1======================== 2ftrace - Function Tracer 3======================== 4 5Copyright 2008 Red Hat Inc. 6 7:Author: Steven Rostedt <srostedt@redhat.com> 8:License: The GNU Free Documentation License, Version 1.2 9 (dual licensed under the GPL v2) 10:Original Reviewers: Elias Oltmanns, Randy Dunlap, Andrew Morton, 11 John Kacur, and David Teigland. 12 13- Written for: 2.6.28-rc2 14- Updated for: 3.10 15- Updated for: 4.13 - Copyright 2017 VMware Inc. Steven Rostedt 16- Converted to rst format - Changbin Du <changbin.du@intel.com> 17 18Introduction 19------------ 20 21Ftrace is an internal tracer designed to help out developers and 22designers of systems to find what is going on inside the kernel. 23It can be used for debugging or analyzing latencies and 24performance issues that take place outside of user-space. 25 26Although ftrace is typically considered the function tracer, it 27is really a framework of several assorted tracing utilities. 28There's latency tracing to examine what occurs between interrupts 29disabled and enabled, as well as for preemption and from a time 30a task is woken to the task is actually scheduled in. 31 32One of the most common uses of ftrace is the event tracing. 33Throughout the kernel is hundreds of static event points that 34can be enabled via the tracefs file system to see what is 35going on in certain parts of the kernel. 36 37See events.txt for more information. 38 39 40Implementation Details 41---------------------- 42 43See :doc:`ftrace-design` for details for arch porters and such. 44 45 46The File System 47--------------- 48 49Ftrace uses the tracefs file system to hold the control files as 50well as the files to display output. 51 52When tracefs is configured into the kernel (which selecting any ftrace 53option will do) the directory /sys/kernel/tracing will be created. To mount 54this directory, you can add to your /etc/fstab file:: 55 56 tracefs /sys/kernel/tracing tracefs defaults 0 0 57 58Or you can mount it at run time with:: 59 60 mount -t tracefs nodev /sys/kernel/tracing 61 62For quicker access to that directory you may want to make a soft link to 63it:: 64 65 ln -s /sys/kernel/tracing /tracing 66 67.. attention:: 68 69 Before 4.1, all ftrace tracing control files were within the debugfs 70 file system, which is typically located at /sys/kernel/debug/tracing. 71 For backward compatibility, when mounting the debugfs file system, 72 the tracefs file system will be automatically mounted at: 73 74 /sys/kernel/debug/tracing 75 76 All files located in the tracefs file system will be located in that 77 debugfs file system directory as well. 78 79.. attention:: 80 81 Any selected ftrace option will also create the tracefs file system. 82 The rest of the document will assume that you are in the ftrace directory 83 (cd /sys/kernel/tracing) and will only concentrate on the files within that 84 directory and not distract from the content with the extended 85 "/sys/kernel/tracing" path name. 86 87That's it! (assuming that you have ftrace configured into your kernel) 88 89After mounting tracefs you will have access to the control and output files 90of ftrace. Here is a list of some of the key files: 91 92 93 Note: all time values are in microseconds. 94 95 current_tracer: 96 97 This is used to set or display the current tracer 98 that is configured. 99 100 available_tracers: 101 102 This holds the different types of tracers that 103 have been compiled into the kernel. The 104 tracers listed here can be configured by 105 echoing their name into current_tracer. 106 107 tracing_on: 108 109 This sets or displays whether writing to the trace 110 ring buffer is enabled. Echo 0 into this file to disable 111 the tracer or 1 to enable it. Note, this only disables 112 writing to the ring buffer, the tracing overhead may 113 still be occurring. 114 115 The kernel function tracing_off() can be used within the 116 kernel to disable writing to the ring buffer, which will 117 set this file to "0". User space can re-enable tracing by 118 echoing "1" into the file. 119 120 Note, the function and event trigger "traceoff" will also 121 set this file to zero and stop tracing. Which can also 122 be re-enabled by user space using this file. 123 124 trace: 125 126 This file holds the output of the trace in a human 127 readable format (described below). Note, tracing is temporarily 128 disabled when the file is open for reading. Once all readers 129 are closed, tracing is re-enabled. 130 131 trace_pipe: 132 133 The output is the same as the "trace" file but this 134 file is meant to be streamed with live tracing. 135 Reads from this file will block until new data is 136 retrieved. Unlike the "trace" file, this file is a 137 consumer. This means reading from this file causes 138 sequential reads to display more current data. Once 139 data is read from this file, it is consumed, and 140 will not be read again with a sequential read. The 141 "trace" file is static, and if the tracer is not 142 adding more data, it will display the same 143 information every time it is read. Unlike the 144 "trace" file, opening this file for reading will not 145 temporarily disable tracing. 146 147 trace_options: 148 149 This file lets the user control the amount of data 150 that is displayed in one of the above output 151 files. Options also exist to modify how a tracer 152 or events work (stack traces, timestamps, etc). 153 154 options: 155 156 This is a directory that has a file for every available 157 trace option (also in trace_options). Options may also be set 158 or cleared by writing a "1" or "0" respectively into the 159 corresponding file with the option name. 160 161 tracing_max_latency: 162 163 Some of the tracers record the max latency. 164 For example, the maximum time that interrupts are disabled. 165 The maximum time is saved in this file. The max trace will also be 166 stored, and displayed by "trace". A new max trace will only be 167 recorded if the latency is greater than the value in this file 168 (in microseconds). 169 170 By echoing in a time into this file, no latency will be recorded 171 unless it is greater than the time in this file. 172 173 tracing_thresh: 174 175 Some latency tracers will record a trace whenever the 176 latency is greater than the number in this file. 177 Only active when the file contains a number greater than 0. 178 (in microseconds) 179 180 buffer_size_kb: 181 182 This sets or displays the number of kilobytes each CPU 183 buffer holds. By default, the trace buffers are the same size 184 for each CPU. The displayed number is the size of the 185 CPU buffer and not total size of all buffers. The 186 trace buffers are allocated in pages (blocks of memory 187 that the kernel uses for allocation, usually 4 KB in size). 188 If the last page allocated has room for more bytes 189 than requested, the rest of the page will be used, 190 making the actual allocation bigger than requested or shown. 191 ( Note, the size may not be a multiple of the page size 192 due to buffer management meta-data. ) 193 194 Buffer sizes for individual CPUs may vary 195 (see "per_cpu/cpu0/buffer_size_kb" below), and if they do 196 this file will show "X". 197 198 buffer_total_size_kb: 199 200 This displays the total combined size of all the trace buffers. 201 202 free_buffer: 203 204 If a process is performing tracing, and the ring buffer should be 205 shrunk "freed" when the process is finished, even if it were to be 206 killed by a signal, this file can be used for that purpose. On close 207 of this file, the ring buffer will be resized to its minimum size. 208 Having a process that is tracing also open this file, when the process 209 exits its file descriptor for this file will be closed, and in doing so, 210 the ring buffer will be "freed". 211 212 It may also stop tracing if disable_on_free option is set. 213 214 tracing_cpumask: 215 216 This is a mask that lets the user only trace on specified CPUs. 217 The format is a hex string representing the CPUs. 218 219 set_ftrace_filter: 220 221 When dynamic ftrace is configured in (see the 222 section below "dynamic ftrace"), the code is dynamically 223 modified (code text rewrite) to disable calling of the 224 function profiler (mcount). This lets tracing be configured 225 in with practically no overhead in performance. This also 226 has a side effect of enabling or disabling specific functions 227 to be traced. Echoing names of functions into this file 228 will limit the trace to only those functions. 229 This influences the tracers "function" and "function_graph" 230 and thus also function profiling (see "function_profile_enabled"). 231 232 The functions listed in "available_filter_functions" are what 233 can be written into this file. 234 235 This interface also allows for commands to be used. See the 236 "Filter commands" section for more details. 237 238 As a speed up, since processing strings can't be quite expensive 239 and requires a check of all functions registered to tracing, instead 240 an index can be written into this file. A number (starting with "1") 241 written will instead select the same corresponding at the line position 242 of the "available_filter_functions" file. 243 244 set_ftrace_notrace: 245 246 This has an effect opposite to that of 247 set_ftrace_filter. Any function that is added here will not 248 be traced. If a function exists in both set_ftrace_filter 249 and set_ftrace_notrace, the function will _not_ be traced. 250 251 set_ftrace_pid: 252 253 Have the function tracer only trace the threads whose PID are 254 listed in this file. 255 256 If the "function-fork" option is set, then when a task whose 257 PID is listed in this file forks, the child's PID will 258 automatically be added to this file, and the child will be 259 traced by the function tracer as well. This option will also 260 cause PIDs of tasks that exit to be removed from the file. 261 262 set_event_pid: 263 264 Have the events only trace a task with a PID listed in this file. 265 Note, sched_switch and sched_wake_up will also trace events 266 listed in this file. 267 268 To have the PIDs of children of tasks with their PID in this file 269 added on fork, enable the "event-fork" option. That option will also 270 cause the PIDs of tasks to be removed from this file when the task 271 exits. 272 273 set_graph_function: 274 275 Functions listed in this file will cause the function graph 276 tracer to only trace these functions and the functions that 277 they call. (See the section "dynamic ftrace" for more details). 278 Note, set_ftrace_filter and set_ftrace_notrace still affects 279 what functions are being traced. 280 281 set_graph_notrace: 282 283 Similar to set_graph_function, but will disable function graph 284 tracing when the function is hit until it exits the function. 285 This makes it possible to ignore tracing functions that are called 286 by a specific function. 287 288 available_filter_functions: 289 290 This lists the functions that ftrace has processed and can trace. 291 These are the function names that you can pass to 292 "set_ftrace_filter", "set_ftrace_notrace", 293 "set_graph_function", or "set_graph_notrace". 294 (See the section "dynamic ftrace" below for more details.) 295 296 dyn_ftrace_total_info: 297 298 This file is for debugging purposes. The number of functions that 299 have been converted to nops and are available to be traced. 300 301 enabled_functions: 302 303 This file is more for debugging ftrace, but can also be useful 304 in seeing if any function has a callback attached to it. 305 Not only does the trace infrastructure use ftrace function 306 trace utility, but other subsystems might too. This file 307 displays all functions that have a callback attached to them 308 as well as the number of callbacks that have been attached. 309 Note, a callback may also call multiple functions which will 310 not be listed in this count. 311 312 If the callback registered to be traced by a function with 313 the "save regs" attribute (thus even more overhead), a 'R' 314 will be displayed on the same line as the function that 315 is returning registers. 316 317 If the callback registered to be traced by a function with 318 the "ip modify" attribute (thus the regs->ip can be changed), 319 an 'I' will be displayed on the same line as the function that 320 can be overridden. 321 322 If the architecture supports it, it will also show what callback 323 is being directly called by the function. If the count is greater 324 than 1 it most likely will be ftrace_ops_list_func(). 325 326 If the callback of the function jumps to a trampoline that is 327 specific to a the callback and not the standard trampoline, 328 its address will be printed as well as the function that the 329 trampoline calls. 330 331 function_profile_enabled: 332 333 When set it will enable all functions with either the function 334 tracer, or if configured, the function graph tracer. It will 335 keep a histogram of the number of functions that were called 336 and if the function graph tracer was configured, it will also keep 337 track of the time spent in those functions. The histogram 338 content can be displayed in the files: 339 340 trace_stat/function<cpu> ( function0, function1, etc). 341 342 trace_stat: 343 344 A directory that holds different tracing stats. 345 346 kprobe_events: 347 348 Enable dynamic trace points. See kprobetrace.txt. 349 350 kprobe_profile: 351 352 Dynamic trace points stats. See kprobetrace.txt. 353 354 max_graph_depth: 355 356 Used with the function graph tracer. This is the max depth 357 it will trace into a function. Setting this to a value of 358 one will show only the first kernel function that is called 359 from user space. 360 361 printk_formats: 362 363 This is for tools that read the raw format files. If an event in 364 the ring buffer references a string, only a pointer to the string 365 is recorded into the buffer and not the string itself. This prevents 366 tools from knowing what that string was. This file displays the string 367 and address for the string allowing tools to map the pointers to what 368 the strings were. 369 370 saved_cmdlines: 371 372 Only the pid of the task is recorded in a trace event unless 373 the event specifically saves the task comm as well. Ftrace 374 makes a cache of pid mappings to comms to try to display 375 comms for events. If a pid for a comm is not listed, then 376 "<...>" is displayed in the output. 377 378 If the option "record-cmd" is set to "0", then comms of tasks 379 will not be saved during recording. By default, it is enabled. 380 381 saved_cmdlines_size: 382 383 By default, 128 comms are saved (see "saved_cmdlines" above). To 384 increase or decrease the amount of comms that are cached, echo 385 in a the number of comms to cache, into this file. 386 387 saved_tgids: 388 389 If the option "record-tgid" is set, on each scheduling context switch 390 the Task Group ID of a task is saved in a table mapping the PID of 391 the thread to its TGID. By default, the "record-tgid" option is 392 disabled. 393 394 snapshot: 395 396 This displays the "snapshot" buffer and also lets the user 397 take a snapshot of the current running trace. 398 See the "Snapshot" section below for more details. 399 400 stack_max_size: 401 402 When the stack tracer is activated, this will display the 403 maximum stack size it has encountered. 404 See the "Stack Trace" section below. 405 406 stack_trace: 407 408 This displays the stack back trace of the largest stack 409 that was encountered when the stack tracer is activated. 410 See the "Stack Trace" section below. 411 412 stack_trace_filter: 413 414 This is similar to "set_ftrace_filter" but it limits what 415 functions the stack tracer will check. 416 417 trace_clock: 418 419 Whenever an event is recorded into the ring buffer, a 420 "timestamp" is added. This stamp comes from a specified 421 clock. By default, ftrace uses the "local" clock. This 422 clock is very fast and strictly per cpu, but on some 423 systems it may not be monotonic with respect to other 424 CPUs. In other words, the local clocks may not be in sync 425 with local clocks on other CPUs. 426 427 Usual clocks for tracing:: 428 429 # cat trace_clock 430 [local] global counter x86-tsc 431 432 The clock with the square brackets around it is the one in effect. 