3111 [003] 1637.254682: hrtimer_try_to_cancel <-hrtimer_cancel
2072 yum-updatesd-3111 [003] 1637.254683: lock_hrtimer_base <-hrtimer_try_to_cancel
2073 yum-updatesd-3111 [003] 1637.254685: fget_light <-do_sys_poll
2074 yum-updatesd-3111 [003] 1637.254686: pipe_poll <-do_sys_poll
2075 # echo > set_ftrace_pid
2079 # TASK-PID CPU# TIMESTAMP FUNCTION
2081 ##### CPU 3 buffer started ####
2082 yum-updatesd-3111 [003] 1701.957688: free_poll_entry <-poll_freewait
2083 yum-updatesd-3111 [003] 1701.957689: remove_wait_queue <-free_poll_entry
2084 yum-updatesd-3111 [003] 1701.957691: fput <-free_poll_entry
2085 yum-updatesd-3111 [003] 1701.957692: audit_syscall_exit <-sysret_audit
2086 yum-updatesd-3111 [003] 1701.957693: path_put <-audit_syscall_exit
2088 If you want to trace a function when executing, you could use
2089 something like this simple program:
2093 #include <sys/types.h>
2094 #include <sys/stat.h>
2100 #define STR(x) _STR(x)
2101 #define MAX_PATH 256
2103 const char *find_tracefs(void)
2105 static char tracefs[MAX_PATH+1];
2106 static int tracefs_found;
2113 if ((fp = fopen("/proc/mounts","r")) == NULL) {
2114 perror("/proc/mounts");
2118 while (fscanf(fp, "%*s %"
2120 "s %99s %*s %*d %*d\n",
2121 tracefs, type) == 2) {
2122 if (strcmp(type, "tracefs") == 0)
2127 if (strcmp(type, "tracefs") != 0) {
2128 fprintf(stderr, "tracefs not mounted");
2132 strcat(tracefs, "/tracing/");
2138 const char *tracing_file(const char *file_name)
2140 static char trace_file[MAX_PATH+1];
2141 snprintf(trace_file, MAX_PATH, "%s/%s", find_tracefs(), file_name);
2145 int main (int argc, char **argv)
2155 ffd = open(tracing_file("current_tracer"), O_WRONLY);
2158 write(ffd, "nop", 3);
2160 fd = open(tracing_file("set_ftrace_pid"), O_WRONLY);
2161 s = sprintf(line, "%d\n", getpid());
2164 write(ffd, "function", 8);
2169 execvp(argv[1], argv+1);
2175 Or this simple script!
2180 tracefs=`sed -ne 's/^tracefs \(.*\) tracefs.*/\1/p' /proc/mounts`
2181 echo nop > $tracefs/tracing/current_tracer
2182 echo 0 > $tracefs/tracing/tracing_on
2183 echo $$ > $tracefs/tracing/set_ftrace_pid
2184 echo function > $tracefs/tracing/current_tracer
2185 echo 1 > $tracefs/tracing/tracing_on
2190 function graph tracer
2191 ---------------------------
2193 This tracer is similar to the function tracer except that it
2194 probes a function on its entry and its exit. This is done by
2195 using a dynamically allocated stack of return addresses in each
2196 task_struct. On function entry the tracer overwrites the return
2197 address of each function traced to set a custom probe. Thus the
2198 original return address is stored on the stack of return address
2201 Probing on both ends of a function leads to special features
2204 - measure of a function's time execution
2205 - having a reliable call stack to draw function calls graph
2207 This tracer is useful in several situations:
2209 - you want to find the reason of a strange kernel behavior and
2210 need to see what happens in detail on any areas (or specific
2213 - you are experiencing weird latencies but it's difficult to
2216 - you want to find quickly which path is taken by a specific
2219 - you just want to peek inside a working kernel and want to see
2222 # tracer: function_graph
2224 # CPU DURATION FUNCTION CALLS
2228 0) | do_sys_open() {
2230 0) | kmem_cache_alloc() {
2231 0) 1.382 us | __might_sleep();
2233 0) | strncpy_from_user() {
2234 0) | might_fault() {
2235 0) 1.389 us | __might_sleep();
2240 0) 0.668 us | _spin_lock();
2241 0) 0.570 us | expand_files();
2242 0) 0.586 us | _spin_unlock();
2245 There are several columns that can be dynamically
2246 enabled/disabled. You can use every combination of options you
2247 want, depending on your needs.
2249 - The cpu number on which the function executed is default
2250 enabled. It is sometimes better to only trace one cpu (see
2251 tracing_cpu_mask file) or you might sometimes see unordered
2252 function calls while cpu tracing switch.
2254 hide: echo nofuncgraph-cpu > trace_options
2255 show: echo funcgraph-cpu > trace_options
2257 - The duration (function's time of execution) is displayed on
2258 the closing bracket line of a function or on the same line
2259 than the current function in case of a leaf one. It is default
2262 hide: echo nofuncgraph-duration > trace_options
2263 show: echo funcgraph-duration > trace_options
2265 - The overhead field precedes the duration field in case of
2266 reached duration thresholds.
