Home | History | Annotate | Download | only in Documentation
      1 perf-script-python(1)
      2 ====================
      3 
      4 NAME
      5 ----
      6 perf-script-python - Process trace data with a Python script
      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 perf script 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.
     20 
     21 A QUICK EXAMPLE
     22 ---------------
     23 
     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 script 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.
     30 
     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 script
     33 scripts via 'perf script -l'.  As such, this script also shows how to
     34 integrate your script into the list of general-purpose 'perf script'
     35 scripts listed by that command.
     36 
     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
     39 of its output (syscall names are not yet supported, they will appear
     40 as numbers):
     41 
     42 ----
     43 syscall events:
     44 
     45 event                                          count
     46 ----------------------------------------  -----------
     47 sys_write                                     455067
     48 sys_getdents                                    4072
     49 sys_close                                       3037
     50 sys_swapoff                                     1769
     51 sys_read                                         923
     52 sys_sched_setparam                               826
     53 sys_open                                         331
     54 sys_newfstat                                     326
     55 sys_mmap                                         217
     56 sys_munmap                                       216
     57 sys_futex                                        141
     58 sys_select                                       102
     59 sys_poll                                          84
     60 sys_setitimer                                     12
     61 sys_writev                                         8
     62 15                                                 8
     63 sys_lseek                                          7
     64 sys_rt_sigprocmask                                 6
     65 sys_wait4                                          3
     66 sys_ioctl                                          3
     67 sys_set_robust_list                                1
     68 sys_exit                                           1
     69 56                                                 1
     70 sys_access                                         1
     71 ----
     72 
     73 Basically our task is to keep a per-syscall tally that gets updated
     74 every time a system call occurs in the system.  Our script will do
     75 that, but first we need to record the data that will be processed by
     76 that script.  Theoretically, there are a couple of ways we could do
     77 that:
     78 
     79 - we could enable every event under the tracing/events/syscalls
     80   directory, but this is over 600 syscalls, well beyond the number
     81   allowable by perf.  These individual syscall events will however be
     82   useful if we want to later use the guidance we get from the
     83   general-purpose scripts to drill down and get more detail about
     84   individual syscalls of interest.
     85 
     86 - we can enable the sys_enter and/or sys_exit syscalls found under
     87   tracing/events/raw_syscalls.  These are called for all syscalls; the
     88   'id' field can be used to distinguish between individual syscall
     89   numbers.
     90 
     91 For this script, we only need to know that a syscall was entered; we
     92 don't care how it exited, so we'll use 'perf record' to record only
     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 data ]
     99 [ perf record: Captured and wrote 56.545 MB perf.data (~2470503 samples) ]
    100 ----
    101 
    102 The options basically say to collect data for every syscall event
    103 system-wide and multiplex the per-cpu output into a single stream.
    104 That single stream will be recorded in a file in the current directory
    105 called perf.data.
    106 
    107 Once we have a perf.data file containing our data, we can use the -g
    108 'perf script' option to generate a Python script that will contain a
    109 callback handler for each event type found in the perf.data trace
    110 stream (for more details, see the STARTER SCRIPTS section).
    111 
    112 ----
    113 # perf script -g python
    114 generated Python script: perf-script.py
    115 
    116 The output file created also in the current directory is named
    117 perf-script.py.  Here's the file in its entirety:
    118 
    119 # perf script event handlers, generated by perf script -g python
    120 # Licensed under the terms of the GNU GPL License version 2
    121 
    122 # The common_* event handler fields are the most useful fields common to
    123 # all events.  They don't necessarily correspond to the 'common_*' fields
    124 # in the format files.  Those fields not available as handler params can
    125 # be retrieved using Python functions of the form common_*(context).
    126 # See the perf-script-python Documentation for the list of available functions.
    127 
    128 import os
    129 import sys
    130 
    131 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
    132 	'/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
    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, context, common_cpu,
    144 	common_secs, common_nsecs, common_pid, common_comm,
    145 	id, args):
    146 		print_header(event_name, common_cpu, common_secs, common_nsecs,
    147 			common_pid, common_comm)
    148 
    149 		print "id=%d, args=%s\n" % \
    150 		(id, args),
    151 
    152 def trace_unhandled(event_name, context, common_cpu, common_secs, common_nsecs,
    153 		common_pid, common_comm):
    154 		print_header(event_name, common_cpu, common_secs, common_nsecs,
    155 		common_pid, common_comm)
    156 
    157 def print_header(event_name, cpu, secs, nsecs, pid, comm):
    158 	print "%-20s %5u %05u.%09u %8u %-20s " % \
    159 	(event_name, cpu, secs, nsecs, pid, comm),
    160 ----
    161 
    162 At the top is a comment block followed by some import statements and a
    163 path append which every perf script script should include.