433 434 local: 435 Default clock, but may not be in sync across CPUs 436 437 global: 438 This clock is in sync with all CPUs but may 439 be a bit slower than the local clock. 440 441 counter: 442 This is not a clock at all, but literally an atomic 443 counter. It counts up one by one, but is in sync 444 with all CPUs. This is useful when you need to 445 know exactly the order events occurred with respect to 446 each other on different CPUs. 447 448 uptime: 449 This uses the jiffies counter and the time stamp 450 is relative to the time since boot up. 451 452 perf: 453 This makes ftrace use the same clock that perf uses. 454 Eventually perf will be able to read ftrace buffers 455 and this will help out in interleaving the data. 456 457 x86-tsc: 458 Architectures may define their own clocks. For 459 example, x86 uses its own TSC cycle clock here. 460 461 ppc-tb: 462 This uses the powerpc timebase register value. 463 This is in sync across CPUs and can also be used 464 to correlate events across hypervisor/guest if 465 tb_offset is known. 466 467 mono: 468 This uses the fast monotonic clock (CLOCK_MONOTONIC) 469 which is monotonic and is subject to NTP rate adjustments. 470 471 mono_raw: 472 This is the raw monotonic clock (CLOCK_MONOTONIC_RAW) 473 which is monotonic but is not subject to any rate adjustments 474 and ticks at the same rate as the hardware clocksource. 475 476 boot: 477 This is the boot clock (CLOCK_BOOTTIME) and is based on the 478 fast monotonic clock, but also accounts for time spent in 479 suspend. Since the clock access is designed for use in 480 tracing in the suspend path, some side effects are possible 481 if clock is accessed after the suspend time is accounted before 482 the fast mono clock is updated. In this case, the clock update 483 appears to happen slightly sooner than it normally would have. 484 Also on 32-bit systems, it's possible that the 64-bit boot offset 485 sees a partial update. These effects are rare and post 486 processing should be able to handle them. See comments in the 487 ktime_get_boot_fast_ns() function for more information. 488 489 To set a clock, simply echo the clock name into this file:: 490 491 # echo global > trace_clock 492 493 trace_marker: 494 495 This is a very useful file for synchronizing user space 496 with events happening in the kernel. Writing strings into 497 this file will be written into the ftrace buffer. 498 499 It is useful in applications to open this file at the start 500 of the application and just reference the file descriptor 501 for the file:: 502 503 void trace_write(const char *fmt, ...) 504 { 505 va_list ap; 506 char buf[256]; 507 int n; 508 509 if (trace_fd < 0) 510 return; 511 512 va_start(ap, fmt); 513 n = vsnprintf(buf, 256, fmt, ap); 514 va_end(ap); 515 516 write(trace_fd, buf, n); 517 } 518 519 start:: 520 521 trace_fd = open("trace_marker", WR_ONLY); 522 523 Note: Writing into the trace_marker file can also initiate triggers 524 that are written into /sys/kernel/tracing/events/ftrace/print/trigger 525 See "Event triggers" in Documentation/trace/events.rst and an 526 example in Documentation/trace/histogram.rst (Section 3.) 527 528 trace_marker_raw: 529 530 This is similar to trace_marker above, but is meant for for binary data 531 to be written to it, where a tool can be used to parse the data 532 from trace_pipe_raw. 533 534 uprobe_events: 535 536 Add dynamic tracepoints in programs. 537 See uprobetracer.txt 538 539 uprobe_profile: 540 541 Uprobe statistics. See uprobetrace.txt 542 543 instances: 544 545 This is a way to make multiple trace buffers where different 546 events can be recorded in different buffers. 547 See "Instances" section below. 548 549 events: 550 551 This is the trace event directory. It holds event tracepoints 552 (also known as static tracepoints) that have been compiled 553 into the kernel. It shows what event tracepoints exist 554 and how they are grouped by system. There are "enable" 555 files at various levels that can enable the tracepoints 556 when a "1" is written to them. 557 558 See events.txt for more information. 559 560 set_event: 561 562 By echoing in the event into this file, will enable that event. 563 564 See events.txt for more information. 565 566 available_events: 567 568 A list of events that can be enabled in tracing. 569 570 See events.txt for more information. 571 572 timestamp_mode: 573 574 Certain tracers may change the timestamp mode used when 575 logging trace events into the event buffer. Events with 576 different modes can coexist within a buffer but the mode in 577 effect when an event is logged determines which timestamp mode 578 is used for that event. The default timestamp mode is 579 'delta'. 580 581 Usual timestamp modes for tracing: 582 583 # cat timestamp_mode 584 [delta] absolute 585 586 The timestamp mode with the square brackets around it is the 587 one in effect. 588 589 delta: Default timestamp mode - timestamp is a delta against 590 a per-buffer timestamp. 591 592 absolute: The timestamp is a full timestamp, not a delta 593 against some other value. As such it takes up more 594 space and is less efficient. 595 596 hwlat_detector: 597 598 Directory for the Hardware Latency Detector. 599 See "Hardware Latency Detector" section below. 600 601 per_cpu: 602 603 This is a directory that contains the trace per_cpu information. 604 605 per_cpu/cpu0/buffer_size_kb: 606 607 The ftrace buffer is defined per_cpu. That is, there's a separate 608 buffer for each CPU to allow writes to be done atomically, 609 and free from cache bouncing. These buffers may have different 610 size buffers. This file is similar to the buffer_size_kb 611 file, but it only displays or sets the buffer size for the 612 specific CPU. (here cpu0). 613 614 per_cpu/cpu0/trace: 615 616 This is similar to the "trace" file, but it will only display 617 the data specific for the CPU. If written to, it only clears 618 the specific CPU buffer. 619 620 per_cpu/cpu0/trace_pipe 621 622 This is similar to the "trace_pipe" file, and is a consuming 623 read, but it will only display (and consume) the data specific 624 for the CPU. 625 626 per_cpu/cpu0/trace_pipe_raw 627 628 For tools that can parse the ftrace ring buffer binary format, 629 the trace_pipe_raw file can be used to extract the data 630 from the ring buffer directly. With the use of the splice() 631 system call, the buffer data can be quickly transferred to 632 a file or to the network where a server is collecting the 633 data. 634 635 Like trace_pipe, this is a consuming reader, where multiple 636 reads will always produce different data. 637 638 per_cpu/cpu0/snapshot: 639 640 This is similar to the main "snapshot" file, but will only 641 snapshot the current CPU (if supported). It only displays 642 the content of the snapshot for a given CPU, and if 643 written to, only clears this CPU buffer. 644 645 per_cpu/cpu0/snapshot_raw: 646 647 Similar to the trace_pipe_raw, but will read the binary format 648 from the snapshot buffer for the given CPU. 649 650 per_cpu/cpu0/stats: 651 652 This displays certain stats about the ring buffer: 653 654 entries: 655 The number of events that are still in the buffer. 656 657 overrun: 658 The number of lost events due to overwriting when 659 the buffer was full. 660 661 commit overrun: 662 Should always be zero. 663 This gets set if so many events happened within a nested 664 event (ring buffer is re-entrant), that it fills the 665 buffer and starts dropping events. 666 667 bytes: 668 Bytes actually read (not overwritten). 669 670 oldest event ts: 671 The oldest timestamp in the buffer 672 673 now ts: 674 The current timestamp 675 676 dropped events: 677 Events lost due to overwrite option being off. 678 679 read events: 680 The number of events read. 681 682The Tracers 683----------- 684 685Here is the list of current tracers that may be configured. 686 687 "function" 688 689 Function call tracer to trace all kernel functions. 690 691 "function_graph" 692 693 Similar to the function tracer except that the 694 function tracer probes the functions on their entry 695 whereas the function graph tracer traces on both entry 696 and exit of the functions. It then provides the ability 697 to draw a graph of function calls similar to C code 698 source. 699 700 "blk" 701 702 The block tracer. The tracer used by the blktrace user 703 application. 704 705 "hwlat" 706 707 The Hardware Latency tracer is used to detect if the hardware 708 produces any latency. See "Hardware Latency Detector" section 709 below. 710 711 "irqsoff" 712 713 Traces the areas that disable interrupts and saves 714 the trace with the longest max latency. 715 See tracing_max_latency. When a new max is recorded, 716 it replaces the old trace. It is best to view this 717 trace with the latency-format option enabled, which 718 happens automatically when the tracer is selected. 719 720 "preemptoff" 721 722 Similar to irqsoff but traces and records the amount of 723 time for which preemption is disabled. 724 725 "preemptirqsoff" 726 727 Similar to irqsoff and preemptoff, but traces and 728 records the largest time for which irqs and/or preemption 729 is disabled. 730 731 "wakeup" 732 733 Traces and records the max latency that it takes for 734 the highest priority task to get scheduled after 735 it has been woken up. 736 Traces all tasks as an average developer would expect. 737 738 "wakeup_rt" 739 740 Traces and records the max latency that it takes for just 741 RT tasks (as the current "wakeup" does). This is useful 742 for those interested in wake up timings of RT tasks. 743 744 "wakeup_dl" 745 746 Traces and records the max latency that it takes for 747 a SCHED_DEADLINE task to be woken (as the "wakeup" and 748 "wakeup_rt" does). 749 750 "mmiotrace" 751 752 A special tracer that is used to trace binary module. 753 It will trace all the calls that a module makes to the 754 hardware. Everything it writes and reads from the I/O 755 as well. 756 757 "branch" 758 759 This tracer can be configured when tracing likely/unlikely 760 calls within the kernel. It will trace when a likely and 761 unlikely branch is hit and if it was correct in its prediction 762 of being correct. 763 764 "nop" 765 766 This is the "trace nothing" tracer. To remove all 767 tracers from tracing simply echo "nop" into 768 current_tracer. 769 770Error conditions 771---------------- 772 773 For most ftrace commands, failure modes are obvious and communicated 774 using standard return codes. 775 776 For other more involved commands, extended error information may be 777 available via the tracing/error_log file. For the commands that 778 support it, reading the tracing/error_log file after an error will 779 display more detailed information about what went wrong, if 780 information is available. The tracing/error_log file is a circular 781 error log displaying a small number (currently, 8) of ftrace errors 782 for the last (8) failed commands. 783 784 The extended error information and usage takes the form shown in 785 this example:: 786 787 # echo xxx > /sys/kernel/debug/tracing/events/sched/sched_wakeup/trigger 788 echo: write error: Invalid argument 789 790 # cat /sys/kernel/debug/tracing/error_log 791 [ 5348.887237] location: error: Couldn't yyy: zzz 792 Command: xxx 793 ^ 794 [ 7517.023364] location: error: Bad rrr: sss 795 Command: ppp qqq 796 ^ 797 798 To clear the error log, echo the empty string into it:: 799 800 # echo > /sys/kernel/debug/tracing/error_log 801 802Examples of using the tracer 803---------------------------- 804 805Here are typical examples of using the tracers when controlling 806them only with the tracefs interface (without using any 807user-land utilities). 808 809Output format: 810-------------- 811 812Here is an example of the output format of the file "trace":: 813 814 # tracer: function 815 # 816 # entries-in-buffer/entries-written: 140080/250280 #P:4 817 # 818 # _-----=> irqs-off 819 # / _----=> need-resched 820 # | / _---=> hardirq/softirq 821 # || / _--=> preempt-depth 822 # ||| / delay 823 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 824 # | | | |||| | | 825 bash-1977 [000] .... 17284.993652: sys_close <-system_call_fastpath 826 bash-1977 [000] .... 17284.993653: __close_fd <-sys_close 827 bash-1977 [000] .... 17284.993653: _raw_spin_lock <-__close_fd 828 sshd-1974 [003] .... 17284.993653: __srcu_read_unlock <-fsnotify 829 bash-1977 [000] .... 17284.993654: add_preempt_count <-_raw_spin_lock 830 bash-1977 [000] ...1 17284.993655: _raw_spin_unlock <-__close_fd 831 bash-1977 [000] ...1 17284.993656: sub_preempt_count <-_raw_spin_unlock 832 bash-1977 [000] .... 17284.993657: filp_close <-__close_fd 833 bash-1977 [000] .... 17284.993657: dnotify_flush <-filp_close 834 sshd-1974 [003] .... 17284.993658: sys_select <-system_call_fastpath 835 .... 836 837A header is printed with the tracer name that is represented by 838the trace. In this case the tracer is "function". Then it shows the 839number of events in the buffer as well as the total number of entries 840that were written. The difference is the number of entries that were 841lost due to the buffer filling up (250280 - 140080 = 110200 events 842lost). 843 844The header explains the content of the events. Task name "bash", the task 845PID "1977", the CPU that it was running on "000", the latency format 846(explained below), the timestamp in <secs>.<usecs> format, the 847function name that was traced "sys_close" and the parent function that 848called this function "system_call_fastpath". The timestamp is the time 849at which the function was entered. 850 851Latency trace format 852-------------------- 853 854When the latency-format option is enabled or when one of the latency 855tracers is set, the trace file gives somewhat more information to see 856why a latency happened. Here is a typical trace:: 857 858 # tracer: irqsoff 859 # 860 # irqsoff latency trace v1.1.5 on 3.8.0-test+ 861 # -------------------------------------------------------------------- 862 # latency: 259 us, #4/4, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 863 # ----------------- 864 # | task: ps-6143 (uid:0 nice:0 policy:0 rt_prio:0) 865 # ----------------- 866 # => started at: __lock_task_sighand 867 # => ended at: _raw_spin_unlock_irqrestore 868 # 869 # 870 # _------=> CPU# 871 # / _-----=> irqs-off 872 # | / _----=> need-resched 873 # || / _---=> hardirq/softirq 874 # ||| / _--=> preempt-depth 875 # |||| / delay 876 # cmd pid ||||| time | caller 877 # \ / ||||| \ | / 878 ps-6143 2d... 0us!: trace_hardirqs_off <-__lock_task_sighand 879 ps-6143 2d..1 259us+: trace_hardirqs_on <-_raw_spin_unlock_irqrestore 880 ps-6143 2d..1 263us+: time_hardirqs_on <-_raw_spin_unlock_irqrestore 881 ps-6143 2d..1 306us : <stack trace> 882 => trace_hardirqs_on_caller 883 => trace_hardirqs_on 884 => _raw_spin_unlock_irqrestore 885 => do_task_stat 886 => proc_tgid_stat 887 => proc_single_show 888 => seq_read 889 => vfs_read 890 => sys_read 891 => system_call_fastpath 892 893 894This shows that the current tracer is "irqsoff" tracing the time 895for which interrupts were disabled. It gives the trace version (which 896never changes) and the version of the kernel upon which this was executed on 897(3.8). Then it displays the max latency in microseconds (259 us). The number 898of trace entries displayed and the total number (both are four: #4/4). 899VP, KP, SP, and HP are always zero and are reserved for later use. 900#P is the number of online CPUs (#P:4). 901 902The task is the process that was running when the latency 903occurred. (ps pid: 6143). 904 905The start and stop (the functions in which the interrupts were 906disabled and enabled respectively) that caused the latencies: 907 908 - __lock_task_sighand is where the interrupts were disabled. 909 - _raw_spin_unlock_irqrestore is where they were enabled again. 910 911The next lines after the header are the trace itself. The header 912explains which is which. 