2268 hide: echo nofuncgraph-overhead > trace_options
2269 show: echo funcgraph-overhead > trace_options
2270 depends on: funcgraph-duration
2274 3) # 1837.709 us | } /* __switch_to */
2275 3) | finish_task_switch() {
2276 3) 0.313 us | _raw_spin_unlock_irq();
2278 3) # 1889.063 us | } /* __schedule */
2279 3) ! 140.417 us | } /* __schedule */
2280 3) # 2034.948 us | } /* schedule */
2281 3) * 33998.59 us | } /* schedule_preempt_disabled */
2285 1) 0.260 us | msecs_to_jiffies();
2286 1) 0.313 us | __rcu_read_unlock();
2289 1) 0.313 us | rcu_bh_qs();
2290 1) 0.313 us | __local_bh_enable();
2292 1) 0.365 us | idle_cpu();
2293 1) | rcu_irq_exit() {
2294 1) 0.417 us | rcu_eqs_enter_common.isra.47();
2298 1) @ 119760.2 us | }
2304 2) 0.417 us | scheduler_ipi();
2312 + means that the function exceeded 10 usecs.
2313 ! means that the function exceeded 100 usecs.
2314 # means that the function exceeded 1000 usecs.
2315 * means that the function exceeded 10 msecs.
2316 @ means that the function exceeded 100 msecs.
2317 $ means that the function exceeded 1 sec.
2320 - The task/pid field displays the thread cmdline and pid which
2321 executed the function. It is default disabled.
2323 hide: echo nofuncgraph-proc > trace_options
2324 show: echo funcgraph-proc > trace_options
2328 # tracer: function_graph
2330 # CPU TASK/PID DURATION FUNCTION CALLS
2332 0) sh-4802 | | d_free() {
2333 0) sh-4802 | | call_rcu() {
2334 0) sh-4802 | | __call_rcu() {
2335 0) sh-4802 | 0.616 us | rcu_process_gp_end();
2336 0) sh-4802 | 0.586 us | check_for_new_grace_period();
2337 0) sh-4802 | 2.899 us | }
2338 0) sh-4802 | 4.040 us | }
2339 0) sh-4802 | 5.151 us | }
2340 0) sh-4802 | + 49.370 us | }
2343 - The absolute time field is an absolute timestamp given by the
2344 system clock since it started. A snapshot of this time is
2345 given on each entry/exit of functions
2347 hide: echo nofuncgraph-abstime > trace_options
2348 show: echo funcgraph-abstime > trace_options
2353 # TIME CPU DURATION FUNCTION CALLS
2355 360.774522 | 1) 0.541 us | }
2356 360.774522 | 1) 4.663 us | }
2357 360.774523 | 1) 0.541 us | __wake_up_bit();
2358 360.774524 | 1) 6.796 us | }
2359 360.774524 | 1) 7.952 us | }
2360 360.774525 | 1) 9.063 us | }
2361 360.774525 | 1) 0.615 us | journal_mark_dirty();
2362 360.774527 | 1) 0.578 us | __brelse();
2363 360.774528 | 1) | reiserfs_prepare_for_journal() {
2364 360.774528 | 1) | unlock_buffer() {
2365 360.774529 | 1) | wake_up_bit() {
2366 360.774529 | 1) | bit_waitqueue() {
2367 360.774530 | 1) 0.594 us | __phys_addr();
2370 The function name is always displayed after the closing bracket
2371 for a function if the start of that function is not in the
2374 Display of the function name after the closing bracket may be
2375 enabled for functions whose start is in the trace buffer,
2376 allowing easier searching with grep for function durations.
2377 It is default disabled.
2379 hide: echo nofuncgraph-tail > trace_options
2380 show: echo funcgraph-tail > trace_options
2382 Example with nofuncgraph-tail (default):
2384 0) | kmem_cache_free() {
2385 0) 0.518 us | __phys_addr();
2389 Example with funcgraph-tail:
2391 0) | kmem_cache_free() {
2392 0) 0.518 us | __phys_addr();
2393 0) 1.757 us | } /* kmem_cache_free() */
2394 0) 2.861 us | } /* putname() */
2396 You can put some comments on specific functions by using
2397 trace_printk() For example, if you want to put a comment inside
2398 the __might_sleep() function, you just have to include
2399 <linux/ftrace.h> and call trace_printk() inside __might_sleep()
2401 trace_printk("I'm a comment!\n")
2405 1) | __might_sleep() {
2406 1) | /* I'm a comment! */
2410 You might find other useful features for this tracer in the
2411 following "dynamic ftrace" section such as tracing only specific
2417 If CONFIG_DYNAMIC_FTRACE is set, the system will run with
2418 virtually no overhead when function tracing is disabled. The way
2419 this works is the mcount function call (placed at the start of
2420 every kernel function, produced by the -pg switch in gcc),
2421 starts of pointing to a simple return. (Enabling FTRACE will
2422 include the -pg switch in the compiling of the kernel.)