    164 
    165 Following that are a couple generated functions, trace_begin() and
    166 trace_end(), which are called at the beginning and the end of the
    167 script respectively (for more details, see the SCRIPT_LAYOUT section
    168 below).
    169 
    170 Following those are the 'event handler' functions generated one for
    171 every event in the 'perf record' output.  The handler functions take
    172 the form subsystem__event_name, and contain named parameters, one for
    173 each field in the event; in this case, there's only one event,
    174 raw_syscalls__sys_enter().  (see the EVENT HANDLERS section below for
    175 more info on event handlers).
    176 
    177 The final couple of functions are, like the begin and end functions,
    178 generated for every script.  The first, trace_unhandled(), is called
    179 every time the script finds an event in the perf.data file that
    180 doesn't correspond to any event handler in the script.  This could
    181 mean either that the record step recorded event types that it wasn't
    182 really interested in, or the script was run against a trace file that
    183 doesn't correspond to the script.
    184 
    185 The script generated by -g option simply prints a line for each
    186 event found in the trace stream i.e. it basically just dumps the event
    187 and its parameter values to stdout.  The print_header() function is
    188 simply a utility function used for that purpose.  Let's rename the
    189 script and run it to see the default output:
    190 
    191 ----
    192 # mv perf-script.py syscall-counts.py
    193 # perf script -s syscall-counts.py
    194 
    195 raw_syscalls__sys_enter     1 00840.847582083     7506 perf                  id=1, args=
    196 raw_syscalls__sys_enter     1 00840.847595764     7506 perf                  id=1, args=
    197 raw_syscalls__sys_enter     1 00840.847620860     7506 perf                  id=1, args=
    198 raw_syscalls__sys_enter     1 00840.847710478     6533 npviewer.bin          id=78, args=
    199 raw_syscalls__sys_enter     1 00840.847719204     6533 npviewer.bin          id=142, args=
    200 raw_syscalls__sys_enter     1 00840.847755445     6533 npviewer.bin          id=3, args=
    201 raw_syscalls__sys_enter     1 00840.847775601     6533 npviewer.bin          id=3, args=
    202 raw_syscalls__sys_enter     1 00840.847781820     6533 npviewer.bin          id=3, args=
    203 .
    204 .
    205 .
    206 ----
    207 
    208 Of course, for this script, we're not interested in printing every
    209 trace event, but rather aggregating it in a useful way.  So we'll get
    210 rid of everything to do with printing as well as the trace_begin() and
    211 trace_unhandled() functions, which we won't be using.  That leaves us
    212 with this minimalistic skeleton:
    213 
    214 ----
    215 import os
    216 import sys
    217 
    218 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
    219 	'/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
    220 
    221 from perf_trace_context import *
    222 from Core import *
    223 
    224 def trace_end():
    225 	print "in trace_end"
    226 
    227 def raw_syscalls__sys_enter(event_name, context, common_cpu,
    228 	common_secs, common_nsecs, common_pid, common_comm,
    229 	id, args):
    230 ----
    231 
    232 In trace_end(), we'll simply print the results, but first we need to
    233 generate some results to print.  To do that we need to have our
    234 sys_enter() handler do the necessary tallying until all events have
    235 been counted.  A hash table indexed by syscall id is a good way to
    236 store that information; every time the sys_enter() handler is called,
    237 we simply increment a count associated with that hash entry indexed by
    238 that syscall id:
    239 
    240 ----
    241   syscalls = autodict()
    242 
    243   try:
    244     syscalls[id] += 1
    245   except TypeError:
    246     syscalls[id] = 1
    247 ----
    248 
    249 The syscalls 'autodict' object is a special kind of Python dictionary
    250 (implemented in Core.py) that implements Perl's 'autovivifying' hashes
    251 in Python i.e. with autovivifying hashes, you can assign nested hash
    252 values without having to go to the trouble of creating intermediate
    253 levels if they don't exist e.g syscalls[comm][pid][id] = 1 will create
    254 the intermediate hash levels and finally assign the value 1 to the
    255 hash entry for 'id' (because the value being assigned isn't a hash
    256 object itself, the initial value is assigned in the TypeError
    257 exception.  Well, there may be a better way to do this in Python but
    258 that's what works for now).