913 914 cmd: The name of the process in the trace. 915 916 pid: The PID of that process. 917 918 CPU#: The CPU which the process was running on. 919 920 irqs-off: 'd' interrupts are disabled. '.' otherwise. 921 .. caution:: If the architecture does not support a way to 922 read the irq flags variable, an 'X' will always 923 be printed here. 924 925 need-resched: 926 - 'N' both TIF_NEED_RESCHED and PREEMPT_NEED_RESCHED is set, 927 - 'n' only TIF_NEED_RESCHED is set, 928 - 'p' only PREEMPT_NEED_RESCHED is set, 929 - '.' otherwise. 930 931 hardirq/softirq: 932 - 'Z' - NMI occurred inside a hardirq 933 - 'z' - NMI is running 934 - 'H' - hard irq occurred inside a softirq. 935 - 'h' - hard irq is running 936 - 's' - soft irq is running 937 - '.' - normal context. 938 939 preempt-depth: The level of preempt_disabled 940 941The above is mostly meaningful for kernel developers. 942 943 time: 944 When the latency-format option is enabled, the trace file 945 output includes a timestamp relative to the start of the 946 trace. This differs from the output when latency-format 947 is disabled, which includes an absolute timestamp. 948 949 delay: 950 This is just to help catch your eye a bit better. And 951 needs to be fixed to be only relative to the same CPU. 952 The marks are determined by the difference between this 953 current trace and the next trace. 954 955 - '$' - greater than 1 second 956 - '@' - greater than 100 millisecond 957 - '*' - greater than 10 millisecond 958 - '#' - greater than 1000 microsecond 959 - '!' - greater than 100 microsecond 960 - '+' - greater than 10 microsecond 961 - ' ' - less than or equal to 10 microsecond. 962 963 The rest is the same as the 'trace' file. 964 965 Note, the latency tracers will usually end with a back trace 966 to easily find where the latency occurred. 967 968trace_options 969------------- 970 971The trace_options file (or the options directory) is used to control 972what gets printed in the trace output, or manipulate the tracers. 973To see what is available, simply cat the file:: 974 975 cat trace_options 976 print-parent 977 nosym-offset 978 nosym-addr 979 noverbose 980 noraw 981 nohex 982 nobin 983 noblock 984 trace_printk 985 annotate 986 nouserstacktrace 987 nosym-userobj 988 noprintk-msg-only 989 context-info 990 nolatency-format 991 record-cmd 992 norecord-tgid 993 overwrite 994 nodisable_on_free 995 irq-info 996 markers 997 noevent-fork 998 function-trace 999 nofunction-fork 1000 nodisplay-graph 1001 nostacktrace 1002 nobranch 1003 1004To disable one of the options, echo in the option prepended with 1005"no":: 1006 1007 echo noprint-parent > trace_options 1008 1009To enable an option, leave off the "no":: 1010 1011 echo sym-offset > trace_options 1012 1013Here are the available options: 1014 1015 print-parent 1016 On function traces, display the calling (parent) 1017 function as well as the function being traced. 1018 :: 1019 1020 print-parent: 1021 bash-4000 [01] 1477.606694: simple_strtoul <-kstrtoul 1022 1023 noprint-parent: 1024 bash-4000 [01] 1477.606694: simple_strtoul 1025 1026 1027 sym-offset 1028 Display not only the function name, but also the 1029 offset in the function. For example, instead of 1030 seeing just "ktime_get", you will see 1031 "ktime_get+0xb/0x20". 1032 :: 1033 1034 sym-offset: 1035 bash-4000 [01] 1477.606694: simple_strtoul+0x6/0xa0 1036 1037 sym-addr 1038 This will also display the function address as well 1039 as the function name. 1040 :: 1041 1042 sym-addr: 1043 bash-4000 [01] 1477.606694: simple_strtoul <c0339346> 1044 1045 verbose 1046 This deals with the trace file when the 1047 latency-format option is enabled. 1048 :: 1049 1050 bash 4000 1 0 00000000 00010a95 [58127d26] 1720.415ms \ 1051 (+0.000ms): simple_strtoul (kstrtoul) 1052 1053 raw 1054 This will display raw numbers. This option is best for 1055 use with user applications that can translate the raw 1056 numbers better than having it done in the kernel. 1057 1058 hex 1059 Similar to raw, but the numbers will be in a hexadecimal format. 1060 1061 bin 1062 This will print out the formats in raw binary. 1063 1064 block 1065 When set, reading trace_pipe will not block when polled. 1066 1067 trace_printk 1068 Can disable trace_printk() from writing into the buffer. 1069 1070 annotate 1071 It is sometimes confusing when the CPU buffers are full 1072 and one CPU buffer had a lot of events recently, thus 1073 a shorter time frame, were another CPU may have only had 1074 a few events, which lets it have older events. When 1075 the trace is reported, it shows the oldest events first, 1076 and it may look like only one CPU ran (the one with the 1077 oldest events). When the annotate option is set, it will 1078 display when a new CPU buffer started:: 1079 1080 <idle>-0 [001] dNs4 21169.031481: wake_up_idle_cpu <-add_timer_on 1081 <idle>-0 [001] dNs4 21169.031482: _raw_spin_unlock_irqrestore <-add_timer_on 1082 <idle>-0 [001] .Ns4 21169.031484: sub_preempt_count <-_raw_spin_unlock_irqrestore 1083 ##### CPU 2 buffer started #### 1084 <idle>-0 [002] .N.1 21169.031484: rcu_idle_exit <-cpu_idle 1085 <idle>-0 [001] .Ns3 21169.031484: _raw_spin_unlock <-clocksource_watchdog 1086 <idle>-0 [001] .Ns3 21169.031485: sub_preempt_count <-_raw_spin_unlock 1087 1088 userstacktrace 1089 This option changes the trace. It records a 1090 stacktrace of the current user space thread after 1091 each trace event. 1092 1093 sym-userobj 1094 when user stacktrace are enabled, look up which 1095 object the address belongs to, and print a 1096 relative address. This is especially useful when 1097 ASLR is on, otherwise you don't get a chance to 1098 resolve the address to object/file/line after 1099 the app is no longer running 1100 1101 The lookup is performed when you read 1102 trace,trace_pipe. Example:: 1103 1104 a.out-1623 [000] 40874.465068: /root/a.out[+0x480] <-/root/a.out[+0 1105 x494] <- /root/a.out[+0x4a8] <- /lib/libc-2.7.so[+0x1e1a6] 1106 1107 1108 printk-msg-only 1109 When set, trace_printk()s will only show the format 1110 and not their parameters (if trace_bprintk() or 1111 trace_bputs() was used to save the trace_printk()). 1112 1113 context-info 1114 Show only the event data. Hides the comm, PID, 1115 timestamp, CPU, and other useful data. 1116 1117 latency-format 1118 This option changes the trace output. When it is enabled, 1119 the trace displays additional information about the 1120 latency, as described in "Latency trace format". 1121 1122 record-cmd 1123 When any event or tracer is enabled, a hook is enabled 1124 in the sched_switch trace point to fill comm cache 1125 with mapped pids and comms. But this may cause some 1126 overhead, and if you only care about pids, and not the 1127 name of the task, disabling this option can lower the 1128 impact of tracing. See "saved_cmdlines". 1129 1130 record-tgid 1131 When any event or tracer is enabled, a hook is enabled 1132 in the sched_switch trace point to fill the cache of 1133 mapped Thread Group IDs (TGID) mapping to pids. See 1134 "saved_tgids". 1135 1136 overwrite 1137 This controls what happens when the trace buffer is 1138 full. If "1" (default), the oldest events are 1139 discarded and overwritten. If "0", then the newest 1140 events are discarded. 1141 (see per_cpu/cpu0/stats for overrun and dropped) 1142 1143 disable_on_free 1144 When the free_buffer is closed, tracing will 1145 stop (tracing_on set to 0). 1146 1147 irq-info 1148 Shows the interrupt, preempt count, need resched data. 1149 When disabled, the trace looks like:: 1150 1151 # tracer: function 1152 # 1153 # entries-in-buffer/entries-written: 144405/9452052 #P:4 1154 # 1155 # TASK-PID CPU# TIMESTAMP FUNCTION 1156 # | | | | | 1157 <idle>-0 [002] 23636.756054: ttwu_do_activate.constprop.89 <-try_to_wake_up 1158 <idle>-0 [002] 23636.756054: activate_task <-ttwu_do_activate.constprop.89 1159 <idle>-0 [002] 23636.756055: enqueue_task <-activate_task 1160 1161 1162 markers 1163 When set, the trace_marker is writable (only by root). 1164 When disabled, the trace_marker will error with EINVAL 1165 on write. 1166 1167 event-fork 1168 When set, tasks with PIDs listed in set_event_pid will have 1169 the PIDs of their children added to set_event_pid when those 1170 tasks fork. Also, when tasks with PIDs in set_event_pid exit, 1171 their PIDs will be removed from the file. 1172 1173 function-trace 1174 The latency tracers will enable function tracing 1175 if this option is enabled (default it is). When 1176 it is disabled, the latency tracers do not trace 1177 functions. This keeps the overhead of the tracer down 1178 when performing latency tests. 1179 1180 function-fork 1181 When set, tasks with PIDs listed in set_ftrace_pid will 1182 have the PIDs of their children added to set_ftrace_pid 1183 when those tasks fork. Also, when tasks with PIDs in 1184 set_ftrace_pid exit, their PIDs will be removed from the 1185 file. 1186 1187 display-graph 1188 When set, the latency tracers (irqsoff, wakeup, etc) will 1189 use function graph tracing instead of function tracing. 1190 1191 stacktrace 1192 When set, a stack trace is recorded after any trace event 1193 is recorded. 1194 1195 branch 1196 Enable branch tracing with the tracer. This enables branch 1197 tracer along with the currently set tracer. Enabling this 1198 with the "nop" tracer is the same as just enabling the 1199 "branch" tracer. 1200 1201.. tip:: Some tracers have their own options. They only appear in this 1202 file when the tracer is active. They always appear in the 1203 options directory. 1204 1205 1206Here are the per tracer options: 1207 1208Options for function tracer: 1209 1210 func_stack_trace 1211 When set, a stack trace is recorded after every 1212 function that is recorded. NOTE! Limit the functions 1213 that are recorded before enabling this, with 1214 "set_ftrace_filter" otherwise the system performance 1215 will be critically degraded. Remember to disable 1216 this option before clearing the function filter. 1217 1218Options for function_graph tracer: 1219 1220 Since the function_graph tracer has a slightly different output 1221 it has its own options to control what is displayed. 1222 1223 funcgraph-overrun 1224 When set, the "overrun" of the graph stack is 1225 displayed after each function traced. The 1226 overrun, is when the stack depth of the calls 1227 is greater than what is reserved for each task. 1228 Each task has a fixed array of functions to 1229 trace in the call graph. If the depth of the 1230 calls exceeds that, the function is not traced. 1231 The overrun is the number of functions missed 1232 due to exceeding this array. 1233 1234 funcgraph-cpu 1235 When set, the CPU number of the CPU where the trace 1236 occurred is displayed. 1237 1238 funcgraph-overhead 1239 When set, if the function takes longer than 1240 A certain amount, then a delay marker is 1241 displayed. See "delay" above, under the 1242 header description. 1243 1244 funcgraph-proc 1245 Unlike other tracers, the process' command line 1246 is not displayed by default, but instead only 1247 when a task is traced in and out during a context 1248 switch. Enabling this options has the command 1249 of each process displayed at every line. 1250 1251 funcgraph-duration 1252 At the end of each function (the return) 1253 the duration of the amount of time in the 1254 function is displayed in microseconds. 1255 1256 funcgraph-abstime 1257 When set, the timestamp is displayed at each line. 1258 1259 funcgraph-irqs 1260 When disabled, functions that happen inside an 1261 interrupt will not be traced. 1262 1263 funcgraph-tail 1264 When set, the return event will include the function 1265 that it represents. By default this is off, and 1266 only a closing curly bracket "}" is displayed for 1267 the return of a function. 1268 1269 sleep-time 1270 When running function graph tracer, to include 1271 the time a task schedules out in its function. 1272 When enabled, it will account time the task has been 1273 scheduled out as part of the function call. 1274 1275 graph-time 1276 When running function profiler with function graph tracer, 1277 to include the time to call nested functions. When this is 1278 not set, the time reported for the function will only 1279 include the time the function itself executed for, not the 1280 time for functions that it called. 1281 1282Options for blk tracer: 1283 1284 blk_classic 1285 Shows a more minimalistic output. 1286 1287 1288irqsoff 1289------- 1290 1291When interrupts are disabled, the CPU can not react to any other 1292external event (besides NMIs and SMIs). This prevents the timer 1293interrupt from triggering or the mouse interrupt from letting 1294the kernel know of a new mouse event. The result is a latency 1295with the reaction time. 1296 1297The irqsoff tracer tracks the time for which interrupts are 1298disabled. When a new maximum latency is hit, the tracer saves 1299the trace leading up to that latency point so that every time a 1300new maximum is reached, the old saved trace is discarded and the 1301new trace is saved. 1302 1303To reset the maximum, echo 0 into tracing_max_latency. Here is 1304an example:: 1305 1306 # echo 0 > options/function-trace 1307 # echo irqsoff > current_tracer 1308 # echo 1 > tracing_on 1309 # echo 0 > tracing_max_latency 1310 # ls -ltr 1311 [...] 1312 # echo 0 > tracing_on 1313 # cat trace 1314 # tracer: irqsoff 1315 # 1316 # irqsoff latency trace v1.1.5 on 3.8.0-test+ 1317 # -------------------------------------------------------------------- 1318 # latency: 16 us, #4/4, CPU#0 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1319 # ----------------- 1320 # | task: swapper/0-0 (uid:0 nice:0 policy:0 rt_prio:0) 1321 # ----------------- 1322 # => started at: run_timer_softirq 1323 # => ended at: run_timer_softirq 1324 # 1325 # 1326 # _------=> CPU# 1327 # / _-----=> irqs-off 1328 # | / _----=> need-resched 1329 # || / _---=> hardirq/softirq 1330 # ||| / _--=> preempt-depth 1331 # |||| / delay 1332 # cmd pid ||||| time | caller 1333 # \ / ||||| \ | / 1334 <idle>-0 0d.s2 0us+: _raw_spin_lock_irq <-run_timer_softirq 1335 <idle>-0 0dNs3 17us : _raw_spin_unlock_irq <-run_timer_softirq 1336 <idle>-0 0dNs3 17us+: trace_hardirqs_on <-run_timer_softirq 1337 <idle>-0 0dNs3 25us : <stack trace> 1338 => _raw_spin_unlock_irq 1339 => run_timer_softirq 1340 => __do_softirq 1341 => call_softirq 1342 => do_softirq 1343 => irq_exit 1344 => smp_apic_timer_interrupt 1345 => apic_timer_interrupt 1346 => rcu_idle_exit 1347 => cpu_idle 1348 => rest_init 1349 => start_kernel 1350 => x86_64_start_reservations 1351 => x86_64_start_kernel 1352 1353Here we see that that we had a latency of 16 microseconds (which is 1354very good). The _raw_spin_lock_irq in run_timer_softirq disabled 1355interrupts. The difference between the 16 and the displayed 1356timestamp 25us occurred because the clock was incremented 1357between the time of recording the max latency and the time of 1358recording the function that had that latency. 1359 1360Note the above example had function-trace not set. If we set 1361function-trace, we get a much larger output:: 1362 1363 with echo 1 > options/function-trace 1364 1365 # tracer: irqsoff 1366 # 1367 # irqsoff latency trace v1.1.5 on 3.8.0-test+ 1368 # -------------------------------------------------------------------- 1369 # latency: 71 us, #168/168, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1370 # ----------------- 1371 # | task: bash-2042 (uid:0 nice:0 policy:0 rt_prio:0) 1372 # ----------------- 1373 # => started at: ata_scsi_queuecmd 1374 # => ended at: ata_scsi_queuecmd 1375 # 1376 # 1377 # _------=> CPU# 1378 # / _-----=> irqs-off 1379 # | / _----=> need-resched 1380 # || / _---=> hardirq/softirq 1381 # ||| / _--=> preempt-depth 1382 # |||| / delay 1383 # cmd pid ||||| time | caller 1384 # \ / ||||| \ | / 1385 bash-2042 3d... 0us : _raw_spin_lock_irqsave <-ata_scsi_queuecmd 1386 bash-2042 3d... 0us : add_preempt_count <-_raw_spin_lock_irqsave 1387 bash-2042 3d..1 1us : ata_scsi_find_dev <-ata_scsi_queuecmd 1388 bash-2042 3d..1 1us : __ata_scsi_find_dev <-ata_scsi_find_dev 1389 bash-2042 3d..1 2us : ata_find_dev.part.14 <-__ata_scsi_find_dev 1390 bash-2042 3d..1 2us : ata_qc_new_init <-__ata_scsi_queuecmd 1391 bash-2042 3d..1 3us : ata_sg_init <-__ata_scsi_queuecmd 1392 bash-2042 3d..1 4us : ata_scsi_rw_xlat <-__ata_scsi_queuecmd 1393 bash-2042 3d..1 4us : ata_build_rw_tf <-ata_scsi_rw_xlat 1394 [...] 