2424 At compile time every C file object is run through the
2425 recordmcount program (located in the scripts directory). This
2426 program will parse the ELF headers in the C object to find all
2427 the locations in the .text section that call mcount. Starting
2428 with gcc verson 4.6, the -mfentry has been added for x86, which
2429 calls "__fentry__" instead of "mcount". Which is called before
2430 the creation of the stack frame.
2432 Note, not all sections are traced. They may be prevented by either
2433 a notrace, or blocked another way and all inline functions are not
2434 traced. Check the "available_filter_functions" file to see what functions
2437 A section called "__mcount_loc" is created that holds
2438 references to all the mcount/fentry call sites in the .text section.
2439 The recordmcount program re-links this section back into the
2440 original object. The final linking stage of the kernel will add all these
2441 references into a single table.
2443 On boot up, before SMP is initialized, the dynamic ftrace code
2444 scans this table and updates all the locations into nops. It
2445 also records the locations, which are added to the
2446 available_filter_functions list. Modules are processed as they
2447 are loaded and before they are executed. When a module is
2448 unloaded, it also removes its functions from the ftrace function
2449 list. This is automatic in the module unload code, and the
2450 module author does not need to worry about it.
2452 When tracing is enabled, the process of modifying the function
2453 tracepoints is dependent on architecture. The old method is to use
2454 kstop_machine to prevent races with the CPUs executing code being
2455 modified (which can cause the CPU to do undesirable things, especially
2456 if the modified code crosses cache (or page) boundaries), and the nops are
2457 patched back to calls. But this time, they do not call mcount
2458 (which is just a function stub). They now call into the ftrace
2461 The new method of modifying the function tracepoints is to place
2462 a breakpoint at the location to be modified, sync all CPUs, modify
2463 the rest of the instruction not covered by the breakpoint. Sync
2464 all CPUs again, and then remove the breakpoint with the finished
2465 version to the ftrace call site.
2467 Some archs do not even need to monkey around with the synchronization,
2468 and can just slap the new code on top of the old without any
2469 problems with other CPUs executing it at the same time.
2471 One special side-effect to the recording of the functions being
2472 traced is that we can now selectively choose which functions we
2473 wish to trace and which ones we want the mcount calls to remain
2476 Two files are used, one for enabling and one for disabling the
2477 tracing of specified functions. They are:
2485 A list of available functions that you can add to these files is
2488 available_filter_functions
2490 # cat available_filter_functions
2499 If I am only interested in sys_nanosleep and hrtimer_interrupt:
2501 # echo sys_nanosleep hrtimer_interrupt > set_ftrace_filter
2502 # echo function > current_tracer
2503 # echo 1 > tracing_on
2505 # echo 0 > tracing_on
2509 # entries-in-buffer/entries-written: 5/5 #P:4
2512 # / _----=> need-resched
2513 # | / _---=> hardirq/softirq
2514 # || / _--=> preempt-depth
2516 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2518 usleep-2665 [001] .... 4186.475355: sys_nanosleep <-system_call_fastpath
2519 <idle>-0 [001] d.h1 4186.475409: hrtimer_interrupt <-smp_apic_timer_interrupt
2520 usleep-2665 [001] d.h1 4186.475426: hrtimer_interrupt <-smp_apic_timer_interrupt
2521 <idle>-0 [003] d.h1 4186.475426: hrtimer_interrupt <-smp_apic_timer_interrupt
2522 <idle>-0 [002] d.h1 4186.475427: hrtimer_interrupt <-smp_apic_timer_interrupt
2524 To see which functions are being traced, you can cat the file:
2526 # cat set_ftrace_filter
2531 Perhaps this is not enough. The filters also allow glob(7) matching.
2533 <match>* - will match functions that begin with <match>
2534 *<match> - will match functions that end with <match>
2535 *<match>* - will match functions that have <match> in it
2536 <match1>*<match2> - will match functions that begin with
2537 <match1> and end with <match2>
2539 Note: It is better to use quotes to enclose the wild cards,
2540 otherwise the shell may expand the parameters into names
2541 of files in the local directory.