    259 
    260 Putting that code into the raw_syscalls__sys_enter() handler, we
    261 effectively end up with a single-level dictionary keyed on syscall id
    262 and having the counts we've tallied as values.
    263 
    264 The print_syscall_totals() function iterates over the entries in the
    265 dictionary and displays a line for each entry containing the syscall
    266 name (the dictonary keys contain the syscall ids, which are passed to
    267 the Util function syscall_name(), which translates the raw syscall
    268 numbers to the corresponding syscall name strings).  The output is
    269 displayed after all the events in the trace have been processed, by
    270 calling the print_syscall_totals() function from the trace_end()
    271 handler called at the end of script processing.
    272 
    273 The final script producing the output shown above is shown in its
    274 entirety below (syscall_name() helper is not yet available, you can
    275 only deal with id's for now):
    276 
    277 ----
    278 import os
    279 import sys
    280 
    281 sys.path.append(os.environ['PERF_EXEC_PATH'] + \
    282 	'/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
    283 
    284 from perf_trace_context import *
    285 from Core import *
    286 from Util import *
    287 
    288 syscalls = autodict()
    289 
    290 def trace_end():
    291 	print_syscall_totals()
    292 
    293 def raw_syscalls__sys_enter(event_name, context, common_cpu,
    294 	common_secs, common_nsecs, common_pid, common_comm,
    295 	id, args):
    296 	try:
    297 		syscalls[id] += 1
    298 	except TypeError:
    299 		syscalls[id] = 1
    300 
    301 def print_syscall_totals():
    302     if for_comm is not None:
    303 	    print "\nsyscall events for %s:\n\n" % (for_comm),
    304     else:
    305 	    print "\nsyscall events:\n\n",
    306 
    307     print "%-40s  %10s\n" % ("event", "count"),
    308     print "%-40s  %10s\n" % ("----------------------------------------", \
    309                                  "-----------"),
    310 
    311     for id, val in sorted(syscalls.iteritems(), key = lambda(k, v): (v, k), \
    312 				  reverse = True):
    313 	    print "%-40s  %10d\n" % (syscall_name(id), val),
    314 ----
    315 
    316 The script can be run just as before:
    317 
    318   # perf script -s syscall-counts.py
    319 
    320 So those are the essential steps in writing and running a script.  The
    321 process can be generalized to any tracepoint or set of tracepoints
    322 you're interested in - basically find the tracepoint(s) you're
    323 interested in by looking at the list of available events shown by
    324 'perf list' and/or look in /sys/kernel/debug/tracing events for
    325 detailed event and field info, record the corresponding trace data
    326 using 'perf record', passing it the list of interesting events,
    327 generate a skeleton script using 'perf script -g python' and modify the
    328 code to aggregate and display it for your particular needs.
    329 
    330 After you've done that you may end up with a general-purpose script
    331 that you want to keep around and have available for future use.  By
    332 writing a couple of very simple shell scripts and putting them in the
    333 right place, you can have your script listed alongside the other
    334 scripts listed by the 'perf script -l' command e.g.:
    335 
    336 ----
    337 root@tropicana:~# perf script -l
    338 List of available trace scripts:
    339   wakeup-latency                       system-wide min/max/avg wakeup latency
    340   rw-by-file <comm>                    r/w activity for a program, by file
    341   rw-by-pid                            system-wide r/w activity
    342 ----
    343 
    344 A nice side effect of doing this is that you also then capture the
    345 probably lengthy 'perf record' command needed to record the events for
    346 the script.
    347 
    348 To have the script appear as a 'built-in' script, you write two simple
    349 scripts, one for recording and one for 'reporting'.
    350 
    351 The 'record' script is a shell script with the same base name as your
    352 script, but with -record appended.  The shell script should be put
    353 into the perf/scripts/python/bin directory in the kernel source tree.
