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 performance
19 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 the
23 tracers that are currently in ftrace include a tracer to trace
24 context switches, the time it takes for a high priority task to
25 run after it was woken up, the time interrupts are disabled, and
26 more (ftrace allows for tracer plugins, which means that the list of
27 tracers can always grow).
33 Ftrace uses the debugfs file system to hold the control files as well
34 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 simplicity
42 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.
53 current_tracer: This is used to set or display the current tracer
56 available_tracers: This holds the different types of tracers that
57 have been compiled into the kernel. The tracers
58 listed here can be configured by echoing their name
61 tracing_enabled: This sets or displays whether the current_tracer
62 is activated and tracing or not. Echo 0 into this
63 file to disable the tracer or 1 to enable it.
65 trace: This file holds the output of the trace in a human readable
66 format (described below).
68 latency_trace: This file shows the same trace but the information
69 is organized more to display possible latencies
70 in the system (described below).
72 trace_pipe: The output is the same as the "trace" file but this
73 file is meant to be streamed with live tracing.
74 Reads from this file will block until new data
75 is retrieved. Unlike the "trace" and "latency_trace"
76 files, this file is a consumer. This means reading
77 from this file causes sequential reads to display
78 more current data. Once data is read from this
79 file, it is consumed, and will not be read
80 again with a sequential read. The "trace" and
81 "latency_trace" files are static, and if the
82 tracer is not adding more data, they will display
83 the same information every time they are read.
85 trace_options: This file lets the user control the amount of data
86 that is displayed in one of the above output
89 trace_max_latency: Some of the tracers record the max latency.
90 For example, the time interrupts are disabled.
91 This time is saved in this file. The max trace
92 will also be stored, and displayed by either
93 "trace" or "latency_trace". A new max trace will
94 only be recorded if the latency is greater than
95 the value in this file. (in microseconds)
97 buffer_size_kb: This sets or displays the number of kilobytes each CPU
98 buffer can hold. The tracer buffers are the same size
99 for each CPU. The displayed number is the size of the
100 CPU buffer and not total size of all buffers. The
101 trace buffers are allocated in pages (blocks of memory
102 that the kernel uses for allocation, usually 4 KB in size).
103 If the last page allocated has room for more bytes
104 than requested, the rest of the page will be used,
105 making the actual allocation bigger than requested.
106 (Note, the size may not be a multiple of the page size due
107 to buffer managment overhead.)
109 This can only be updated when the current_tracer
112 tracing_cpumask: This is a mask that lets the user only trace
113 on specified CPUS. The format is a hex string
114 representing the CPUS.
116 set_ftrace_filter: When dynamic ftrace is configured in (see the
117 section below "dynamic ftrace"), the code is dynamically
118 modified (code text rewrite) to disable calling of the
119 function profiler (mcount). This lets tracing be configured
120 in with practically no overhead in performance. This also
121 has a side effect of enabling or disabling specific functions
122 to be traced. Echoing names of functions into this file
123 will limit the trace to only those functions.
125 set_ftrace_notrace: This has an effect opposite to that of
126 set_ftrace_filter. Any function that is added here will not
127 be traced. If a function exists in both set_ftrace_filter
128 and set_ftrace_notrace, the function will _not_ be traced.
130 set_ftrace_pid: Have the function tracer only trace a single thread.
132 set_graph_function: Select the function where the trace have to start
133 with the function graph tracer (See the section
134 "dynamic ftrace" for more details).
136 available_filter_functions: This lists the functions that ftrace
137 has processed and can trace. These are the function
138 names that you can pass to "set_ftrace_filter" or
139 "set_ftrace_notrace". (See the section "dynamic ftrace"
140 below for more details.)
146 Here is the list of current tracers that may be configured.
148 function - function tracer that uses mcount to trace all functions.
150 function_graph_tracer - similar to the function tracer except that the
151 function tracer probes the functions on their entry whereas the
152 function graph tracer traces on both entry and exit of the
153 functions. It then provides the ability to draw a graph of
154 function calls like a primitive C code source.
156 sched_switch - traces the context switches between tasks.
158 irqsoff - traces the areas that disable interrupts and saves
159 the trace with the longest max latency.
160 See tracing_max_latency. When a new max is recorded,
161 it replaces the old trace. It is best to view this
162 trace via the latency_trace file.
164 preemptoff - Similar to irqsoff but traces and records the amount of
165 time for which preemption is disabled.
167 preemptirqsoff - Similar to irqsoff and preemptoff, but traces and
168 records the largest time for which irqs and/or preemption
171 wakeup - Traces and records the max latency that it takes for
172 the highest priority task to get scheduled after
173 it has been woken up.
175 nop - This is not a tracer. To remove all tracers from tracing
176 simply echo "nop" into current_tracer.
178 hw-branch-tracer - traces branches on all cpu's in a circular buffer.
181 Examples of using the tracer
182 ----------------------------
184 Here are typical examples of using the tracers when controlling them only
185 with the debugfs interface (without using any user-land utilities).
190 Here is an example of the output format of the file "trace"
195 # TASK-PID CPU# TIMESTAMP FUNCTION
197 bash-4251 [01] 10152.583854: path_put <-path_walk
198 bash-4251 [01] 10152.583855: dput <-path_put
199 bash-4251 [01] 10152.583855: _atomic_dec_and_lock <-dput
202 A header is printed with the tracer name that is represented by the trace.
203 In this case the tracer is "function". Then a header showing the format. Task
204 name "bash", the task PID "4251", the CPU that it was running on
205 "01", the timestamp in <secs>.<usecs> format, the function name that was
206 traced "path_put" and the parent function that called this function
207 "path_walk". The timestamp is the time at which the function was
210 The sched_switch tracer also includes tracing of task wakeups and
213 ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 2916:115:S
214 ksoftirqd/1-7 [01] 1453.070013: 7:115:R + 10:115:S
215 ksoftirqd/1-7 [01] 1453.070013: 7:115:R ==> 10:115:R
216 events/1-10 [01] 1453.070013: 10:115:S ==> 2916:115:R
217 kondemand/1-2916 [01] 1453.070013: 2916:115:S ==> 7:115:R
218 ksoftirqd/1-7 [01] 1453.070013: 7:115:S ==> 0:140:R
220 Wake ups are represented by a "+" and the context switches are shown as
221 "==>". The format is:
225 Previous task Next Task
227 <pid>:<prio>:<state> ==> <pid>:<prio>:<state>
231 Current task Task waking up
233 <pid>:<prio>:<state> + <pid>:<prio>:<state>
235 The prio is the internal kernel priority, which is the inverse of the
236 priority that is usually displayed by user-space tools. Zero represents
237 the highest priority (99). Prio 100 starts the "nice" priorities with
238 100 being equal to nice -20 and 139 being nice 19. The prio "140" is
239 reserved for the idle task which is the lowest priority thread (pid 0).
245 For traces that display latency times, the latency_trace file gives
246 somewhat more information to see why a latency happened. Here is a typical
251 irqsoff latency trace v1.1.5 on 2.6.26-rc8
252 --------------------------------------------------------------------
253 latency: 97 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
255 | task: swapper-0 (uid:0 nice:0 policy:0 rt_prio:0)
257 => started at: apic_timer_interrupt
258 => ended at: do_softirq
261 # / _-----=> irqs-off
262 # | / _----=> need-resched
263 # || / _---=> hardirq/softirq
264 # ||| / _--=> preempt-depth
267 # cmd pid ||||| time | caller
269 <idle>-0 0d..1 0us+: trace_hardirqs_off_thunk (apic_timer_interrupt)
270 <idle>-0 0d.s. 97us : __do_softirq (do_softirq)
271 <idle>-0 0d.s1 98us : trace_hardirqs_on (do_softirq)
275 This shows that the current tracer is "irqsoff" tracing the time for which
276 interrupts were disabled. It gives the trace version and the version
277 of the kernel upon which this was executed on (2.6.26-rc8). Then it displays
278 the max latency in microsecs (97 us). The number of trace entries displayed
279 and the total number recorded (both are three: #3/3). The type of
280 preemption that was used (PREEMPT). VP, KP, SP, and HP are always zero
281 and are reserved for later use. #P is the number of online CPUS (#P:2).