1395 bash-2042 3d..1 67us : delay_tsc <-__delay 1396 bash-2042 3d..1 67us : add_preempt_count <-delay_tsc 1397 bash-2042 3d..2 67us : sub_preempt_count <-delay_tsc 1398 bash-2042 3d..1 67us : add_preempt_count <-delay_tsc 1399 bash-2042 3d..2 68us : sub_preempt_count <-delay_tsc 1400 bash-2042 3d..1 68us+: ata_bmdma_start <-ata_bmdma_qc_issue 1401 bash-2042 3d..1 71us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd 1402 bash-2042 3d..1 71us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd 1403 bash-2042 3d..1 72us+: trace_hardirqs_on <-ata_scsi_queuecmd 1404 bash-2042 3d..1 120us : <stack trace> 1405 => _raw_spin_unlock_irqrestore 1406 => ata_scsi_queuecmd 1407 => scsi_dispatch_cmd 1408 => scsi_request_fn 1409 => __blk_run_queue_uncond 1410 => __blk_run_queue 1411 => blk_queue_bio 1412 => generic_make_request 1413 => submit_bio 1414 => submit_bh 1415 => __ext3_get_inode_loc 1416 => ext3_iget 1417 => ext3_lookup 1418 => lookup_real 1419 => __lookup_hash 1420 => walk_component 1421 => lookup_last 1422 => path_lookupat 1423 => filename_lookup 1424 => user_path_at_empty 1425 => user_path_at 1426 => vfs_fstatat 1427 => vfs_stat 1428 => sys_newstat 1429 => system_call_fastpath 1430 1431 1432Here we traced a 71 microsecond latency. But we also see all the 1433functions that were called during that time. Note that by 1434enabling function tracing, we incur an added overhead. This 1435overhead may extend the latency times. But nevertheless, this 1436trace has provided some very helpful debugging information. 1437 1438If we prefer function graph output instead of function, we can set 1439display-graph option:: 1440 1441 with echo 1 > options/display-graph 1442 1443 # tracer: irqsoff 1444 # 1445 # irqsoff latency trace v1.1.5 on 4.20.0-rc6+ 1446 # -------------------------------------------------------------------- 1447 # latency: 3751 us, #274/274, CPU#0 | (M:desktop VP:0, KP:0, SP:0 HP:0 #P:4) 1448 # ----------------- 1449 # | task: bash-1507 (uid:0 nice:0 policy:0 rt_prio:0) 1450 # ----------------- 1451 # => started at: free_debug_processing 1452 # => ended at: return_to_handler 1453 # 1454 # 1455 # _-----=> irqs-off 1456 # / _----=> need-resched 1457 # | / _---=> hardirq/softirq 1458 # || / _--=> preempt-depth 1459 # ||| / 1460 # REL TIME CPU TASK/PID |||| DURATION FUNCTION CALLS 1461 # | | | | |||| | | | | | | 1462 0 us | 0) bash-1507 | d... | 0.000 us | _raw_spin_lock_irqsave(); 1463 0 us | 0) bash-1507 | d..1 | 0.378 us | do_raw_spin_trylock(); 1464 1 us | 0) bash-1507 | d..2 | | set_track() { 1465 2 us | 0) bash-1507 | d..2 | | save_stack_trace() { 1466 2 us | 0) bash-1507 | d..2 | | __save_stack_trace() { 1467 3 us | 0) bash-1507 | d..2 | | __unwind_start() { 1468 3 us | 0) bash-1507 | d..2 | | get_stack_info() { 1469 3 us | 0) bash-1507 | d..2 | 0.351 us | in_task_stack(); 1470 4 us | 0) bash-1507 | d..2 | 1.107 us | } 1471 [...] 1472 3750 us | 0) bash-1507 | d..1 | 0.516 us | do_raw_spin_unlock(); 1473 3750 us | 0) bash-1507 | d..1 | 0.000 us | _raw_spin_unlock_irqrestore(); 1474 3764 us | 0) bash-1507 | d..1 | 0.000 us | tracer_hardirqs_on(); 1475 bash-1507 0d..1 3792us : <stack trace> 1476 => free_debug_processing 1477 => __slab_free 1478 => kmem_cache_free 1479 => vm_area_free 1480 => remove_vma 1481 => exit_mmap 1482 => mmput 1483 => flush_old_exec 1484 => load_elf_binary 1485 => search_binary_handler 1486 => __do_execve_file.isra.32 1487 => __x64_sys_execve 1488 => do_syscall_64 1489 => entry_SYSCALL_64_after_hwframe 1490 1491preemptoff 1492---------- 1493 1494When preemption is disabled, we may be able to receive 1495interrupts but the task cannot be preempted and a higher 1496priority task must wait for preemption to be enabled again 1497before it can preempt a lower priority task. 1498 1499The preemptoff tracer traces the places that disable preemption. 1500Like the irqsoff tracer, it records the maximum latency for 1501which preemption was disabled. The control of preemptoff tracer 1502is much like the irqsoff tracer. 1503:: 1504 1505 # echo 0 > options/function-trace 1506 # echo preemptoff > current_tracer 1507 # echo 1 > tracing_on 1508 # echo 0 > tracing_max_latency 1509 # ls -ltr 1510 [...] 1511 # echo 0 > tracing_on 1512 # cat trace 1513 # tracer: preemptoff 1514 # 1515 # preemptoff latency trace v1.1.5 on 3.8.0-test+ 1516 # -------------------------------------------------------------------- 1517 # latency: 46 us, #4/4, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1518 # ----------------- 1519 # | task: sshd-1991 (uid:0 nice:0 policy:0 rt_prio:0) 1520 # ----------------- 1521 # => started at: do_IRQ 1522 # => ended at: do_IRQ 1523 # 1524 # 1525 # _------=> CPU# 1526 # / _-----=> irqs-off 1527 # | / _----=> need-resched 1528 # || / _---=> hardirq/softirq 1529 # ||| / _--=> preempt-depth 1530 # |||| / delay 1531 # cmd pid ||||| time | caller 1532 # \ / ||||| \ | / 1533 sshd-1991 1d.h. 0us+: irq_enter <-do_IRQ 1534 sshd-1991 1d..1 46us : irq_exit <-do_IRQ 1535 sshd-1991 1d..1 47us+: trace_preempt_on <-do_IRQ 1536 sshd-1991 1d..1 52us : <stack trace> 1537 => sub_preempt_count 1538 => irq_exit 1539 => do_IRQ 1540 => ret_from_intr 1541 1542 1543This has some more changes. Preemption was disabled when an 1544interrupt came in (notice the 'h'), and was enabled on exit. 1545But we also see that interrupts have been disabled when entering 1546the preempt off section and leaving it (the 'd'). We do not know if 1547interrupts were enabled in the mean time or shortly after this 1548was over. 1549:: 1550 1551 # tracer: preemptoff 1552 # 1553 # preemptoff latency trace v1.1.5 on 3.8.0-test+ 1554 # -------------------------------------------------------------------- 1555 # latency: 83 us, #241/241, CPU#1 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1556 # ----------------- 1557 # | task: bash-1994 (uid:0 nice:0 policy:0 rt_prio:0) 1558 # ----------------- 1559 # => started at: wake_up_new_task 1560 # => ended at: task_rq_unlock 1561 # 1562 # 1563 # _------=> CPU# 1564 # / _-----=> irqs-off 1565 # | / _----=> need-resched 1566 # || / _---=> hardirq/softirq 1567 # ||| / _--=> preempt-depth 1568 # |||| / delay 1569 # cmd pid ||||| time | caller 1570 # \ / ||||| \ | / 1571 bash-1994 1d..1 0us : _raw_spin_lock_irqsave <-wake_up_new_task 1572 bash-1994 1d..1 0us : select_task_rq_fair <-select_task_rq 1573 bash-1994 1d..1 1us : __rcu_read_lock <-select_task_rq_fair 1574 bash-1994 1d..1 1us : source_load <-select_task_rq_fair 1575 bash-1994 1d..1 1us : source_load <-select_task_rq_fair 1576 [...] 1577 bash-1994 1d..1 12us : irq_enter <-smp_apic_timer_interrupt 1578 bash-1994 1d..1 12us : rcu_irq_enter <-irq_enter 1579 bash-1994 1d..1 13us : add_preempt_count <-irq_enter 1580 bash-1994 1d.h1 13us : exit_idle <-smp_apic_timer_interrupt 1581 bash-1994 1d.h1 13us : hrtimer_interrupt <-smp_apic_timer_interrupt 1582 bash-1994 1d.h1 13us : _raw_spin_lock <-hrtimer_interrupt 1583 bash-1994 1d.h1 14us : add_preempt_count <-_raw_spin_lock 1584 bash-1994 1d.h2 14us : ktime_get_update_offsets <-hrtimer_interrupt 1585 [...] 1586 bash-1994 1d.h1 35us : lapic_next_event <-clockevents_program_event 1587 bash-1994 1d.h1 35us : irq_exit <-smp_apic_timer_interrupt 1588 bash-1994 1d.h1 36us : sub_preempt_count <-irq_exit 1589 bash-1994 1d..2 36us : do_softirq <-irq_exit 1590 bash-1994 1d..2 36us : __do_softirq <-call_softirq 1591 bash-1994 1d..2 36us : __local_bh_disable <-__do_softirq 1592 bash-1994 1d.s2 37us : add_preempt_count <-_raw_spin_lock_irq 1593 bash-1994 1d.s3 38us : _raw_spin_unlock <-run_timer_softirq 1594 bash-1994 1d.s3 39us : sub_preempt_count <-_raw_spin_unlock 1595 bash-1994 1d.s2 39us : call_timer_fn <-run_timer_softirq 1596 [...] 1597 bash-1994 1dNs2 81us : cpu_needs_another_gp <-rcu_process_callbacks 1598 bash-1994 1dNs2 82us : __local_bh_enable <-__do_softirq 1599 bash-1994 1dNs2 82us : sub_preempt_count <-__local_bh_enable 1600 bash-1994 1dN.2 82us : idle_cpu <-irq_exit 1601 bash-1994 1dN.2 83us : rcu_irq_exit <-irq_exit 1602 bash-1994 1dN.2 83us : sub_preempt_count <-irq_exit 1603 bash-1994 1.N.1 84us : _raw_spin_unlock_irqrestore <-task_rq_unlock 1604 bash-1994 1.N.1 84us+: trace_preempt_on <-task_rq_unlock 1605 bash-1994 1.N.1 104us : <stack trace> 1606 => sub_preempt_count 1607 => _raw_spin_unlock_irqrestore 1608 => task_rq_unlock 1609 => wake_up_new_task 1610 => do_fork 1611 => sys_clone 1612 => stub_clone 1613 1614 1615The above is an example of the preemptoff trace with 1616function-trace set. Here we see that interrupts were not disabled 1617the entire time. The irq_enter code lets us know that we entered 1618an interrupt 'h'. Before that, the functions being traced still 1619show that it is not in an interrupt, but we can see from the 1620functions themselves that this is not the case. 1621 1622preemptirqsoff 1623-------------- 1624 1625Knowing the locations that have interrupts disabled or 1626preemption disabled for the longest times is helpful. But 1627sometimes we would like to know when either preemption and/or 1628interrupts are disabled. 1629 1630Consider the following code:: 1631 1632 local_irq_disable(); 1633 call_function_with_irqs_off(); 1634 preempt_disable(); 1635 call_function_with_irqs_and_preemption_off(); 1636 local_irq_enable(); 1637 call_function_with_preemption_off(); 1638 preempt_enable(); 1639 1640The irqsoff tracer will record the total length of 1641call_function_with_irqs_off() and 1642call_function_with_irqs_and_preemption_off(). 1643 1644The preemptoff tracer will record the total length of 1645call_function_with_irqs_and_preemption_off() and 1646call_function_with_preemption_off(). 1647 1648But neither will trace the time that interrupts and/or 1649preemption is disabled. This total time is the time that we can 1650not schedule. To record this time, use the preemptirqsoff 1651tracer. 1652 1653Again, using this trace is much like the irqsoff and preemptoff 1654tracers. 1655:: 1656 1657 # echo 0 > options/function-trace 1658 # echo preemptirqsoff > current_tracer 1659 # echo 1 > tracing_on 1660 # echo 0 > tracing_max_latency 1661 # ls -ltr 1662 [...] 1663 # echo 0 > tracing_on 1664 # cat trace 1665 # tracer: preemptirqsoff 1666 # 1667 # preemptirqsoff latency trace v1.1.5 on 3.8.0-test+ 1668 # -------------------------------------------------------------------- 1669 # latency: 100 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1670 # ----------------- 1671 # | task: ls-2230 (uid:0 nice:0 policy:0 rt_prio:0) 1672 # ----------------- 1673 # => started at: ata_scsi_queuecmd 1674 # => ended at: ata_scsi_queuecmd 1675 # 1676 # 1677 # _------=> CPU# 1678 # / _-----=> irqs-off 1679 # | / _----=> need-resched 1680 # || / _---=> hardirq/softirq 1681 # ||| / _--=> preempt-depth 1682 # |||| / delay 1683 # cmd pid ||||| time | caller 1684 # \ / ||||| \ | / 1685 ls-2230 3d... 0us+: _raw_spin_lock_irqsave <-ata_scsi_queuecmd 1686 ls-2230 3...1 100us : _raw_spin_unlock_irqrestore <-ata_scsi_queuecmd 1687 ls-2230 3...1 101us+: trace_preempt_on <-ata_scsi_queuecmd 1688 ls-2230 3...1 111us : <stack trace> 1689 => sub_preempt_count 1690 => _raw_spin_unlock_irqrestore 1691 => ata_scsi_queuecmd 1692 => scsi_dispatch_cmd 1693 => scsi_request_fn 1694 => __blk_run_queue_uncond 1695 => __blk_run_queue 1696 => blk_queue_bio 1697 => generic_make_request 1698 => submit_bio 1699 => submit_bh 1700 => ext3_bread 1701 => ext3_dir_bread 1702 => htree_dirblock_to_tree 1703 => ext3_htree_fill_tree 1704 => ext3_readdir 1705 => vfs_readdir 1706 => sys_getdents 1707 => system_call_fastpath 1708 1709 1710The trace_hardirqs_off_thunk is called from assembly on x86 when 1711interrupts are disabled in the assembly code. Without the 1712function tracing, we do not know if interrupts were enabled 1713within the preemption points. We do see that it started with 1714preemption enabled. 1715 1716Here is a trace with function-trace set:: 1717 1718 # tracer: preemptirqsoff 1719 # 1720 # preemptirqsoff latency trace v1.1.5 on 3.8.0-test+ 1721 # -------------------------------------------------------------------- 1722 # latency: 161 us, #339/339, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1723 # ----------------- 1724 # | task: ls-2269 (uid:0 nice:0 policy:0 rt_prio:0) 1725 # ----------------- 1726 # => started at: schedule 1727 # => ended at: mutex_unlock 1728 # 1729 # 1730 # _------=> CPU# 1731 # / _-----=> irqs-off 1732 # | / _----=> need-resched 1733 # || / _---=> hardirq/softirq 1734 # ||| / _--=> preempt-depth 1735 # |||| / delay 1736 # cmd pid ||||| time | caller 1737 # \ / ||||| \ | / 1738 kworker/-59 3...1 0us : __schedule <-schedule 1739 kworker/-59 3d..1 0us : rcu_preempt_qs <-rcu_note_context_switch 1740 kworker/-59 3d..1 1us : add_preempt_count <-_raw_spin_lock_irq 1741 kworker/-59 3d..2 1us : deactivate_task <-__schedule 1742 kworker/-59 3d..2 1us : dequeue_task <-deactivate_task 1743 kworker/-59 3d..2 2us : update_rq_clock <-dequeue_task 1744 kworker/-59 3d..2 2us : dequeue_task_fair <-dequeue_task 1745 kworker/-59 3d..2 2us : update_curr <-dequeue_task_fair 1746 kworker/-59 3d..2 2us : update_min_vruntime <-update_curr 1747 kworker/-59 3d..2 3us : cpuacct_charge <-update_curr 1748 kworker/-59 3d..2 3us : __rcu_read_lock <-cpuacct_charge 1749 kworker/-59 3d..2 3us : __rcu_read_unlock <-cpuacct_charge 1750 kworker/-59 3d..2 3us : update_cfs_rq_blocked_load <-dequeue_task_fair 1751 kworker/-59 3d..2 4us : clear_buddies <-dequeue_task_fair 1752 kworker/-59 3d..2 4us : account_entity_dequeue <-dequeue_task_fair 1753 kworker/-59 3d..2 4us : update_min_vruntime <-dequeue_task_fair 1754 kworker/-59 3d..2 4us : update_cfs_shares <-dequeue_task_fair 1755 kworker/-59 3d..2 5us : hrtick_update <-dequeue_task_fair 1756 kworker/-59 3d..2 5us : wq_worker_sleeping <-__schedule 1757 kworker/-59 3d..2 5us : kthread_data <-wq_worker_sleeping 1758 kworker/-59 3d..2 5us : put_prev_task_fair <-__schedule 1759 kworker/-59 3d..2 6us : pick_next_task_fair <-pick_next_task 1760 kworker/-59 3d..2 6us : clear_buddies <-pick_next_task_fair 1761 kworker/-59 3d..2 6us : set_next_entity <-pick_next_task_fair 1762 kworker/-59 3d..2 6us : update_stats_wait_end <-set_next_entity 1763 ls-2269 3d..2 7us : finish_task_switch <-__schedule 1764 ls-2269 3d..2 7us : _raw_spin_unlock_irq <-finish_task_switch 1765 ls-2269 3d..2 8us : do_IRQ <-ret_from_intr 1766 ls-2269 3d..2 8us : irq_enter <-do_IRQ 1767 ls-2269 3d..2 8us : rcu_irq_enter <-irq_enter 1768 ls-2269 3d..2 9us : add_preempt_count <-irq_enter 1769 ls-2269 3d.h2 9us : exit_idle <-do_IRQ 1770 [...] 1771 ls-2269 3d.h3 20us : sub_preempt_count <-_raw_spin_unlock 1772 ls-2269 3d.h2 20us : irq_exit <-do_IRQ 1773 ls-2269 3d.h2 21us : sub_preempt_count <-irq_exit 1774 ls-2269 3d..3 21us : do_softirq <-irq_exit 1775 ls-2269 3d..3 21us : __do_softirq <-call_softirq 1776 ls-2269 3d..3 21us+: __local_bh_disable <-__do_softirq 1777 ls-2269 3d.s4 29us : sub_preempt_count <-_local_bh_enable_ip 1778 ls-2269 3d.s5 29us : sub_preempt_count <-_local_bh_enable_ip 1779 ls-2269 3d.s5 31us : do_IRQ <-ret_from_intr 1780 ls-2269 3d.s5 31us : irq_enter <-do_IRQ 1781 ls-2269 3d.s5 31us : rcu_irq_enter <-irq_enter 1782 [...] 1783 ls-2269 3d.s5 31us : rcu_irq_enter <-irq_enter 1784 ls-2269 3d.s5 32us : add_preempt_count <-irq_enter 1785 ls-2269 3d.H5 32us : exit_idle <-do_IRQ 1786 ls-2269 3d.H5 32us : handle_irq <-do_IRQ 1787 ls-2269 3d.H5 32us : irq_to_desc <-handle_irq 1788 ls-2269 3d.H5 33us : handle_fasteoi_irq <-handle_irq 1789 [...] 