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 Reviewers: Elias Oltmanns, Randy Dunlap, Andrew Morton,
9 John Kacur, and David Teigland.
11 Written for: 2.6.28-rc2
16 Ftrace is an internal tracer designed to help out developers and
17 designers of systems to find what is going on inside the kernel.
18 It can be used for debugging or analyzing latencies and
19 performance issues that take place outside of user-space.
21 Although ftrace is the function tracer, it also includes an
22 infrastructure that allows for other types of tracing. Some of
23 the tracers that are currently in ftrace include a tracer to
24 trace context switches, the time it takes for a high priority
25 task to run after it was woken up, the time interrupts are
26 disabled, and more (ftrace allows for tracer plugins, which
27 means that the list of tracers can always grow).
33 Ftrace uses the debugfs file system to hold the control files as
34 well as the files to display output.
36 To mount the debugfs system:
39 # mount -t debugfs nodev /debug
41 ( Note: it is more common to mount at /sys/kernel/debug, but for
42 simplicity this document will use /debug)
44 That's it! (assuming that you have ftrace configured into your kernel)
46 After mounting the debugfs, you can see a directory called
47 "tracing". This directory contains the control and output files
48 of ftrace. Here is a list of some of the key files:
51 Note: all time values are in microseconds.
55 This is used to set or display the current tracer
60 This holds the different types of tracers that
61 have been compiled into the kernel. The
62 tracers listed here can be configured by
63 echoing their name into current_tracer.
67 This sets or displays whether the current_tracer
68 is activated and tracing or not. Echo 0 into this
69 file to disable the tracer or 1 to enable it.
73 This file holds the output of the trace in a human
74 readable format (described below).
78 This file shows the same trace but the information
79 is organized more to display possible latencies
80 in the system (described below).
84 The output is the same as the "trace" file but this
85 file is meant to be streamed with live tracing.
86 Reads from this file will block until new data
87 is retrieved. Unlike the "trace" and "latency_trace"
88 files, this file is a consumer. This means reading
89 from this file causes sequential reads to display
90 more current data. Once data is read from this
91 file, it is consumed, and will not be read
92 again with a sequential read. The "trace" and
93 "latency_trace" files are static, and if the
94 tracer is not adding more data, they will display
95 the same information every time they are read.
99 This file lets the user control the amount of data
100 that is displayed in one of the above output
105 Some of the tracers record the max latency.
106 For example, the time interrupts are disabled.
107 This time is saved in this file. The max trace
108 will also be stored, and displayed by either
109 "trace" or "latency_trace". A new max trace will
110 only be recorded if the latency is greater than
111 the value in this file. (in microseconds)
115 This sets or displays the number of kilobytes each CPU
116 buffer can hold. The tracer buffers are the same size
117 for each CPU. The displayed number is the size of the
118 CPU buffer and not total size of all buffers. The
119 trace buffers are allocated in pages (blocks of memory
120 that the kernel uses for allocation, usually 4 KB in size).
121 If the last page allocated has room for more bytes
122 than requested, the rest of the page will be used,
123 making the actual allocation bigger than requested.
124 ( Note, the size may not be a multiple of the page size
125 due to buffer managment overhead. )
127 This can only be updated when the current_tracer
132 This is a mask that lets the user only trace
133 on specified CPUS. The format is a hex string
134 representing the CPUS.
138 When dynamic ftrace is configured in (see the
139 section below "dynamic ftrace"), the code is dynamically
140 modified (code text rewrite) to disable calling of the
141 function profiler (mcount). This lets tracing be configured
142 in with practically no overhead in performance. This also
143 has a side effect of enabling or disabling specific functions
144 to be traced. Echoing names of functions into this file
145 will limit the trace to only those functions.
149 This has an effect opposite to that of
150 set_ftrace_filter. Any function that is added here will not
151 be traced. If a function exists in both set_ftrace_filter
152 and set_ftrace_notrace, the function will _not_ be traced.
156 Have the function tracer only trace a single thread.
160 Set a "trigger" function where tracing should start
161 with the function graph tracer (See the section
162 "dynamic ftrace" for more details).
164 available_filter_functions:
166 This lists the functions that ftrace
167 has processed and can trace. These are the function
168 names that you can pass to "set_ftrace_filter" or
169 "set_ftrace_notrace". (See the section "dynamic ftrace"
170 below for more details.)
176 Here is the list of current tracers that may be configured.
180 Function call tracer to trace all kernel functions.
182 "function_graph_tracer"
184 Similar to the function tracer except that the
185 function tracer probes the functions on their entry
186 whereas the function graph tracer traces on both entry
187 and exit of the functions. It then provides the ability
188 to draw a graph of function calls similar to C code
193 Traces the context switches and wakeups between tasks.
197 Traces the areas that disable interrupts and saves
198 the trace with the longest max latency.
199 See tracing_max_latency. When a new max is recorded,
200 it replaces the old trace. It is best to view this
201 trace via the latency_trace file.
205 Similar to irqsoff but traces and records the amount of
206 time for which preemption is disabled.
210 Similar to irqsoff and preemptoff, but traces and
211 records the largest time for which irqs and/or preemption
216 Traces and records the max latency that it takes for
217 the highest priority task to get scheduled after
218 it has been woken up.
222 Uses the BTS CPU feature on x86 CPUs to traces all
227 This is the "trace nothing" tracer. To remove all
228 tracers from tracing simply echo "nop" into
232 Examples of using the tracer
233 ----------------------------
235 Here are typical examples of using the tracers when controlling
236 them only with the debugfs interface (without using any
237 user-land utilities).
242 Here is an example of the output format of the file "trace"
247 # TASK-PID CPU# TIMESTAMP FUNCTION
249 bash-4251 [01] 10152.583854: path_put <-path_walk
250 bash-4251 [01] 10152.583855: dput <-path_put
251 bash-4251 [01] 10152.583855: _atomic_dec_and_lock <-dput
254 A header is printed with the tracer name that is represented by
255 the trace. In this case the tracer is "function". Then a header
256 showing the format. Task name "bash", the task PID "4251", the
257 CPU that it was running on "01", the timestamp in <secs>.<usecs>
258 format, the function name that was traced "path_put" and the
259 parent function that called this function "path_walk". The
260 timestamp is the time at which the function was entered.
262 The sched_switch tracer also includes tracing of task wakeups
263 and context switches.
265 ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 2916:115:S
266 ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 10:115:S
267 ksoftirqd/1-7 [01] 1453.070013: 7:115:R ==> 10:115:R
268 events/1-10 [01] 1453.070013: 10:115:S ==> 2916:115:R
269 kondemand/1-2916 [01] 1453.070013: 2916:115:S ==> 7:115:R
270 ksoftirqd/1-7 [01] 1453.070013: 7:115:S ==> 0:140:R
272 Wake ups are represented by a "+" and the context switches are
273 shown as "==>". The format is:
277 Previous task Next Task
279 <pid>:<prio>:<state> ==> <pid>:<prio>:<state>
283 Current task Task waking up
285 <pid>:<prio>:<state> + <pid>:<prio>:<state>
287 The prio is the internal kernel priority, which is the inverse
288 of the priority that is usually displayed by user-space tools.
289 Zero represents the highest priority (99). Prio 100 starts the
290 "nice" priorities with 100 being equal to nice -20 and 139 being
291 nice 19. The prio "140" is reserved for the idle task which is
292 the lowest priority thread (pid 0).
298 For traces that display latency times, the latency_trace file
299 gives somewhat more information to see why a latency happened.
300 Here is a typical trace.
304 irqsoff latency trace v1.1.5 on 2.6.26-rc8
305 --------------------------------------------------------------------
306 latency: 97 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
308 | task: swapper-0 (uid:0 nice:0 policy:0 rt_prio:0)
310 => started at: apic_timer_interrupt
311 => ended at: do_softirq
314 # / _-----=> irqs-off
315 # | / _----=> need-resched
316 # || / _---=> hardirq/softirq
317 # ||| / _--=> preempt-depth
320 # cmd pid ||||| time | caller
322 <idle>-0 0d..1 0us+: trace_hardirqs_off_thunk (apic_timer_interrupt)
323 <idle>-0 0d.s. 97us : __do_softirq (do_softirq)
324 <idle>-0 0d.s1 98us : trace_hardirqs_on (do_softirq)
327 This shows that the current tracer is "irqsoff" tracing the time
328 for which interrupts were disabled. It gives the trace version
329 and the version of the kernel upon which this was executed on
330 (2.6.26-rc8). Then it displays the max latency in microsecs (97
331 us). The number of trace entries displayed and the total number
332 recorded (both are three: #3/3). The type of preemption that was
333 used (PREEMPT). VP, KP, SP, and HP are always zero and are
334 reserved for later use. #P is the number of online CPUS (#P:2).
