1 perf-script-python(1)
2 ====================
3
4 NAME
5 ----
6 perf-script-python - Process trace data with a
7
8 SYNOPSIS
9 --------
10 [verse]
11 'perf script' [-s [Python]:script[.py] ]
12
13 DESCRIPTION
14 -----------
15
16 This perf script option is used to process per
17 built-in Python interpreter. It reads and pro
18 displays the results of the trace analysis imp
19 Python script, if any.
20
21 A QUICK EXAMPLE
22 ---------------
23
24 This section shows the process, start to finis
25 Python script that aggregates and extracts use
26 raw perf script stream. You can avoid reading
27 document if an example is enough for you; the
28 provides more details on each step and lists t
29 available to script writers.
30
31 This example actually details the steps that w
32 'syscall-counts' script you see when you list
33 scripts via 'perf script -l'. As such, this s
34 integrate your script into the list of general
35 scripts listed by that command.
36
37 The syscall-counts script is a simple script,
38 basic ideas necessary to create a useful scrip
39 of its output (syscall names are not yet suppo
40 as numbers):
41
42 ----
43 syscall events:
44
45 event
46 ---------------------------------------- ----
47 sys_write
48 sys_getdents
49 sys_close
50 sys_swapoff
51 sys_read
52 sys_sched_setparam
53 sys_open
54 sys_newfstat
55 sys_mmap
56 sys_munmap
57 sys_futex
58 sys_select
59 sys_poll
60 sys_setitimer
61 sys_writev
62 15
63 sys_lseek
64 sys_rt_sigprocmask
65 sys_wait4
66 sys_ioctl
67 sys_set_robust_list
68 sys_exit
69 56
70 sys_access
71 ----
72
73 Basically our task is to keep a per-syscall ta
74 every time a system call occurs in the system.
75 that, but first we need to record the data tha
76 that script. Theoretically, there are a coupl
77 that:
78
79 - we could enable every event under the tracin
80 directory, but this is over 600 syscalls, we
81 allowable by perf. These individual syscall
82 useful if we want to later use the guidance
83 general-purpose scripts to drill down and ge
84 individual syscalls of interest.
85
86 - we can enable the sys_enter and/or sys_exit
87 tracing/events/raw_syscalls. These are call
88 'id' field can be used to distinguish betwee
89 numbers.
90
91 For this script, we only need to know that a s
92 don't care how it exited, so we'll use 'perf r
93 the sys_enter events:
94
95 ----
96 # perf record -a -e raw_syscalls:sys_enter
97
98 ^C[ perf record: Woken up 1 times to write dat
99 [ perf record: Captured and wrote 56.545 MB pe
100 ----
101
102 The options basically say to collect data for
103 system-wide and multiplex the per-cpu output i
104 That single stream will be recorded in a file
105 called perf.data.
106
107 Once we have a perf.data file containing our d
108 'perf script' option to generate a Python scri
109 callback handler for each event type found in
110 stream (for more details, see the STARTER SCRI
111
112 ----
113 # perf script -g python
114 generated Python script: perf-script.py
115
116 The output file created also in the current di
117 perf-script.py. Here's the file in its entire
118
119 # perf script event handlers, generated by per
120 # Licensed under the terms of the GNU GPL Lice
121
122 # The common_* event handler fields are the mo
123 # all events. They don't necessarily correspo
124 # in the format files. Those fields not avail
125 # be retrieved using Python functions of the f
126 # See the perf-script-python Documentation for
127
128 import os
129 import sys
130
131 sys.path.append(os.environ['PERF_EXEC_PATH'] +
132 '/scripts/python/Perf-Trace-Util/lib/P
133
134 from perf_trace_context import *
135 from Core import *
136
137 def trace_begin():
138 print "in trace_begin"
139
140 def trace_end():
141 print "in trace_end"
142
143 def raw_syscalls__sys_enter(event_name, contex
144 common_secs, common_nsecs, common_pid,
145 id, args):
146 print_header(event_name, commo
147 common_pid, common_com
148
149 print "id=%d, args=%s\n" % \
150 (id, args),
151
152 def trace_unhandled(event_name, context, event
153 print ' '.join(['%s=%s'%(k,str
154
155 def print_header(event_name, cpu, secs, nsecs,
156 print "%-20s %5u %05u.%09u %8u %-20s "
157 (event_name, cpu, secs, nsecs, pid, co
158 ----
159
160 At the top is a comment block followed by some
161 path append which every perf script script sho
162
163 Following that are a couple generated function
164 trace_end(), which are called at the beginning
165 script respectively (for more details, see the
166 below).