283 The task is the process that was running when the latency occurred.
286 The start and stop (the functions in which the interrupts were disabled and
287 enabled respectively) that caused the latencies:
289 apic_timer_interrupt is where the interrupts were disabled.
290 do_softirq is where they were enabled again.
292 The next lines after the header are the trace itself. The header
293 explains which is which.
295 cmd: The name of the process in the trace.
297 pid: The PID of that process.
299 CPU#: The CPU which the process was running on.
301 irqs-off: 'd' interrupts are disabled. '.' otherwise.
302 Note: If the architecture does not support a way to
303 read the irq flags variable, an 'X' will always
306 need-resched: 'N' task need_resched is set, '.' otherwise.
309 'H' - hard irq occurred inside a softirq.
310 'h' - hard irq is running
311 's' - soft irq is running
312 '.' - normal context.
314 preempt-depth: The level of preempt_disabled
316 The above is mostly meaningful for kernel developers.
318 time: This differs from the trace file output. The trace file output
319 includes an absolute timestamp. The timestamp used by the
320 latency_trace file is relative to the start of the trace.
322 delay: This is just to help catch your eye a bit better. And
323 needs to be fixed to be only relative to the same CPU.
324 The marks are determined by the difference between this
325 current trace and the next trace.
326 '!' - greater than preempt_mark_thresh (default 100)
327 '+' - greater than 1 microsecond
328 ' ' - less than or equal to 1 microsecond.
330 The rest is the same as the 'trace' file.
336 The trace_options file is used to control what gets printed in the trace
337 output. To see what is available, simply cat the file:
339 cat /debug/tracing/trace_options
340 print-parent nosym-offset nosym-addr noverbose noraw nohex nobin \
341 noblock nostacktrace nosched-tree nouserstacktrace nosym-userobj
343 To disable one of the options, echo in the option prepended with "no".
345 echo noprint-parent > /debug/tracing/trace_options
347 To enable an option, leave off the "no".
349 echo sym-offset > /debug/tracing/trace_options
351 Here are the available options:
353 print-parent - On function traces, display the calling function
354 as well as the function being traced.
357 bash-4000 [01] 1477.606694: simple_strtoul <-strict_strtoul
360 bash-4000 [01] 1477.606694: simple_strtoul
363 sym-offset - Display not only the function name, but also the offset
364 in the function. For example, instead of seeing just
365 "ktime_get", you will see "ktime_get+0xb/0x20".
368 bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0
370 sym-addr - this will also display the function address as well as
374 bash-4000 [01] 1477.606694: simple_strtoul <c0339346>
376 verbose - This deals with the latency_trace file.
378 bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \
379 (+0.000ms): simple_strtoul (strict_strtoul)
381 raw - This will display raw numbers. This option is best for use with
382 user applications that can translate the raw numbers better than
383 having it done in the kernel.
385 hex - Similar to raw, but the numbers will be in a hexadecimal format.
387 bin - This will print out the formats in raw binary.
389 block - TBD (needs update)
391 stacktrace - This is one of the options that changes the trace itself.
392 When a trace is recorded, so is the stack of functions.
393 This allows for back traces of trace sites.
395 userstacktrace - This option changes the trace.
396 It records a stacktrace of the current userspace thread.
398 sym-userobj - when user stacktrace are enabled, look up which object the
399 address belongs to, and print a relative address
400 This is especially useful when ASLR is on, otherwise you don't
401 get a chance to resolve the address to object/file/line after the app is no
404 The lookup is performed when you read trace,trace_pipe,latency_trace. Example:
406 a.out-1623 [000] 40874.465068: /root/a.out[+0x480] <-/root/a.out[+0
407 x494] <- /root/a.out[+0x4a8] <- /lib/libc-2.7.so[+0x1e1a6]
409 sched-tree - TBD (any users??)
415 This tracer simply records schedule switches. Here is an example
418 # echo sched_switch > /debug/tracing/current_tracer
419 # echo 1 > /debug/tracing/tracing_enabled
421 # echo 0 > /debug/tracing/tracing_enabled
422 # cat /debug/tracing/trace
424 # tracer: sched_switch
426 # TASK-PID CPU# TIMESTAMP FUNCTION
428 bash-3997 [01] 240.132281: 3997:120:R + 4055:120:R
429 bash-3997 [01] 240.132284: 3997:120:R ==> 4055:120:R
430 sleep-4055 [01] 240.132371: 4055:120:S ==> 3997:120:R
431 bash-3997 [01] 240.132454: 3997:120:R + 4055:120:S
432 bash-3997 [01] 240.132457: 3997:120:R ==> 4055:120:R
433 sleep-4055 [01] 240.132460: 4055:120:D ==> 3997:120:R
434 bash-3997 [01] 240.132463: 3997:120:R + 4055:120:D
435 bash-3997 [01] 240.132465: 3997:120:R ==> 4055:120:R
436 <idle>-0 [00] 240.132589: 0:140:R + 4:115:S
437 <idle>-0 [00] 240.132591: 0:140:R ==> 4:115:R
438 ksoftirqd/0-4 [00] 240.132595: 4:115:S ==> 0:140:R
439 <idle>-0 [00] 240.132598: 0:140:R + 4:115:S
440 <idle>-0 [00] 240.132599: 0:140:R ==> 4:115:R
441 ksoftirqd/0-4 [00] 240.132603: 4:115:S ==> 0:140:R
442 sleep-4055 [01] 240.133058: 4055:120:S ==> 3997:120:R
446 As we have discussed previously about this format, the header shows
447 the name of the trace and points to the options. The "FUNCTION"
448 is a misnomer since here it represents the wake ups and context
451 The sched_switch file only lists the wake ups (represented with '+')
452 and context switches ('==>') with the previous task or current task
453 first followed by the next task or task waking up. The format for both
454 of these is PID:KERNEL-PRIO:TASK-STATE. Remember that the KERNEL-PRIO
455 is the inverse of the actual priority with zero (0) being the highest
456 priority and the nice values starting at 100 (nice -20). Below is
457 a quick chart to map the kernel priority to user land priorities.
459 Kernel priority: 0 to 99 ==> user RT priority 99 to 0
460 Kernel priority: 100 to 139 ==> user nice -20 to 19
461 Kernel priority: 140 ==> idle task priority
465 R - running : wants to run, may not actually be running
466 S - sleep : process is waiting to be woken up (handles signals)
467 D - disk sleep (uninterruptible sleep) : process must be woken up
469 T - stopped : process suspended
470 t - traced : process is being traced (with something like gdb)
471 Z - zombie : process waiting to be cleaned up
478 The following tracers (listed below) give different output depending
479 on whether or not the sysctl ftrace_enabled is set. To set ftrace_enabled,
480 one can either use the sysctl function or set it via the proc
481 file system interface.
483 sysctl kernel.ftrace_enabled=1
487 echo 1 > /proc/sys/kernel/ftrace_enabled
489 To disable ftrace_enabled simply replace the '1' with '0' in
492 When ftrace_enabled is set the tracers will also record the functions
493 that are within the trace. The descriptions of the tracers
494 will also show an example with ftrace enabled.
500 When interrupts are disabled, the CPU can not react to any other
501 external event (besides NMIs and SMIs). This prevents the timer
502 interrupt from triggering or the mouse interrupt from letting the
503 kernel know of a new mouse event. The result is a latency with the
506 The irqsoff tracer tracks the time for which interrupts are disabled.
507 When a new maximum latency is hit, the tracer saves the trace leading up
508 to that latency point so that every time a new maximum is reached, the old
509 saved trace is discarded and the new trace is saved.
511 To reset the maximum, echo 0 into tracing_max_latency. Here is an
514 # echo irqsoff > /debug/tracing/current_tracer
515 # echo 0 > /debug/tracing/tracing_max_latency
516 # echo 1 > /debug/tracing/tracing_enabled
519 # echo 0 > /debug/tracing/tracing_enabled
520 # cat /debug/tracing/latency_trace
523 irqsoff latency trace v1.1.5 on 2.6.26
524 --------------------------------------------------------------------
525 latency: 12 us, #3/3, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
527 | task: bash-3730 (uid:0 nice:0 policy:0 rt_prio:0)
529 => started at: sys_setpgid
530 => ended at: sys_setpgid
533 # / _-----=> irqs-off
534 # | / _----=> need-resched
535 # || / _---=> hardirq/softirq
536 # ||| / _--=> preempt-depth
539 # cmd pid ||||| time | caller
541 bash-3730 1d... 0us : _write_lock_irq (sys_setpgid)
542 bash-3730 1d..1 1us+: _write_unlock_irq (sys_setpgid)
543 bash-3730 1d..2 14us : trace_hardirqs_on (sys_setpgid)
546 Here we see that that we had a latency of 12 microsecs (which is
547 very good). The _write_lock_irq in sys_setpgid disabled interrupts.