336 The task is the process that was running when the latency
337 occurred. (swapper pid: 0).
339 The start and stop (the functions in which the interrupts were
340 disabled and enabled respectively) that caused the latencies:
342 apic_timer_interrupt is where the interrupts were disabled.
343 do_softirq is where they were enabled again.
345 The next lines after the header are the trace itself. The header
346 explains which is which.
348 cmd: The name of the process in the trace.
350 pid: The PID of that process.
352 CPU#: The CPU which the process was running on.
354 irqs-off: 'd' interrupts are disabled. '.' otherwise.
355 Note: If the architecture does not support a way to
356 read the irq flags variable, an 'X' will always
359 need-resched: 'N' task need_resched is set, '.' otherwise.
362 'H' - hard irq occurred inside a softirq.
363 'h' - hard irq is running
364 's' - soft irq is running
365 '.' - normal context.
367 preempt-depth: The level of preempt_disabled
369 The above is mostly meaningful for kernel developers.
371 time: This differs from the trace file output. The trace file output
372 includes an absolute timestamp. The timestamp used by the
373 latency_trace file is relative to the start of the trace.
375 delay: This is just to help catch your eye a bit better. And
376 needs to be fixed to be only relative to the same CPU.
377 The marks are determined by the difference between this
378 current trace and the next trace.
379 '!' - greater than preempt_mark_thresh (default 100)
380 '+' - greater than 1 microsecond
381 ' ' - less than or equal to 1 microsecond.
383 The rest is the same as the 'trace' file.
389 The trace_options file is used to control what gets printed in
390 the trace output. To see what is available, simply cat the file:
392 cat /debug/tracing/trace_options
393 print-parent nosym-offset nosym-addr noverbose noraw nohex nobin \
394 noblock nostacktrace nosched-tree nouserstacktrace nosym-userobj
396 To disable one of the options, echo in the option prepended with
399 echo noprint-parent > /debug/tracing/trace_options
401 To enable an option, leave off the "no".
403 echo sym-offset > /debug/tracing/trace_options
405 Here are the available options:
407 print-parent - On function traces, display the calling (parent)
408 function as well as the function being traced.
411 bash-4000 [01] 1477.606694: simple_strtoul <-strict_strtoul
414 bash-4000 [01] 1477.606694: simple_strtoul
417 sym-offset - Display not only the function name, but also the
418 offset in the function. For example, instead of
419 seeing just "ktime_get", you will see
420 "ktime_get+0xb/0x20".
423 bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0
425 sym-addr - this will also display the function address as well
426 as the function name.
429 bash-4000 [01] 1477.606694: simple_strtoul <c0339346>
431 verbose - This deals with the latency_trace file.
433 bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \
434 (+0.000ms): simple_strtoul (strict_strtoul)
436 raw - This will display raw numbers. This option is best for
437 use with user applications that can translate the raw
438 numbers better than having it done in the kernel.
440 hex - Similar to raw, but the numbers will be in a hexadecimal
443 bin - This will print out the formats in raw binary.
445 block - TBD (needs update)
447 stacktrace - This is one of the options that changes the trace
448 itself. When a trace is recorded, so is the stack
449 of functions. This allows for back traces of
452 userstacktrace - This option changes the trace. It records a
453 stacktrace of the current userspace thread.
455 sym-userobj - when user stacktrace are enabled, look up which
456 object the address belongs to, and print a
457 relative address. This is especially useful when
458 ASLR is on, otherwise you don't get a chance to
459 resolve the address to object/file/line after
460 the app is no longer running
462 The lookup is performed when you read
463 trace,trace_pipe,latency_trace. Example:
465 a.out-1623 [000] 40874.465068: /root/a.out[+0x480] <-/root/a.out[+0
466 x494] <- /root/a.out[+0x4a8] <- /lib/libc-2.7.so[+0x1e1a6]
468 sched-tree - trace all tasks that are on the runqueue, at
469 every scheduling event. Will add overhead if
470 there's a lot of tasks running at once.
476 This tracer simply records schedule switches. Here is an example
479 # echo sched_switch > /debug/tracing/current_tracer
480 # echo 1 > /debug/tracing/tracing_enabled
482 # echo 0 > /debug/tracing/tracing_enabled
483 # cat /debug/tracing/trace
485 # tracer: sched_switch
487 # TASK-PID CPU# TIMESTAMP FUNCTION
489 bash-3997 [01] 240.132281: 3997:120:R + 4055:120:R
490 bash-3997 [01] 240.132284: 3997:120:R ==> 4055:120:R
491 sleep-4055 [01] 240.132371: 4055:120:S ==> 3997:120:R
492 bash-3997 [01] 240.132454: 3997:120:R + 4055:120:S
493 bash-3997 [01] 240.132457: 3997:120:R ==> 4055:120:R
494 sleep-4055 [01] 240.132460: 4055:120:D ==> 3997:120:R
495 bash-3997 [01] 240.132463: 3997:120:R + 4055:120:D
496 bash-3997 [01] 240.132465: 3997:120:R ==> 4055:120:R
497 <idle>-0 [00] 240.132589: 0:140:R + 4:115:S
498 <idle>-0 [00] 240.132591: 0:140:R ==> 4:115:R
499 ksoftirqd/0-4 [00] 240.132595: 4:115:S ==> 0:140:R
500 <idle>-0 [00] 240.132598: 0:140:R + 4:115:S
501 <idle>-0 [00] 240.132599: 0:140:R ==> 4:115:R
502 ksoftirqd/0-4 [00] 240.132603: 4:115:S ==> 0:140:R
503 sleep-4055 [01] 240.133058: 4055:120:S ==> 3997:120:R
507 As we have discussed previously about this format, the header
508 shows the name of the trace and points to the options. The
509 "FUNCTION" is a misnomer since here it represents the wake ups
510 and context switches.
512 The sched_switch file only lists the wake ups (represented with
513 '+') and context switches ('==>') with the previous task or
514 current task first followed by the next task or task waking up.
515 The format for both of these is PID:KERNEL-PRIO:TASK-STATE.
516 Remember that the KERNEL-PRIO is the inverse of the actual
517 priority with zero (0) being the highest priority and the nice
518 values starting at 100 (nice -20). Below is a quick chart to map
519 the kernel priority to user land priorities.
521 Kernel priority: 0 to 99 ==> user RT priority 99 to 0
522 Kernel priority: 100 to 139 ==> user nice -20 to 19
523 Kernel priority: 140 ==> idle task priority
527 R - running : wants to run, may not actually be running
528 S - sleep : process is waiting to be woken up (handles signals)
529 D - disk sleep (uninterruptible sleep) : process must be woken up
531 T - stopped : process suspended
532 t - traced : process is being traced (with something like gdb)
533 Z - zombie : process waiting to be cleaned up
540 The following tracers (listed below) give different output
541 depending on whether or not the sysctl ftrace_enabled is set. To
542 set ftrace_enabled, one can either use the sysctl function or
543 set it via the proc file system interface.
545 sysctl kernel.ftrace_enabled=1
549 echo 1 > /proc/sys/kernel/ftrace_enabled
551 To disable ftrace_enabled simply replace the '1' with '0' in the
554 When ftrace_enabled is set the tracers will also record the
555 functions that are within the trace. The descriptions of the
556 tracers will also show an example with ftrace enabled.
562 When interrupts are disabled, the CPU can not react to any other
563 external event (besides NMIs and SMIs). This prevents the timer
564 interrupt from triggering or the mouse interrupt from letting
565 the kernel know of a new mouse event. The result is a latency
566 with the reaction time.