2543 # echo 'hrtimer_*' > set_ftrace_filter
2549 # entries-in-buffer/entries-written: 897/897 #P:4
2552 # / _----=> need-resched
2553 # | / _---=> hardirq/softirq
2554 # || / _--=> preempt-depth
2556 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2558 <idle>-0 [003] dN.1 4228.547803: hrtimer_cancel <-tick_nohz_idle_exit
2559 <idle>-0 [003] dN.1 4228.547804: hrtimer_try_to_cancel <-hrtimer_cancel
2560 <idle>-0 [003] dN.2 4228.547805: hrtimer_force_reprogram <-__remove_hrtimer
2561 <idle>-0 [003] dN.1 4228.547805: hrtimer_forward <-tick_nohz_idle_exit
2562 <idle>-0 [003] dN.1 4228.547805: hrtimer_start_range_ns <-hrtimer_start_expires.constprop.11
2563 <idle>-0 [003] d..1 4228.547858: hrtimer_get_next_event <-get_next_timer_interrupt
2564 <idle>-0 [003] d..1 4228.547859: hrtimer_start <-__tick_nohz_idle_enter
2565 <idle>-0 [003] d..2 4228.547860: hrtimer_force_reprogram <-__rem
2567 Notice that we lost the sys_nanosleep.
2569 # cat set_ftrace_filter
2574 hrtimer_try_to_cancel
2578 hrtimer_force_reprogram
2579 hrtimer_get_next_event
2583 hrtimer_get_remaining
2585 hrtimer_init_sleeper
2588 This is because the '>' and '>>' act just like they do in bash.
2589 To rewrite the filters, use '>'
2590 To append to the filters, use '>>'
2592 To clear out a filter so that all functions will be recorded
2595 # echo > set_ftrace_filter
2596 # cat set_ftrace_filter
2599 Again, now we want to append.
2601 # echo sys_nanosleep > set_ftrace_filter
2602 # cat set_ftrace_filter
2604 # echo 'hrtimer_*' >> set_ftrace_filter
2605 # cat set_ftrace_filter
2610 hrtimer_try_to_cancel
2614 hrtimer_force_reprogram
2615 hrtimer_get_next_event
2620 hrtimer_get_remaining
2622 hrtimer_init_sleeper
2625 The set_ftrace_notrace prevents those functions from being
2628 # echo '*preempt*' '*lock*' > set_ftrace_notrace
2634 # entries-in-buffer/entries-written: 39608/39608 #P:4
2637 # / _----=> need-resched
2638 # | / _---=> hardirq/softirq
2639 # || / _--=> preempt-depth
2641 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2643 bash-1994 [000] .... 4342.324896: file_ra_state_init <-do_dentry_open
2644 bash-1994 [000] .... 4342.324897: open_check_o_direct <-do_last
2645 bash-1994 [000] .... 4342.324897: ima_file_check <-do_last
2646 bash-1994 [000] .... 4342.324898: process_measurement <-ima_file_check
2647 bash-1994 [000] .... 4342.324898: ima_get_action <-process_measurement
2648 bash-1994 [000] .... 4342.324898: ima_match_policy <-ima_get_action
2649 bash-1994 [000] .... 4342.324899: do_truncate <-do_last
2650 bash-1994 [000] .... 4342.324899: should_remove_suid <-do_truncate
2651 bash-1994 [000] .... 4342.324899: notify_change <-do_truncate
2652 bash-1994 [000] .... 4342.324900: current_fs_time <-notify_change
2653 bash-1994 [000] .... 4342.324900: current_kernel_time <-current_fs_time
2654 bash-1994 [000] .... 4342.324900: timespec_trunc <-current_fs_time
2656 We can see that there's no more lock or preempt tracing.
2659 Dynamic ftrace with the function graph tracer
2660 ---------------------------------------------
2662 Although what has been explained above concerns both the
2663 function tracer and the function-graph-tracer, there are some
2664 special features only available in the function-graph tracer.
2666 If you want to trace only one function and all of its children,
2667 you just have to echo its name into set_graph_function:
2669 echo __do_fault > set_graph_function
2671 will produce the following "expanded" trace of the __do_fault()
2675 0) | filemap_fault() {
2676 0) | find_lock_page() {
2677 0) 0.804 us | find_get_page();
2678 0) | __might_sleep() {
2682 0) 0.653 us | _spin_lock();
2683 0) 0.578 us | page_add_file_rmap();
2684 0) 0.525 us | native_set_pte_at();
2685 0) 0.585 us | _spin_unlock();
2686 0) | unlock_page() {
2687 0) 0.541 us | page_waitqueue();
2688 0) 0.639 us | __wake_up_bit();
2692 0) | filemap_fault() {
2693 0) | find_lock_page() {
2694 0) 0.698 us | find_get_page();
2695 0) | __might_sleep() {
2699 0) 0.631 us | _spin_lock();
2700 0) 0.571 us | page_add_file_rmap();
2701 0) 0.526 us | native_set_pte_at();
2702 0) 0.586 us | _spin_unlock();
2703 0) | unlock_page() {
2704 0) 0.533 us | page_waitqueue();
2705 0) 0.638 us | __wake_up_bit();
2709 You can also expand several functions at once:
2711 echo sys_open > set_graph_function
2712 echo sys_close >> set_graph_function
2714 Now if you want to go back to trace all functions you can clear
2715 this special filter via:
2717 echo > set_graph_function
2723 Note, the proc sysctl ftrace_enable is a big on/off switch for the
2724 function tracer. By default it is enabled (when function tracing is
2725 enabled in the kernel). If it is disabled, all function tracing is
2726 disabled. This includes not only the function tracers for ftrace, but
2727 also for any other uses (perf, kprobes, stack tracing, profiling, etc).