548 The difference between the 12 and the displayed timestamp 14us occurred
549 because the clock was incremented between the time of recording the max
550 latency and the time of recording the function that had that latency.
552 Note the above example had ftrace_enabled not set. If we set the
553 ftrace_enabled, we get a much larger output:
557 irqsoff latency trace v1.1.5 on 2.6.26-rc8
558 --------------------------------------------------------------------
559 latency: 50 us, #101/101, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
561 | task: ls-4339 (uid:0 nice:0 policy:0 rt_prio:0)
563 => started at: __alloc_pages_internal
564 => ended at: __alloc_pages_internal
567 # / _-----=> irqs-off
568 # | / _----=> need-resched
569 # || / _---=> hardirq/softirq
570 # ||| / _--=> preempt-depth
573 # cmd pid ||||| time | caller
575 ls-4339 0...1 0us+: get_page_from_freelist (__alloc_pages_internal)
576 ls-4339 0d..1 3us : rmqueue_bulk (get_page_from_freelist)
577 ls-4339 0d..1 3us : _spin_lock (rmqueue_bulk)
578 ls-4339 0d..1 4us : add_preempt_count (_spin_lock)
579 ls-4339 0d..2 4us : __rmqueue (rmqueue_bulk)
580 ls-4339 0d..2 5us : __rmqueue_smallest (__rmqueue)
581 ls-4339 0d..2 5us : __mod_zone_page_state (__rmqueue_smallest)
582 ls-4339 0d..2 6us : __rmqueue (rmqueue_bulk)
583 ls-4339 0d..2 6us : __rmqueue_smallest (__rmqueue)
584 ls-4339 0d..2 7us : __mod_zone_page_state (__rmqueue_smallest)
585 ls-4339 0d..2 7us : __rmqueue (rmqueue_bulk)
586 ls-4339 0d..2 8us : __rmqueue_smallest (__rmqueue)
588 ls-4339 0d..2 46us : __rmqueue_smallest (__rmqueue)
589 ls-4339 0d..2 47us : __mod_zone_page_state (__rmqueue_smallest)
590 ls-4339 0d..2 47us : __rmqueue (rmqueue_bulk)
591 ls-4339 0d..2 48us : __rmqueue_smallest (__rmqueue)
592 ls-4339 0d..2 48us : __mod_zone_page_state (__rmqueue_smallest)
593 ls-4339 0d..2 49us : _spin_unlock (rmqueue_bulk)
594 ls-4339 0d..2 49us : sub_preempt_count (_spin_unlock)
595 ls-4339 0d..1 50us : get_page_from_freelist (__alloc_pages_internal)
596 ls-4339 0d..2 51us : trace_hardirqs_on (__alloc_pages_internal)
600 Here we traced a 50 microsecond latency. But we also see all the
601 functions that were called during that time. Note that by enabling
602 function tracing, we incur an added overhead. This overhead may
603 extend the latency times. But nevertheless, this trace has provided
604 some very helpful debugging information.
610 When preemption is disabled, we may be able to receive interrupts but
611 the task cannot be preempted and a higher priority task must wait
612 for preemption to be enabled again before it can preempt a lower
615 The preemptoff tracer traces the places that disable preemption.
616 Like the irqsoff tracer, it records the maximum latency for which preemption
617 was disabled. The control of preemptoff tracer is much like the irqsoff
620 # echo preemptoff > /debug/tracing/current_tracer
621 # echo 0 > /debug/tracing/tracing_max_latency
622 # echo 1 > /debug/tracing/tracing_enabled
625 # echo 0 > /debug/tracing/tracing_enabled
626 # cat /debug/tracing/latency_trace
629 preemptoff latency trace v1.1.5 on 2.6.26-rc8
630 --------------------------------------------------------------------
631 latency: 29 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
633 | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
635 => started at: do_IRQ
636 => ended at: __do_softirq
639 # / _-----=> irqs-off
640 # | / _----=> need-resched
641 # || / _---=> hardirq/softirq
642 # ||| / _--=> preempt-depth
645 # cmd pid ||||| time | caller
647 sshd-4261 0d.h. 0us+: irq_enter (do_IRQ)
648 sshd-4261 0d.s. 29us : _local_bh_enable (__do_softirq)
649 sshd-4261 0d.s1 30us : trace_preempt_on (__do_softirq)
652 This has some more changes. Preemption was disabled when an interrupt
653 came in (notice the 'h'), and was enabled while doing a softirq.
654 (notice the 's'). But we also see that interrupts have been disabled
655 when entering the preempt off section and leaving it (the 'd').
656 We do not know if interrupts were enabled in the mean time.
660 preemptoff latency trace v1.1.5 on 2.6.26-rc8
661 --------------------------------------------------------------------
662 latency: 63 us, #87/87, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
664 | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
666 => started at: remove_wait_queue
667 => ended at: __do_softirq
670 # / _-----=> irqs-off
671 # | / _----=> need-resched
672 # || / _---=> hardirq/softirq
673 # ||| / _--=> preempt-depth
676 # cmd pid ||||| time | caller
678 sshd-4261 0d..1 0us : _spin_lock_irqsave (remove_wait_queue)
679 sshd-4261 0d..1 1us : _spin_unlock_irqrestore (remove_wait_queue)
680 sshd-4261 0d..1 2us : do_IRQ (common_interrupt)
681 sshd-4261 0d..1 2us : irq_enter (do_IRQ)
682 sshd-4261 0d..1 2us : idle_cpu (irq_enter)
683 sshd-4261 0d..1 3us : add_preempt_count (irq_enter)
684 sshd-4261 0d.h1 3us : idle_cpu (irq_enter)
685 sshd-4261 0d.h. 4us : handle_fasteoi_irq (do_IRQ)
687 sshd-4261 0d.h. 12us : add_preempt_count (_spin_lock)
688 sshd-4261 0d.h1 12us : ack_ioapic_quirk_irq (handle_fasteoi_irq)
689 sshd-4261 0d.h1 13us : move_native_irq (ack_ioapic_quirk_irq)
690 sshd-4261 0d.h1 13us : _spin_unlock (handle_fasteoi_irq)
691 sshd-4261 0d.h1 14us : sub_preempt_count (_spin_unlock)
692 sshd-4261 0d.h1 14us : irq_exit (do_IRQ)
693 sshd-4261 0d.h1 15us : sub_preempt_count (irq_exit)
694 sshd-4261 0d..2 15us : do_softirq (irq_exit)
695 sshd-4261 0d... 15us : __do_softirq (do_softirq)
696 sshd-4261 0d... 16us : __local_bh_disable (__do_softirq)
697 sshd-4261 0d... 16us+: add_preempt_count (__local_bh_disable)
698 sshd-4261 0d.s4 20us : add_preempt_count (__local_bh_disable)
699 sshd-4261 0d.s4 21us : sub_preempt_count (local_bh_enable)
700 sshd-4261 0d.s5 21us : sub_preempt_count (local_bh_enable)
702 sshd-4261 0d.s6 41us : add_preempt_count (__local_bh_disable)
703 sshd-4261 0d.s6 42us : sub_preempt_count (local_bh_enable)
704 sshd-4261 0d.s7 42us : sub_preempt_count (local_bh_enable)
705 sshd-4261 0d.s5 43us : add_preempt_count (__local_bh_disable)
706 sshd-4261 0d.s5 43us : sub_preempt_count (local_bh_enable_ip)
707 sshd-4261 0d.s6 44us : sub_preempt_count (local_bh_enable_ip)
708 sshd-4261 0d.s5 44us : add_preempt_count (__local_bh_disable)
709 sshd-4261 0d.s5 45us : sub_preempt_count (local_bh_enable)
711 sshd-4261 0d.s. 63us : _local_bh_enable (__do_softirq)
712 sshd-4261 0d.s1 64us : trace_preempt_on (__do_softirq)
715 The above is an example of the preemptoff trace with ftrace_enabled
716 set. Here we see that interrupts were disabled the entire time.