568 The irqsoff tracer tracks the time for which interrupts are
569 disabled. When a new maximum latency is hit, the tracer saves
570 the trace leading up to that latency point so that every time a
571 new maximum is reached, the old saved trace is discarded and the
574 To reset the maximum, echo 0 into tracing_max_latency. Here is
577 # echo irqsoff > /debug/tracing/current_tracer
578 # echo 0 > /debug/tracing/tracing_max_latency
579 # echo 1 > /debug/tracing/tracing_enabled
582 # echo 0 > /debug/tracing/tracing_enabled
583 # cat /debug/tracing/latency_trace
586 irqsoff latency trace v1.1.5 on 2.6.26
587 --------------------------------------------------------------------
588 latency: 12 us, #3/3, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
590 | task: bash-3730 (uid:0 nice:0 policy:0 rt_prio:0)
592 => started at: sys_setpgid
593 => ended at: sys_setpgid
596 # / _-----=> irqs-off
597 # | / _----=> need-resched
598 # || / _---=> hardirq/softirq
599 # ||| / _--=> preempt-depth
602 # cmd pid ||||| time | caller
604 bash-3730 1d... 0us : _write_lock_irq (sys_setpgid)
605 bash-3730 1d..1 1us+: _write_unlock_irq (sys_setpgid)
606 bash-3730 1d..2 14us : trace_hardirqs_on (sys_setpgid)
609 Here we see that that we had a latency of 12 microsecs (which is
610 very good). The _write_lock_irq in sys_setpgid disabled
611 interrupts. The difference between the 12 and the displayed
612 timestamp 14us occurred because the clock was incremented
613 between the time of recording the max latency and the time of
614 recording the function that had that latency.
616 Note the above example had ftrace_enabled not set. If we set the
617 ftrace_enabled, we get a much larger output:
621 irqsoff latency trace v1.1.5 on 2.6.26-rc8
622 --------------------------------------------------------------------
623 latency: 50 us, #101/101, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
625 | task: ls-4339 (uid:0 nice:0 policy:0 rt_prio:0)
627 => started at: __alloc_pages_internal
628 => ended at: __alloc_pages_internal
631 # / _-----=> irqs-off
632 # | / _----=> need-resched
633 # || / _---=> hardirq/softirq
634 # ||| / _--=> preempt-depth
637 # cmd pid ||||| time | caller
639 ls-4339 0...1 0us+: get_page_from_freelist (__alloc_pages_internal)
640 ls-4339 0d..1 3us : rmqueue_bulk (get_page_from_freelist)
641 ls-4339 0d..1 3us : _spin_lock (rmqueue_bulk)
642 ls-4339 0d..1 4us : add_preempt_count (_spin_lock)
643 ls-4339 0d..2 4us : __rmqueue (rmqueue_bulk)
644 ls-4339 0d..2 5us : __rmqueue_smallest (__rmqueue)
645 ls-4339 0d..2 5us : __mod_zone_page_state (__rmqueue_smallest)
646 ls-4339 0d..2 6us : __rmqueue (rmqueue_bulk)
647 ls-4339 0d..2 6us : __rmqueue_smallest (__rmqueue)
648 ls-4339 0d..2 7us : __mod_zone_page_state (__rmqueue_smallest)
649 ls-4339 0d..2 7us : __rmqueue (rmqueue_bulk)
650 ls-4339 0d..2 8us : __rmqueue_smallest (__rmqueue)
652 ls-4339 0d..2 46us : __rmqueue_smallest (__rmqueue)
653 ls-4339 0d..2 47us : __mod_zone_page_state (__rmqueue_smallest)
654 ls-4339 0d..2 47us : __rmqueue (rmqueue_bulk)
655 ls-4339 0d..2 48us : __rmqueue_smallest (__rmqueue)
656 ls-4339 0d..2 48us : __mod_zone_page_state (__rmqueue_smallest)
657 ls-4339 0d..2 49us : _spin_unlock (rmqueue_bulk)
658 ls-4339 0d..2 49us : sub_preempt_count (_spin_unlock)
659 ls-4339 0d..1 50us : get_page_from_freelist (__alloc_pages_internal)
660 ls-4339 0d..2 51us : trace_hardirqs_on (__alloc_pages_internal)
664 Here we traced a 50 microsecond latency. But we also see all the
665 functions that were called during that time. Note that by
666 enabling function tracing, we incur an added overhead. This
667 overhead may extend the latency times. But nevertheless, this
668 trace has provided some very helpful debugging information.
674 When preemption is disabled, we may be able to receive
675 interrupts but the task cannot be preempted and a higher
676 priority task must wait for preemption to be enabled again
677 before it can preempt a lower priority task.
679 The preemptoff tracer traces the places that disable preemption.
680 Like the irqsoff tracer, it records the maximum latency for
681 which preemption was disabled. The control of preemptoff tracer
682 is much like the irqsoff tracer.
684 # echo preemptoff > /debug/tracing/current_tracer
685 # echo 0 > /debug/tracing/tracing_max_latency
686 # echo 1 > /debug/tracing/tracing_enabled
689 # echo 0 > /debug/tracing/tracing_enabled
690 # cat /debug/tracing/latency_trace
693 preemptoff latency trace v1.1.5 on 2.6.26-rc8
694 --------------------------------------------------------------------
695 latency: 29 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
697 | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
699 => started at: do_IRQ
700 => ended at: __do_softirq
703 # / _-----=> irqs-off
704 # | / _----=> need-resched
705 # || / _---=> hardirq/softirq
706 # ||| / _--=> preempt-depth
709 # cmd pid ||||| time | caller
711 sshd-4261 0d.h. 0us+: irq_enter (do_IRQ)
712 sshd-4261 0d.s. 29us : _local_bh_enable (__do_softirq)
713 sshd-4261 0d.s1 30us : trace_preempt_on (__do_softirq)
716 This has some more changes. Preemption was disabled when an
717 interrupt came in (notice the 'h'), and was enabled while doing
718 a softirq. (notice the 's'). But we also see that interrupts
719 have been disabled when entering the preempt off section and
720 leaving it (the 'd'). We do not know if interrupts were enabled
725 preemptoff latency trace v1.1.5 on 2.6.26-rc8
726 --------------------------------------------------------------------
727 latency: 63 us, #87/87, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
729 | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
731 => started at: remove_wait_queue
732 => ended at: __do_softirq
735 # / _-----=> irqs-off
736 # | / _----=> need-resched
737 # || / _---=> hardirq/softirq
738 # ||| / _--=> preempt-depth
741 # cmd pid ||||| time | caller
743 sshd-4261 0d..1 0us : _spin_lock_irqsave (remove_wait_queue)
744 sshd-4261 0d..1 1us : _spin_unlock_irqrestore (remove_wait_queue)
745 sshd-4261 0d..1 2us : do_IRQ (common_interrupt)
746 sshd-4261 0d..1 2us : irq_enter (do_IRQ)
747 sshd-4261 0d..1 2us : idle_cpu (irq_enter)
748 sshd-4261 0d..1 3us : add_preempt_count (irq_enter)
749 sshd-4261 0d.h1 3us : idle_cpu (irq_enter)
750 sshd-4261 0d.h. 4us : handle_fasteoi_irq (do_IRQ)
752 sshd-4261 0d.h. 12us : add_preempt_count (_spin_lock)
753 sshd-4261 0d.h1 12us : ack_ioapic_quirk_irq (handle_fasteoi_irq)
754 sshd-4261 0d.h1 13us : move_native_irq (ack_ioapic_quirk_irq)
755 sshd-4261 0d.h1 13us : _spin_unlock (handle_fasteoi_irq)
756 sshd-4261 0d.h1 14us : sub_preempt_count (_spin_unlock)
757 sshd-4261 0d.h1 14us : irq_exit (do_IRQ)
758 sshd-4261 0d.h1 15us : sub_preempt_count (irq_exit)
759 sshd-4261 0d..2 15us : do_softirq (irq_exit)
760 sshd-4261 0d... 15us : __do_softirq (do_softirq)
761 sshd-4261 0d... 16us : __local_bh_disable (__do_softirq)
762 sshd-4261 0d... 16us+: add_preempt_count (__local_bh_disable)
763 sshd-4261 0d.s4 20us : add_preempt_count (__local_bh_disable)
764 sshd-4261 0d.s4 21us : sub_preempt_count (local_bh_enable)
765 sshd-4261 0d.s5 21us : sub_preempt_count (local_bh_enable)
767 sshd-4261 0d.s6 41us : add_preempt_count (__local_bh_disable)
768 sshd-4261 0d.s6 42us : sub_preempt_count (local_bh_enable)
769 sshd-4261 0d.s7 42us : sub_preempt_count (local_bh_enable)
770 sshd-4261 0d.s5 43us : add_preempt_count (__local_bh_disable)
771 sshd-4261 0d.s5 43us : sub_preempt_count (local_bh_enable_ip)
772 sshd-4261 0d.s6 44us : sub_preempt_count (local_bh_enable_ip)
773 sshd-4261 0d.s5 44us : add_preempt_count (__local_bh_disable)
774 sshd-4261 0d.s5 45us : sub_preempt_count (local_bh_enable)
776 sshd-4261 0d.s. 63us : _local_bh_enable (__do_softirq)
777 sshd-4261 0d.s1 64us : trace_preempt_on (__do_softirq)
780 The above is an example of the preemptoff trace with
781 ftrace_enabled set. Here we see that interrupts were disabled
782 the entire time. The irq_enter code lets us know that we entered
783 an interrupt 'h'. Before that, the functions being traced still
784 show that it is not in an interrupt, but we can see from the
785 functions themselves that this is not the case.