1790 ls-2269 3d.s5 158us : _raw_spin_unlock_irqrestore <-rtl8139_poll 1791 ls-2269 3d.s3 158us : net_rps_action_and_irq_enable.isra.65 <-net_rx_action 1792 ls-2269 3d.s3 159us : __local_bh_enable <-__do_softirq 1793 ls-2269 3d.s3 159us : sub_preempt_count <-__local_bh_enable 1794 ls-2269 3d..3 159us : idle_cpu <-irq_exit 1795 ls-2269 3d..3 159us : rcu_irq_exit <-irq_exit 1796 ls-2269 3d..3 160us : sub_preempt_count <-irq_exit 1797 ls-2269 3d... 161us : __mutex_unlock_slowpath <-mutex_unlock 1798 ls-2269 3d... 162us+: trace_hardirqs_on <-mutex_unlock 1799 ls-2269 3d... 186us : <stack trace> 1800 => __mutex_unlock_slowpath 1801 => mutex_unlock 1802 => process_output 1803 => n_tty_write 1804 => tty_write 1805 => vfs_write 1806 => sys_write 1807 => system_call_fastpath 1808 1809This is an interesting trace. It started with kworker running and 1810scheduling out and ls taking over. But as soon as ls released the 1811rq lock and enabled interrupts (but not preemption) an interrupt 1812triggered. When the interrupt finished, it started running softirqs. 1813But while the softirq was running, another interrupt triggered. 1814When an interrupt is running inside a softirq, the annotation is 'H'. 1815 1816 1817wakeup 1818------ 1819 1820One common case that people are interested in tracing is the 1821time it takes for a task that is woken to actually wake up. 1822Now for non Real-Time tasks, this can be arbitrary. But tracing 1823it none the less can be interesting. 1824 1825Without function tracing:: 1826 1827 # echo 0 > options/function-trace 1828 # echo wakeup > current_tracer 1829 # echo 1 > tracing_on 1830 # echo 0 > tracing_max_latency 1831 # chrt -f 5 sleep 1 1832 # echo 0 > tracing_on 1833 # cat trace 1834 # tracer: wakeup 1835 # 1836 # wakeup latency trace v1.1.5 on 3.8.0-test+ 1837 # -------------------------------------------------------------------- 1838 # latency: 15 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1839 # ----------------- 1840 # | task: kworker/3:1H-312 (uid:0 nice:-20 policy:0 rt_prio:0) 1841 # ----------------- 1842 # 1843 # _------=> CPU# 1844 # / _-----=> irqs-off 1845 # | / _----=> need-resched 1846 # || / _---=> hardirq/softirq 1847 # ||| / _--=> preempt-depth 1848 # |||| / delay 1849 # cmd pid ||||| time | caller 1850 # \ / ||||| \ | / 1851 <idle>-0 3dNs7 0us : 0:120:R + [003] 312:100:R kworker/3:1H 1852 <idle>-0 3dNs7 1us+: ttwu_do_activate.constprop.87 <-try_to_wake_up 1853 <idle>-0 3d..3 15us : __schedule <-schedule 1854 <idle>-0 3d..3 15us : 0:120:R ==> [003] 312:100:R kworker/3:1H 1855 1856The tracer only traces the highest priority task in the system 1857to avoid tracing the normal circumstances. Here we see that 1858the kworker with a nice priority of -20 (not very nice), took 1859just 15 microseconds from the time it woke up, to the time it 1860ran. 1861 1862Non Real-Time tasks are not that interesting. A more interesting 1863trace is to concentrate only on Real-Time tasks. 1864 1865wakeup_rt 1866--------- 1867 1868In a Real-Time environment it is very important to know the 1869wakeup time it takes for the highest priority task that is woken 1870up to the time that it executes. This is also known as "schedule 1871latency". I stress the point that this is about RT tasks. It is 1872also important to know the scheduling latency of non-RT tasks, 1873but the average schedule latency is better for non-RT tasks. 1874Tools like LatencyTop are more appropriate for such 1875measurements. 1876 1877Real-Time environments are interested in the worst case latency. 1878That is the longest latency it takes for something to happen, 1879and not the average. We can have a very fast scheduler that may 1880only have a large latency once in a while, but that would not 1881work well with Real-Time tasks. The wakeup_rt tracer was designed 1882to record the worst case wakeups of RT tasks. Non-RT tasks are 1883not recorded because the tracer only records one worst case and 1884tracing non-RT tasks that are unpredictable will overwrite the 1885worst case latency of RT tasks (just run the normal wakeup 1886tracer for a while to see that effect). 1887 1888Since this tracer only deals with RT tasks, we will run this 1889slightly differently than we did with the previous tracers. 1890Instead of performing an 'ls', we will run 'sleep 1' under 1891'chrt' which changes the priority of the task. 1892:: 1893 1894 # echo 0 > options/function-trace 1895 # echo wakeup_rt > current_tracer 1896 # echo 1 > tracing_on 1897 # echo 0 > tracing_max_latency 1898 # chrt -f 5 sleep 1 1899 # echo 0 > tracing_on 1900 # cat trace 1901 # tracer: wakeup 1902 # 1903 # tracer: wakeup_rt 1904 # 1905 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+ 1906 # -------------------------------------------------------------------- 1907 # latency: 5 us, #4/4, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1908 # ----------------- 1909 # | task: sleep-2389 (uid:0 nice:0 policy:1 rt_prio:5) 1910 # ----------------- 1911 # 1912 # _------=> CPU# 1913 # / _-----=> irqs-off 1914 # | / _----=> need-resched 1915 # || / _---=> hardirq/softirq 1916 # ||| / _--=> preempt-depth 1917 # |||| / delay 1918 # cmd pid ||||| time | caller 1919 # \ / ||||| \ | / 1920 <idle>-0 3d.h4 0us : 0:120:R + [003] 2389: 94:R sleep 1921 <idle>-0 3d.h4 1us+: ttwu_do_activate.constprop.87 <-try_to_wake_up 1922 <idle>-0 3d..3 5us : __schedule <-schedule 1923 <idle>-0 3d..3 5us : 0:120:R ==> [003] 2389: 94:R sleep 1924 1925 1926Running this on an idle system, we see that it only took 5 microseconds 1927to perform the task switch. Note, since the trace point in the schedule 1928is before the actual "switch", we stop the tracing when the recorded task 1929is about to schedule in. This may change if we add a new marker at the 1930end of the scheduler. 1931 1932Notice that the recorded task is 'sleep' with the PID of 2389 1933and it has an rt_prio of 5. This priority is user-space priority 1934and not the internal kernel priority. The policy is 1 for 1935SCHED_FIFO and 2 for SCHED_RR. 1936 1937Note, that the trace data shows the internal priority (99 - rtprio). 1938:: 1939 1940 <idle>-0 3d..3 5us : 0:120:R ==> [003] 2389: 94:R sleep 1941 1942The 0:120:R means idle was running with a nice priority of 0 (120 - 120) 1943and in the running state 'R'. The sleep task was scheduled in with 19442389: 94:R. That is the priority is the kernel rtprio (99 - 5 = 94) 1945and it too is in the running state. 1946 1947Doing the same with chrt -r 5 and function-trace set. 1948:: 1949 1950 echo 1 > options/function-trace 1951 1952 # tracer: wakeup_rt 1953 # 1954 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+ 1955 # -------------------------------------------------------------------- 1956 # latency: 29 us, #85/85, CPU#3 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 1957 # ----------------- 1958 # | task: sleep-2448 (uid:0 nice:0 policy:1 rt_prio:5) 1959 # ----------------- 1960 # 1961 # _------=> CPU# 1962 # / _-----=> irqs-off 1963 # | / _----=> need-resched 1964 # || / _---=> hardirq/softirq 1965 # ||| / _--=> preempt-depth 1966 # |||| / delay 1967 # cmd pid ||||| time | caller 1968 # \ / ||||| \ | / 1969 <idle>-0 3d.h4 1us+: 0:120:R + [003] 2448: 94:R sleep 1970 <idle>-0 3d.h4 2us : ttwu_do_activate.constprop.87 <-try_to_wake_up 1971 <idle>-0 3d.h3 3us : check_preempt_curr <-ttwu_do_wakeup 1972 <idle>-0 3d.h3 3us : resched_curr <-check_preempt_curr 1973 <idle>-0 3dNh3 4us : task_woken_rt <-ttwu_do_wakeup 1974 <idle>-0 3dNh3 4us : _raw_spin_unlock <-try_to_wake_up 1975 <idle>-0 3dNh3 4us : sub_preempt_count <-_raw_spin_unlock 1976 <idle>-0 3dNh2 5us : ttwu_stat <-try_to_wake_up 1977 <idle>-0 3dNh2 5us : _raw_spin_unlock_irqrestore <-try_to_wake_up 1978 <idle>-0 3dNh2 6us : sub_preempt_count <-_raw_spin_unlock_irqrestore 1979 <idle>-0 3dNh1 6us : _raw_spin_lock <-__run_hrtimer 1980 <idle>-0 3dNh1 6us : add_preempt_count <-_raw_spin_lock 1981 <idle>-0 3dNh2 7us : _raw_spin_unlock <-hrtimer_interrupt 1982 <idle>-0 3dNh2 7us : sub_preempt_count <-_raw_spin_unlock 1983 <idle>-0 3dNh1 7us : tick_program_event <-hrtimer_interrupt 1984 <idle>-0 3dNh1 7us : clockevents_program_event <-tick_program_event 1985 <idle>-0 3dNh1 8us : ktime_get <-clockevents_program_event 1986 <idle>-0 3dNh1 8us : lapic_next_event <-clockevents_program_event 1987 <idle>-0 3dNh1 8us : irq_exit <-smp_apic_timer_interrupt 1988 <idle>-0 3dNh1 9us : sub_preempt_count <-irq_exit 1989 <idle>-0 3dN.2 9us : idle_cpu <-irq_exit 1990 <idle>-0 3dN.2 9us : rcu_irq_exit <-irq_exit 1991 <idle>-0 3dN.2 10us : rcu_eqs_enter_common.isra.45 <-rcu_irq_exit 1992 <idle>-0 3dN.2 10us : sub_preempt_count <-irq_exit 1993 <idle>-0 3.N.1 11us : rcu_idle_exit <-cpu_idle 1994 <idle>-0 3dN.1 11us : rcu_eqs_exit_common.isra.43 <-rcu_idle_exit 1995 <idle>-0 3.N.1 11us : tick_nohz_idle_exit <-cpu_idle 1996 <idle>-0 3dN.1 12us : menu_hrtimer_cancel <-tick_nohz_idle_exit 1997 <idle>-0 3dN.1 12us : ktime_get <-tick_nohz_idle_exit 1998 <idle>-0 3dN.1 12us : tick_do_update_jiffies64 <-tick_nohz_idle_exit 1999 <idle>-0 3dN.1 13us : cpu_load_update_nohz <-tick_nohz_idle_exit 2000 <idle>-0 3dN.1 13us : _raw_spin_lock <-cpu_load_update_nohz 2001 <idle>-0 3dN.1 13us : add_preempt_count <-_raw_spin_lock 2002 <idle>-0 3dN.2 13us : __cpu_load_update <-cpu_load_update_nohz 2003 <idle>-0 3dN.2 14us : sched_avg_update <-__cpu_load_update 2004 <idle>-0 3dN.2 14us : _raw_spin_unlock <-cpu_load_update_nohz 2005 <idle>-0 3dN.2 14us : sub_preempt_count <-_raw_spin_unlock 2006 <idle>-0 3dN.1 15us : calc_load_nohz_stop <-tick_nohz_idle_exit 2007 <idle>-0 3dN.1 15us : touch_softlockup_watchdog <-tick_nohz_idle_exit 2008 <idle>-0 3dN.1 15us : hrtimer_cancel <-tick_nohz_idle_exit 2009 <idle>-0 3dN.1 15us : hrtimer_try_to_cancel <-hrtimer_cancel 2010 <idle>-0 3dN.1 16us : lock_hrtimer_base.isra.18 <-hrtimer_try_to_cancel 2011 <idle>-0 3dN.1 16us : _raw_spin_lock_irqsave <-lock_hrtimer_base.isra.18 2012 <idle>-0 3dN.1 16us : add_preempt_count <-_raw_spin_lock_irqsave 2013 <idle>-0 3dN.2 17us : __remove_hrtimer <-remove_hrtimer.part.16 2014 <idle>-0 3dN.2 17us : hrtimer_force_reprogram <-__remove_hrtimer 2015 <idle>-0 3dN.2 17us : tick_program_event <-hrtimer_force_reprogram 2016 <idle>-0 3dN.2 18us : clockevents_program_event <-tick_program_event 2017 <idle>-0 3dN.2 18us : ktime_get <-clockevents_program_event 2018 <idle>-0 3dN.2 18us : lapic_next_event <-clockevents_program_event 2019 <idle>-0 3dN.2 19us : _raw_spin_unlock_irqrestore <-hrtimer_try_to_cancel 2020 <idle>-0 3dN.2 19us : sub_preempt_count <-_raw_spin_unlock_irqrestore 2021 <idle>-0 3dN.1 19us : hrtimer_forward <-tick_nohz_idle_exit 2022 <idle>-0 3dN.1 20us : ktime_add_safe <-hrtimer_forward 2023 <idle>-0 3dN.1 20us : ktime_add_safe <-hrtimer_forward 2024 <idle>-0 3dN.1 20us : hrtimer_start_range_ns <-hrtimer_start_expires.constprop.11 2025 <idle>-0 3dN.1 20us : __hrtimer_start_range_ns <-hrtimer_start_range_ns 2026 <idle>-0 3dN.1 21us : lock_hrtimer_base.isra.18 <-__hrtimer_start_range_ns 2027 <idle>-0 3dN.1 21us : _raw_spin_lock_irqsave <-lock_hrtimer_base.isra.18 2028 <idle>-0 3dN.1 21us : add_preempt_count <-_raw_spin_lock_irqsave 2029 <idle>-0 3dN.2 22us : ktime_add_safe <-__hrtimer_start_range_ns 2030 <idle>-0 3dN.2 22us : enqueue_hrtimer <-__hrtimer_start_range_ns 2031 <idle>-0 3dN.2 22us : tick_program_event <-__hrtimer_start_range_ns 2032 <idle>-0 3dN.2 23us : clockevents_program_event <-tick_program_event 2033 <idle>-0 3dN.2 23us : ktime_get <-clockevents_program_event 2034 <idle>-0 3dN.2 23us : lapic_next_event <-clockevents_program_event 2035 <idle>-0 3dN.2 24us : _raw_spin_unlock_irqrestore <-__hrtimer_start_range_ns 2036 <idle>-0 3dN.2 24us : sub_preempt_count <-_raw_spin_unlock_irqrestore 2037 <idle>-0 3dN.1 24us : account_idle_ticks <-tick_nohz_idle_exit 2038 <idle>-0 3dN.1 24us : account_idle_time <-account_idle_ticks 2039 <idle>-0 3.N.1 25us : sub_preempt_count <-cpu_idle 2040 <idle>-0 3.N.. 25us : schedule <-cpu_idle 2041 <idle>-0 3.N.. 25us : __schedule <-preempt_schedule 2042 <idle>-0 3.N.. 26us : add_preempt_count <-__schedule 2043 <idle>-0 3.N.1 26us : rcu_note_context_switch <-__schedule 2044 <idle>-0 3.N.1 26us : rcu_sched_qs <-rcu_note_context_switch 2045 <idle>-0 3dN.1 27us : rcu_preempt_qs <-rcu_note_context_switch 2046 <idle>-0 3.N.1 27us : _raw_spin_lock_irq <-__schedule 2047 <idle>-0 3dN.1 27us : add_preempt_count <-_raw_spin_lock_irq 2048 <idle>-0 3dN.2 28us : put_prev_task_idle <-__schedule 2049 <idle>-0 3dN.2 28us : pick_next_task_stop <-pick_next_task 2050 <idle>-0 3dN.2 28us : pick_next_task_rt <-pick_next_task 2051 <idle>-0 3dN.2 29us : dequeue_pushable_task <-pick_next_task_rt 2052 <idle>-0 3d..3 29us : __schedule <-preempt_schedule 2053 <idle>-0 3d..3 30us : 0:120:R ==> [003] 2448: 94:R sleep 2054 2055This isn't that big of a trace, even with function tracing enabled, 2056so I included the entire trace. 2057 2058The interrupt went off while when the system was idle. Somewhere 2059before task_woken_rt() was called, the NEED_RESCHED flag was set, 2060this is indicated by the first occurrence of the 'N' flag. 2061 2062Latency tracing and events 2063-------------------------- 2064As function tracing can induce a much larger latency, but without 2065seeing what happens within the latency it is hard to know what 2066caused it. There is a middle ground, and that is with enabling 2067events. 2068:: 2069 2070 # echo 0 > options/function-trace 2071 # echo wakeup_rt > current_tracer 2072 # echo 1 > events/enable 2073 # echo 1 > tracing_on 2074 # echo 0 > tracing_max_latency 2075 # chrt -f 5 sleep 1 2076 # echo 0 > tracing_on 2077 # cat trace 2078 # tracer: wakeup_rt 2079 # 2080 # wakeup_rt latency trace v1.1.5 on 3.8.0-test+ 2081 # -------------------------------------------------------------------- 2082 # latency: 6 us, #12/12, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) 2083 # ----------------- 2084 # | task: sleep-5882 (uid:0 nice:0 policy:1 rt_prio:5) 2085 # ----------------- 2086 # 2087 # _------=> CPU# 2088 # / _-----=> irqs-off 2089 # | / _----=> need-resched 2090 # || / _---=> hardirq/softirq 2091 # ||| / _--=> preempt-depth 2092 # |||| / delay 2093 # cmd pid ||||| time | caller 2094 # \ / ||||| \ | / 2095 <idle>-0 2d.h4 0us : 0:120:R + [002] 5882: 94:R sleep 2096 <idle>-0 2d.h4 0us : ttwu_do_activate.constprop.87 <-try_to_wake_up 2097 <idle>-0 2d.h4 1us : sched_wakeup: comm=sleep pid=5882 prio=94 success=1 target_cpu=002 2098 <idle>-0 2dNh2 1us : hrtimer_expire_exit: hrtimer=ffff88007796feb8 2099 <idle>-0 2.N.2 2us : power_end: cpu_id=2 2100 <idle>-0 2.N.2 3us : cpu_idle: state=4294967295 cpu_id=2 2101 <idle>-0 2dN.3 4us : hrtimer_cancel: hrtimer=ffff88007d50d5e0 2102 <idle>-0 2dN.3 4us : hrtimer_start: hrtimer=ffff88007d50d5e0 function=tick_sched_timer expires=34311211000000 softexpires=34311211000000 2103 <idle>-0 2.N.2 5us : rcu_utilization: Start context switch 2104 <idle>-0 2.N.2 5us : rcu_utilization: End context switch 2105 <idle>-0 2d..3 6us : __schedule <-schedule 2106 <idle>-0 2d..3 6us : 0:120:R ==> [002] 5882: 94:R sleep 2107 2108 2109Hardware Latency Detector 2110------------------------- 2111 2112The hardware latency detector is executed by enabling the "hwlat" tracer. 2113 2114NOTE, this tracer will affect the performance of the system as it will 2115periodically make a CPU constantly busy with interrupts disabled. 2116:: 2117 2118 # echo hwlat > current_tracer 2119 # sleep 100 2120 # cat trace 2121 # tracer: hwlat 2122 # 2123 # _-----=> irqs-off 2124 # / _----=> need-resched 2125 # | / _---=> hardirq/softirq 2126 # || / _--=> preempt-depth 2127 # ||| / delay 2128 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 2129 # | | | |||| | | 2130 <...>-3638 [001] d... 19452.055471: #1 inner/outer(us): 12/14 ts:1499801089.066141940 2131 <...>-3638 [003] d... 19454.071354: #2 inner/outer(us): 11/9 ts:1499801091.082164365 2132 <...>-3638 [002] dn.. 19461.126852: #3 inner/outer(us): 12/9 ts:1499801098.138150062 2133 <...>-3638 [001] d... 19488.340960: #4 inner/outer(us): 8/12 ts:1499801125.354139633 2134 <...>-3638 [003] d... 19494.388553: #5 inner/outer(us): 8/12 ts:1499801131.402150961 2135 <...>-3638 [003] d... 19501.283419: #6 inner/outer(us): 0/12 ts:1499801138.297435289 nmi-total:4 nmi-count:1 2136 2137 2138The above output is somewhat the same in the header. All events will have 2139interrupts disabled 'd'. Under the FUNCTION title there is: 2140 2141 #1 2142 This is the count of events recorded that were greater than the 2143 tracing_threshold (See below). 