717 The irq_enter code lets us know that we entered an interrupt 'h'.
718 Before that, the functions being traced still show that it is not
719 in an interrupt, but we can see from the functions themselves that
720 this is not the case.
722 Notice that __do_softirq when called does not have a preempt_count.
723 It may seem that we missed a preempt enabling. What really happened
724 is that the preempt count is held on the thread's stack and we
725 switched to the softirq stack (4K stacks in effect). The code
726 does not copy the preempt count, but because interrupts are disabled,
727 we do not need to worry about it. Having a tracer like this is good
728 for letting people know what really happens inside the kernel.
734 Knowing the locations that have interrupts disabled or preemption
735 disabled for the longest times is helpful. But sometimes we would
736 like to know when either preemption and/or interrupts are disabled.
738 Consider the following code:
741 call_function_with_irqs_off();
743 call_function_with_irqs_and_preemption_off();
745 call_function_with_preemption_off();
748 The irqsoff tracer will record the total length of
749 call_function_with_irqs_off() and
750 call_function_with_irqs_and_preemption_off().
752 The preemptoff tracer will record the total length of
753 call_function_with_irqs_and_preemption_off() and
754 call_function_with_preemption_off().
756 But neither will trace the time that interrupts and/or preemption
757 is disabled. This total time is the time that we can not schedule.
758 To record this time, use the preemptirqsoff tracer.
760 Again, using this trace is much like the irqsoff and preemptoff tracers.
762 # echo preemptirqsoff > /debug/tracing/current_tracer
763 # echo 0 > /debug/tracing/tracing_max_latency
764 # echo 1 > /debug/tracing/tracing_enabled
767 # echo 0 > /debug/tracing/tracing_enabled
768 # cat /debug/tracing/latency_trace
769 # tracer: preemptirqsoff
771 preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8
772 --------------------------------------------------------------------
773 latency: 293 us, #3/3, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
775 | task: ls-4860 (uid:0 nice:0 policy:0 rt_prio:0)
777 => started at: apic_timer_interrupt
778 => ended at: __do_softirq
781 # / _-----=> irqs-off
782 # | / _----=> need-resched
783 # || / _---=> hardirq/softirq
784 # ||| / _--=> preempt-depth
787 # cmd pid ||||| time | caller
789 ls-4860 0d... 0us!: trace_hardirqs_off_thunk (apic_timer_interrupt)
790 ls-4860 0d.s. 294us : _local_bh_enable (__do_softirq)
791 ls-4860 0d.s1 294us : trace_preempt_on (__do_softirq)
795 The trace_hardirqs_off_thunk is called from assembly on x86 when
796 interrupts are disabled in the assembly code. Without the function
797 tracing, we do not know if interrupts were enabled within the preemption
798 points. We do see that it started with preemption enabled.
800 Here is a trace with ftrace_enabled set:
803 # tracer: preemptirqsoff
805 preemptirqsoff latency trace v1.1.5 on 2.6.26-rc8
806 --------------------------------------------------------------------
807 latency: 105 us, #183/183, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
809 | task: sshd-4261 (uid:0 nice:0 policy:0 rt_prio:0)
811 => started at: write_chan
812 => ended at: __do_softirq
815 # / _-----=> irqs-off
816 # | / _----=> need-resched
817 # || / _---=> hardirq/softirq
818 # ||| / _--=> preempt-depth
821 # cmd pid ||||| time | caller
823 ls-4473 0.N.. 0us : preempt_schedule (write_chan)
824 ls-4473 0dN.1 1us : _spin_lock (schedule)
825 ls-4473 0dN.1 2us : add_preempt_count (_spin_lock)
826 ls-4473 0d..2 2us : put_prev_task_fair (schedule)
828 ls-4473 0d..2 13us : set_normalized_timespec (ktime_get_ts)
829 ls-4473 0d..2 13us : __switch_to (schedule)
830 sshd-4261 0d..2 14us : finish_task_switch (schedule)
831 sshd-4261 0d..2 14us : _spin_unlock_irq (finish_task_switch)
832 sshd-4261 0d..1 15us : add_preempt_count (_spin_lock_irqsave)
833 sshd-4261 0d..2 16us : _spin_unlock_irqrestore (hrtick_set)
834 sshd-4261 0d..2 16us : do_IRQ (common_interrupt)
835 sshd-4261 0d..2 17us : irq_enter (do_IRQ)
836 sshd-4261 0d..2 17us : idle_cpu (irq_enter)
837 sshd-4261 0d..2 18us : add_preempt_count (irq_enter)
838 sshd-4261 0d.h2 18us : idle_cpu (irq_enter)
839 sshd-4261 0d.h. 18us : handle_fasteoi_irq (do_IRQ)
840 sshd-4261 0d.h. 19us : _spin_lock (handle_fasteoi_irq)
841 sshd-4261 0d.h. 19us : add_preempt_count (_spin_lock)
842 sshd-4261 0d.h1 20us : _spin_unlock (handle_fasteoi_irq)
843 sshd-4261 0d.h1 20us : sub_preempt_count (_spin_unlock)
845 sshd-4261 0d.h1 28us : _spin_unlock (handle_fasteoi_irq)
846 sshd-4261 0d.h1 29us : sub_preempt_count (_spin_unlock)
847 sshd-4261 0d.h2 29us : irq_exit (do_IRQ)
848 sshd-4261 0d.h2 29us : sub_preempt_count (irq_exit)
849 sshd-4261 0d..3 30us : do_softirq (irq_exit)
850 sshd-4261 0d... 30us : __do_softirq (do_softirq)
851 sshd-4261 0d... 31us : __local_bh_disable (__do_softirq)
852 sshd-4261 0d... 31us+: add_preempt_count (__local_bh_disable)
853 sshd-4261 0d.s4 34us : add_preempt_count (__local_bh_disable)
855 sshd-4261 0d.s3 43us : sub_preempt_count (local_bh_enable_ip)
856 sshd-4261 0d.s4 44us : sub_preempt_count (local_bh_enable_ip)
857 sshd-4261 0d.s3 44us : smp_apic_timer_interrupt (apic_timer_interrupt)
858 sshd-4261 0d.s3 45us : irq_enter (smp_apic_timer_interrupt)
859 sshd-4261 0d.s3 45us : idle_cpu (irq_enter)
860 sshd-4261 0d.s3 46us : add_preempt_count (irq_enter)
861 sshd-4261 0d.H3 46us : idle_cpu (irq_enter)
862 sshd-4261 0d.H3 47us : hrtimer_interrupt (smp_apic_timer_interrupt)
863 sshd-4261 0d.H3 47us : ktime_get (hrtimer_interrupt)
865 sshd-4261 0d.H3 81us : tick_program_event (hrtimer_interrupt)
866 sshd-4261 0d.H3 82us : ktime_get (tick_program_event)
867 sshd-4261 0d.H3 82us : ktime_get_ts (ktime_get)
868 sshd-4261 0d.H3 83us : getnstimeofday (ktime_get_ts)
869 sshd-4261 0d.H3 83us : set_normalized_timespec (ktime_get_ts)
870 sshd-4261 0d.H3 84us : clockevents_program_event (tick_program_event)
871 sshd-4261 0d.H3 84us : lapic_next_event (clockevents_program_event)
872 sshd-4261 0d.H3 85us : irq_exit (smp_apic_timer_interrupt)
873 sshd-4261 0d.H3 85us : sub_preempt_count (irq_exit)
874 sshd-4261 0d.s4 86us : sub_preempt_count (irq_exit)
875 sshd-4261 0d.s3 86us : add_preempt_count (__local_bh_disable)
877 sshd-4261 0d.s1 98us : sub_preempt_count (net_rx_action)
878 sshd-4261 0d.s. 99us : add_preempt_count (_spin_lock_irq)
879 sshd-4261 0d.s1 99us+: _spin_unlock_irq (run_timer_softirq)
880 sshd-4261 0d.s. 104us : _local_bh_enable (__do_softirq)
881 sshd-4261 0d.s. 104us : sub_preempt_count (_local_bh_enable)
882 sshd-4261 0d.s. 105us : _local_bh_enable (__do_softirq)
883 sshd-4261 0d.s1 105us : trace_preempt_on (__do_softirq)
886 This is a very interesting trace. It started with the preemption of
887 the ls task. We see that the task had the "need_resched" bit set
888 via the 'N' in the trace. Interrupts were disabled before the spin_lock
889 at the beginning of the trace. We see that a schedule took place to run
890 sshd. When the interrupts were enabled, we took an interrupt.
