6 perf-trace-python - Process trace data with a Python script
11 'perf trace' [-s [Python]:script[.py] ]
16 This perf trace option is used to process perf trace data using perf's
17 built-in Python interpreter. It reads and processes the input file and
18 displays the results of the trace analysis implemented in the given
19 Python script, if any.
24 This section shows the process, start to finish, of creating a working
25 Python script that aggregates and extracts useful information from a
26 raw perf trace stream. You can avoid reading the rest of this
27 document if an example is enough for you; the rest of the document
28 provides more details on each step and lists the library functions
29 available to script writers.
31 This example actually details the steps that were used to create the
32 'syscall-counts' script you see when you list the available perf trace
33 scripts via 'perf trace -l'. As such, this script also shows how to
34 integrate your script into the list of general-purpose 'perf trace'
35 scripts listed by that command.
37 The syscall-counts script is a simple script, but demonstrates all the
38 basic ideas necessary to create a useful script. Here's an example
45 ---------------------------------------- -----------
51 sys_sched_setparam 826
72 Basically our task is to keep a per-syscall tally that gets updated
73 every time a system call occurs in the system. Our script will do
74 that, but first we need to record the data that will be processed by
75 that script. Theoretically, there are a couple of ways we could do
78 - we could enable every event under the tracing/events/syscalls
79 directory, but this is over 600 syscalls, well beyond the number
80 allowable by perf. These individual syscall events will however be
81 useful if we want to later use the guidance we get from the
82 general-purpose scripts to drill down and get more detail about
83 individual syscalls of interest.
85 - we can enable the sys_enter and/or sys_exit syscalls found under
86 tracing/events/raw_syscalls. These are called for all syscalls; the
87 'id' field can be used to distinguish between individual syscall
90 For this script, we only need to know that a syscall was entered; we
91 don't care how it exited, so we'll use 'perf record' to record only
95 # perf record -c 1 -f -a -M -R -e raw_syscalls:sys_enter
97 ^C[ perf record: Woken up 1 times to write data ]
98 [ perf record: Captured and wrote 56.545 MB perf.data (~2470503 samples) ]
101 The options basically say to collect data for every syscall event
102 system-wide and multiplex the per-cpu output into a single stream.
103 That single stream will be recorded in a file in the current directory
106 Once we have a perf.data file containing our data, we can use the -g
107 'perf trace' option to generate a Python script that will contain a
108 callback handler for each event type found in the perf.data trace
109 stream (for more details, see the STARTER SCRIPTS section).
112 # perf trace -g python
113 generated Python script: perf-trace.py
115 The output file created also in the current directory is named
116 perf-trace.py. Here's the file in its entirety:
118 # perf trace event handlers, generated by perf trace -g python
119 # Licensed under the terms of the GNU GPL License version 2
121 # The common_* event handler fields are the most useful fields common to
122 # all events. They don't necessarily correspond to the 'common_*' fields
123 # in the format files. Those fields not available as handler params can
124 # be retrieved using Python functions of the form common_*(context).
125 # See the perf-trace-python Documentation for the list of available functions.
130 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
131 '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
133 from perf_trace_context import *
137 print "in trace_begin"
142 def raw_syscalls__sys_enter(event_name, context, common_cpu,
143 common_secs, common_nsecs, common_pid, common_comm,
145 print_header(event_name, common_cpu, common_secs, common_nsecs,
146 common_pid, common_comm)
148 print "id=%d, args=%s\n" % \
151 def trace_unhandled(event_name, context, common_cpu, common_secs, common_nsecs,
152 common_pid, common_comm):
153 print_header(event_name, common_cpu, common_secs, common_nsecs,
154 common_pid, common_comm)
156 def print_header(event_name, cpu, secs, nsecs, pid, comm):
157 print "%-20s %5u %05u.%09u %8u %-20s " % \
158 (event_name, cpu, secs, nsecs, pid, comm),
161 At the top is a comment block followed by some import statements and a
162 path append which every perf trace script should include.
164 Following that are a couple generated functions, trace_begin() and
165 trace_end(), which are called at the beginning and the end of the
166 script respectively (for more details, see the SCRIPT_LAYOUT section
169 Following those are the 'event handler' functions generated one for
170 every event in the 'perf record' output. The handler functions take
171 the form subsystem__event_name, and contain named parameters, one for
172 each field in the event; in this case, there's only one event,
173 raw_syscalls__sys_enter(). (see the EVENT HANDLERS section below for
174 more info on event handlers).
176 The final couple of functions are, like the begin and end functions,
177 generated for every script. The first, trace_unhandled(), is called
178 every time the script finds an event in the perf.data file that
179 doesn't correspond to any event handler in the script. This could
180 mean either that the record step recorded event types that it wasn't
181 really interested in, or the script was run against a trace file that
182 doesn't correspond to the script.