2144 2145 inner/outer(us): 12/14 2146 2147 This shows two numbers as "inner latency" and "outer latency". The test 2148 runs in a loop checking a timestamp twice. The latency detected within 2149 the two timestamps is the "inner latency" and the latency detected 2150 after the previous timestamp and the next timestamp in the loop is 2151 the "outer latency". 2152 2153 ts:1499801089.066141940 2154 2155 The absolute timestamp that the event happened. 2156 2157 nmi-total:4 nmi-count:1 2158 2159 On architectures that support it, if an NMI comes in during the 2160 test, the time spent in NMI is reported in "nmi-total" (in 2161 microseconds). 2162 2163 All architectures that have NMIs will show the "nmi-count" if an 2164 NMI comes in during the test. 2165 2166hwlat files: 2167 2168 tracing_threshold 2169 This gets automatically set to "10" to represent 10 2170 microseconds. This is the threshold of latency that 2171 needs to be detected before the trace will be recorded. 2172 2173 Note, when hwlat tracer is finished (another tracer is 2174 written into "current_tracer"), the original value for 2175 tracing_threshold is placed back into this file. 2176 2177 hwlat_detector/width 2178 The length of time the test runs with interrupts disabled. 2179 2180 hwlat_detector/window 2181 The length of time of the window which the test 2182 runs. That is, the test will run for "width" 2183 microseconds per "window" microseconds 2184 2185 tracing_cpumask 2186 When the test is started. A kernel thread is created that 2187 runs the test. This thread will alternate between CPUs 2188 listed in the tracing_cpumask between each period 2189 (one "window"). To limit the test to specific CPUs 2190 set the mask in this file to only the CPUs that the test 2191 should run on. 2192 2193function 2194-------- 2195 2196This tracer is the function tracer. Enabling the function tracer 2197can be done from the debug file system. Make sure the 2198ftrace_enabled is set; otherwise this tracer is a nop. 2199See the "ftrace_enabled" section below. 2200:: 2201 2202 # sysctl kernel.ftrace_enabled=1 2203 # echo function > current_tracer 2204 # echo 1 > tracing_on 2205 # usleep 1 2206 # echo 0 > tracing_on 2207 # cat trace 2208 # tracer: function 2209 # 2210 # entries-in-buffer/entries-written: 24799/24799 #P:4 2211 # 2212 # _-----=> irqs-off 2213 # / _----=> need-resched 2214 # | / _---=> hardirq/softirq 2215 # || / _--=> preempt-depth 2216 # ||| / delay 2217 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 2218 # | | | |||| | | 2219 bash-1994 [002] .... 3082.063030: mutex_unlock <-rb_simple_write 2220 bash-1994 [002] .... 3082.063031: __mutex_unlock_slowpath <-mutex_unlock 2221 bash-1994 [002] .... 3082.063031: __fsnotify_parent <-fsnotify_modify 2222 bash-1994 [002] .... 3082.063032: fsnotify <-fsnotify_modify 2223 bash-1994 [002] .... 3082.063032: __srcu_read_lock <-fsnotify 2224 bash-1994 [002] .... 3082.063032: add_preempt_count <-__srcu_read_lock 2225 bash-1994 [002] ...1 3082.063032: sub_preempt_count <-__srcu_read_lock 2226 bash-1994 [002] .... 3082.063033: __srcu_read_unlock <-fsnotify 2227 [...] 2228 2229 2230Note: function tracer uses ring buffers to store the above 2231entries. The newest data may overwrite the oldest data. 2232Sometimes using echo to stop the trace is not sufficient because 2233the tracing could have overwritten the data that you wanted to 2234record. For this reason, it is sometimes better to disable 2235tracing directly from a program. This allows you to stop the 2236tracing at the point that you hit the part that you are 2237interested in. To disable the tracing directly from a C program, 2238something like following code snippet can be used:: 2239 2240 int trace_fd; 2241 [...] 2242 int main(int argc, char *argv[]) { 2243 [...] 2244 trace_fd = open(tracing_file("tracing_on"), O_WRONLY); 2245 [...] 2246 if (condition_hit()) { 2247 write(trace_fd, "0", 1); 2248 } 2249 [...] 2250 } 2251 2252 2253Single thread tracing 2254--------------------- 2255 2256By writing into set_ftrace_pid you can trace a 2257single thread. For example:: 2258 2259 # cat set_ftrace_pid 2260 no pid 2261 # echo 3111 > set_ftrace_pid 2262 # cat set_ftrace_pid 2263 3111 2264 # echo function > current_tracer 2265 # cat trace | head 2266 # tracer: function 2267 # 2268 # TASK-PID CPU# TIMESTAMP FUNCTION 2269 # | | | | | 2270 yum-updatesd-3111 [003] 1637.254676: finish_task_switch <-thread_return 2271 yum-updatesd-3111 [003] 1637.254681: hrtimer_cancel <-schedule_hrtimeout_range 2272 yum-updatesd-3111 [003] 1637.254682: hrtimer_try_to_cancel <-hrtimer_cancel 2273 yum-updatesd-3111 [003] 1637.254683: lock_hrtimer_base <-hrtimer_try_to_cancel 2274 yum-updatesd-3111 [003] 1637.254685: fget_light <-do_sys_poll 2275 yum-updatesd-3111 [003] 1637.254686: pipe_poll <-do_sys_poll 2276 # echo > set_ftrace_pid 2277 # cat trace |head 2278 # tracer: function 2279 # 2280 # TASK-PID CPU# TIMESTAMP FUNCTION 2281 # | | | | | 2282 ##### CPU 3 buffer started #### 2283 yum-updatesd-3111 [003] 1701.957688: free_poll_entry <-poll_freewait 2284 yum-updatesd-3111 [003] 1701.957689: remove_wait_queue <-free_poll_entry 2285 yum-updatesd-3111 [003] 1701.957691: fput <-free_poll_entry 2286 yum-updatesd-3111 [003] 1701.957692: audit_syscall_exit <-sysret_audit 2287 yum-updatesd-3111 [003] 1701.957693: path_put <-audit_syscall_exit 2288 2289If you want to trace a function when executing, you could use 2290something like this simple program. 2291:: 2292 2293 #include <stdio.h> 2294 #include <stdlib.h> 2295 #include <sys/types.h> 2296 #include <sys/stat.h> 2297 #include <fcntl.h> 2298 #include <unistd.h> 2299 #include <string.h> 2300 2301 #define _STR(x) #x 2302 #define STR(x) _STR(x) 2303 #define MAX_PATH 256 2304 2305 const char *find_tracefs(void) 2306 { 2307 static char tracefs[MAX_PATH+1]; 2308 static int tracefs_found; 2309 char type[100]; 2310 FILE *fp; 2311 2312 if (tracefs_found) 2313 return tracefs; 2314 2315 if ((fp = fopen("/proc/mounts","r")) == NULL) { 2316 perror("/proc/mounts"); 2317 return NULL; 2318 } 2319 2320 while (fscanf(fp, "%*s %" 2321 STR(MAX_PATH) 2322 "s %99s %*s %*d %*d\n", 2323 tracefs, type) == 2) { 2324 if (strcmp(type, "tracefs") == 0) 2325 break; 2326 } 2327 fclose(fp); 2328 2329 if (strcmp(type, "tracefs") != 0) { 2330 fprintf(stderr, "tracefs not mounted"); 2331 return NULL; 2332 } 2333 2334 strcat(tracefs, "/tracing/"); 2335 tracefs_found = 1; 2336 2337 return tracefs; 2338 } 2339 2340 const char *tracing_file(const char *file_name) 2341 { 2342 static char trace_file[MAX_PATH+1]; 2343 snprintf(trace_file, MAX_PATH, "%s/%s", find_tracefs(), file_name); 2344 return trace_file; 2345 } 2346 2347 int main (int argc, char **argv) 2348 { 2349 if (argc < 1) 2350 exit(-1); 2351 2352 if (fork() > 0) { 2353 int fd, ffd; 2354 char line[64]; 2355 int s; 2356 2357 ffd = open(tracing_file("current_tracer"), O_WRONLY); 2358 if (ffd < 0) 2359 exit(-1); 2360 write(ffd, "nop", 3); 2361 2362 fd = open(tracing_file("set_ftrace_pid"), O_WRONLY); 2363 s = sprintf(line, "%d\n", getpid()); 2364 write(fd, line, s); 2365 2366 write(ffd, "function", 8); 2367 2368 close(fd); 2369 close(ffd); 2370 2371 execvp(argv[1], argv+1); 2372 } 2373 2374 return 0; 2375 } 2376 2377Or this simple script! 2378:: 2379 2380 #!/bin/bash 2381 2382 tracefs=`sed -ne 's/^tracefs \(.*\) tracefs.*/\1/p' /proc/mounts` 2383 echo nop > $tracefs/tracing/current_tracer 2384 echo 0 > $tracefs/tracing/tracing_on 2385 echo $$ > $tracefs/tracing/set_ftrace_pid 2386 echo function > $tracefs/tracing/current_tracer 2387 echo 1 > $tracefs/tracing/tracing_on 2388 exec "$@" 2389 2390 2391function graph tracer 2392--------------------------- 2393 2394This tracer is similar to the function tracer except that it 2395probes a function on its entry and its exit. This is done by 2396using a dynamically allocated stack of return addresses in each 2397task_struct. On function entry the tracer overwrites the return 2398address of each function traced to set a custom probe. Thus the 2399original return address is stored on the stack of return address 2400in the task_struct. 2401 2402Probing on both ends of a function leads to special features 2403such as: 2404 2405- measure of a function's time execution 2406- having a reliable call stack to draw function calls graph 2407 2408This tracer is useful in several situations: 2409 2410- you want to find the reason of a strange kernel behavior and 2411 need to see what happens in detail on any areas (or specific 2412 ones). 2413 2414- you are experiencing weird latencies but it's difficult to 2415 find its origin. 2416 2417- you want to find quickly which path is taken by a specific 2418 function 2419 2420- you just want to peek inside a working kernel and want to see 2421 what happens there. 2422 2423:: 2424 2425 # tracer: function_graph 2426 # 2427 # CPU DURATION FUNCTION CALLS 2428 # | | | | | | | 2429 2430 0) | sys_open() { 2431 0) | do_sys_open() { 2432 0) | getname() { 2433 0) | kmem_cache_alloc() { 2434 0) 1.382 us | __might_sleep(); 2435 0) 2.478 us | } 2436 0) | strncpy_from_user() { 2437 0) | might_fault() { 2438 0) 1.389 us | __might_sleep(); 2439 0) 2.553 us | } 2440 0) 3.807 us | } 2441 0) 7.876 us | } 2442 0) | alloc_fd() { 2443 0) 0.668 us | _spin_lock(); 2444 0) 0.570 us | expand_files(); 2445 0) 0.586 us | _spin_unlock(); 2446 2447 2448There are several columns that can be dynamically 2449enabled/disabled. You can use every combination of options you 2450want, depending on your needs. 2451 2452- The cpu number on which the function executed is default 2453 enabled. It is sometimes better to only trace one cpu (see 2454 tracing_cpu_mask file) or you might sometimes see unordered 2455 function calls while cpu tracing switch. 2456 2457 - hide: echo nofuncgraph-cpu > trace_options 2458 - show: echo funcgraph-cpu > trace_options 2459 2460- The duration (function's time of execution) is displayed on 2461 the closing bracket line of a function or on the same line 2462 than the current function in case of a leaf one. It is default 2463 enabled. 2464 2465 - hide: echo nofuncgraph-duration > trace_options 2466 - show: echo funcgraph-duration > trace_options 2467 2468- The overhead field precedes the duration field in case of 2469 reached duration thresholds. 2470 2471 - hide: echo nofuncgraph-overhead > trace_options 2472 - show: echo funcgraph-overhead > trace_options 2473 - depends on: funcgraph-duration 2474 2475 ie:: 2476 2477 3) # 1837.709 us | } /* __switch_to */ 2478 3) | finish_task_switch() { 2479 3) 0.313 us | _raw_spin_unlock_irq(); 2480 3) 3.177 us | } 2481 3) # 1889.063 us | } /* __schedule */ 2482 3) ! 140.417 us | } /* __schedule */ 2483 3) # 2034.948 us | } /* schedule */ 2484 3) * 33998.59 us | } /* schedule_preempt_disabled */ 2485 2486 [...] 2487 2488 1) 0.260 us | msecs_to_jiffies(); 2489 1) 0.313 us | __rcu_read_unlock(); 2490 1) + 61.770 us | } 2491 1) + 64.479 us | } 2492 1) 0.313 us | rcu_bh_qs(); 2493 1) 0.313 us | __local_bh_enable(); 2494 1) ! 217.240 us | } 2495 1) 0.365 us | idle_cpu(); 2496 1) | rcu_irq_exit() { 2497 1) 0.417 us | rcu_eqs_enter_common.isra.47(); 2498 1) 3.125 us | } 2499 1) ! 227.812 us | } 2500 1) ! 457.395 us | } 2501 1) @ 119760.2 us | } 2502 2503 [...] 2504 2505 2) | handle_IPI() { 2506 1) 6.979 us | } 2507 2) 0.417 us | scheduler_ipi(); 2508 1) 9.791 us | } 2509 1) + 12.917 us | } 2510 2) 3.490 us | } 2511 1) + 15.729 us | } 2512 1) + 18.542 us | } 2513 2) $ 3594274 us | } 2514 2515Flags:: 2516 2517 + means that the function exceeded 10 usecs. 2518 ! means that the function exceeded 100 usecs. 2519 # means that the function exceeded 1000 usecs. 2520 * means that the function exceeded 10 msecs. 2521 @ means that the function exceeded 100 msecs. 2522 $ means that the function exceeded 1 sec. 2523 2524 2525- The task/pid field displays the thread cmdline and pid which 2526 executed the function. It is default disabled. 2527 2528 - hide: echo nofuncgraph-proc > trace_options 2529 - show: echo funcgraph-proc > trace_options 2530 2531 ie:: 2532 2533 # tracer: function_graph 2534 # 2535 # CPU TASK/PID DURATION FUNCTION CALLS 2536 # | | | | | | | | | 2537 0) sh-4802 | | d_free() { 2538 0) sh-4802 | | call_rcu() { 2539 0) sh-4802 | | __call_rcu() { 2540 0) sh-4802 | 0.616 us | rcu_process_gp_end(); 2541 0) sh-4802 | 0.586 us | check_for_new_grace_period(); 2542 0) sh-4802 | 2.899 us | } 2543 0) sh-4802 | 4.040 us | } 2544 0) sh-4802 | 5.151 us | } 2545 0) sh-4802 | + 49.370 us | } 2546 2547 2548- The absolute time field is an absolute timestamp given by the 2549 system clock since it started. A snapshot of this time is 2550 given on each entry/exit of functions 2551 2552 - hide: echo nofuncgraph-abstime > trace_options 2553 - show: echo funcgraph-abstime > trace_options 2554 2555 ie:: 2556 2557 # 2558 # TIME CPU DURATION FUNCTION CALLS 2559 # | | | | | | | | 2560 360.774522 | 1) 0.541 us | } 2561 360.774522 | 1) 4.663 us | } 2562 360.774523 | 1) 0.541 us | __wake_up_bit(); 2563 360.774524 | 1) 6.796 us | } 2564 360.774524 | 1) 7.952 us | } 2565 360.774525 | 1) 9.063 us | } 2566 360.774525 | 1) 0.615 us | journal_mark_dirty(); 2567 360.774527 | 1) 0.578 us | __brelse(); 2568 360.774528 | 1) | reiserfs_prepare_for_journal() { 2569 360.774528 | 1) | unlock_buffer() { 2570 360.774529 | 1) | wake_up_bit() { 2571 360.774529 | 1) | bit_waitqueue() { 2572 360.774530 | 1) 0.594 us | __phys_addr(); 2573 2574 2575The function name is always displayed after the closing bracket 2576for a function if the start of that function is not in the 2577trace buffer. 2578 2579Display of the function name after the closing bracket may be 2580enabled for functions whose start is in the trace buffer, 2581allowing easier searching with grep for function durations. 2582It is default disabled. 2583 2584 - hide: echo nofuncgraph-tail > trace_options 2585 - show: echo funcgraph-tail > trace_options 2586 2587 Example with nofuncgraph-tail (default):: 2588 2589 0) | putname() { 2590 0) | kmem_cache_free() { 2591 0) 0.518 us | __phys_addr(); 2592 0) 1.757 us | } 2593 0) 2.861 us | } 2594 2595 Example with funcgraph-tail:: 2596 2597 0) | putname() { 2598 0) | kmem_cache_free() { 2599 0) 0.518 us | __phys_addr(); 2600 0) 1.757 us | } /* kmem_cache_free() */ 2601 0) 2.861 us | } /* putname() */ 2602 2603You can put some comments on specific functions by using 2604trace_printk() For example, if you want to put a comment inside 2605the __might_sleep() function, you just have to include 2606<linux/ftrace.h> and call trace_printk() inside __might_sleep():: 2607 2608 trace_printk("I'm a comment!\n") 2609 2610will produce:: 2611 2612 1) | __might_sleep() { 2613 1) | /* I'm a comment! */ 2614 1) 1.449 us | } 2615 2616 2617You might find other useful features for this tracer in the 2618following "dynamic ftrace" section such as tracing only specific 2619functions or tasks. 2620 2621dynamic ftrace 2622-------------- 2623 2624If CONFIG_DYNAMIC_FTRACE is set, the system will run with 2625virtually no overhead when function tracing is disabled. The way 2626this works is the mcount function call (placed at the start of 2627every kernel function, produced by the -pg switch in gcc), 2628starts of pointing to a simple return. (Enabling FTRACE will 2629include the -pg switch in the compiling of the kernel.) 2630 2631At compile time every C file object is run through the 2632recordmcount program (located in the scripts directory). This 2633program will parse the ELF headers in the C object to find all 2634the locations in the .text section that call mcount. Starting 2635with gcc version 4.6, the -mfentry has been added for x86, which 2636calls "__fentry__" instead of "mcount". Which is called before 2637the creation of the stack frame. 2638 2639Note, not all sections are traced. They may be prevented by either 2640a notrace, or blocked another way and all inline functions are not 2641traced. Check the "available_filter_functions" file to see what functions 2642can be traced. 2643 2644A section called "__mcount_loc" is created that holds 2645references to all the mcount/fentry call sites in the .text section. 