891 On return from the interrupt handler, the softirq ran. We took another
892 interrupt while running the softirq as we see from the capital 'H'.
898 In a Real-Time environment it is very important to know the wakeup
899 time it takes for the highest priority task that is woken up to the
900 time that it executes. This is also known as "schedule latency".
901 I stress the point that this is about RT tasks. It is also important
902 to know the scheduling latency of non-RT tasks, but the average
903 schedule latency is better for non-RT tasks. Tools like
904 LatencyTop are more appropriate for such measurements.
906 Real-Time environments are interested in the worst case latency.
907 That is the longest latency it takes for something to happen, and
908 not the average. We can have a very fast scheduler that may only
909 have a large latency once in a while, but that would not work well
910 with Real-Time tasks. The wakeup tracer was designed to record
911 the worst case wakeups of RT tasks. Non-RT tasks are not recorded
912 because the tracer only records one worst case and tracing non-RT
913 tasks that are unpredictable will overwrite the worst case latency
916 Since this tracer only deals with RT tasks, we will run this slightly
917 differently than we did with the previous tracers. Instead of performing
918 an 'ls', we will run 'sleep 1' under 'chrt' which changes the
919 priority of the task.
921 # echo wakeup > /debug/tracing/current_tracer
922 # echo 0 > /debug/tracing/tracing_max_latency
923 # echo 1 > /debug/tracing/tracing_enabled
925 # echo 0 > /debug/tracing/tracing_enabled
926 # cat /debug/tracing/latency_trace
929 wakeup latency trace v1.1.5 on 2.6.26-rc8
930 --------------------------------------------------------------------
931 latency: 4 us, #2/2, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
933 | task: sleep-4901 (uid:0 nice:0 policy:1 rt_prio:5)
937 # / _-----=> irqs-off
938 # | / _----=> need-resched
939 # || / _---=> hardirq/softirq
940 # ||| / _--=> preempt-depth
943 # cmd pid ||||| time | caller
945 <idle>-0 1d.h4 0us+: try_to_wake_up (wake_up_process)
946 <idle>-0 1d..4 4us : schedule (cpu_idle)
950 Running this on an idle system, we see that it only took 4 microseconds
951 to perform the task switch. Note, since the trace marker in the
952 schedule is before the actual "switch", we stop the tracing when
953 the recorded task is about to schedule in. This may change if
954 we add a new marker at the end of the scheduler.
956 Notice that the recorded task is 'sleep' with the PID of 4901 and it
957 has an rt_prio of 5. This priority is user-space priority and not
958 the internal kernel priority. The policy is 1 for SCHED_FIFO and 2
961 Doing the same with chrt -r 5 and ftrace_enabled set.
965 wakeup latency trace v1.1.5 on 2.6.26-rc8
966 --------------------------------------------------------------------
967 latency: 50 us, #60/60, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:2)
969 | task: sleep-4068 (uid:0 nice:0 policy:2 rt_prio:5)
973 # / _-----=> irqs-off
974 # | / _----=> need-resched
975 # || / _---=> hardirq/softirq
976 # ||| / _--=> preempt-depth
979 # cmd pid ||||| time | caller
981 ksoftirq-7 1d.H3 0us : try_to_wake_up (wake_up_process)
982 ksoftirq-7 1d.H4 1us : sub_preempt_count (marker_probe_cb)
983 ksoftirq-7 1d.H3 2us : check_preempt_wakeup (try_to_wake_up)
984 ksoftirq-7 1d.H3 3us : update_curr (check_preempt_wakeup)
985 ksoftirq-7 1d.H3 4us : calc_delta_mine (update_curr)
986 ksoftirq-7 1d.H3 5us : __resched_task (check_preempt_wakeup)
987 ksoftirq-7 1d.H3 6us : task_wake_up_rt (try_to_wake_up)
988 ksoftirq-7 1d.H3 7us : _spin_unlock_irqrestore (try_to_wake_up)
990 ksoftirq-7 1d.H2 17us : irq_exit (smp_apic_timer_interrupt)
991 ksoftirq-7 1d.H2 18us : sub_preempt_count (irq_exit)
992 ksoftirq-7 1d.s3 19us : sub_preempt_count (irq_exit)
993 ksoftirq-7 1..s2 20us : rcu_process_callbacks (__do_softirq)
995 ksoftirq-7 1..s2 26us : __rcu_process_callbacks (rcu_process_callbacks)
996 ksoftirq-7 1d.s2 27us : _local_bh_enable (__do_softirq)
997 ksoftirq-7 1d.s2 28us : sub_preempt_count (_local_bh_enable)
998 ksoftirq-7 1.N.3 29us : sub_preempt_count (ksoftirqd)
999 ksoftirq-7 1.N.2 30us : _cond_resched (ksoftirqd)
1000 ksoftirq-7 1.N.2 31us : __cond_resched (_cond_resched)
1001 ksoftirq-7 1.N.2 32us : add_preempt_count (__cond_resched)
1002 ksoftirq-7 1.N.2 33us : schedule (__cond_resched)
1003 ksoftirq-7 1.N.2 33us : add_preempt_count (schedule)
1004 ksoftirq-7 1.N.3 34us : hrtick_clear (schedule)
1005 ksoftirq-7 1dN.3 35us : _spin_lock (schedule)
1006 ksoftirq-7 1dN.3 36us : add_preempt_count (_spin_lock)
1007 ksoftirq-7 1d..4 37us : put_prev_task_fair (schedule)
1008 ksoftirq-7 1d..4 38us : update_curr (put_prev_task_fair)
1010 ksoftirq-7 1d..5 47us : _spin_trylock (tracing_record_cmdline)
1011 ksoftirq-7 1d..5 48us : add_preempt_count (_spin_trylock)
1012 ksoftirq-7 1d..6 49us : _spin_unlock (tracing_record_cmdline)
1013 ksoftirq-7 1d..6 49us : sub_preempt_count (_spin_unlock)
1014 ksoftirq-7 1d..4 50us : schedule (__cond_resched)
1016 The interrupt went off while running ksoftirqd. This task runs at
1017 SCHED_OTHER. Why did not we see the 'N' set early? This may be
1018 a harmless bug with x86_32 and 4K stacks. On x86_32 with 4K stacks
1019 configured, the interrupt and softirq run with their own stack.
1020 Some information is held on the top of the task's stack (need_resched
1021 and preempt_count are both stored there). The setting of the NEED_RESCHED
1022 bit is done directly to the task's stack, but the reading of the
1023 NEED_RESCHED is done by looking at the current stack, which in this case
1024 is the stack for the hard interrupt. This hides the fact that NEED_RESCHED
1025 has been set. We do not see the 'N' until we switch back to the task's
1031 This tracer is the function tracer. Enabling the function tracer
1032 can be done from the debug file system. Make sure the ftrace_enabled is
1033 set; otherwise this tracer is a nop.