2729 Please disable this with care.
2731 This can be disable (and enabled) with:
2733 sysctl kernel.ftrace_enabled=0
2734 sysctl kernel.ftrace_enabled=1
2738 echo 0 > /proc/sys/kernel/ftrace_enabled
2739 echo 1 > /proc/sys/kernel/ftrace_enabled
2745 A few commands are supported by the set_ftrace_filter interface.
2746 Trace commands have the following format:
2748 <function>:<command>:<parameter>
2750 The following commands are supported:
2753 This command enables function filtering per module. The
2754 parameter defines the module. For example, if only the write*
2755 functions in the ext3 module are desired, run:
2757 echo 'write*:mod:ext3' > set_ftrace_filter
2759 This command interacts with the filter in the same way as
2760 filtering based on function names. Thus, adding more functions
2761 in a different module is accomplished by appending (>>) to the
2762 filter file. Remove specific module functions by prepending
2765 echo '!writeback*:mod:ext3' >> set_ftrace_filter
2767 Mod command supports module globbing. Disable tracing for all
2768 functions except a specific module:
2770 echo '!*:mod:!ext3' >> set_ftrace_filter
2772 Disable tracing for all modules, but still trace kernel:
2774 echo '!*:mod:*' >> set_ftrace_filter
2776 Enable filter only for kernel:
2778 echo '*write*:mod:!*' >> set_ftrace_filter
2780 Enable filter for module globbing:
2782 echo '*write*:mod:*snd*' >> set_ftrace_filter
2785 These commands turn tracing on and off when the specified
2786 functions are hit. The parameter determines how many times the
2787 tracing system is turned on and off. If unspecified, there is
2788 no limit. For example, to disable tracing when a schedule bug
2789 is hit the first 5 times, run:
2791 echo '__schedule_bug:traceoff:5' > set_ftrace_filter
2793 To always disable tracing when __schedule_bug is hit:
2795 echo '__schedule_bug:traceoff' > set_ftrace_filter
2797 These commands are cumulative whether or not they are appended
2798 to set_ftrace_filter. To remove a command, prepend it by '!'
2799 and drop the parameter:
2801 echo '!__schedule_bug:traceoff:0' > set_ftrace_filter
2803 The above removes the traceoff command for __schedule_bug
2804 that have a counter. To remove commands without counters:
2806 echo '!__schedule_bug:traceoff' > set_ftrace_filter
2809 Will cause a snapshot to be triggered when the function is hit.
2811 echo 'native_flush_tlb_others:snapshot' > set_ftrace_filter
2813 To only snapshot once:
2815 echo 'native_flush_tlb_others:snapshot:1' > set_ftrace_filter
2817 To remove the above commands:
2819 echo '!native_flush_tlb_others:snapshot' > set_ftrace_filter
2820 echo '!native_flush_tlb_others:snapshot:0' > set_ftrace_filter
2822 - enable_event/disable_event
2823 These commands can enable or disable a trace event. Note, because
2824 function tracing callbacks are very sensitive, when these commands
2825 are registered, the trace point is activated, but disabled in
2826 a "soft" mode. That is, the tracepoint will be called, but
2827 just will not be traced. The event tracepoint stays in this mode
2828 as long as there's a command that triggers it.
2830 echo 'try_to_wake_up:enable_event:sched:sched_switch:2' > \
2835 <function>:enable_event:<system>:<event>[:count]
2836 <function>:disable_event:<system>:<event>[:count]
2838 To remove the events commands:
2841 echo '!try_to_wake_up:enable_event:sched:sched_switch:0' > \
2843 echo '!schedule:disable_event:sched:sched_switch' > \
2847 When the function is hit, it will dump the contents of the ftrace
2848 ring buffer to the console. This is useful if you need to debug
2849 something, and want to dump the trace when a certain function
2850 is hit. Perhaps its a function that is called before a tripple
2851 fault happens and does not allow you to get a regular dump.
2854 When the function is hit, it will dump the contents of the ftrace
2855 ring buffer for the current CPU to the console. Unlike the "dump"
2856 command, it only prints out the contents of the ring buffer for the
2857 CPU that executed the function that triggered the dump.
2862 The trace_pipe outputs the same content as the trace file, but
2863 the effect on the tracing is different. Every read from
2864 trace_pipe is consumed. This means that subsequent reads will be
2865 different. The trace is live.