2646The recordmcount program re-links this section back into the 2647original object. The final linking stage of the kernel will add all these 2648references into a single table. 2649 2650On boot up, before SMP is initialized, the dynamic ftrace code 2651scans this table and updates all the locations into nops. It 2652also records the locations, which are added to the 2653available_filter_functions list. Modules are processed as they 2654are loaded and before they are executed. When a module is 2655unloaded, it also removes its functions from the ftrace function 2656list. This is automatic in the module unload code, and the 2657module author does not need to worry about it. 2658 2659When tracing is enabled, the process of modifying the function 2660tracepoints is dependent on architecture. The old method is to use 2661kstop_machine to prevent races with the CPUs executing code being 2662modified (which can cause the CPU to do undesirable things, especially 2663if the modified code crosses cache (or page) boundaries), and the nops are 2664patched back to calls. But this time, they do not call mcount 2665(which is just a function stub). They now call into the ftrace 2666infrastructure. 2667 2668The new method of modifying the function tracepoints is to place 2669a breakpoint at the location to be modified, sync all CPUs, modify 2670the rest of the instruction not covered by the breakpoint. Sync 2671all CPUs again, and then remove the breakpoint with the finished 2672version to the ftrace call site. 2673 2674Some archs do not even need to monkey around with the synchronization, 2675and can just slap the new code on top of the old without any 2676problems with other CPUs executing it at the same time. 2677 2678One special side-effect to the recording of the functions being 2679traced is that we can now selectively choose which functions we 2680wish to trace and which ones we want the mcount calls to remain 2681as nops. 2682 2683Two files are used, one for enabling and one for disabling the 2684tracing of specified functions. They are: 2685 2686 set_ftrace_filter 2687 2688and 2689 2690 set_ftrace_notrace 2691 2692A list of available functions that you can add to these files is 2693listed in: 2694 2695 available_filter_functions 2696 2697:: 2698 2699 # cat available_filter_functions 2700 put_prev_task_idle 2701 kmem_cache_create 2702 pick_next_task_rt 2703 get_online_cpus 2704 pick_next_task_fair 2705 mutex_lock 2706 [...] 2707 2708If I am only interested in sys_nanosleep and hrtimer_interrupt:: 2709 2710 # echo sys_nanosleep hrtimer_interrupt > set_ftrace_filter 2711 # echo function > current_tracer 2712 # echo 1 > tracing_on 2713 # usleep 1 2714 # echo 0 > tracing_on 2715 # cat trace 2716 # tracer: function 2717 # 2718 # entries-in-buffer/entries-written: 5/5 #P:4 2719 # 2720 # _-----=> irqs-off 2721 # / _----=> need-resched 2722 # | / _---=> hardirq/softirq 2723 # || / _--=> preempt-depth 2724 # ||| / delay 2725 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 2726 # | | | |||| | | 2727 usleep-2665 [001] .... 4186.475355: sys_nanosleep <-system_call_fastpath 2728 <idle>-0 [001] d.h1 4186.475409: hrtimer_interrupt <-smp_apic_timer_interrupt 2729 usleep-2665 [001] d.h1 4186.475426: hrtimer_interrupt <-smp_apic_timer_interrupt 2730 <idle>-0 [003] d.h1 4186.475426: hrtimer_interrupt <-smp_apic_timer_interrupt 2731 <idle>-0 [002] d.h1 4186.475427: hrtimer_interrupt <-smp_apic_timer_interrupt 2732 2733To see which functions are being traced, you can cat the file: 2734:: 2735 2736 # cat set_ftrace_filter 2737 hrtimer_interrupt 2738 sys_nanosleep 2739 2740 2741Perhaps this is not enough. The filters also allow glob(7) matching. 2742 2743 ``<match>*`` 2744 will match functions that begin with <match> 2745 ``*<match>`` 2746 will match functions that end with <match> 2747 ``*<match>*`` 2748 will match functions that have <match> in it 2749 ``<match1>*<match2>`` 2750 will match functions that begin with <match1> and end with <match2> 2751 2752.. note:: 2753 It is better to use quotes to enclose the wild cards, 2754 otherwise the shell may expand the parameters into names 2755 of files in the local directory. 2756 2757:: 2758 2759 # echo 'hrtimer_*' > set_ftrace_filter 2760 2761Produces:: 2762 2763 # tracer: function 2764 # 2765 # entries-in-buffer/entries-written: 897/897 #P:4 2766 # 2767 # _-----=> irqs-off 2768 # / _----=> need-resched 2769 # | / _---=> hardirq/softirq 2770 # || / _--=> preempt-depth 2771 # ||| / delay 2772 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 2773 # | | | |||| | | 2774 <idle>-0 [003] dN.1 4228.547803: hrtimer_cancel <-tick_nohz_idle_exit 2775 <idle>-0 [003] dN.1 4228.547804: hrtimer_try_to_cancel <-hrtimer_cancel 2776 <idle>-0 [003] dN.2 4228.547805: hrtimer_force_reprogram <-__remove_hrtimer 2777 <idle>-0 [003] dN.1 4228.547805: hrtimer_forward <-tick_nohz_idle_exit 2778 <idle>-0 [003] dN.1 4228.547805: hrtimer_start_range_ns <-hrtimer_start_expires.constprop.11 2779 <idle>-0 [003] d..1 4228.547858: hrtimer_get_next_event <-get_next_timer_interrupt 2780 <idle>-0 [003] d..1 4228.547859: hrtimer_start <-__tick_nohz_idle_enter 2781 <idle>-0 [003] d..2 4228.547860: hrtimer_force_reprogram <-__rem 2782 2783Notice that we lost the sys_nanosleep. 2784:: 2785 2786 # cat set_ftrace_filter 2787 hrtimer_run_queues 2788 hrtimer_run_pending 2789 hrtimer_init 2790 hrtimer_cancel 2791 hrtimer_try_to_cancel 2792 hrtimer_forward 2793 hrtimer_start 2794 hrtimer_reprogram 2795 hrtimer_force_reprogram 2796 hrtimer_get_next_event 2797 hrtimer_interrupt 2798 hrtimer_nanosleep 2799 hrtimer_wakeup 2800 hrtimer_get_remaining 2801 hrtimer_get_res 2802 hrtimer_init_sleeper 2803 2804 2805This is because the '>' and '>>' act just like they do in bash. 2806To rewrite the filters, use '>' 2807To append to the filters, use '>>' 2808 2809To clear out a filter so that all functions will be recorded 2810again:: 2811 2812 # echo > set_ftrace_filter 2813 # cat set_ftrace_filter 2814 # 2815 2816Again, now we want to append. 2817 2818:: 2819 2820 # echo sys_nanosleep > set_ftrace_filter 2821 # cat set_ftrace_filter 2822 sys_nanosleep 2823 # echo 'hrtimer_*' >> set_ftrace_filter 2824 # cat set_ftrace_filter 2825 hrtimer_run_queues 2826 hrtimer_run_pending 2827 hrtimer_init 2828 hrtimer_cancel 2829 hrtimer_try_to_cancel 2830 hrtimer_forward 2831 hrtimer_start 2832 hrtimer_reprogram 2833 hrtimer_force_reprogram 2834 hrtimer_get_next_event 2835 hrtimer_interrupt 2836 sys_nanosleep 2837 hrtimer_nanosleep 2838 hrtimer_wakeup 2839 hrtimer_get_remaining 2840 hrtimer_get_res 2841 hrtimer_init_sleeper 2842 2843 2844The set_ftrace_notrace prevents those functions from being 2845traced. 2846:: 2847 2848 # echo '*preempt*' '*lock*' > set_ftrace_notrace 2849 2850Produces:: 2851 2852 # tracer: function 2853 # 2854 # entries-in-buffer/entries-written: 39608/39608 #P:4 2855 # 2856 # _-----=> irqs-off 2857 # / _----=> need-resched 2858 # | / _---=> hardirq/softirq 2859 # || / _--=> preempt-depth 2860 # ||| / delay 2861 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 2862 # | | | |||| | | 2863 bash-1994 [000] .... 4342.324896: file_ra_state_init <-do_dentry_open 2864 bash-1994 [000] .... 4342.324897: open_check_o_direct <-do_last 2865 bash-1994 [000] .... 4342.324897: ima_file_check <-do_last 2866 bash-1994 [000] .... 4342.324898: process_measurement <-ima_file_check 2867 bash-1994 [000] .... 4342.324898: ima_get_action <-process_measurement 2868 bash-1994 [000] .... 4342.324898: ima_match_policy <-ima_get_action 2869 bash-1994 [000] .... 4342.324899: do_truncate <-do_last 2870 bash-1994 [000] .... 4342.324899: should_remove_suid <-do_truncate 2871 bash-1994 [000] .... 4342.324899: notify_change <-do_truncate 2872 bash-1994 [000] .... 4342.324900: current_fs_time <-notify_change 2873 bash-1994 [000] .... 4342.324900: current_kernel_time <-current_fs_time 2874 bash-1994 [000] .... 4342.324900: timespec_trunc <-current_fs_time 2875 2876We can see that there's no more lock or preempt tracing. 2877 2878Selecting function filters via index 2879------------------------------------ 2880 2881Because processing of strings is expensive (the address of the function 2882needs to be looked up before comparing to the string being passed in), 2883an index can be used as well to enable functions. This is useful in the 2884case of setting thousands of specific functions at a time. By passing 2885in a list of numbers, no string processing will occur. Instead, the function 2886at the specific location in the internal array (which corresponds to the 2887functions in the "available_filter_functions" file), is selected. 2888 2889:: 2890 2891 # echo 1 > set_ftrace_filter 2892 2893Will select the first function listed in "available_filter_functions" 2894 2895:: 2896 2897 # head -1 available_filter_functions 2898 trace_initcall_finish_cb 2899 2900 # cat set_ftrace_filter 2901 trace_initcall_finish_cb 2902 2903 # head -50 available_filter_functions | tail -1 2904 x86_pmu_commit_txn 2905 2906 # echo 1 50 > set_ftrace_filter 2907 # cat set_ftrace_filter 2908 trace_initcall_finish_cb 2909 x86_pmu_commit_txn 2910 2911Dynamic ftrace with the function graph tracer 2912--------------------------------------------- 2913 2914Although what has been explained above concerns both the 2915function tracer and the function-graph-tracer, there are some 2916special features only available in the function-graph tracer. 2917 2918If you want to trace only one function and all of its children, 2919you just have to echo its name into set_graph_function:: 2920 2921 echo __do_fault > set_graph_function 2922 2923will produce the following "expanded" trace of the __do_fault() 2924function:: 2925 2926 0) | __do_fault() { 2927 0) | filemap_fault() { 2928 0) | find_lock_page() { 2929 0) 0.804 us | find_get_page(); 2930 0) | __might_sleep() { 2931 0) 1.329 us | } 2932 0) 3.904 us | } 2933 0) 4.979 us | } 2934 0) 0.653 us | _spin_lock(); 2935 0) 0.578 us | page_add_file_rmap(); 2936 0) 0.525 us | native_set_pte_at(); 2937 0) 0.585 us | _spin_unlock(); 2938 0) | unlock_page() { 2939 0) 0.541 us | page_waitqueue(); 2940 0) 0.639 us | __wake_up_bit(); 2941 0) 2.786 us | } 2942 0) + 14.237 us | } 2943 0) | __do_fault() { 2944 0) | filemap_fault() { 2945 0) | find_lock_page() { 2946 0) 0.698 us | find_get_page(); 2947 0) | __might_sleep() { 2948 0) 1.412 us | } 2949 0) 3.950 us | } 2950 0) 5.098 us | } 2951 0) 0.631 us | _spin_lock(); 2952 0) 0.571 us | page_add_file_rmap(); 2953 0) 0.526 us | native_set_pte_at(); 2954 0) 0.586 us | _spin_unlock(); 2955 0) | unlock_page() { 2956 0) 0.533 us | page_waitqueue(); 2957 0) 0.638 us | __wake_up_bit(); 2958 0) 2.793 us | } 2959 0) + 14.012 us | } 2960 2961You can also expand several functions at once:: 2962 2963 echo sys_open > set_graph_function 2964 echo sys_close >> set_graph_function 2965 2966Now if you want to go back to trace all functions you can clear 2967this special filter via:: 2968 2969 echo > set_graph_function 2970 2971 2972ftrace_enabled 2973-------------- 2974 2975Note, the proc sysctl ftrace_enable is a big on/off switch for the 2976function tracer. By default it is enabled (when function tracing is 2977enabled in the kernel). If it is disabled, all function tracing is 2978disabled. This includes not only the function tracers for ftrace, but 2979also for any other uses (perf, kprobes, stack tracing, profiling, etc). 2980 2981Please disable this with care. 2982 2983This can be disable (and enabled) with:: 2984 2985 sysctl kernel.ftrace_enabled=0 2986 sysctl kernel.ftrace_enabled=1 2987 2988 or 2989 2990 echo 0 > /proc/sys/kernel/ftrace_enabled 2991 echo 1 > /proc/sys/kernel/ftrace_enabled 2992 2993 2994Filter commands 2995--------------- 2996 2997A few commands are supported by the set_ftrace_filter interface. 2998Trace commands have the following format:: 2999 3000 <function>:<command>:<parameter> 3001 3002The following commands are supported: 3003 3004- mod: 3005 This command enables function filtering per module. The 3006 parameter defines the module. For example, if only the write* 3007 functions in the ext3 module are desired, run: 3008 3009 echo 'write*:mod:ext3' > set_ftrace_filter 3010 3011 This command interacts with the filter in the same way as 3012 filtering based on function names. Thus, adding more functions 3013 in a different module is accomplished by appending (>>) to the 3014 filter file. Remove specific module functions by prepending 3015 '!':: 3016 3017 echo '!writeback*:mod:ext3' >> set_ftrace_filter 3018 3019 Mod command supports module globbing. Disable tracing for all 3020 functions except a specific module:: 3021 3022 echo '!*:mod:!ext3' >> set_ftrace_filter 3023 3024 Disable tracing for all modules, but still trace kernel:: 3025 3026 echo '!*:mod:*' >> set_ftrace_filter 3027 3028 Enable filter only for kernel:: 3029 3030 echo '*write*:mod:!*' >> set_ftrace_filter 3031 3032 Enable filter for module globbing:: 3033 3034 echo '*write*:mod:*snd*' >> set_ftrace_filter 3035 3036- traceon/traceoff: 3037 These commands turn tracing on and off when the specified 3038 functions are hit. The parameter determines how many times the 3039 tracing system is turned on and off. If unspecified, there is 3040 no limit. For example, to disable tracing when a schedule bug 3041 is hit the first 5 times, run:: 3042 3043 echo '__schedule_bug:traceoff:5' > set_ftrace_filter 3044 3045 To always disable tracing when __schedule_bug is hit:: 3046 3047 echo '__schedule_bug:traceoff' > set_ftrace_filter 3048 3049 These commands are cumulative whether or not they are appended 3050 to set_ftrace_filter. To remove a command, prepend it by '!' 3051 and drop the parameter:: 3052 3053 echo '!__schedule_bug:traceoff:0' > set_ftrace_filter 3054 3055 The above removes the traceoff command for __schedule_bug 3056 that have a counter. To remove commands without counters:: 3057 3058 echo '!__schedule_bug:traceoff' > set_ftrace_filter 3059 3060- snapshot: 3061 Will cause a snapshot to be triggered when the function is hit. 3062 :: 3063 3064 echo 'native_flush_tlb_others:snapshot' > set_ftrace_filter 3065 3066 To only snapshot once: 3067 :: 3068 3069 echo 'native_flush_tlb_others:snapshot:1' > set_ftrace_filter 3070 3071 To remove the above commands:: 3072 3073 echo '!native_flush_tlb_others:snapshot' > set_ftrace_filter 3074 echo '!native_flush_tlb_others:snapshot:0' > set_ftrace_filter 3075 3076- enable_event/disable_event: 3077 These commands can enable or disable a trace event. Note, because 3078 function tracing callbacks are very sensitive, when these commands 3079 are registered, the trace point is activated, but disabled in 3080 a "soft" mode. That is, the tracepoint will be called, but 3081 just will not be traced. The event tracepoint stays in this mode 3082 as long as there's a command that triggers it. 3083 :: 3084 3085 echo 'try_to_wake_up:enable_event:sched:sched_switch:2' > \ 3086 set_ftrace_filter 3087 3088 The format is:: 3089 3090 <function>:enable_event:<system>:<event>[:count] 3091 <function>:disable_event:<system>:<event>[:count] 3092 3093 To remove the events commands:: 3094 3095 echo '!try_to_wake_up:enable_event:sched:sched_switch:0' > \ 3096 set_ftrace_filter 3097 echo '!schedule:disable_event:sched:sched_switch' > \ 3098 set_ftrace_filter 3099 3100- dump: 3101 When the function is hit, it will dump the contents of the ftrace 3102 ring buffer to the console. This is useful if you need to debug 3103 something, and want to dump the trace when a certain function 3104 is hit. Perhaps it's a function that is called before a triple 3105 fault happens and does not allow you to get a regular dump. 3106 3107- cpudump: 3108 When the function is hit, it will dump the contents of the ftrace 3109 ring buffer for the current CPU to the console. Unlike the "dump" 3110 command, it only prints out the contents of the ring buffer for the 3111 CPU that executed the function that triggered the dump. 3112 3113- stacktrace: 3114 When the function is hit, a stack trace is recorded. 