1035 # sysctl kernel.ftrace_enabled=1
1036 # echo function > /debug/tracing/current_tracer
1037 # echo 1 > /debug/tracing/tracing_enabled
1039 # echo 0 > /debug/tracing/tracing_enabled
1040 # cat /debug/tracing/trace
1043 # TASK-PID CPU# TIMESTAMP FUNCTION
1045 bash-4003 [00] 123.638713: finish_task_switch <-schedule
1046 bash-4003 [00] 123.638714: _spin_unlock_irq <-finish_task_switch
1047 bash-4003 [00] 123.638714: sub_preempt_count <-_spin_unlock_irq
1048 bash-4003 [00] 123.638715: hrtick_set <-schedule
1049 bash-4003 [00] 123.638715: _spin_lock_irqsave <-hrtick_set
1050 bash-4003 [00] 123.638716: add_preempt_count <-_spin_lock_irqsave
1051 bash-4003 [00] 123.638716: _spin_unlock_irqrestore <-hrtick_set
1052 bash-4003 [00] 123.638717: sub_preempt_count <-_spin_unlock_irqrestore
1053 bash-4003 [00] 123.638717: hrtick_clear <-hrtick_set
1054 bash-4003 [00] 123.638718: sub_preempt_count <-schedule
1055 bash-4003 [00] 123.638718: sub_preempt_count <-preempt_schedule
1056 bash-4003 [00] 123.638719: wait_for_completion <-__stop_machine_run
1057 bash-4003 [00] 123.638719: wait_for_common <-wait_for_completion
1058 bash-4003 [00] 123.638720: _spin_lock_irq <-wait_for_common
1059 bash-4003 [00] 123.638720: add_preempt_count <-_spin_lock_irq
1063 Note: function tracer uses ring buffers to store the above entries.
1064 The newest data may overwrite the oldest data. Sometimes using echo to
1065 stop the trace is not sufficient because the tracing could have overwritten
1066 the data that you wanted to record. For this reason, it is sometimes better to
1067 disable tracing directly from a program. This allows you to stop the
1068 tracing at the point that you hit the part that you are interested in.
1069 To disable the tracing directly from a C program, something like following
1070 code snippet can be used:
1074 int main(int argc, char *argv[]) {
1076 trace_fd = open("/debug/tracing/tracing_enabled", O_WRONLY);
1078 if (condition_hit()) {
1079 write(trace_fd, "0", 1);
1084 Note: Here we hard coded the path name. The debugfs mount is not
1085 guaranteed to be at /debug (and is more commonly at /sys/kernel/debug).
1086 For simple one time traces, the above is sufficent. For anything else,
1087 a search through /proc/mounts may be needed to find where the debugfs
1088 file-system is mounted.
1091 Single thread tracing
1092 ---------------------
1094 By writing into /debug/tracing/set_ftrace_pid you can trace a
1095 single thread. For example:
1097 # cat /debug/tracing/set_ftrace_pid
1099 # echo 3111 > /debug/tracing/set_ftrace_pid
1100 # cat /debug/tracing/set_ftrace_pid
1102 # echo function > /debug/tracing/current_tracer
1103 # cat /debug/tracing/trace | head
1106 # TASK-PID CPU# TIMESTAMP FUNCTION
1108 yum-updatesd-3111 [003] 1637.254676: finish_task_switch <-thread_return
1109 yum-updatesd-3111 [003] 1637.254681: hrtimer_cancel <-schedule_hrtimeout_range
1110 yum-updatesd-3111 [003] 1637.254682: hrtimer_try_to_cancel <-hrtimer_cancel
1111 yum-updatesd-3111 [003] 1637.254683: lock_hrtimer_base <-hrtimer_try_to_cancel
1112 yum-updatesd-3111 [003] 1637.254685: fget_light <-do_sys_poll
1113 yum-updatesd-3111 [003] 1637.254686: pipe_poll <-do_sys_poll
1114 # echo -1 > /debug/tracing/set_ftrace_pid
1115 # cat /debug/tracing/trace |head
1118 # TASK-PID CPU# TIMESTAMP FUNCTION
1120 ##### CPU 3 buffer started ####
1121 yum-updatesd-3111 [003] 1701.957688: free_poll_entry <-poll_freewait
1122 yum-updatesd-3111 [003] 1701.957689: remove_wait_queue <-free_poll_entry
1123 yum-updatesd-3111 [003] 1701.957691: fput <-free_poll_entry
1124 yum-updatesd-3111 [003] 1701.957692: audit_syscall_exit <-sysret_audit
1125 yum-updatesd-3111 [003] 1701.957693: path_put <-audit_syscall_exit
1127 If you want to trace a function when executing, you could use
1128 something like this simple program:
1132 #include <sys/types.h>
1133 #include <sys/stat.h>
1137 int main (int argc, char **argv)
1147 ffd = open("/debug/tracing/current_tracer", O_WRONLY);
1150 write(ffd, "nop", 3);
1152 fd = open("/debug/tracing/set_ftrace_pid", O_WRONLY);
1153 s = sprintf(line, "%d\n", getpid());
1156 write(ffd, "function", 8);
1161 execvp(argv[1], argv+1);
1168 hw-branch-tracer (x86 only)
1169 ---------------------------
1171 This tracer uses the x86 last branch tracing hardware feature to
1172 collect a branch trace on all cpus with relatively low overhead.
1174 The tracer uses a fixed-size circular buffer per cpu and only
1175 traces ring 0 branches. The trace file dumps that buffer in the
1178 # tracer: hw-branch-tracer
1181 0 scheduler_tick+0xb5/0x1bf <- task_tick_idle+0x5/0x6
1182 2 run_posix_cpu_timers+0x2b/0x72a <- run_posix_cpu_timers+0x25/0x72a
1183 0 scheduler_tick+0x139/0x1bf <- scheduler_tick+0xed/0x1bf
1184 0 scheduler_tick+0x17c/0x1bf <- scheduler_tick+0x148/0x1bf
1185 2 run_posix_cpu_timers+0x9e/0x72a <- run_posix_cpu_timers+0x5e/0x72a
1186 0 scheduler_tick+0x1b6/0x1bf <- scheduler_tick+0x1aa/0x1bf
1189 The tracer may be used to dump the trace for the oops'ing cpu on a
1190 kernel oops into the system log. To enable this, ftrace_dump_on_oops
1191 must be set. To set ftrace_dump_on_oops, one can either use the sysctl
1192 function or set it via the proc system interface.
1194 sysctl kernel.ftrace_dump_on_oops=1
1198 echo 1 > /proc/sys/kernel/ftrace_dump_on_oops
1201 Here's an example of such a dump after a null pointer dereference in a
1204 [57848.105921] BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
1205 [57848.106019] IP: [<ffffffffa0000006>] open+0x6/0x14 [oops]
1206 [57848.106019] PGD 2354e9067 PUD 2375e7067 PMD 0
1207 [57848.106019] Oops: 0002 [#1] SMP
1208 [57848.106019] last sysfs file: /sys/devices/pci0000:00/0000:00:1e.0/0000:20:05.0/local_cpus
1209 [57848.106019] Dumping ftrace buffer:
1210 [57848.106019] ---------------------------------
1212 [57848.106019] 0 chrdev_open+0xe6/0x165 <- cdev_put+0x23/0x24
1213 [57848.106019] 0 chrdev_open+0x117/0x165 <- chrdev_open+0xfa/0x165
1214 [57848.106019] 0 chrdev_open+0x120/0x165 <- chrdev_open+0x11c/0x165
1215 [57848.106019] 0 chrdev_open+0x134/0x165 <- chrdev_open+0x12b/0x165
1216 [57848.106019] 0 open+0x0/0x14 [oops] <- chrdev_open+0x144/0x165
1217 [57848.106019] 0 page_fault+0x0/0x30 <- open+0x6/0x14 [oops]
1218 [57848.106019] 0 error_entry+0x0/0x5b <- page_fault+0x4/0x30
1219 [57848.106019] 0 error_kernelspace+0x0/0x31 <- error_entry+0x59/0x5b
1220 [57848.106019] 0 error_sti+0x0/0x1 <- error_kernelspace+0x2d/0x31
1221 [57848.106019] 0 page_fault+0x9/0x30 <- error_sti+0x0/0x1
1222 [57848.106019] 0 do_page_fault+0x0/0x881 <- page_fault+0x1a/0x30
1224 [57848.106019] 0 do_page_fault+0x66b/0x881 <- is_prefetch+0x1ee/0x1f2
1225 [57848.106019] 0 do_page_fault+0x6e0/0x881 <- do_page_fault+0x67a/0x881
1226 [57848.106019] 0 oops_begin+0x0/0x96 <- do_page_fault+0x6e0/0x881
1227 [57848.106019] 0 trace_hw_branch_oops+0x0/0x2d <- oops_begin+0x9/0x96
1229 [57848.106019] 0 ds_suspend_bts+0x2a/0xe3 <- ds_suspend_bts+0x1a/0xe3
1230 [57848.106019] ---------------------------------
1231 [57848.106019] CPU 0
1232 [57848.106019] Modules linked in: oops
1233 [57848.106019] Pid: 5542, comm: cat Tainted: G W 2.6.28 #23
1234 [57848.106019] RIP: 0010:[<ffffffffa0000006>] [<ffffffffa0000006>] open+0x6/0x14 [oops]
1235 [57848.106019] RSP: 0018:ffff880235457d48 EFLAGS: 00010246
1239 function graph tracer
1240 ---------------------------
1242 This tracer is similar to the function tracer except that it probes
1243 a function on its entry and its exit.