787 Notice that __do_softirq when called does not have a
788 preempt_count. It may seem that we missed a preempt enabling.
789 What really happened is that the preempt count is held on the
790 thread's stack and we switched to the softirq stack (4K stacks
791 in effect). The code does not copy the preempt count, but
792 because interrupts are disabled, we do not need to worry about
793 it. Having a tracer like this is good for letting people know
794 what really happens inside the kernel.
800 Knowing the locations that have interrupts disabled or
801 preemption disabled for the longest times is helpful. But
802 sometimes we would like to know when either preemption and/or
803 interrupts are disabled.
805 Consider the following code:
808 call_function_with_irqs_off();
810 call_function_with_irqs_and_preemption_off();
812 call_function_with_preemption_off();
815 The irqsoff tracer will record the total length of
816 call_function_with_irqs_off() and
817 call_function_with_irqs_and_preemption_off().
819 The preemptoff tracer will record the total length of
820 call_function_with_irqs_and_preemption_off() and
821 call_function_with_preemption_off().
823 But neither will trace the time that interrupts and/or
824 preemption is disabled. This total time is the time that we can
825 not schedule. To record this time, use the preemptirqsoff
828 Again, using this trace is much like the irqsoff and preemptoff
831 # echo preemptirqsoff > /debug/tracing/current_tracer
832 # echo 0 > /debug/tracing/tracing_max_latency
833 # echo 1 > /debug/tracing/tracing_enabled
836 # echo 0 > /debug/tracing/tracing_enabled
837 # cat /debug/tracing/latency_trace
838 # tracer: preemptirqsoff
840 preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8
841 --------------------------------------------------------------------
842 latency: 293 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
844 | task: ls-4860 (uid:0 nice:0 policy:0 rt_prio:0)
846 => started at: apic_timer_interrupt
847 => ended at: __do_softirq
850 # / _-----=> irqs-off
851 # | / _----=> need-resched
852 # || / _---=> hardirq/softirq
853 # ||| / _--=> preempt-depth
856 # cmd pid ||||| time | caller
858 ls-4860 0d... 0us!: trace_hardirqs_off_thunk (apic_timer_interrupt)
859 ls-4860 0d.s. 294us : _local_bh_enable (__do_softirq)
860 ls-4860 0d.s1 294us : trace_preempt_on (__do_softirq)
864 The trace_hardirqs_off_thunk is called from assembly on x86 when
865 interrupts are disabled in the assembly code. Without the
866 function tracing, we do not know if interrupts were enabled
867 within the preemption points. We do see that it started with
870 Here is a trace with ftrace_enabled set:
873 # tracer: preemptirqsoff
875 preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8
876 --------------------------------------------------------------------
877 latency: 105 us, #183/183, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
879 | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
881 => started at: write_chan
882 => ended at: __do_softirq
885 # / _-----=> irqs-off
886 # | / _----=> need-resched
887 # || / _---=> hardirq/softirq
888 # ||| / _--=> preempt-depth
891 # cmd pid ||||| time | caller
893 ls-4473 0.N.. 0us : preempt_schedule (write_chan)
894 ls-4473 0dN.1 1us : _spin_lock (schedule)
895 ls-4473 0dN.1 2us : add_preempt_count (_spin_lock)
896 ls-4473 0d..2 2us : put_prev_task_fair (schedule)
898 ls-4473 0d..2 13us : set_normalized_timespec (ktime_get_ts)
899 ls-4473 0d..2 13us : __switch_to (schedule)
900 sshd-4261 0d..2 14us : finish_task_switch (schedule)
901 sshd-4261 0d..2 14us : _spin_unlock_irq (finish_task_switch)
902 sshd-4261 0d..1 15us : add_preempt_count (_spin_lock_irqsave)
903 sshd-4261 0d..2 16us : _spin_unlock_irqrestore (hrtick_set)
904 sshd-4261 0d..2 16us : do_IRQ (common_interrupt)
905 sshd-4261 0d..2 17us : irq_enter (do_IRQ)
906 sshd-4261 0d..2 17us : idle_cpu (irq_enter)
907 sshd-4261 0d..2 18us : add_preempt_count (irq_enter)
908 sshd-4261 0d.h2 18us : idle_cpu (irq_enter)
909 sshd-4261 0d.h. 18us : handle_fasteoi_irq (do_IRQ)
910 sshd-4261 0d.h. 19us : _spin_lock (handle_fasteoi_irq)
911 sshd-4261 0d.h. 19us : add_preempt_count (_spin_lock)
912 sshd-4261 0d.h1 20us : _spin_unlock (handle_fasteoi_irq)
913 sshd-4261 0d.h1 20us : sub_preempt_count (_spin_unlock)
915 sshd-4261 0d.h1 28us : _spin_unlock (handle_fasteoi_irq)
916 sshd-4261 0d.h1 29us : sub_preempt_count (_spin_unlock)
917 sshd-4261 0d.h2 29us : irq_exit (do_IRQ)
918 sshd-4261 0d.h2 29us : sub_preempt_count (irq_exit)
919 sshd-4261 0d..3 30us : do_softirq (irq_exit)
920 sshd-4261 0d... 30us : __do_softirq (do_softirq)
921 sshd-4261 0d... 31us : __local_bh_disable (__do_softirq)
922 sshd-4261 0d... 31us+: add_preempt_count (__local_bh_disable)
923 sshd-4261 0d.s4 34us : add_preempt_count (__local_bh_disable)
925 sshd-4261 0d.s3 43us : sub_preempt_count (local_bh_enable_ip)
926 sshd-4261 0d.s4 44us : sub_preempt_count (local_bh_enable_ip)
927 sshd-4261 0d.s3 44us : smp_apic_timer_interrupt (apic_timer_interrupt)
928 sshd-4261 0d.s3 45us : irq_enter (smp_apic_timer_interrupt)
929 sshd-4261 0d.s3 45us : idle_cpu (irq_enter)
930 sshd-4261 0d.s3 46us : add_preempt_count (irq_enter)
931 sshd-4261 0d.H3 46us : idle_cpu (irq_enter)
932 sshd-4261 0d.H3 47us : hrtimer_interrupt (smp_apic_timer_interrupt)
933 sshd-4261 0d.H3 47us : ktime_get (hrtimer_interrupt)
935 sshd-4261 0d.H3 81us : tick_program_event (hrtimer_interrupt)
936 sshd-4261 0d.H3 82us : ktime_get (tick_program_event)
937 sshd-4261 0d.H3 82us : ktime_get_ts (ktime_get)
938 sshd-4261 0d.H3 83us : getnstimeofday (ktime_get_ts)
939 sshd-4261 0d.H3 83us : set_normalized_timespec (ktime_get_ts)
940 sshd-4261 0d.H3 84us : clockevents_program_event (tick_program_event)
941 sshd-4261 0d.H3 84us : lapic_next_event (clockevents_program_event)
942 sshd-4261 0d.H3 85us : irq_exit (smp_apic_timer_interrupt)
943 sshd-4261 0d.H3 85us : sub_preempt_count (irq_exit)
944 sshd-4261 0d.s4 86us : sub_preempt_count (irq_exit)
945 sshd-4261 0d.s3 86us : add_preempt_count (__local_bh_disable)
947 sshd-4261 0d.s1 98us : sub_preempt_count (net_rx_action)
948 sshd-4261 0d.s. 99us : add_preempt_count (_spin_lock_irq)
949 sshd-4261 0d.s1 99us+: _spin_unlock_irq (run_timer_softirq)
950 sshd-4261 0d.s. 104us : _local_bh_enable (__do_softirq)
951 sshd-4261 0d.s. 104us : sub_preempt_count (_local_bh_enable)
952 sshd-4261 0d.s. 105us : _local_bh_enable (__do_softirq)
953 sshd-4261 0d.s1 105us : trace_preempt_on (__do_softirq)
956 This is a very interesting trace. It started with the preemption
957 of the ls task. We see that the task had the "need_resched" bit
958 set via the 'N' in the trace. Interrupts were disabled before
959 the spin_lock at the beginning of the trace. We see that a
960 schedule took place to run sshd. When the interrupts were
961 enabled, we took an interrupt. On return from the interrupt
962 handler, the softirq ran. We took another interrupt while
963 running the softirq as we see from the capital 'H'.