184 The script generated by -g option option simply prints a line for each
185 event found in the trace stream i.e. it basically just dumps the event
186 and its parameter values to stdout. The print_header() function is
187 simply a utility function used for that purpose. Let's rename the
188 script and run it to see the default output:
191 # mv perf-trace.py syscall-counts.py
192 # perf trace -s syscall-counts.py
194 raw_syscalls__sys_enter 1 00840.847582083 7506 perf id=1, args=
195 raw_syscalls__sys_enter 1 00840.847595764 7506 perf id=1, args=
196 raw_syscalls__sys_enter 1 00840.847620860 7506 perf id=1, args=
197 raw_syscalls__sys_enter 1 00840.847710478 6533 npviewer.bin id=78, args=
198 raw_syscalls__sys_enter 1 00840.847719204 6533 npviewer.bin id=142, args=
199 raw_syscalls__sys_enter 1 00840.847755445 6533 npviewer.bin id=3, args=
200 raw_syscalls__sys_enter 1 00840.847775601 6533 npviewer.bin id=3, args=
201 raw_syscalls__sys_enter 1 00840.847781820 6533 npviewer.bin id=3, args=
207 Of course, for this script, we're not interested in printing every
208 trace event, but rather aggregating it in a useful way. So we'll get
209 rid of everything to do with printing as well as the trace_begin() and
210 trace_unhandled() functions, which we won't be using. That leaves us
211 with this minimalistic skeleton:
217 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
218 '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
220 from perf_trace_context import *
226 def raw_syscalls__sys_enter(event_name, context, common_cpu,
227 common_secs, common_nsecs, common_pid, common_comm,
231 In trace_end(), we'll simply print the results, but first we need to
232 generate some results to print. To do that we need to have our
233 sys_enter() handler do the necessary tallying until all events have
234 been counted. A hash table indexed by syscall id is a good way to
235 store that information; every time the sys_enter() handler is called,
236 we simply increment a count associated with that hash entry indexed by
240 syscalls = autodict()
248 The syscalls 'autodict' object is a special kind of Python dictionary
249 (implemented in Core.py) that implements Perl's 'autovivifying' hashes
250 in Python i.e. with autovivifying hashes, you can assign nested hash
251 values without having to go to the trouble of creating intermediate
252 levels if they don't exist e.g syscalls[comm][pid][id] = 1 will create
253 the intermediate hash levels and finally assign the value 1 to the
254 hash entry for 'id' (because the value being assigned isn't a hash
255 object itself, the initial value is assigned in the TypeError
256 exception. Well, there may be a better way to do this in Python but
257 that's what works for now).
259 Putting that code into the raw_syscalls__sys_enter() handler, we
260 effectively end up with a single-level dictionary keyed on syscall id
261 and having the counts we've tallied as values.
263 The print_syscall_totals() function iterates over the entries in the
264 dictionary and displays a line for each entry containing the syscall
265 name (the dictonary keys contain the syscall ids, which are passed to
266 the Util function syscall_name(), which translates the raw syscall
267 numbers to the corresponding syscall name strings). The output is
268 displayed after all the events in the trace have been processed, by
269 calling the print_syscall_totals() function from the trace_end()
270 handler called at the end of script processing.
272 The final script producing the output shown above is shown in its
279 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
280 '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
282 from perf_trace_context import *
286 syscalls = autodict()
289 print_syscall_totals()
291 def raw_syscalls__sys_enter(event_name, context, common_cpu,
292 common_secs, common_nsecs, common_pid, common_comm,
299 def print_syscall_totals():
300 if for_comm is not None:
301 print "\nsyscall events for %s:\n\n" % (for_comm),
303 print "\nsyscall events:\n\n",
305 print "%-40s %10s\n" % ("event", "count"),
306 print "%-40s %10s\n" % ("----------------------------------------", \
309 for id, val in sorted(syscalls.iteritems(), key = lambda(k, v): (v, k), \
311 print "%-40s %10d\n" % (syscall_name(id), val),
314 The script can be run just as before:
316 # perf trace -s syscall-counts.py
318 So those are the essential steps in writing and running a script. The
319 process can be generalized to any tracepoint or set of tracepoints
320 you're interested in - basically find the tracepoint(s) you're
321 interested in by looking at the list of available events shown by
322 'perf list' and/or look in /sys/kernel/debug/tracing events for
323 detailed event and field info, record the corresponding trace data
324 using 'perf record', passing it the list of interesting events,
325 generate a skeleton script using 'perf trace -g python' and modify the
326 code to aggregate and display it for your particular needs.