1244 This is done by setting a dynamically allocated stack of return addresses on each
1245 task_struct. Then the tracer overwrites the return address of each function traced
1246 to set a custom probe. Thus the original return address is stored on the stack of return
1247 address in the task_struct.
1249 Probing on both extremities of a function leads to special features such as
1251 _ measure of function's time execution
1252 _ having a reliable call stack to draw function calls graph
1254 This tracer is useful in several situations:
1256 _ you want to find the reason of a strange kernel behavior and need to see
1257 what happens in detail on any areas (or specific ones).
1258 _ you are experiencing weird latencies but it's difficult to find its origin.
1259 _ you want to find quickly which path is taken by a specific function
1260 _ you just want to see what happens inside your kernel
1262 # tracer: function_graph
1264 # CPU DURATION FUNCTION CALLS
1268 0) | do_sys_open() {
1270 0) | kmem_cache_alloc() {
1271 0) 1.382 us | __might_sleep();
1273 0) | strncpy_from_user() {
1274 0) | might_fault() {
1275 0) 1.389 us | __might_sleep();
1280 0) 0.668 us | _spin_lock();
1281 0) 0.570 us | expand_files();
1282 0) 0.586 us | _spin_unlock();
1285 There are several columns that can be dynamically enabled/disabled.
1286 You can use every combination of options you want, depending on your needs.
1288 _ The cpu number on which the function executed is default enabled.
1289 It is sometimes better to only trace one cpu (see tracing_cpu_mask file)
1290 or you might sometimes see unordered function calls while cpu tracing switch.
1292 hide: echo nofuncgraph-cpu > /debug/tracing/trace_options
1293 show: echo funcgraph-cpu > /debug/tracing/trace_options
1295 _ The duration (function's time of execution) is displayed on the closing bracket
1296 line of a function or on the same line than the current function in case of a leaf
1297 one. It is default enabled.
1299 hide: echo nofuncgraph-duration > /debug/tracing/trace_options
1300 show: echo funcgraph-duration > /debug/tracing/trace_options
1302 _ The overhead field precedes the duration one in case of reached duration thresholds.
1304 hide: echo nofuncgraph-overhead > /debug/tracing/trace_options
1305 show: echo funcgraph-overhead > /debug/tracing/trace_options
1306 depends on: funcgraph-duration
1311 0) 0.646 us | _spin_lock_irqsave();
1312 0) 0.684 us | _spin_unlock_irqrestore();
1314 0) 0.548 us | fput();
1320 0) | kmem_cache_free() {
1321 0) 0.518 us | __phys_addr();
1327 + means that the function exceeded 10 usecs.
1328 ! means that the function exceeded 100 usecs.
1331 _ The task/pid field displays the thread cmdline and pid which executed the function.
1332 It is default disabled.
1334 hide: echo nofuncgraph-proc > /debug/tracing/trace_options
1335 show: echo funcgraph-proc > /debug/tracing/trace_options
1339 # tracer: function_graph
1341 # CPU TASK/PID DURATION FUNCTION CALLS
1343 0) sh-4802 | | d_free() {
1344 0) sh-4802 | | call_rcu() {
1345 0) sh-4802 | | __call_rcu() {
1346 0) sh-4802 | 0.616 us | rcu_process_gp_end();
1347 0) sh-4802 | 0.586 us | check_for_new_grace_period();
1348 0) sh-4802 | 2.899 us | }
1349 0) sh-4802 | 4.040 us | }
1350 0) sh-4802 | 5.151 us | }
1351 0) sh-4802 | + 49.370 us | }
1354 _ The absolute time field is an absolute timestamp given by the clock since
1355 it started. A snapshot of this time is given on each entry/exit of functions
1357 hide: echo nofuncgraph-abstime > /debug/tracing/trace_options
1358 show: echo funcgraph-abstime > /debug/tracing/trace_options
1363 # TIME CPU DURATION FUNCTION CALLS
1365 360.774522 | 1) 0.541 us | }
1366 360.774522 | 1) 4.663 us | }
1367 360.774523 | 1) 0.541 us | __wake_up_bit();
1368 360.774524 | 1) 6.796 us | }
1369 360.774524 | 1) 7.952 us | }
1370 360.774525 | 1) 9.063 us | }
1371 360.774525 | 1) 0.615 us | journal_mark_dirty();
1372 360.774527 | 1) 0.578 us | __brelse();
1373 360.774528 | 1) | reiserfs_prepare_for_journal() {
1374 360.774528 | 1) | unlock_buffer() {
1375 360.774529 | 1) | wake_up_bit() {
1376 360.774529 | 1) | bit_waitqueue() {
1377 360.774530 | 1) 0.594 us | __phys_addr();
1380 You can put some comments on specific functions by using ftrace_printk()
1381 For example, if you want to put a comment inside the __might_sleep() function,
1382 you just have to include <linux/ftrace.h> and call ftrace_printk() inside __might_sleep()
1384 ftrace_printk("I'm a comment!\n")
1388 1) | __might_sleep() {
1389 1) | /* I'm a comment! */
1393 You might find other useful features for this tracer on the "dynamic ftrace"
1394 section such as tracing only specific functions or tasks.
1399 If CONFIG_DYNAMIC_FTRACE is set, the system will run with
1400 virtually no overhead when function tracing is disabled. The way
1401 this works is the mcount function call (placed at the start of
1402 every kernel function, produced by the -pg switch in gcc), starts
1403 of pointing to a simple return. (Enabling FTRACE will include the
1404 -pg switch in the compiling of the kernel.)
1406 At compile time every C file object is run through the
1407 recordmcount.pl script (located in the scripts directory). This
1408 script will process the C object using objdump to find all the
1409 locations in the .text section that call mcount. (Note, only
1410 the .text section is processed, since processing other sections
1411 like .init.text may cause races due to those sections being freed).
1413 A new section called "__mcount_loc" is created that holds references
1414 to all the mcount call sites in the .text section. This section is
1415 compiled back into the original object. The final linker will add
1416 all these references into a single table.
1418 On boot up, before SMP is initialized, the dynamic ftrace code
1419 scans this table and updates all the locations into nops. It also
1420 records the locations, which are added to the available_filter_functions
1421 list. Modules are processed as they are loaded and before they are
1422 executed. When a module is unloaded, it also removes its functions from
1423 the ftrace function list. This is automatic in the module unload
1424 code, and the module author does not need to worry about it.
1426 When tracing is enabled, kstop_machine is called to prevent races
1427 with the CPUS executing code being modified (which can cause the
1428 CPU to do undesireable things), and the nops are patched back
1429 to calls. But this time, they do not call mcount (which is just
1430 a function stub). They now call into the ftrace infrastructure.
1432 One special side-effect to the recording of the functions being
1433 traced is that we can now selectively choose which functions we
1434 wish to trace and which ones we want the mcount calls to remain as
1437 Two files are used, one for enabling and one for disabling the tracing
1438 of specified functions. They are:
1446 A list of available functions that you can add to these files is listed
1449 available_filter_functions
1451 # cat /debug/tracing/available_filter_functions
1460 If I am only interested in sys_nanosleep and hrtimer_interrupt:
1462 # echo sys_nanosleep hrtimer_interrupt \
1463 > /debug/tracing/set_ftrace_filter
1464 # echo ftrace > /debug/tracing/current_tracer
1465 # echo 1 > /debug/tracing/tracing_enabled
1467 # echo 0 > /debug/tracing/tracing_enabled
1468 # cat /debug/tracing/trace
1471 # TASK-PID CPU# TIMESTAMP FUNCTION
1473 usleep-4134 [00] 1317.070017: hrtimer_interrupt <-smp_apic_timer_interrupt
1474 usleep-4134 [00] 1317.070111: sys_nanosleep <-syscall_call
1475 <idle>-0 [00] 1317.070115: hrtimer_interrupt <-smp_apic_timer_interrupt
1477 To see which functions are being traced, you can cat the file:
1479 # cat /debug/tracing/set_ftrace_filter
1484 Perhaps this is not enough. The filters also allow simple wild cards.
1485 Only the following are currently available
1487 <match>* - will match functions that begin with <match>
1488 *<match> - will match functions that end with <match>
1489 *<match>* - will match functions that have <match> in it
1491 These are the only wild cards which are supported.
1493 <match>*<match> will not work.