969 In a Real-Time environment it is very important to know the
970 wakeup time it takes for the highest priority task that is woken
971 up to the time that it executes. This is also known as "schedule
972 latency". I stress the point that this is about RT tasks. It is
973 also important to know the scheduling latency of non-RT tasks,
974 but the average schedule latency is better for non-RT tasks.
975 Tools like LatencyTop are more appropriate for such
978 Real-Time environments are interested in the worst case latency.
979 That is the longest latency it takes for something to happen,
980 and not the average. We can have a very fast scheduler that may
981 only have a large latency once in a while, but that would not
982 work well with Real-Time tasks. The wakeup tracer was designed
983 to record the worst case wakeups of RT tasks. Non-RT tasks are
984 not recorded because the tracer only records one worst case and
985 tracing non-RT tasks that are unpredictable will overwrite the
986 worst case latency of RT tasks.
988 Since this tracer only deals with RT tasks, we will run this
989 slightly differently than we did with the previous tracers.
990 Instead of performing an 'ls', we will run 'sleep 1' under
991 'chrt' which changes the priority of the task.
993 # echo wakeup > /debug/tracing/current_tracer
994 # echo 0 > /debug/tracing/tracing_max_latency
995 # echo 1 > /debug/tracing/tracing_enabled
997 # echo 0 > /debug/tracing/tracing_enabled
998 # cat /debug/tracing/latency_trace
1001 wakeup latency trace v1.1.5 on 2.6.26-rc8
1002 --------------------------------------------------------------------
1003 latency: 4 us, #2/2, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
1005 | task: sleep-4901 (uid:0 nice:0 policy:1 rt_prio:5)
1009 # / _-----=> irqs-off
1010 # | / _----=> need-resched
1011 # || / _---=> hardirq/softirq
1012 # ||| / _--=> preempt-depth
1015 # cmd pid ||||| time | caller
1017 <idle>-0 1d.h4 0us+: try_to_wake_up (wake_up_process)
1018 <idle>-0 1d..4 4us : schedule (cpu_idle)
1021 Running this on an idle system, we see that it only took 4
1022 microseconds to perform the task switch. Note, since the trace
1023 marker in the schedule is before the actual "switch", we stop
1024 the tracing when the recorded task is about to schedule in. This
1025 may change if we add a new marker at the end of the scheduler.
1027 Notice that the recorded task is 'sleep' with the PID of 4901
1028 and it has an rt_prio of 5. This priority is user-space priority
1029 and not the internal kernel priority. The policy is 1 for
1030 SCHED_FIFO and 2 for SCHED_RR.
1032 Doing the same with chrt -r 5 and ftrace_enabled set.
1036 wakeup latency trace v1.1.5 on 2.6.26-rc8
1037 --------------------------------------------------------------------
1038 latency: 50 us, #60/60, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
1040 | task: sleep-4068 (uid:0 nice:0 policy:2 rt_prio:5)
1044 # / _-----=> irqs-off
1045 # | / _----=> need-resched
1046 # || / _---=> hardirq/softirq
1047 # ||| / _--=> preempt-depth
1050 # cmd pid ||||| time | caller
1052 ksoftirq-7 1d.H3 0us : try_to_wake_up (wake_up_process)
1053 ksoftirq-7 1d.H4 1us : sub_preempt_count (marker_probe_cb)
1054 ksoftirq-7 1d.H3 2us : check_preempt_wakeup (try_to_wake_up)
1055 ksoftirq-7 1d.H3 3us : update_curr (check_preempt_wakeup)
1056 ksoftirq-7 1d.H3 4us : calc_delta_mine (update_curr)
1057 ksoftirq-7 1d.H3 5us : __resched_task (check_preempt_wakeup)
1058 ksoftirq-7 1d.H3 6us : task_wake_up_rt (try_to_wake_up)
1059 ksoftirq-7 1d.H3 7us : _spin_unlock_irqrestore (try_to_wake_up)
1061 ksoftirq-7 1d.H2 17us : irq_exit (smp_apic_timer_interrupt)
1062 ksoftirq-7 1d.H2 18us : sub_preempt_count (irq_exit)
1063 ksoftirq-7 1d.s3 19us : sub_preempt_count (irq_exit)
1064 ksoftirq-7 1..s2 20us : rcu_process_callbacks (__do_softirq)
1066 ksoftirq-7 1..s2 26us : __rcu_process_callbacks (rcu_process_callbacks)
1067 ksoftirq-7 1d.s2 27us : _local_bh_enable (__do_softirq)
1068 ksoftirq-7 1d.s2 28us : sub_preempt_count (_local_bh_enable)
1069 ksoftirq-7 1.N.3 29us : sub_preempt_count (ksoftirqd)
1070 ksoftirq-7 1.N.2 30us : _cond_resched (ksoftirqd)
1071 ksoftirq-7 1.N.2 31us : __cond_resched (_cond_resched)
1072 ksoftirq-7 1.N.2 32us : add_preempt_count (__cond_resched)
1073 ksoftirq-7 1.N.2 33us : schedule (__cond_resched)
1074 ksoftirq-7 1.N.2 33us : add_preempt_count (schedule)
1075 ksoftirq-7 1.N.3 34us : hrtick_clear (schedule)
1076 ksoftirq-7 1dN.3 35us : _spin_lock (schedule)
1077 ksoftirq-7 1dN.3 36us : add_preempt_count (_spin_lock)
1078 ksoftirq-7 1d..4 37us : put_prev_task_fair (schedule)
1079 ksoftirq-7 1d..4 38us : update_curr (put_prev_task_fair)
1081 ksoftirq-7 1d..5 47us : _spin_trylock (tracing_record_cmdline)
1082 ksoftirq-7 1d..5 48us : add_preempt_count (_spin_trylock)
1083 ksoftirq-7 1d..6 49us : _spin_unlock (tracing_record_cmdline)
1084 ksoftirq-7 1d..6 49us : sub_preempt_count (_spin_unlock)
1085 ksoftirq-7 1d..4 50us : schedule (__cond_resched)
1087 The interrupt went off while running ksoftirqd. This task runs
1088 at SCHED_OTHER. Why did not we see the 'N' set early? This may
1089 be a harmless bug with x86_32 and 4K stacks. On x86_32 with 4K
1090 stacks configured, the interrupt and softirq run with their own
1091 stack. Some information is held on the top of the task's stack
1092 (need_resched and preempt_count are both stored there). The
1093 setting of the NEED_RESCHED bit is done directly to the task's
1094 stack, but the reading of the NEED_RESCHED is done by looking at
1095 the current stack, which in this case is the stack for the hard
1096 interrupt. This hides the fact that NEED_RESCHED has been set.
1097 We do not see the 'N' until we switch back to the task's
1103 This tracer is the function tracer. Enabling the function tracer
1104 can be done from the debug file system. Make sure the
1105 ftrace_enabled is set; otherwise this tracer is a nop.