167
168 Following those are the 'event handler' functi
169 every event in the 'perf record' output. The
170 the form subsystem\__event_name, and contain n
171 each field in the event; in this case, there's
172 raw_syscalls__sys_enter(). (see the EVENT HAN
173 more info on event handlers).
174
175 The final couple of functions are, like the be
176 generated for every script. The first, trace_
177 every time the script finds an event in the pe
178 doesn't correspond to any event handler in the
179 mean either that the record step recorded even
180 really interested in, or the script was run ag
181 doesn't correspond to the script.
182
183 The script generated by -g option simply print
184 event found in the trace stream i.e. it basica
185 and its parameter values to stdout. The print
186 simply a utility function used for that purpos
187 script and run it to see the default output:
188
189 ----
190 # mv perf-script.py syscall-counts.py
191 # perf script -s syscall-counts.py
192
193 raw_syscalls__sys_enter 1 00840.847582083
194 raw_syscalls__sys_enter 1 00840.847595764
195 raw_syscalls__sys_enter 1 00840.847620860
196 raw_syscalls__sys_enter 1 00840.847710478
197 raw_syscalls__sys_enter 1 00840.847719204
198 raw_syscalls__sys_enter 1 00840.847755445
199 raw_syscalls__sys_enter 1 00840.847775601
200 raw_syscalls__sys_enter 1 00840.847781820
201 .
202 .
203 .
204 ----
205
206 Of course, for this script, we're not interest
207 trace event, but rather aggregating it in a us
208 rid of everything to do with printing as well
209 trace_unhandled() functions, which we won't be
210 with this minimalistic skeleton:
211
212 ----
213 import os
214 import sys
215
216 sys.path.append(os.environ['PERF_EXEC_PATH'] +
217 '/scripts/python/Perf-Trace-Util/lib/P
218
219 from perf_trace_context import *
220 from Core import *
221
222 def trace_end():
223 print "in trace_end"
224
225 def raw_syscalls__sys_enter(event_name, contex
226 common_secs, common_nsecs, common_pid,
227 id, args):
228 ----
229
230 In trace_end(), we'll simply print the results
231 generate some results to print. To do that we
232 sys_enter() handler do the necessary tallying
233 been counted. A hash table indexed by syscall
234 store that information; every time the sys_ent
235 we simply increment a count associated with th
236 that syscall id:
237
238 ----
239 syscalls = autodict()
240
241 try:
242 syscalls[id] += 1
243 except TypeError:
244 syscalls[id] = 1
245 ----
246
247 The syscalls 'autodict' object is a special ki
248 (implemented in Core.py) that implements Perl'
249 in Python i.e. with autovivifying hashes, you
250 values without having to go to the trouble of
251 levels if they don't exist e.g syscalls[comm][
252 the intermediate hash levels and finally assig
253 hash entry for 'id' (because the value being a
254 object itself, the initial value is assigned i
255 exception. Well, there may be a better way to
256 that's what works for now).
257
258 Putting that code into the raw_syscalls__sys_e
259 effectively end up with a single-level diction
260 and having the counts we've tallied as values.