1495 Note: It is better to use quotes to enclose the wild cards, otherwise
1496 the shell may expand the parameters into names of files in the local
1499 # echo 'hrtimer_*' > /debug/tracing/set_ftrace_filter
1505 # TASK-PID CPU# TIMESTAMP FUNCTION
1507 bash-4003 [00] 1480.611794: hrtimer_init <-copy_process
1508 bash-4003 [00] 1480.611941: hrtimer_start <-hrtick_set
1509 bash-4003 [00] 1480.611956: hrtimer_cancel <-hrtick_clear
1510 bash-4003 [00] 1480.611956: hrtimer_try_to_cancel <-hrtimer_cancel
1511 <idle>-0 [00] 1480.612019: hrtimer_get_next_event <-get_next_timer_interrupt
1512 <idle>-0 [00] 1480.612025: hrtimer_get_next_event <-get_next_timer_interrupt
1513 <idle>-0 [00] 1480.612032: hrtimer_get_next_event <-get_next_timer_interrupt
1514 <idle>-0 [00] 1480.612037: hrtimer_get_next_event <-get_next_timer_interrupt
1515 <idle>-0 [00] 1480.612382: hrtimer_get_next_event <-get_next_timer_interrupt
1518 Notice that we lost the sys_nanosleep.
1520 # cat /debug/tracing/set_ftrace_filter
1525 hrtimer_try_to_cancel
1529 hrtimer_force_reprogram
1530 hrtimer_get_next_event
1534 hrtimer_get_remaining
1536 hrtimer_init_sleeper
1539 This is because the '>' and '>>' act just like they do in bash.
1540 To rewrite the filters, use '>'
1541 To append to the filters, use '>>'
1543 To clear out a filter so that all functions will be recorded again:
1545 # echo > /debug/tracing/set_ftrace_filter
1546 # cat /debug/tracing/set_ftrace_filter
1549 Again, now we want to append.
1551 # echo sys_nanosleep > /debug/tracing/set_ftrace_filter
1552 # cat /debug/tracing/set_ftrace_filter
1554 # echo 'hrtimer_*' >> /debug/tracing/set_ftrace_filter
1555 # cat /debug/tracing/set_ftrace_filter
1560 hrtimer_try_to_cancel
1564 hrtimer_force_reprogram
1565 hrtimer_get_next_event
1570 hrtimer_get_remaining
1572 hrtimer_init_sleeper
1575 The set_ftrace_notrace prevents those functions from being traced.
1577 # echo '*preempt*' '*lock*' > /debug/tracing/set_ftrace_notrace
1583 # TASK-PID CPU# TIMESTAMP FUNCTION
1585 bash-4043 [01] 115.281644: finish_task_switch <-schedule
1586 bash-4043 [01] 115.281645: hrtick_set <-schedule
1587 bash-4043 [01] 115.281645: hrtick_clear <-hrtick_set
1588 bash-4043 [01] 115.281646: wait_for_completion <-__stop_machine_run
1589 bash-4043 [01] 115.281647: wait_for_common <-wait_for_completion
1590 bash-4043 [01] 115.281647: kthread_stop <-stop_machine_run
1591 bash-4043 [01] 115.281648: init_waitqueue_head <-kthread_stop
1592 bash-4043 [01] 115.281648: wake_up_process <-kthread_stop
1593 bash-4043 [01] 115.281649: try_to_wake_up <-wake_up_process
1595 We can see that there's no more lock or preempt tracing.
1598 * Dynamic ftrace with the function graph tracer *
1601 Although what has been explained above concerns both the function tracer and
1602 the function_graph_tracer, the following concerns only the latter.
1604 If you want to trace only one function and all of its childs, you just have
1605 to echo its name on set_graph_function:
1607 echo __do_fault > set_graph_function
1609 will produce the following:
1612 0) | filemap_fault() {
1613 0) | find_lock_page() {
1614 0) 0.804 us | find_get_page();
1615 0) | __might_sleep() {
1619 0) 0.653 us | _spin_lock();
1620 0) 0.578 us | page_add_file_rmap();
1621 0) 0.525 us | native_set_pte_at();
1622 0) 0.585 us | _spin_unlock();
1623 0) | unlock_page() {
1624 0) 0.541 us | page_waitqueue();
1625 0) 0.639 us | __wake_up_bit();
1629 0) | filemap_fault() {
1630 0) | find_lock_page() {
1631 0) 0.698 us | find_get_page();
1632 0) | __might_sleep() {
1636 0) 0.631 us | _spin_lock();
1637 0) 0.571 us | page_add_file_rmap();
1638 0) 0.526 us | native_set_pte_at();
1639 0) 0.586 us | _spin_unlock();
1640 0) | unlock_page() {
1641 0) 0.533 us | page_waitqueue();
1642 0) 0.638 us | __wake_up_bit();
1646 You can also select several functions:
1648 echo sys_open > set_graph_function
1649 echo sys_close >> set_graph_function
1651 Now if you want to go back to trace all functions
1653 echo > set_graph_function
1659 The trace_pipe outputs the same content as the trace file, but the effect
1660 on the tracing is different. Every read from trace_pipe is consumed.
1661 This means that subsequent reads will be different. The trace
1664 # echo function > /debug/tracing/current_tracer
1665 # cat /debug/tracing/trace_pipe > /tmp/trace.out &
1667 # echo 1 > /debug/tracing/tracing_enabled
1669 # echo 0 > /debug/tracing/tracing_enabled
1670 # cat /debug/tracing/trace
1673 # TASK-PID CPU# TIMESTAMP FUNCTION
1677 # cat /tmp/trace.out
1678 bash-4043 [00] 41.267106: finish_task_switch <-schedule
1679 bash-4043 [00] 41.267106: hrtick_set <-schedule
1680 bash-4043 [00] 41.267107: hrtick_clear <-hrtick_set
1681 bash-4043 [00] 41.267108: wait_for_completion <-__stop_machine_run
1682 bash-4043 [00] 41.267108: wait_for_common <-wait_for_completion
1683 bash-4043 [00] 41.267109: kthread_stop <-stop_machine_run
1684 bash-4043 [00] 41.267109: init_waitqueue_head <-kthread_stop
1685 bash-4043 [00] 41.267110: wake_up_process <-kthread_stop
1686 bash-4043 [00] 41.267110: try_to_wake_up <-wake_up_process
1687 bash-4043 [00] 41.267111: select_task_rq_rt <-try_to_wake_up
1690 Note, reading the trace_pipe file will block until more input is added.
1691 By changing the tracer, trace_pipe will issue an EOF. We needed
1692 to set the function tracer _before_ we "cat" the trace_pipe file.
1698 Having too much or not enough data can be troublesome in diagnosing
1699 an issue in the kernel. The file buffer_size_kb is used to modify
1700 the size of the internal trace buffers. The number listed
1701 is the number of entries that can be recorded per CPU. To know
1702 the full size, multiply the number of possible CPUS with the
1705 # cat /debug/tracing/buffer_size_kb
1706 1408 (units kilobytes)
1708 Note, to modify this, you must have tracing completely disabled. To do that,
1709 echo "nop" into the current_tracer. If the current_tracer is not set
1710 to "nop", an EINVAL error will be returned.
1712 # echo nop > /debug/tracing/current_tracer
1713 # echo 10000 > /debug/tracing/buffer_size_kb
1714 # cat /debug/tracing/buffer_size_kb
1715 10000 (units kilobytes)
1717 The number of pages which will be allocated is limited to a percentage
1718 of available memory. Allocating too much will produce an error.
1720 # echo 1000000000000 > /debug/tracing/buffer_size_kb
1721 -bash: echo: write error: Cannot allocate memory
1722 # cat /debug/tracing/buffer_size_kb