1107 # sysctl kernel.ftrace_enabled=1
1108 # echo function > /debug/tracing/current_tracer
1109 # echo 1 > /debug/tracing/tracing_enabled
1111 # echo 0 > /debug/tracing/tracing_enabled
1112 # cat /debug/tracing/trace
1115 # TASK-PID CPU# TIMESTAMP FUNCTION
1117 bash-4003 [00] 123.638713: finish_task_switch <-schedule
1118 bash-4003 [00] 123.638714: _spin_unlock_irq <-finish_task_switch
1119 bash-4003 [00] 123.638714: sub_preempt_count <-_spin_unlock_irq
1120 bash-4003 [00] 123.638715: hrtick_set <-schedule
1121 bash-4003 [00] 123.638715: _spin_lock_irqsave <-hrtick_set
1122 bash-4003 [00] 123.638716: add_preempt_count <-_spin_lock_irqsave
1123 bash-4003 [00] 123.638716: _spin_unlock_irqrestore <-hrtick_set
1124 bash-4003 [00] 123.638717: sub_preempt_count <-_spin_unlock_irqrestore
1125 bash-4003 [00] 123.638717: hrtick_clear <-hrtick_set
1126 bash-4003 [00] 123.638718: sub_preempt_count <-schedule
1127 bash-4003 [00] 123.638718: sub_preempt_count <-preempt_schedule
1128 bash-4003 [00] 123.638719: wait_for_completion <-__stop_machine_run
1129 bash-4003 [00] 123.638719: wait_for_common <-wait_for_completion
1130 bash-4003 [00] 123.638720: _spin_lock_irq <-wait_for_common
1131 bash-4003 [00] 123.638720: add_preempt_count <-_spin_lock_irq
1135 Note: function tracer uses ring buffers to store the above
1136 entries. The newest data may overwrite the oldest data.
1137 Sometimes using echo to stop the trace is not sufficient because
1138 the tracing could have overwritten the data that you wanted to
1139 record. For this reason, it is sometimes better to disable
1140 tracing directly from a program. This allows you to stop the
1141 tracing at the point that you hit the part that you are
1142 interested in. To disable the tracing directly from a C program,
1143 something like following code snippet can be used:
1147 int main(int argc, char *argv[]) {
1149 trace_fd = open("/debug/tracing/tracing_enabled", O_WRONLY);
1151 if (condition_hit()) {
1152 write(trace_fd, "0", 1);
1157 Note: Here we hard coded the path name. The debugfs mount is not
1158 guaranteed to be at /debug (and is more commonly at
1159 /sys/kernel/debug). For simple one time traces, the above is
1160 sufficent. For anything else, a search through /proc/mounts may
1161 be needed to find where the debugfs file-system is mounted.
1164 Single thread tracing
1165 ---------------------
1167 By writing into /debug/tracing/set_ftrace_pid you can trace a
1168 single thread. For example:
1170 # cat /debug/tracing/set_ftrace_pid
1172 # echo 3111 > /debug/tracing/set_ftrace_pid
1173 # cat /debug/tracing/set_ftrace_pid
1175 # echo function > /debug/tracing/current_tracer
1176 # cat /debug/tracing/trace | head
1179 # TASK-PID CPU# TIMESTAMP FUNCTION
1181 yum-updatesd-3111 [003] 1637.254676: finish_task_switch <-thread_return
1182 yum-updatesd-3111 [003] 1637.254681: hrtimer_cancel <-schedule_hrtimeout_range
1183 yum-updatesd-3111 [003] 1637.254682: hrtimer_try_to_cancel <-hrtimer_cancel
1184 yum-updatesd-3111 [003] 1637.254683: lock_hrtimer_base <-hrtimer_try_to_cancel
1185 yum-updatesd-3111 [003] 1637.254685: fget_light <-do_sys_poll
1186 yum-updatesd-3111 [003] 1637.254686: pipe_poll <-do_sys_poll
1187 # echo -1 > /debug/tracing/set_ftrace_pid
1188 # cat /debug/tracing/trace |head
1191 # TASK-PID CPU# TIMESTAMP FUNCTION
1193 ##### CPU 3 buffer started ####
1194 yum-updatesd-3111 [003] 1701.957688: free_poll_entry <-poll_freewait
1195 yum-updatesd-3111 [003] 1701.957689: remove_wait_queue <-free_poll_entry
1196 yum-updatesd-3111 [003] 1701.957691: fput <-free_poll_entry
1197 yum-updatesd-3111 [003] 1701.957692: audit_syscall_exit <-sysret_audit
1198 yum-updatesd-3111 [003] 1701.957693: path_put <-audit_syscall_exit
1200 If you want to trace a function when executing, you could use
1201 something like this simple program:
1205 #include <sys/types.h>
1206 #include <sys/stat.h>
1210 int main (int argc, char **argv)
1220 ffd = open("/debug/tracing/current_tracer", O_WRONLY);
1223 write(ffd, "nop", 3);
1225 fd = open("/debug/tracing/set_ftrace_pid", O_WRONLY);
1226 s = sprintf(line, "%d\n", getpid());
1229 write(ffd, "function", 8);
1234 execvp(argv[1], argv+1);
1241 hw-branch-tracer (x86 only)
1242 ---------------------------
1244 This tracer uses the x86 last branch tracing hardware feature to
1245 collect a branch trace on all cpus with relatively low overhead.
1247 The tracer uses a fixed-size circular buffer per cpu and only
1248 traces ring 0 branches. The trace file dumps that buffer in the
1251 # tracer: hw-branch-tracer
1254 0 scheduler_tick+0xb5/0x1bf <- task_tick_idle+0x5/0x6
1255 2 run_posix_cpu_timers+0x2b/0x72a <- run_posix_cpu_timers+0x25/0x72a
1256 0 scheduler_tick+0x139/0x1bf <- scheduler_tick+0xed/0x1bf
1257 0 scheduler_tick+0x17c/0x1bf <- scheduler_tick+0x148/0x1bf
1258 2 run_posix_cpu_timers+0x9e/0x72a <- run_posix_cpu_timers+0x5e/0x72a
1259 0 scheduler_tick+0x1b6/0x1bf <- scheduler_tick+0x1aa/0x1bf
1262 The tracer may be used to dump the trace for the oops'ing cpu on
1263 a kernel oops into the system log. To enable this,
1264 ftrace_dump_on_oops must be set. To set ftrace_dump_on_oops, one
1265 can either use the sysctl function or set it via the proc system
1268 sysctl kernel.ftrace_dump_on_oops=1
1272 echo 1 > /proc/sys/kernel/ftrace_dump_on_oops
1275 Here's an example of such a dump after a null pointer
1276 dereference in a kernel module:
1278 [57848.105921] BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
1279 [57848.106019] IP: [<ffffffffa0000006>] open+0x6/0x14 [oops]
1280 [57848.106019] PGD 2354e9067 PUD 2375e7067 PMD 0
1281 [57848.106019] Oops: 0002 [#1] SMP
1282 [57848.106019] last sysfs file: /sys/devices/pci0000:00/0000:00:1e.0/0000:20:05.0/local_cpus
1283 [57848.106019] Dumping ftrace buffer:
1284 [57848.106019] ---------------------------------
1286 [57848.106019] 0 chrdev_open+0xe6/0x165 <- cdev_put+0x23/0x24
1287 [57848.106019] 0 chrdev_open+0x117/0x165 <- chrdev_open+0xfa/0x165
1288 [57848.106019] 0 chrdev_open+0x120/0x165 <- chrdev_open+0x11c/0x165
1289 [57848.106019] 0 chrdev_open+0x134/0x165 <- chrdev_open+0x12b/0x165
1290 [57848.106019] 0 open+0x0/0x14 [oops] <- chrdev_open+0x144/0x165
1291 [57848.106019] 0 page_fault+0x0/0x30 <- open+0x6/0x14 [oops]
1292 [57848.106019] 0 error_entry+0x0/0x5b <- page_fault+0x4/0x30
1293 [57848.106019] 0 error_kernelspace+0x0/0x31 <- error_entry+0x59/0x5b
1294 [57848.106019] 0 error_sti+0x0/0x1 <- error_kernelspace+0x2d/0x31
1295 [57848.106019] 0 page_fault+0x9/0x30 <- error_sti+0x0/0x1
1296 [57848.106019] 0 do_page_fault+0x0/0x881 <- page_fault+0x1a/0x30
1298 [57848.106019] 0 do_page_fault+0x66b/0x881 <- is_prefetch+0x1ee/0x1f2
1299 [57848.106019] 0 do_page_fault+0x6e0/0x881 <- do_page_fault+0x67a/0x881
1300 [57848.106019] 0 oops_begin+0x0/0x96 <- do_page_fault+0x6e0/0x881
1301 [57848.106019] 0 trace_hw_branch_oops+0x0/0x2d <- oops_begin+0x9/0x96
1303 [57848.106019] 0 ds_suspend_bts+0x2a/0xe3 <- ds_suspend_bts+0x1a/0xe3
1304 [57848.106019] ---------------------------------
1305 [57848.106019] CPU 0
1306 [57848.106019] Modules linked in: oops
1307 [57848.106019] Pid: 5542, comm: cat Tainted: G W 2.6.28 #23
1308 [57848.106019] RIP: 0010:[<ffffffffa0000006>] [<ffffffffa0000006>] open+0x6/0x14 [oops]
1309 [57848.106019] RSP: 0018:ffff880235457d48 EFLAGS: 00010246
1313 function graph tracer
1314 ---------------------------
1316 This tracer is similar to the function tracer except that it
1317 probes a function on its entry and its exit. This is done by
1318 using a dynamically allocated stack of return addresses in each
1319 task_struct. On function entry the tracer overwrites the return
1320 address of each function traced to set a custom probe. Thus the
1321 original return address is stored on the stack of return address
1324 Probing on both ends of a function leads to special features
1327 - measure of a function's time execution
1328 - having a reliable call stack to draw function calls graph
1330 This tracer is useful in several situations:
1332 - you want to find the reason of a strange kernel behavior and
1333 need to see what happens in detail on any areas (or specific
1336 - you are experiencing weird latencies but it's difficult to
1339 - you want to find quickly which path is taken by a specific
1342 - you just want to peek inside a working kernel and want to see
1345 # tracer: function_graph
1347 # CPU DURATION FUNCTION CALLS
1351 0) | do_sys_open() {
1353 0) | kmem_cache_alloc() {
1354 0) 1.382 us | __might_sleep();
1356 0) | strncpy_from_user() {
1357 0) | might_fault() {
1358 0) 1.389 us | __might_sleep();
1363 0) 0.668 us | _spin_lock();
1364 0) 0.570 us | expand_files();
1365 0) 0.586 us | _spin_unlock();
1368 There are several columns that can be dynamically
1369 enabled/disabled. You can use every combination of options you
1370 want, depending on your needs.