328 After you've done that you may end up with a general-purpose script
329 that you want to keep around and have available for future use. By
330 writing a couple of very simple shell scripts and putting them in the
331 right place, you can have your script listed alongside the other
332 scripts listed by the 'perf trace -l' command e.g.:
335 root@tropicana:~# perf trace -l
336 List of available trace scripts:
337 workqueue-stats workqueue stats (ins/exe/create/destroy)
338 wakeup-latency system-wide min/max/avg wakeup latency
339 rw-by-file <comm> r/w activity for a program, by file
340 rw-by-pid system-wide r/w activity
343 A nice side effect of doing this is that you also then capture the
344 probably lengthy 'perf record' command needed to record the events for
347 To have the script appear as a 'built-in' script, you write two simple
348 scripts, one for recording and one for 'reporting'.
350 The 'record' script is a shell script with the same base name as your
351 script, but with -record appended. The shell script should be put
352 into the perf/scripts/python/bin directory in the kernel source tree.
353 In that script, you write the 'perf record' command-line needed for
357 # cat kernel-source/tools/perf/scripts/python/bin/syscall-counts-record
360 perf record -c 1 -f -a -M -R -e raw_syscalls:sys_enter
363 The 'report' script is also a shell script with the same base name as
364 your script, but with -report appended. It should also be located in
365 the perf/scripts/python/bin directory. In that script, you write the
366 'perf trace -s' command-line needed for running your script:
369 # cat kernel-source/tools/perf/scripts/python/bin/syscall-counts-report
372 # description: system-wide syscall counts
373 perf trace -s ~/libexec/perf-core/scripts/python/syscall-counts.py
376 Note that the location of the Python script given in the shell script
377 is in the libexec/perf-core/scripts/python directory - this is where
378 the script will be copied by 'make install' when you install perf.
379 For the installation to install your script there, your script needs
380 to be located in the perf/scripts/python directory in the kernel
384 # ls -al kernel-source/tools/perf/scripts/python
386 root@tropicana:/home/trz/src/tip# ls -al tools/perf/scripts/python
388 drwxr-xr-x 4 trz trz 4096 2010-01-26 22:30 .
389 drwxr-xr-x 4 trz trz 4096 2010-01-26 22:29 ..
390 drwxr-xr-x 2 trz trz 4096 2010-01-26 22:29 bin
391 -rw-r--r-- 1 trz trz 2548 2010-01-26 22:29 check-perf-trace.py
392 drwxr-xr-x 3 trz trz 4096 2010-01-26 22:49 Perf-Trace-Util
393 -rw-r--r-- 1 trz trz 1462 2010-01-26 22:30 syscall-counts.py
396 Once you've done that (don't forget to do a new 'make install',
397 otherwise your script won't show up at run-time), 'perf trace -l'
398 should show a new entry for your script:
401 root@tropicana:~# perf trace -l
402 List of available trace scripts:
403 workqueue-stats workqueue stats (ins/exe/create/destroy)
404 wakeup-latency system-wide min/max/avg wakeup latency
405 rw-by-file <comm> r/w activity for a program, by file
406 rw-by-pid system-wide r/w activity
407 syscall-counts system-wide syscall counts
410 You can now perform the record step via 'perf trace record':
412 # perf trace record syscall-counts
414 and display the output using 'perf trace report':
416 # perf trace report syscall-counts
421 You can quickly get started writing a script for a particular set of
422 trace data by generating a skeleton script using 'perf trace -g
423 python' in the same directory as an existing perf.data trace file.
424 That will generate a starter script containing a handler for each of
425 the event types in the trace file; it simply prints every available
426 field for each event in the trace file.
428 You can also look at the existing scripts in
429 ~/libexec/perf-core/scripts/python for typical examples showing how to
430 do basic things like aggregate event data, print results, etc. Also,
431 the check-perf-trace.py script, while not interesting for its results,
432 attempts to exercise all of the main scripting features.
437 When perf trace is invoked using a trace script, a user-defined
438 'handler function' is called for each event in the trace. If there's
439 no handler function defined for a given event type, the event is
440 ignored (or passed to a 'trace_handled' function, see below) and the
441 next event is processed.
443 Most of the event's field values are passed as arguments to the
444 handler function; some of the less common ones aren't - those are
445 available as calls back into the perf executable (see below).