2867 # echo function > current_tracer
2868 # cat trace_pipe > /tmp/trace.out &
2870 # echo 1 > tracing_on
2872 # echo 0 > tracing_on
2876 # entries-in-buffer/entries-written: 0/0 #P:4
2879 # / _----=> need-resched
2880 # | / _---=> hardirq/softirq
2881 # || / _--=> preempt-depth
2883 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
2887 # cat /tmp/trace.out
2888 bash-1994 [000] .... 5281.568961: mutex_unlock <-rb_simple_write
2889 bash-1994 [000] .... 5281.568963: __mutex_unlock_slowpath <-mutex_unlock
2890 bash-1994 [000] .... 5281.568963: __fsnotify_parent <-fsnotify_modify
2891 bash-1994 [000] .... 5281.568964: fsnotify <-fsnotify_modify
2892 bash-1994 [000] .... 5281.568964: __srcu_read_lock <-fsnotify
2893 bash-1994 [000] .... 5281.568964: add_preempt_count <-__srcu_read_lock
2894 bash-1994 [000] ...1 5281.568965: sub_preempt_count <-__srcu_read_lock
2895 bash-1994 [000] .... 5281.568965: __srcu_read_unlock <-fsnotify
2896 bash-1994 [000] .... 5281.568967: sys_dup2 <-system_call_fastpath
2899 Note, reading the trace_pipe file will block until more input is
2905 Having too much or not enough data can be troublesome in
2906 diagnosing an issue in the kernel. The file buffer_size_kb is
2907 used to modify the size of the internal trace buffers. The
2908 number listed is the number of entries that can be recorded per
2909 CPU. To know the full size, multiply the number of possible CPUs
2910 with the number of entries.
2912 # cat buffer_size_kb
2913 1408 (units kilobytes)
2915 Or simply read buffer_total_size_kb
2917 # cat buffer_total_size_kb
2920 To modify the buffer, simple echo in a number (in 1024 byte segments).
2922 # echo 10000 > buffer_size_kb
2923 # cat buffer_size_kb
2924 10000 (units kilobytes)
2926 It will try to allocate as much as possible. If you allocate too
2927 much, it can cause Out-Of-Memory to trigger.
2929 # echo 1000000000000 > buffer_size_kb
2930 -bash: echo: write error: Cannot allocate memory
2931 # cat buffer_size_kb
2934 The per_cpu buffers can be changed individually as well:
2936 # echo 10000 > per_cpu/cpu0/buffer_size_kb
2937 # echo 100 > per_cpu/cpu1/buffer_size_kb
2939 When the per_cpu buffers are not the same, the buffer_size_kb
2940 at the top level will just show an X
2942 # cat buffer_size_kb
2945 This is where the buffer_total_size_kb is useful:
2947 # cat buffer_total_size_kb
2950 Writing to the top level buffer_size_kb will reset all the buffers
2951 to be the same again.
2955 CONFIG_TRACER_SNAPSHOT makes a generic snapshot feature
2956 available to all non latency tracers. (Latency tracers which
2957 record max latency, such as "irqsoff" or "wakeup", can't use
2958 this feature, since those are already using the snapshot
2959 mechanism internally.)
2961 Snapshot preserves a current trace buffer at a particular point
2962 in time without stopping tracing. Ftrace swaps the current
2963 buffer with a spare buffer, and tracing continues in the new
2964 current (=previous spare) buffer.
2966 The following tracefs files in "tracing" are related to this
2971 This is used to take a snapshot and to read the output
2972 of the snapshot. Echo 1 into this file to allocate a
2973 spare buffer and to take a snapshot (swap), then read
2974 the snapshot from this file in the same format as
2975 "trace" (described above in the section "The File
2976 System"). Both reads snapshot and tracing are executable
2977 in parallel. When the spare buffer is allocated, echoing
2978 0 frees it, and echoing else (positive) values clear the
2980 More details are shown in the table below.
2982 status\input | 0 | 1 | else |
2983 --------------+------------+------------+------------+
2984 not allocated |(do nothing)| alloc+swap |(do nothing)|
2985 --------------+------------+------------+------------+
2986 allocated | free | swap | clear |
2987 --------------+------------+------------+------------+
2989 Here is an example of using the snapshot feature.
2991 # echo 1 > events/sched/enable
2996 # entries-in-buffer/entries-written: 71/71 #P:8
2999 # / _----=> need-resched
3000 # | / _---=> hardirq/softirq
3001 # || / _--=> preempt-depth
3003 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3005 <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
3006 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
3008 <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
3013 # entries-in-buffer/entries-written: 77/77 #P:8
3016 # / _----=> need-resched
3017 # | / _---=> hardirq/softirq
3018 # || / _--=> preempt-depth
3020 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3022 <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
3023 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
3027 If you try to use this snapshot feature when current tracer is
3028 one of the latency tracers, you will get the following results.
3030 # echo wakeup > current_tracer
3032 bash: echo: write error: Device or resource busy
3034 cat: snapshot: Device or resource busy
3039 In the tracefs tracing directory is a directory called "instances".
3040 This directory can have new directories created inside of it using
3041 mkdir, and removing directories with rmdir. The directory created
3042 with mkdir in this directory will already contain files and other
3043 directories after it is created.