3115 3116trace_pipe 3117---------- 3118 3119The trace_pipe outputs the same content as the trace file, but 3120the effect on the tracing is different. Every read from 3121trace_pipe is consumed. This means that subsequent reads will be 3122different. The trace is live. 3123:: 3124 3125 # echo function > current_tracer 3126 # cat trace_pipe > /tmp/trace.out & 3127 [1] 4153 3128 # echo 1 > tracing_on 3129 # usleep 1 3130 # echo 0 > tracing_on 3131 # cat trace 3132 # tracer: function 3133 # 3134 # entries-in-buffer/entries-written: 0/0 #P:4 3135 # 3136 # _-----=> irqs-off 3137 # / _----=> need-resched 3138 # | / _---=> hardirq/softirq 3139 # || / _--=> preempt-depth 3140 # ||| / delay 3141 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 3142 # | | | |||| | | 3143 3144 # 3145 # cat /tmp/trace.out 3146 bash-1994 [000] .... 5281.568961: mutex_unlock <-rb_simple_write 3147 bash-1994 [000] .... 5281.568963: __mutex_unlock_slowpath <-mutex_unlock 3148 bash-1994 [000] .... 5281.568963: __fsnotify_parent <-fsnotify_modify 3149 bash-1994 [000] .... 5281.568964: fsnotify <-fsnotify_modify 3150 bash-1994 [000] .... 5281.568964: __srcu_read_lock <-fsnotify 3151 bash-1994 [000] .... 5281.568964: add_preempt_count <-__srcu_read_lock 3152 bash-1994 [000] ...1 5281.568965: sub_preempt_count <-__srcu_read_lock 3153 bash-1994 [000] .... 5281.568965: __srcu_read_unlock <-fsnotify 3154 bash-1994 [000] .... 5281.568967: sys_dup2 <-system_call_fastpath 3155 3156 3157Note, reading the trace_pipe file will block until more input is 3158added. This is contrary to the trace file. If any process opened 3159the trace file for reading, it will actually disable tracing and 3160prevent new entries from being added. The trace_pipe file does 3161not have this limitation. 3162 3163trace entries 3164------------- 3165 3166Having too much or not enough data can be troublesome in 3167diagnosing an issue in the kernel. The file buffer_size_kb is 3168used to modify the size of the internal trace buffers. The 3169number listed is the number of entries that can be recorded per 3170CPU. To know the full size, multiply the number of possible CPUs 3171with the number of entries. 3172:: 3173 3174 # cat buffer_size_kb 3175 1408 (units kilobytes) 3176 3177Or simply read buffer_total_size_kb 3178:: 3179 3180 # cat buffer_total_size_kb 3181 5632 3182 3183To modify the buffer, simple echo in a number (in 1024 byte segments). 3184:: 3185 3186 # echo 10000 > buffer_size_kb 3187 # cat buffer_size_kb 3188 10000 (units kilobytes) 3189 3190It will try to allocate as much as possible. If you allocate too 3191much, it can cause Out-Of-Memory to trigger. 3192:: 3193 3194 # echo 1000000000000 > buffer_size_kb 3195 -bash: echo: write error: Cannot allocate memory 3196 # cat buffer_size_kb 3197 85 3198 3199The per_cpu buffers can be changed individually as well: 3200:: 3201 3202 # echo 10000 > per_cpu/cpu0/buffer_size_kb 3203 # echo 100 > per_cpu/cpu1/buffer_size_kb 3204 3205When the per_cpu buffers are not the same, the buffer_size_kb 3206at the top level will just show an X 3207:: 3208 3209 # cat buffer_size_kb 3210 X 3211 3212This is where the buffer_total_size_kb is useful: 3213:: 3214 3215 # cat buffer_total_size_kb 3216 12916 3217 3218Writing to the top level buffer_size_kb will reset all the buffers 3219to be the same again. 3220 3221Snapshot 3222-------- 3223CONFIG_TRACER_SNAPSHOT makes a generic snapshot feature 3224available to all non latency tracers. (Latency tracers which 3225record max latency, such as "irqsoff" or "wakeup", can't use 3226this feature, since those are already using the snapshot 3227mechanism internally.) 3228 3229Snapshot preserves a current trace buffer at a particular point 3230in time without stopping tracing. Ftrace swaps the current 3231buffer with a spare buffer, and tracing continues in the new 3232current (=previous spare) buffer. 3233 3234The following tracefs files in "tracing" are related to this 3235feature: 3236 3237 snapshot: 3238 3239 This is used to take a snapshot and to read the output 3240 of the snapshot. Echo 1 into this file to allocate a 3241 spare buffer and to take a snapshot (swap), then read 3242 the snapshot from this file in the same format as 3243 "trace" (described above in the section "The File 3244 System"). Both reads snapshot and tracing are executable 3245 in parallel. When the spare buffer is allocated, echoing 3246 0 frees it, and echoing else (positive) values clear the 3247 snapshot contents. 3248 More details are shown in the table below. 3249 3250 +--------------+------------+------------+------------+ 3251 |status\\input | 0 | 1 | else | 3252 +==============+============+============+============+ 3253 |not allocated |(do nothing)| alloc+swap |(do nothing)| 3254 +--------------+------------+------------+------------+ 3255 |allocated | free | swap | clear | 3256 +--------------+------------+------------+------------+ 3257 3258Here is an example of using the snapshot feature. 3259:: 3260 3261 # echo 1 > events/sched/enable 3262 # echo 1 > snapshot 3263 # cat snapshot 3264 # tracer: nop 3265 # 3266 # entries-in-buffer/entries-written: 71/71 #P:8 3267 # 3268 # _-----=> irqs-off 3269 # / _----=> need-resched 3270 # | / _---=> hardirq/softirq 3271 # || / _--=> preempt-depth 3272 # ||| / delay 3273 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 3274 # | | | |||| | | 3275 <idle>-0 [005] d... 2440.603828: sched_switch: prev_comm=swapper/5 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=snapshot-test-2 next_pid=2242 next_prio=120 3276 sleep-2242 [005] d... 2440.603846: sched_switch: prev_comm=snapshot-test-2 prev_pid=2242 prev_prio=120 prev_state=R ==> next_comm=kworker/5:1 next_pid=60 next_prio=120 3277 [...] 3278 <idle>-0 [002] d... 2440.707230: sched_switch: prev_comm=swapper/2 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=snapshot-test-2 next_pid=2229 next_prio=120 3279 3280 # cat trace 3281 # tracer: nop 3282 # 3283 # entries-in-buffer/entries-written: 77/77 #P:8 3284 # 3285 # _-----=> irqs-off 3286 # / _----=> need-resched 3287 # | / _---=> hardirq/softirq 3288 # || / _--=> preempt-depth 3289 # ||| / delay 3290 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 3291 # | | | |||| | | 3292 <idle>-0 [007] d... 2440.707395: sched_switch: prev_comm=swapper/7 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=snapshot-test-2 next_pid=2243 next_prio=120 3293 snapshot-test-2-2229 [002] d... 2440.707438: sched_switch: prev_comm=snapshot-test-2 prev_pid=2229 prev_prio=120 prev_state=S ==> next_comm=swapper/2 next_pid=0 next_prio=120 3294 [...] 3295 3296 3297If you try to use this snapshot feature when current tracer is 3298one of the latency tracers, you will get the following results. 3299:: 3300 3301 # echo wakeup > current_tracer 3302 # echo 1 > snapshot 3303 bash: echo: write error: Device or resource busy 3304 # cat snapshot 3305 cat: snapshot: Device or resource busy 3306 3307 3308Instances 3309--------- 3310In the tracefs tracing directory is a directory called "instances". 3311This directory can have new directories created inside of it using 3312mkdir, and removing directories with rmdir. The directory created 3313with mkdir in this directory will already contain files and other 3314directories after it is created. 3315:: 3316 3317 # mkdir instances/foo 3318 # ls instances/foo 3319 buffer_size_kb buffer_total_size_kb events free_buffer per_cpu 3320 set_event snapshot trace trace_clock trace_marker trace_options 3321 trace_pipe tracing_on 3322 3323As you can see, the new directory looks similar to the tracing directory 3324itself. In fact, it is very similar, except that the buffer and 3325events are agnostic from the main director, or from any other 3326instances that are created. 3327 3328The files in the new directory work just like the files with the 3329same name in the tracing directory except the buffer that is used 3330is a separate and new buffer. The files affect that buffer but do not 3331affect the main buffer with the exception of trace_options. Currently, 3332the trace_options affect all instances and the top level buffer 3333the same, but this may change in future releases. That is, options 3334may become specific to the instance they reside in. 3335 3336Notice that none of the function tracer files are there, nor is 3337current_tracer and available_tracers. This is because the buffers 3338can currently only have events enabled for them. 3339:: 3340 3341 # mkdir instances/foo 3342 # mkdir instances/bar 3343 # mkdir instances/zoot 3344 # echo 100000 > buffer_size_kb 3345 # echo 1000 > instances/foo/buffer_size_kb 3346 # echo 5000 > instances/bar/per_cpu/cpu1/buffer_size_kb 3347 # echo function > current_trace 3348 # echo 1 > instances/foo/events/sched/sched_wakeup/enable 3349 # echo 1 > instances/foo/events/sched/sched_wakeup_new/enable 3350 # echo 1 > instances/foo/events/sched/sched_switch/enable 3351 # echo 1 > instances/bar/events/irq/enable 3352 # echo 1 > instances/zoot/events/syscalls/enable 3353 # cat trace_pipe 3354 CPU:2 [LOST 11745 EVENTS] 3355 bash-2044 [002] .... 10594.481032: _raw_spin_lock_irqsave <-get_page_from_freelist 3356 bash-2044 [002] d... 10594.481032: add_preempt_count <-_raw_spin_lock_irqsave 3357 bash-2044 [002] d..1 10594.481032: __rmqueue <-get_page_from_freelist 3358 bash-2044 [002] d..1 10594.481033: _raw_spin_unlock <-get_page_from_freelist 3359 bash-2044 [002] d..1 10594.481033: sub_preempt_count <-_raw_spin_unlock 3360 bash-2044 [002] d... 10594.481033: get_pageblock_flags_group <-get_pageblock_migratetype 3361 bash-2044 [002] d... 10594.481034: __mod_zone_page_state <-get_page_from_freelist 3362 bash-2044 [002] d... 10594.481034: zone_statistics <-get_page_from_freelist 3363 bash-2044 [002] d... 10594.481034: __inc_zone_state <-zone_statistics 3364 bash-2044 [002] d... 10594.481034: __inc_zone_state <-zone_statistics 3365 bash-2044 [002] .... 10594.481035: arch_dup_task_struct <-copy_process 3366 [...] 3367 3368 # cat instances/foo/trace_pipe 3369 bash-1998 [000] d..4 136.676759: sched_wakeup: comm=kworker/0:1 pid=59 prio=120 success=1 target_cpu=000 3370 bash-1998 [000] dN.4 136.676760: sched_wakeup: comm=bash pid=1998 prio=120 success=1 target_cpu=000 3371 <idle>-0 [003] d.h3 136.676906: sched_wakeup: comm=rcu_preempt pid=9 prio=120 success=1 target_cpu=003 3372 <idle>-0 [003] d..3 136.676909: sched_switch: prev_comm=swapper/3 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=rcu_preempt next_pid=9 next_prio=120 3373 rcu_preempt-9 [003] d..3 136.676916: sched_switch: prev_comm=rcu_preempt prev_pid=9 prev_prio=120 prev_state=S ==> next_comm=swapper/3 next_pid=0 next_prio=120 3374 bash-1998 [000] d..4 136.677014: sched_wakeup: comm=kworker/0:1 pid=59 prio=120 success=1 target_cpu=000 3375 bash-1998 [000] dN.4 136.677016: sched_wakeup: comm=bash pid=1998 prio=120 success=1 target_cpu=000 3376 bash-1998 [000] d..3 136.677018: sched_switch: prev_comm=bash prev_pid=1998 prev_prio=120 prev_state=R+ ==> next_comm=kworker/0:1 next_pid=59 next_prio=120 3377 kworker/0:1-59 [000] d..4 136.677022: sched_wakeup: comm=sshd pid=1995 prio=120 success=1 target_cpu=001 3378 kworker/0:1-59 [000] d..3 136.677025: sched_switch: prev_comm=kworker/0:1 prev_pid=59 prev_prio=120 prev_state=S ==> next_comm=bash next_pid=1998 next_prio=120 3379 [...] 3380 3381 # cat instances/bar/trace_pipe 3382 migration/1-14 [001] d.h3 138.732674: softirq_raise: vec=3 [action=NET_RX] 3383 <idle>-0 [001] dNh3 138.732725: softirq_raise: vec=3 [action=NET_RX] 3384 bash-1998 [000] d.h1 138.733101: softirq_raise: vec=1 [action=TIMER] 3385 bash-1998 [000] d.h1 138.733102: softirq_raise: vec=9 [action=RCU] 3386 bash-1998 [000] ..s2 138.733105: softirq_entry: vec=1 [action=TIMER] 3387 bash-1998 [000] ..s2 138.733106: softirq_exit: vec=1 [action=TIMER] 3388 bash-1998 [000] ..s2 138.733106: softirq_entry: vec=9 [action=RCU] 3389 bash-1998 [000] ..s2 138.733109: softirq_exit: vec=9 [action=RCU] 3390 sshd-1995 [001] d.h1 138.733278: irq_handler_entry: irq=21 name=uhci_hcd:usb4 3391 sshd-1995 [001] d.h1 138.733280: irq_handler_exit: irq=21 ret=unhandled 3392 sshd-1995 [001] d.h1 138.733281: irq_handler_entry: irq=21 name=eth0 3393 sshd-1995 [001] d.h1 138.733283: irq_handler_exit: irq=21 ret=handled 3394 [...] 3395 3396 # cat instances/zoot/trace 3397 # tracer: nop 3398 # 3399 # entries-in-buffer/entries-written: 18996/18996 #P:4 3400 # 3401 # _-----=> irqs-off 3402 # / _----=> need-resched 3403 # | / _---=> hardirq/softirq 3404 # || / _--=> preempt-depth 3405 # ||| / delay 3406 # TASK-PID CPU# |||| TIMESTAMP FUNCTION 3407 # | | | |||| | | 3408 bash-1998 [000] d... 140.733501: sys_write -> 0x2 3409 bash-1998 [000] d... 140.733504: sys_dup2(oldfd: a, newfd: 1) 3410 bash-1998 [000] d... 140.733506: sys_dup2 -> 0x1 3411 bash-1998 [000] d... 140.733508: sys_fcntl(fd: a, cmd: 1, arg: 0) 3412 bash-1998 [000] d... 140.733509: sys_fcntl -> 0x1 3413 bash-1998 [000] d... 140.733510: sys_close(fd: a) 3414 bash-1998 [000] d... 140.733510: sys_close -> 0x0 3415 bash-1998 [000] d... 140.733514: sys_rt_sigprocmask(how: 0, nset: 0, oset: 6e2768, sigsetsize: 8) 3416 bash-1998 [000] d... 140.733515: sys_rt_sigprocmask -> 0x0 3417 bash-1998 [000] d... 140.733516: sys_rt_sigaction(sig: 2, act: 7fff718846f0, oact: 7fff71884650, sigsetsize: 8) 3418 bash-1998 [000] d... 140.733516: sys_rt_sigaction -> 0x0 3419 3420You can see that the trace of the top most trace buffer shows only 3421the function tracing. The foo instance displays wakeups and task 3422switches. 3423 3424To remove the instances, simply delete their directories: 3425:: 3426 3427 # rmdir instances/foo 3428 # rmdir instances/bar 3429 # rmdir instances/zoot 3430 3431Note, if a process has a trace file open in one of the instance 3432directories, the rmdir will fail with EBUSY. 3433 3434 3435Stack trace 3436----------- 3437Since the kernel has a fixed sized stack, it is important not to 3438waste it in functions. A kernel developer must be conscience of 3439what they allocate on the stack. If they add too much, the system 3440can be in danger of a stack overflow, and corruption will occur, 3441usually leading to a system panic. 3442 3443There are some tools that check this, usually with interrupts 3444periodically checking usage. But if you can perform a check 3445at every function call that will become very useful. As ftrace provides 3446a function tracer, it makes it convenient to check the stack size 3447at every function call. This is enabled via the stack tracer. 3448 3449CONFIG_STACK_TRACER enables the ftrace stack tracing functionality. 3450To enable it, write a '1' into /proc/sys/kernel/stack_tracer_enabled. 3451:: 3452 3453 # echo 1 > /proc/sys/kernel/stack_tracer_enabled 3454 3455You can also enable it from the kernel command line to trace 3456the stack size of the kernel during boot up, by adding "stacktrace" 3457to the kernel command line parameter. 3458 3459After running it for a few minutes, the output looks like: 3460:: 3461 3462 # cat stack_max_size 3463 2928 3464 3465 # cat stack_trace 3466 Depth Size Location (18 entries) 3467 ----- ---- -------- 3468 0) 2928 224 update_sd_lb_stats+0xbc/0x4ac 3469 1) 2704 160 find_busiest_group+0x31/0x1f1 3470 2) 2544 256 load_balance+0xd9/0x662 3471 3) 2288 80 idle_balance+0xbb/0x130 3472 4) 2208 128 __schedule+0x26e/0x5b9 3473 5) 2080 16 schedule+0x64/0x66 3474 6) 2064 128 schedule_timeout+0x34/0xe0 3475 7) 1936 112 wait_for_common+0x97/0xf1 3476 8) 1824 16 wait_for_completion+0x1d/0x1f 3477 9) 1808 128 flush_work+0xfe/0x119 3478 10) 1680 16 tty_flush_to_ldisc+0x1e/0x20 3479 11) 1664 48 input_available_p+0x1d/0x5c 3480 12) 1616 48 n_tty_poll+0x6d/0x134 3481 13) 1568 64 tty_poll+0x64/0x7f 3482 14) 1504 880 do_select+0x31e/0x511 3483 15) 624 400 core_sys_select+0x177/0x216 3484 16) 224 96 sys_select+0x91/0xb9 3485 17) 128 128 system_call_fastpath+0x16/0x1b 3486 3487Note, if -mfentry is being used by gcc, functions get traced before 3488they set up the stack frame. This means that leaf level functions 3489are not tested by the stack tracer when -mfentry is used. 3490 3491Currently, -mfentry is used by gcc 4.6.0 and above on x86 only. 3492 3493More 3494---- 3495More details can be found in the source code, in the `kernel/trace/*.c` files. 3496