1372 - The cpu number on which the function executed is default
1373 enabled. It is sometimes better to only trace one cpu (see
1374 tracing_cpu_mask file) or you might sometimes see unordered
1375 function calls while cpu tracing switch.
1377 hide: echo nofuncgraph-cpu > /debug/tracing/trace_options
1378 show: echo funcgraph-cpu > /debug/tracing/trace_options
1380 - The duration (function's time of execution) is displayed on
1381 the closing bracket line of a function or on the same line
1382 than the current function in case of a leaf one. It is default
1385 hide: echo nofuncgraph-duration > /debug/tracing/trace_options
1386 show: echo funcgraph-duration > /debug/tracing/trace_options
1388 - The overhead field precedes the duration field in case of
1389 reached duration thresholds.
1391 hide: echo nofuncgraph-overhead > /debug/tracing/trace_options
1392 show: echo funcgraph-overhead > /debug/tracing/trace_options
1393 depends on: funcgraph-duration
1398 0) 0.646 us | _spin_lock_irqsave();
1399 0) 0.684 us | _spin_unlock_irqrestore();
1401 0) 0.548 us | fput();
1407 0) | kmem_cache_free() {
1408 0) 0.518 us | __phys_addr();
1414 + means that the function exceeded 10 usecs.
1415 ! means that the function exceeded 100 usecs.
1418 - The task/pid field displays the thread cmdline and pid which
1419 executed the function. It is default disabled.
1421 hide: echo nofuncgraph-proc > /debug/tracing/trace_options
1422 show: echo funcgraph-proc > /debug/tracing/trace_options
1426 # tracer: function_graph
1428 # CPU TASK/PID DURATION FUNCTION CALLS
1430 0) sh-4802 | | d_free() {
1431 0) sh-4802 | | call_rcu() {
1432 0) sh-4802 | | __call_rcu() {
1433 0) sh-4802 | 0.616 us | rcu_process_gp_end();
1434 0) sh-4802 | 0.586 us | check_for_new_grace_period();
1435 0) sh-4802 | 2.899 us | }
1436 0) sh-4802 | 4.040 us | }
1437 0) sh-4802 | 5.151 us | }
1438 0) sh-4802 | + 49.370 us | }
1441 - The absolute time field is an absolute timestamp given by the
1442 system clock since it started. A snapshot of this time is
1443 given on each entry/exit of functions
1445 hide: echo nofuncgraph-abstime > /debug/tracing/trace_options
1446 show: echo funcgraph-abstime > /debug/tracing/trace_options
1451 # TIME CPU DURATION FUNCTION CALLS
1453 360.774522 | 1) 0.541 us | }
1454 360.774522 | 1) 4.663 us | }
1455 360.774523 | 1) 0.541 us | __wake_up_bit();
1456 360.774524 | 1) 6.796 us | }
1457 360.774524 | 1) 7.952 us | }
1458 360.774525 | 1) 9.063 us | }
1459 360.774525 | 1) 0.615 us | journal_mark_dirty();
1460 360.774527 | 1) 0.578 us | __brelse();
1461 360.774528 | 1) | reiserfs_prepare_for_journal() {
1462 360.774528 | 1) | unlock_buffer() {
1463 360.774529 | 1) | wake_up_bit() {
1464 360.774529 | 1) | bit_waitqueue() {
1465 360.774530 | 1) 0.594 us | __phys_addr();
1468 You can put some comments on specific functions by using
1469 trace_printk() For example, if you want to put a comment inside
1470 the __might_sleep() function, you just have to include
1471 <linux/ftrace.h> and call trace_printk() inside __might_sleep()
1473 trace_printk("I'm a comment!\n")
1477 1) | __might_sleep() {
1478 1) | /* I'm a comment! */
1482 You might find other useful features for this tracer in the
1483 following "dynamic ftrace" section such as tracing only specific
1489 If CONFIG_DYNAMIC_FTRACE is set, the system will run with
1490 virtually no overhead when function tracing is disabled. The way
1491 this works is the mcount function call (placed at the start of
1492 every kernel function, produced by the -pg switch in gcc),
1493 starts of pointing to a simple return. (Enabling FTRACE will
1494 include the -pg switch in the compiling of the kernel.)
1496 At compile time every C file object is run through the
1497 recordmcount.pl script (located in the scripts directory). This
1498 script will process the C object using objdump to find all the
1499 locations in the .text section that call mcount. (Note, only the
1500 .text section is processed, since processing other sections like
1501 .init.text may cause races due to those sections being freed).
1503 A new section called "__mcount_loc" is created that holds
1504 references to all the mcount call sites in the .text section.
1505 This section is compiled back into the original object. The
1506 final linker will add all these references into a single table.
1508 On boot up, before SMP is initialized, the dynamic ftrace code
1509 scans this table and updates all the locations into nops. It
1510 also records the locations, which are added to the
1511 available_filter_functions list. Modules are processed as they
1512 are loaded and before they are executed. When a module is
1513 unloaded, it also removes its functions from the ftrace function
1514 list. This is automatic in the module unload code, and the
1515 module author does not need to worry about it.
1517 When tracing is enabled, kstop_machine is called to prevent
1518 races with the CPUS executing code being modified (which can
1519 cause the CPU to do undesireable things), and the nops are
1520 patched back to calls. But this time, they do not call mcount
1521 (which is just a function stub). They now call into the ftrace
1524 One special side-effect to the recording of the functions being
1525 traced is that we can now selectively choose which functions we
1526 wish to trace and which ones we want the mcount calls to remain
1529 Two files are used, one for enabling and one for disabling the
1530 tracing of specified functions. They are:
1538 A list of available functions that you can add to these files is
1541 available_filter_functions
1543 # cat /debug/tracing/available_filter_functions
1552 If I am only interested in sys_nanosleep and hrtimer_interrupt:
1554 # echo sys_nanosleep hrtimer_interrupt \
1555 > /debug/tracing/set_ftrace_filter
1556 # echo ftrace > /debug/tracing/current_tracer
1557 # echo 1 > /debug/tracing/tracing_enabled
1559 # echo 0 > /debug/tracing/tracing_enabled
1560 # cat /debug/tracing/trace
1563 # TASK-PID CPU# TIMESTAMP FUNCTION
1565 usleep-4134 [00] 1317.070017: hrtimer_interrupt <-smp_apic_timer_interrupt
1566 usleep-4134 [00] 1317.070111: sys_nanosleep <-syscall_call
1567 <idle>-0 [00] 1317.070115: hrtimer_interrupt <-smp_apic_timer_interrupt
1569 To see which functions are being traced, you can cat the file:
1571 # cat /debug/tracing/set_ftrace_filter
1576 Perhaps this is not enough. The filters also allow simple wild
1577 cards. Only the following are currently available
1579 <match>* - will match functions that begin with <match>
1580 *<match> - will match functions that end with <match>
1581 *<match>* - will match functions that have <match> in it
1583 These are the only wild cards which are supported.
1585 <match>*<match> will not work.