3045 # mkdir instances/foo
3047 buffer_size_kb buffer_total_size_kb events free_buffer per_cpu
3048 set_event snapshot trace trace_clock trace_marker trace_options
3049 trace_pipe tracing_on
3051 As you can see, the new directory looks similar to the tracing directory
3052 itself. In fact, it is very similar, except that the buffer and
3053 events are agnostic from the main director, or from any other
3054 instances that are created.
3056 The files in the new directory work just like the files with the
3057 same name in the tracing directory except the buffer that is used
3058 is a separate and new buffer. The files affect that buffer but do not
3059 affect the main buffer with the exception of trace_options. Currently,
3060 the trace_options affect all instances and the top level buffer
3061 the same, but this may change in future releases. That is, options
3062 may become specific to the instance they reside in.
3064 Notice that none of the function tracer files are there, nor is
3065 current_tracer and available_tracers. This is because the buffers
3066 can currently only have events enabled for them.
3068 # mkdir instances/foo
3069 # mkdir instances/bar
3070 # mkdir instances/zoot
3071 # echo 100000 > buffer_size_kb
3072 # echo 1000 > instances/foo/buffer_size_kb
3073 # echo 5000 > instances/bar/per_cpu/cpu1/buffer_size_kb
3074 # echo function > current_trace
3075 # echo 1 > instances/foo/events/sched/sched_wakeup/enable
3076 # echo 1 > instances/foo/events/sched/sched_wakeup_new/enable
3077 # echo 1 > instances/foo/events/sched/sched_switch/enable
3078 # echo 1 > instances/bar/events/irq/enable
3079 # echo 1 > instances/zoot/events/syscalls/enable
3081 CPU:2 [LOST 11745 EVENTS]
3082 bash-2044 [002] .... 10594.481032: _raw_spin_lock_irqsave <-get_page_from_freelist
3083 bash-2044 [002] d... 10594.481032: add_preempt_count <-_raw_spin_lock_irqsave
3084 bash-2044 [002] d..1 10594.481032: __rmqueue <-get_page_from_freelist
3085 bash-2044 [002] d..1 10594.481033: _raw_spin_unlock <-get_page_from_freelist
3086 bash-2044 [002] d..1 10594.481033: sub_preempt_count <-_raw_spin_unlock
3087 bash-2044 [002] d... 10594.481033: get_pageblock_flags_group <-get_pageblock_migratetype
3088 bash-2044 [002] d... 10594.481034: __mod_zone_page_state <-get_page_from_freelist
3089 bash-2044 [002] d... 10594.481034: zone_statistics <-get_page_from_freelist
3090 bash-2044 [002] d... 10594.481034: __inc_zone_state <-zone_statistics
3091 bash-2044 [002] d... 10594.481034: __inc_zone_state <-zone_statistics
3092 bash-2044 [002] .... 10594.481035: arch_dup_task_struct <-copy_process
3095 # cat instances/foo/trace_pipe
3096 bash-1998 [000] d..4 136.676759: sched_wakeup: comm=kworker/0:1 pid=59 prio=120 success=1 target_cpu=000
3097 bash-1998 [000] dN.4 136.676760: sched_wakeup: comm=bash pid=1998 prio=120 success=1 target_cpu=000
3098 <idle>-0 [003] d.h3 136.676906: sched_wakeup: comm=rcu_preempt pid=9 prio=120 success=1 target_cpu=003
3099 <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
3100 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
3101 bash-1998 [000] d..4 136.677014: sched_wakeup: comm=kworker/0:1 pid=59 prio=120 success=1 target_cpu=000
3102 bash-1998 [000] dN.4 136.677016: sched_wakeup: comm=bash pid=1998 prio=120 success=1 target_cpu=000
3103 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
3104 kworker/0:1-59 [000] d..4 136.677022: sched_wakeup: comm=sshd pid=1995 prio=120 success=1 target_cpu=001
3105 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
3108 # cat instances/bar/trace_pipe
3109 migration/1-14 [001] d.h3 138.732674: softirq_raise: vec=3 [action=NET_RX]
3110 <idle>-0 [001] dNh3 138.732725: softirq_raise: vec=3 [action=NET_RX]
3111 bash-1998 [000] d.h1 138.733101: softirq_raise: vec=1 [action=TIMER]
3112 bash-1998 [000] d.h1 138.733102: softirq_raise: vec=9 [action=RCU]
3113 bash-1998 [000] ..s2 138.733105: softirq_entry: vec=1 [action=TIMER]
3114 bash-1998 [000] ..s2 138.733106: softirq_exit: vec=1 [action=TIMER]
3115 bash-1998 [000] ..s2 138.733106: softirq_entry: vec=9 [action=RCU]