261
262 The print_syscall_totals() function iterates o
263 dictionary and displays a line for each entry
264 name (the dictionary keys contain the syscall
265 the Util function syscall_name(), which transl
266 numbers to the corresponding syscall name stri
267 displayed after all the events in the trace ha
268 calling the print_syscall_totals() function fr
269 handler called at the end of script processing
270
271 The final script producing the output shown ab
272 entirety below (syscall_name() helper is not y
273 only deal with id's for now):
274
275 ----
276 import os
277 import sys
278
279 sys.path.append(os.environ['PERF_EXEC_PATH'] +
280 '/scripts/python/Perf-Trace-Util/lib/P
281
282 from perf_trace_context import *
283 from Core import *
284 from Util import *
285
286 syscalls = autodict()
287
288 def trace_end():
289 print_syscall_totals()
290
291 def raw_syscalls__sys_enter(event_name, contex
292 common_secs, common_nsecs, common_pid,
293 id, args):
294 try:
295 syscalls[id] += 1
296 except TypeError:
297 syscalls[id] = 1
298
299 def print_syscall_totals():
300 if for_comm is not None:
301 print "\nsyscall events for %s:\n\
302 else:
303 print "\nsyscall events:\n\n",
304
305 print "%-40s %10s\n" % ("event", "count")
306 print "%-40s %10s\n" % ("----------------
307 "-----------"
308
309 for id, val in sorted(syscalls.iteritems()
310 reverse = Tr
311 print "%-40s %10d\n" % (syscall_n
312 ----
313
314 The script can be run just as before:
315
316 # perf script -s syscall-counts.py
317
318 So those are the essential steps in writing an
319 process can be generalized to any tracepoint o
320 you're interested in - basically find the trac
321 interested in by looking at the list of availa
322 'perf list' and/or look in /sys/kernel/tracing
323 detailed event and field info, record the corr
324 using 'perf record', passing it the list of in
325 generate a skeleton script using 'perf script
326 code to aggregate and display it for your part
327
328 After you've done that you may end up with a g
329 that you want to keep around and have availabl
330 writing a couple of very simple shell scripts
331 right place, you can have your script listed a
332 scripts listed by the 'perf script -l' command
333
334 ----
335 # perf script -l
336 List of available trace scripts:
337 wakeup-latency system-
338 rw-by-file <comm> r/w act
339 rw-by-pid system-
340 ----
341
342 A nice side effect of doing this is that you a
343 probably lengthy 'perf record' command needed
344 the script.
345
346 To have the script appear as a 'built-in' scri
347 scripts, one for recording and one for 'report
348
349 The 'record' script is a shell script with the
350 script, but with -record appended. The shell
351 into the perf/scripts/python/bin directory in
352 In that script, you write the 'perf record' co
353 your script:
354
355 ----
356 # cat kernel-source/tools/perf/scripts/python/
357
358 #!/bin/bash
359 perf record -a -e raw_syscalls:sys_enter
360 ----
361
362 The 'report' script is also a shell script wit
363 your script, but with -report appended. It sh
364 the perf/scripts/python/bin directory. In tha
365 'perf script -s' command-line needed for runni
366
367 ----
368 # cat kernel-source/tools/perf/scripts/python/
369
370 #!/bin/bash
371 # description: system-wide syscall counts
372 perf script -s ~/libexec/perf-core/scripts/pyt
373 ----
374
375 Note that the location of the Python script gi
376 is in the libexec/perf-core/scripts/python dir
377 the script will be copied by 'make install' wh
378 For the installation to install your script th
379 to be located in the perf/scripts/python direc
380 source tree:
381
382 ----
383 # ls -al kernel-source/tools/perf/scripts/pyth
384 total 32
385 drwxr-xr-x 4 trz trz 4096 2010-01-26 22:30 .
386 drwxr-xr-x 4 trz trz 4096 2010-01-26 22:29 ..