    354 In that script, you write the 'perf record' command-line needed for
    355 your script:
    356 
    357 ----
    358 # cat kernel-source/tools/perf/scripts/python/bin/syscall-counts-record
    359 
    360 #!/bin/bash
    361 perf record -a -e raw_syscalls:sys_enter
    362 ----
    363 
    364 The 'report' script is also a shell script with the same base name as
    365 your script, but with -report appended.  It should also be located in
    366 the perf/scripts/python/bin directory.  In that script, you write the
    367 'perf script -s' command-line needed for running your script:
    368 
    369 ----
    370 # cat kernel-source/tools/perf/scripts/python/bin/syscall-counts-report
    371 
    372 #!/bin/bash
    373 # description: system-wide syscall counts
    374 perf script -s ~/libexec/perf-core/scripts/python/syscall-counts.py
    375 ----
    376 
    377 Note that the location of the Python script given in the shell script
    378 is in the libexec/perf-core/scripts/python directory - this is where
    379 the script will be copied by 'make install' when you install perf.
    380 For the installation to install your script there, your script needs
    381 to be located in the perf/scripts/python directory in the kernel
    382 source tree:
    383 
    384 ----
    385 # ls -al kernel-source/tools/perf/scripts/python
    386 
    387 root@tropicana:/home/trz/src/tip# ls -al tools/perf/scripts/python
    388 total 32
    389 drwxr-xr-x 4 trz trz 4096 2010-01-26 22:30 .
    390 drwxr-xr-x 4 trz trz 4096 2010-01-26 22:29 ..
    391 drwxr-xr-x 2 trz trz 4096 2010-01-26 22:29 bin
    392 -rw-r--r-- 1 trz trz 2548 2010-01-26 22:29 check-perf-script.py
    393 drwxr-xr-x 3 trz trz 4096 2010-01-26 22:49 Perf-Trace-Util
    394 -rw-r--r-- 1 trz trz 1462 2010-01-26 22:30 syscall-counts.py
    395 ----
    396 
    397 Once you've done that (don't forget to do a new 'make install',
    398 otherwise your script won't show up at run-time), 'perf script -l'
    399 should show a new entry for your script:
    400 
    401 ----
    402 root@tropicana:~# perf script -l
    403 List of available trace scripts:
    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
    408 ----
    409 
    410 You can now perform the record step via 'perf script record':
    411 
    412   # perf script record syscall-counts
    413 
    414 and display the output using 'perf script report':
    415 
    416   # perf script report syscall-counts
    417 
    418 STARTER SCRIPTS
    419 ---------------
    420 
    421 You can quickly get started writing a script for a particular set of
    422 trace data by generating a skeleton script using 'perf script -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.
    427 
    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-script.py script, while not interesting for its results,
    432 attempts to exercise all of the main scripting features.
    433 
    434 EVENT HANDLERS
    435 --------------
    436 
    437 When perf script 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.
    442 
    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).
    446 
    447 As an example, the following perf record command can be used to record
    448 all sched_wakeup events in the system:
    449 
    450  # perf record -a -e sched:sched_wakeup
    451 
    452 Traces meant to be processed using a script should be recorded with
    453 the above option: -a to enable system-wide collection.
    454 
    455 The format file for the sched_wakep event defines the following fields
    456 (see /sys/kernel/debug/tracing/events/sched/sched_wakeup/format):
    457 
    458 ----
    459  format:
    460         field:unsigned short common_type;
    461         field:unsigned char common_flags;
    462         field:unsigned char common_preempt_count;
    463         field:int common_pid;
    464 
    465         field:char comm[TASK_COMM_LEN];
    466         field:pid_t pid;
    467         field:int prio;
    468         field:int success;
    469         field:int target_cpu;
    470 ----
    471 
    472 The handler function for this event would be defined as:
    473 
    474 ----
    475 def sched__sched_wakeup(event_name, context, common_cpu, common_secs,
    476        common_nsecs, common_pid, common_comm,
    477        comm, pid, prio, success, target_cpu):
    478        pass
    479 ----
    480 
    481 The handler function takes the form subsystem__event_name.
    482 
    483 The common_* arguments in the handler's argument list are the set of
    484 arguments passed to all event handlers; some of the fields correspond
    485 to the common_* fields in the format file, but some are synthesized,
    486 and some of the common_* fields aren't common enough to to be passed
    487 to every event as arguments but are available as library functions.
    488 
    489 Here's a brief description of each of the invariant event args:
    490 
    491  event_name 	  	    the name of the event as text
    492  context		    an opaque 'cookie' used in calls back into perf
    493  common_cpu		    the cpu the event occurred on
    494  common_secs		    the secs portion of the event timestamp
    495  common_nsecs		    the nsecs portion of the event timestamp
    496  common_pid		    the pid of the current task
    497  common_comm		    the name of the current process
    498 
    499 All of the remaining fields in the event's format file have
    500 counterparts as handler function arguments of the same name, as can be
    501 seen in the example above.