3116 bash-1998 [000] ..s2 138.733109: softirq_exit: vec=9 [action=RCU]
3117 sshd-1995 [001] d.h1 138.733278: irq_handler_entry: irq=21 name=uhci_hcd:usb4
3118 sshd-1995 [001] d.h1 138.733280: irq_handler_exit: irq=21 ret=unhandled
3119 sshd-1995 [001] d.h1 138.733281: irq_handler_entry: irq=21 name=eth0
3120 sshd-1995 [001] d.h1 138.733283: irq_handler_exit: irq=21 ret=handled
3123 # cat instances/zoot/trace
3126 # entries-in-buffer/entries-written: 18996/18996 #P:4
3129 # / _----=> need-resched
3130 # | / _---=> hardirq/softirq
3131 # || / _--=> preempt-depth
3133 # TASK-PID CPU# |||| TIMESTAMP FUNCTION
3135 bash-1998 [000] d... 140.733501: sys_write -> 0x2
3136 bash-1998 [000] d... 140.733504: sys_dup2(oldfd: a, newfd: 1)
3137 bash-1998 [000] d... 140.733506: sys_dup2 -> 0x1
3138 bash-1998 [000] d... 140.733508: sys_fcntl(fd: a, cmd: 1, arg: 0)
3139 bash-1998 [000] d... 140.733509: sys_fcntl -> 0x1
3140 bash-1998 [000] d... 140.733510: sys_close(fd: a)
3141 bash-1998 [000] d... 140.733510: sys_close -> 0x0
3142 bash-1998 [000] d... 140.733514: sys_rt_sigprocmask(how: 0, nset: 0, oset: 6e2768, sigsetsize: 8)
3143 bash-1998 [000] d... 140.733515: sys_rt_sigprocmask -> 0x0
3144 bash-1998 [000] d... 140.733516: sys_rt_sigaction(sig: 2, act: 7fff718846f0, oact: 7fff71884650, sigsetsize: 8)
3145 bash-1998 [000] d... 140.733516: sys_rt_sigaction -> 0x0
3147 You can see that the trace of the top most trace buffer shows only
3148 the function tracing. The foo instance displays wakeups and task
3151 To remove the instances, simply delete their directories:
3153 # rmdir instances/foo
3154 # rmdir instances/bar
3155 # rmdir instances/zoot
3157 Note, if a process has a trace file open in one of the instance
3158 directories, the rmdir will fail with EBUSY.
3163 Since the kernel has a fixed sized stack, it is important not to
3164 waste it in functions. A kernel developer must be conscience of
3165 what they allocate on the stack. If they add too much, the system
3166 can be in danger of a stack overflow, and corruption will occur,
3167 usually leading to a system panic.
3169 There are some tools that check this, usually with interrupts
3170 periodically checking usage. But if you can perform a check
3171 at every function call that will become very useful. As ftrace provides
3172 a function tracer, it makes it convenient to check the stack size
3173 at every function call. This is enabled via the stack tracer.
3175 CONFIG_STACK_TRACER enables the ftrace stack tracing functionality.
3176 To enable it, write a '1' into /proc/sys/kernel/stack_tracer_enabled.
3178 # echo 1 > /proc/sys/kernel/stack_tracer_enabled
3180 You can also enable it from the kernel command line to trace
3181 the stack size of the kernel during boot up, by adding "stacktrace"
3182 to the kernel command line parameter.
3184 After running it for a few minutes, the output looks like:
3186 # cat stack_max_size
3190 Depth Size Location (18 entries)
3192 0) 2928 224 update_sd_lb_stats+0xbc/0x4ac
3193 1) 2704 160 find_busiest_group+0x31/0x1f1
3194 2) 2544 256 load_balance+0xd9/0x662
3195 3) 2288 80 idle_balance+0xbb/0x130
3196 4) 2208 128 __schedule+0x26e/0x5b9
3197 5) 2080 16 schedule+0x64/0x66
3198 6) 2064 128 schedule_timeout+0x34/0xe0
3199 7) 1936 112 wait_for_common+0x97/0xf1
3200 8) 1824 16 wait_for_completion+0x1d/0x1f
3201 9) 1808 128 flush_work+0xfe/0x119
3202 10) 1680 16 tty_flush_to_ldisc+0x1e/0x20
3203 11) 1664 48 input_available_p+0x1d/0x5c
3204 12) 1616 48 n_tty_poll+0x6d/0x134
3205 13) 1568 64 tty_poll+0x64/0x7f
3206 14) 1504 880 do_select+0x31e/0x511
3207 15) 624 400 core_sys_select+0x177/0x216
3208 16) 224 96 sys_select+0x91/0xb9
3209 17) 128 128 system_call_fastpath+0x16/0x1b
3211 Note, if -mfentry is being used by gcc, functions get traced before
3212 they set up the stack frame. This means that leaf level functions
3213 are not tested by the stack tracer when -mfentry is used.
3215 Currently, -mfentry is used by gcc 4.6.0 and above on x86 only.
3219 More details can be found in the source code, in the
3220 kernel/trace/*.c files.
projects / karo-tx-linux.git / blob