387 drwxr-xr-x 2 trz trz 4096 2010-01-26 22:29 bin
388 -rw-r--r-- 1 trz trz 2548 2010-01-26 22:29 che
389 drwxr-xr-x 3 trz trz 4096 2010-01-26 22:49 Per
390 -rw-r--r-- 1 trz trz 1462 2010-01-26 22:30 sys
391 ----
392
393 Once you've done that (don't forget to do a ne
394 otherwise your script won't show up at run-tim
395 should show a new entry for your script:
396
397 ----
398 # perf script -l
399 List of available trace scripts:
400 wakeup-latency system-
401 rw-by-file <comm> r/w act
402 rw-by-pid system-
403 syscall-counts system-
404 ----
405
406 You can now perform the record step via 'perf
407
408 # perf script record syscall-counts
409
410 and display the output using 'perf script repo
411
412 # perf script report syscall-counts
413
414 STARTER SCRIPTS
415 ---------------
416
417 You can quickly get started writing a script f
418 trace data by generating a skeleton script usi
419 python' in the same directory as an existing p
420 That will generate a starter script containing
421 the event types in the trace file; it simply p
422 field for each event in the trace file.
423
424 You can also look at the existing scripts in
425 ~/libexec/perf-core/scripts/python for typical
426 do basic things like aggregate event data, pri
427 the check-perf-script.py script, while not int
428 attempts to exercise all of the main scripting
429
430 EVENT HANDLERS
431 --------------
432
433 When perf script is invoked using a trace scri
434 'handler function' is called for each event in
435 no handler function defined for a given event
436 ignored (or passed to a 'trace_unhandled' func
437 next event is processed.
438
439 Most of the event's field values are passed as
440 handler function; some of the less common ones
441 available as calls back into the perf executab
442
443 As an example, the following perf record comma
444 all sched_wakeup events in the system:
445
446 # perf record -a -e sched:sched_wakeup
447
448 Traces meant to be processed using a script sh
449 the above option: -a to enable system-wide col
450
451 The format file for the sched_wakeup event def
452 (see /sys/kernel/tracing/events/sched/sched_wa
453
454 ----
455 format:
456 field:unsigned short common_type;
457 field:unsigned char common_flags;
458 field:unsigned char common_preempt_cou
459 field:int common_pid;
460
461 field:char comm[TASK_COMM_LEN];
462 field:pid_t pid;
463 field:int prio;
464 field:int success;
465 field:int target_cpu;
466 ----
467
468 The handler function for this event would be d
469
470 ----
471 def sched__sched_wakeup(event_name, context, c
472 common_nsecs, common_pid, common_comm,
473 comm, pid, prio, success, target_cpu):
474 pass
475 ----
476
477 The handler function takes the form subsystem_
478
479 The common_* arguments in the handler's argume
480 arguments passed to all event handlers; some o
481 to the common_* fields in the format file, but
482 and some of the common_* fields aren't common
483 to every event as arguments but are available
484
485 Here's a brief description of each of the inva
486
487 event_name the name of the ev
488 context an opaque 'cookie'
489 common_cpu the cpu the event
490 common_secs the secs portion o
491 common_nsecs the nsecs portion
492 common_pid the pid of the cur
493 common_comm the name of the cu
494
495 All of the remaining fields in the event's for
496 counterparts as handler function arguments of
497 seen in the example above.
498
499 The above provides the basics needed to direct
500 every event in a trace, which covers 90% of wh
501 write a useful trace script. The sections bel
502
503 SCRIPT LAYOUT
504 -------------
505
506 Every perf script Python script should start b
507 module search path and 'import'ing a few suppo
508 descriptions below):
509
510 ----
511 import os
512 import sys
513
514 sys.path.append(os.environ['PERF_EXEC_PATH']
515 '/scripts/python/Perf-Trace-Util
516
517 from perf_trace_context import *
518 from Core import *
519 ----
520
521 The rest of the script can contain handler fun
522 functions in any order.