447 As an example, the following perf record command can be used to record
448 all sched_wakeup events in the system:
450 # perf record -c 1 -f -a -M -R -e sched:sched_wakeup
452 Traces meant to be processed using a script should be recorded with
453 the above options: -c 1 says to sample every event, -a to enable
454 system-wide collection, -M to multiplex the output, and -R to collect
457 The format file for the sched_wakep event defines the following fields
458 (see /sys/kernel/debug/tracing/events/sched/sched_wakeup/format):
462 field:unsigned short common_type;
463 field:unsigned char common_flags;
464 field:unsigned char common_preempt_count;
465 field:int common_pid;
466 field:int common_lock_depth;
468 field:char comm[TASK_COMM_LEN];
472 field:int target_cpu;
475 The handler function for this event would be defined as:
478 def sched__sched_wakeup(event_name, context, common_cpu, common_secs,
479 common_nsecs, common_pid, common_comm,
480 comm, pid, prio, success, target_cpu):
484 The handler function takes the form subsystem__event_name.
486 The common_* arguments in the handler's argument list are the set of
487 arguments passed to all event handlers; some of the fields correspond
488 to the common_* fields in the format file, but some are synthesized,
489 and some of the common_* fields aren't common enough to to be passed
490 to every event as arguments but are available as library functions.
492 Here's a brief description of each of the invariant event args:
494 event_name the name of the event as text
495 context an opaque 'cookie' used in calls back into perf
496 common_cpu the cpu the event occurred on
497 common_secs the secs portion of the event timestamp
498 common_nsecs the nsecs portion of the event timestamp
499 common_pid the pid of the current task
500 common_comm the name of the current process
502 All of the remaining fields in the event's format file have
503 counterparts as handler function arguments of the same name, as can be
504 seen in the example above.
506 The above provides the basics needed to directly access every field of
507 every event in a trace, which covers 90% of what you need to know to
508 write a useful trace script. The sections below cover the rest.
513 Every perf trace Python script should start by setting up a Python
514 module search path and 'import'ing a few support modules (see module
521 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
522 '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
524 from perf_trace_context import *
528 The rest of the script can contain handler functions and support
529 functions in any order.
531 Aside from the event handler functions discussed above, every script
532 can implement a set of optional functions:
534 *trace_begin*, if defined, is called before any event is processed and
535 gives scripts a chance to do setup tasks:
542 *trace_end*, if defined, is called after all events have been
543 processed and gives scripts a chance to do end-of-script tasks, such
551 *trace_unhandled*, if defined, is called after for any event that
552 doesn't have a handler explicitly defined for it. The standard set
553 of common arguments are passed into it:
556 def trace_unhandled(event_name, context, common_cpu, common_secs,
557 common_nsecs, common_pid, common_comm):
561 The remaining sections provide descriptions of each of the available
562 built-in perf trace Python modules and their associated functions.
564 AVAILABLE MODULES AND FUNCTIONS
565 -------------------------------
567 The following sections describe the functions and variables available
568 via the various perf trace Python modules. To use the functions and
569 variables from the given module, add the corresponding 'from XXXX
570 import' line to your perf trace script.
575 These functions provide some essential functions to user scripts.
577 The *flag_str* and *symbol_str* functions provide human-readable
578 strings for flag and symbolic fields. These correspond to the strings
579 and values parsed from the 'print fmt' fields of the event format
582 flag_str(event_name, field_name, field_value) - returns the string represention corresponding to field_value for the flag field field_name of event event_name
583 symbol_str(event_name, field_name, field_value) - returns the string represention corresponding to field_value for the symbolic field field_name of event event_name
585 The *autodict* function returns a special special kind of Python
586 dictionary that implements Perl's 'autovivifying' hashes in Python
587 i.e. with autovivifying hashes, you can assign nested hash values
588 without having to go to the trouble of creating intermediate levels if
591 autodict() - returns an autovivifying dictionary instance
594 perf_trace_context Module
595 ~~~~~~~~~~~~~~~~~~~~~~~~~
597 Some of the 'common' fields in the event format file aren't all that
598 common, but need to be made accessible to user scripts nonetheless.
600 perf_trace_context defines a set of functions that can be used to
601 access this data in the context of the current event. Each of these
602 functions expects a context variable, which is the same as the
603 context variable passed into every event handler as the second
606 common_pc(context) - returns common_preempt count for the current event
607 common_flags(context) - returns common_flags for the current event
608 common_lock_depth(context) - returns common_lock_depth for the current event
613 Various utility functions for use with perf trace:
615 nsecs(secs, nsecs) - returns total nsecs given secs/nsecs pair
616 nsecs_secs(nsecs) - returns whole secs portion given nsecs
617 nsecs_nsecs(nsecs) - returns nsecs remainder given nsecs
618 nsecs_str(nsecs) - returns printable string in the form secs.nsecs
619 avg(total, n) - returns average given a sum and a total number of values
620 syscall_name(id) - returns the syscall name for the specified syscall_nr
624 linkperf:perf-trace[1]