    502 
    503 The above provides the basics needed to directly access every field of
    504 every event in a trace, which covers 90% of what you need to know to
    505 write a useful trace script.  The sections below cover the rest.
    506 
    507 SCRIPT LAYOUT
    508 -------------
    509 
    510 Every perf script Python script should start by setting up a Python
    511 module search path and 'import'ing a few support modules (see module
    512 descriptions below):
    513 
    514 ----
    515  import os
    516  import sys
    517 
    518  sys.path.append(os.environ['PERF_EXEC_PATH'] + \
    519 	      '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
    520 
    521  from perf_trace_context import *
    522  from Core import *
    523 ----
    524 
    525 The rest of the script can contain handler functions and support
    526 functions in any order.
    527 
    528 Aside from the event handler functions discussed above, every script
    529 can implement a set of optional functions:
    530 
    531 *trace_begin*, if defined, is called before any event is processed and
    532 gives scripts a chance to do setup tasks:
    533 
    534 ----
    535 def trace_begin:
    536     pass
    537 ----
    538 
    539 *trace_end*, if defined, is called after all events have been
    540  processed and gives scripts a chance to do end-of-script tasks, such
    541  as display results:
    542 
    543 ----
    544 def trace_end:
    545     pass
    546 ----
    547 
    548 *trace_unhandled*, if defined, is called after for any event that
    549  doesn't have a handler explicitly defined for it.  The standard set
    550  of common arguments are passed into it:
    551 
    552 ----
    553 def trace_unhandled(event_name, context, common_cpu, common_secs,
    554         common_nsecs, common_pid, common_comm):
    555     pass
    556 ----
    557 
    558 The remaining sections provide descriptions of each of the available
    559 built-in perf script Python modules and their associated functions.
    560 
    561 AVAILABLE MODULES AND FUNCTIONS
    562 -------------------------------
    563 
    564 The following sections describe the functions and variables available
    565 via the various perf script Python modules.  To use the functions and
    566 variables from the given module, add the corresponding 'from XXXX
    567 import' line to your perf script script.
    568 
    569 Core.py Module
    570 ~~~~~~~~~~~~~~
    571 
    572 These functions provide some essential functions to user scripts.
    573 
    574 The *flag_str* and *symbol_str* functions provide human-readable
    575 strings for flag and symbolic fields.  These correspond to the strings
    576 and values parsed from the 'print fmt' fields of the event format
    577 files:
    578 
    579   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
    580   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
    581 
    582 The *autodict* function returns a special kind of Python
    583 dictionary that implements Perl's 'autovivifying' hashes in Python
    584 i.e. with autovivifying hashes, you can assign nested hash values
    585 without having to go to the trouble of creating intermediate levels if
    586 they don't exist.
    587 
    588   autodict() - returns an autovivifying dictionary instance
    589 
    590 
    591 perf_trace_context Module
    592 ~~~~~~~~~~~~~~~~~~~~~~~~~
    593 
    594 Some of the 'common' fields in the event format file aren't all that
    595 common, but need to be made accessible to user scripts nonetheless.
    596 
    597 perf_trace_context defines a set of functions that can be used to
    598 access this data in the context of the current event.  Each of these
    599 functions expects a context variable, which is the same as the
    600 context variable passed into every event handler as the second
    601 argument.
    602 
    603  common_pc(context) - returns common_preempt count for the current event
    604  common_flags(context) - returns common_flags for the current event
    605  common_lock_depth(context) - returns common_lock_depth for the current event
    606 
    607 Util.py Module
    608 ~~~~~~~~~~~~~~
    609 
    610 Various utility functions for use with perf script:
    611 
    612   nsecs(secs, nsecs) - returns total nsecs given secs/nsecs pair
    613   nsecs_secs(nsecs) - returns whole secs portion given nsecs
    614   nsecs_nsecs(nsecs) - returns nsecs remainder given nsecs
    615   nsecs_str(nsecs) - returns printable string in the form secs.nsecs
    616   avg(total, n) - returns average given a sum and a total number of values
    617 
    618 SEE ALSO
    619 --------
    620 linkperf:perf-script[1]
    621