1587 Note: It is better to use quotes to enclose the wild cards,
1588 otherwise the shell may expand the parameters into names
1589 of files in the local directory.
1591 # echo 'hrtimer_*' > /debug/tracing/set_ftrace_filter
1597 # TASK-PID CPU# TIMESTAMP FUNCTION
1599 bash-4003 [00] 1480.611794: hrtimer_init <-copy_process
1600 bash-4003 [00] 1480.611941: hrtimer_start <-hrtick_set
1601 bash-4003 [00] 1480.611956: hrtimer_cancel <-hrtick_clear
1602 bash-4003 [00] 1480.611956: hrtimer_try_to_cancel <-hrtimer_cancel
1603 <idle>-0 [00] 1480.612019: hrtimer_get_next_event <-get_next_timer_interrupt
1604 <idle>-0 [00] 1480.612025: hrtimer_get_next_event <-get_next_timer_interrupt
1605 <idle>-0 [00] 1480.612032: hrtimer_get_next_event <-get_next_timer_interrupt
1606 <idle>-0 [00] 1480.612037: hrtimer_get_next_event <-get_next_timer_interrupt
1607 <idle>-0 [00] 1480.612382: hrtimer_get_next_event <-get_next_timer_interrupt
1610 Notice that we lost the sys_nanosleep.
1612 # cat /debug/tracing/set_ftrace_filter
1617 hrtimer_try_to_cancel
1621 hrtimer_force_reprogram
1622 hrtimer_get_next_event
1626 hrtimer_get_remaining
1628 hrtimer_init_sleeper
1631 This is because the '>' and '>>' act just like they do in bash.
1632 To rewrite the filters, use '>'
1633 To append to the filters, use '>>'
1635 To clear out a filter so that all functions will be recorded
1638 # echo > /debug/tracing/set_ftrace_filter
1639 # cat /debug/tracing/set_ftrace_filter
1642 Again, now we want to append.
1644 # echo sys_nanosleep > /debug/tracing/set_ftrace_filter
1645 # cat /debug/tracing/set_ftrace_filter
1647 # echo 'hrtimer_*' >> /debug/tracing/set_ftrace_filter
1648 # cat /debug/tracing/set_ftrace_filter
1653 hrtimer_try_to_cancel
1657 hrtimer_force_reprogram
1658 hrtimer_get_next_event
1663 hrtimer_get_remaining
1665 hrtimer_init_sleeper
1668 The set_ftrace_notrace prevents those functions from being
1671 # echo '*preempt*' '*lock*' > /debug/tracing/set_ftrace_notrace
1677 # TASK-PID CPU# TIMESTAMP FUNCTION
1679 bash-4043 [01] 115.281644: finish_task_switch <-schedule
1680 bash-4043 [01] 115.281645: hrtick_set <-schedule
1681 bash-4043 [01] 115.281645: hrtick_clear <-hrtick_set
1682 bash-4043 [01] 115.281646: wait_for_completion <-__stop_machine_run
1683 bash-4043 [01] 115.281647: wait_for_common <-wait_for_completion
1684 bash-4043 [01] 115.281647: kthread_stop <-stop_machine_run
1685 bash-4043 [01] 115.281648: init_waitqueue_head <-kthread_stop
1686 bash-4043 [01] 115.281648: wake_up_process <-kthread_stop
1687 bash-4043 [01] 115.281649: try_to_wake_up <-wake_up_process
1689 We can see that there's no more lock or preempt tracing.
1692 Dynamic ftrace with the function graph tracer
1693 ---------------------------------------------
1695 Although what has been explained above concerns both the
1696 function tracer and the function-graph-tracer, there are some
1697 special features only available in the function-graph tracer.
1699 If you want to trace only one function and all of its children,
1700 you just have to echo its name into set_graph_function:
1702 echo __do_fault > set_graph_function
1704 will produce the following "expanded" trace of the __do_fault()
1708 0) | filemap_fault() {
1709 0) | find_lock_page() {
1710 0) 0.804 us | find_get_page();
1711 0) | __might_sleep() {
1715 0) 0.653 us | _spin_lock();
1716 0) 0.578 us | page_add_file_rmap();
1717 0) 0.525 us | native_set_pte_at();
1718 0) 0.585 us | _spin_unlock();
1719 0) | unlock_page() {
1720 0) 0.541 us | page_waitqueue();
1721 0) 0.639 us | __wake_up_bit();
1725 0) | filemap_fault() {
1726 0) | find_lock_page() {
1727 0) 0.698 us | find_get_page();
1728 0) | __might_sleep() {
1732 0) 0.631 us | _spin_lock();
1733 0) 0.571 us | page_add_file_rmap();
1734 0) 0.526 us | native_set_pte_at();
1735 0) 0.586 us | _spin_unlock();
1736 0) | unlock_page() {
1737 0) 0.533 us | page_waitqueue();
1738 0) 0.638 us | __wake_up_bit();
1742 You can also expand several functions at once:
1744 echo sys_open > set_graph_function
1745 echo sys_close >> set_graph_function
1747 Now if you want to go back to trace all functions you can clear
1748 this special filter via:
1750 echo > set_graph_function
1756 The trace_pipe outputs the same content as the trace file, but
1757 the effect on the tracing is different. Every read from
1758 trace_pipe is consumed. This means that subsequent reads will be
1759 different. The trace is live.
1761 # echo function > /debug/tracing/current_tracer
1762 # cat /debug/tracing/trace_pipe > /tmp/trace.out &
1764 # echo 1 > /debug/tracing/tracing_enabled
1766 # echo 0 > /debug/tracing/tracing_enabled
1767 # cat /debug/tracing/trace
1770 # TASK-PID CPU# TIMESTAMP FUNCTION
1774 # cat /tmp/trace.out
1775 bash-4043 [00] 41.267106: finish_task_switch <-schedule
1776 bash-4043 [00] 41.267106: hrtick_set <-schedule
1777 bash-4043 [00] 41.267107: hrtick_clear <-hrtick_set
1778 bash-4043 [00] 41.267108: wait_for_completion <-__stop_machine_run
1779 bash-4043 [00] 41.267108: wait_for_common <-wait_for_completion
1780 bash-4043 [00] 41.267109: kthread_stop <-stop_machine_run
1781 bash-4043 [00] 41.267109: init_waitqueue_head <-kthread_stop
1782 bash-4043 [00] 41.267110: wake_up_process <-kthread_stop
1783 bash-4043 [00] 41.267110: try_to_wake_up <-wake_up_process
1784 bash-4043 [00] 41.267111: select_task_rq_rt <-try_to_wake_up
1787 Note, reading the trace_pipe file will block until more input is
1788 added. By changing the tracer, trace_pipe will issue an EOF. We
1789 needed to set the function tracer _before_ we "cat" the
1796 Having too much or not enough data can be troublesome in
1797 diagnosing an issue in the kernel. The file buffer_size_kb is
1798 used to modify the size of the internal trace buffers. The
1799 number listed is the number of entries that can be recorded per
1800 CPU. To know the full size, multiply the number of possible CPUS
1801 with the number of entries.
1803 # cat /debug/tracing/buffer_size_kb
1804 1408 (units kilobytes)
1806 Note, to modify this, you must have tracing completely disabled.
1807 To do that, echo "nop" into the current_tracer. If the
1808 current_tracer is not set to "nop", an EINVAL error will be
1811 # echo nop > /debug/tracing/current_tracer
1812 # echo 10000 > /debug/tracing/buffer_size_kb
1813 # cat /debug/tracing/buffer_size_kb
1814 10000 (units kilobytes)
1816 The number of pages which will be allocated is limited to a
1817 percentage of available memory. Allocating too much will produce
1820 # echo 1000000000000 > /debug/tracing/buffer_size_kb
1821 -bash: echo: write error: Cannot allocate memory
1822 # cat /debug/tracing/buffer_size_kb
1827 More details can be found in the source code, in the
1828 kernel/tracing/*.c files.