523
524 Aside from the event handler functions discuss
525 can implement a set of optional functions:
526
527 *trace_begin*, if defined, is called before an
528 gives scripts a chance to do setup tasks:
529
530 ----
531 def trace_begin():
532 pass
533 ----
534
535 *trace_end*, if defined, is called after all e
536 processed and gives scripts a chance to do en
537 as display results:
538
539 ----
540 def trace_end():
541 pass
542 ----
543
544 *trace_unhandled*, if defined, is called after
545 doesn't have a handler explicitly defined for
546 of common arguments are passed into it:
547
548 ----
549 def trace_unhandled(event_name, context, event
550 pass
551 ----
552
553 *process_event*, if defined, is called for any
554
555 ----
556 def process_event(param_dict):
557 pass
558 ----
559
560 *context_switch*, if defined, is called for an
561
562 ----
563 def context_switch(ts, cpu, pid, tid, np_pid,
564 pass
565 ----
566
567 *auxtrace_error*, if defined, is called for an
568
569 ----
570 def auxtrace_error(typ, code, cpu, pid, tid, i
571 pass
572 ----
573
574 The remaining sections provide descriptions of
575 built-in perf script Python modules and their
576
577 AVAILABLE MODULES AND FUNCTIONS
578 -------------------------------
579
580 The following sections describe the functions
581 via the various perf script Python modules. T
582 variables from the given module, add the corre
583 import' line to your perf script script.
584
585 Core.py Module
586 ~~~~~~~~~~~~~~
587
588 These functions provide some essential functio
589
590 The *flag_str* and *symbol_str* functions prov
591 strings for flag and symbolic fields. These c
592 and values parsed from the 'print fmt' fields
593 files:
594
595 flag_str(event_name, field_name, field_value
596 symbol_str(event_name, field_name, field_val
597
598 The *autodict* function returns a special kind
599 dictionary that implements Perl's 'autovivifyi
600 i.e. with autovivifying hashes, you can assign
601 without having to go to the trouble of creatin
602 they don't exist.
603
604 autodict() - returns an autovivifying dictio
605
606
607 perf_trace_context Module
608 ~~~~~~~~~~~~~~~~~~~~~~~~~
609
610 Some of the 'common' fields in the event forma
611 common, but need to be made accessible to user
612
613 perf_trace_context defines a set of functions
614 access this data in the context of the current
615 functions expects a context variable, which is
616 context variable passed into every tracepoint
617 argument. For non-tracepoint events, the conte
618 as perf_trace_context.perf_script_context .
619
620 common_pc(context) - returns common_preempt c
621 common_flags(context) - returns common_flags
622 common_lock_depth(context) - returns common_l
623 perf_sample_insn(context) - returns the machi
624 perf_set_itrace_options(context, itrace_optio
625 perf_sample_srcline(context) - returns source
626 perf_sample_srccode(context) - returns source
627
628
629 Util.py Module
630 ~~~~~~~~~~~~~~
631
632 Various utility functions for use with perf sc
633
634 nsecs(secs, nsecs) - returns total nsecs giv
635 nsecs_secs(nsecs) - returns whole secs porti
636 nsecs_nsecs(nsecs) - returns nsecs remainder
637 nsecs_str(nsecs) - returns printable string
638 avg(total, n) - returns average given a sum
639
640 SUPPORTED FIELDS
641 ----------------
642
643 Currently supported fields:
644
645 ev_name, comm, id, stream_id, pid, tid, cpu, i
646 addr, symbol, symoff, dso, time_enabled, time_
647 brstack, brstacksym, datasrc, datasrc_decode,
648 weight, transaction, raw_buf, attr, cpumode.
649
650 Fields that may also be present:
651
652 flags - sample flags
653 flags_disp - sample flags display
654 insn_cnt - instruction count for determining
655 cyc_cnt - cycle count for determining IPC
656 addr_correlates_sym - addr can correlate to a
657 addr_dso - addr dso
658 addr_symbol - addr symbol
659 addr_symoff - addr symbol offset
660
661 Some fields have sub items:
662
663 brstack:
664 from, to, from_dsoname, to_dsoname, mispre
665 predicted, in_tx, abort, cycles.
666
667 brstacksym:
668 items: from, to, pred, in_tx, abort (conve
669
670 For example,
671 We can use this code to print brstack "from",
672
673 if 'brstack' in dict:
674 for entry in dict['brstack']:
675 print "from %s, to %s, cycles
676
677 SEE ALSO
678 --------
679 linkperf:perf-script[1]
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.