Re: perf tools event parsing code

[Jiri Olsa <jolsa@redhat.com>]

On Fri, Jul 31, 2015 at 05:55:32PM -0300, Arnaldo Carvalho de Melo wrote:
> Trying to figure out why this cset, by Wangnan:
> 
> https://git.kernel.org/cgit/linux/kernel/git/acme/linux.git/commit/?h=perf/ebpf&id=42cb6deef348f6afece0ad6fd7b034bfb532000e
> 
> doesn't make the routine he associated with parsing '--event bpf_program.o',
> ended up in this:

that change looks ok to me, but this
backtrac eis from pmu cpu lookup/parsing, please check below
> 
> #0  parse_events__scan_bytes (yybytes=yybytes@entry=0x7fffffff8480 "event=0x2e,umask=0x4f\n", _yybytes_len=22, yyscanner=0x18e4c30)
>     at util/parse-events-flex.c:3312
> #1  0x00000000004addb3 in parse_events__scan_string (yystr=<optimized out>, yystr@entry=0x7fffffff8480 "event=0x2e,umask=0x4f\n", yyscanner=<optimized out>)
>     at util/parse-events-flex.c:3274
> #2  0x000000000048a5e7 in parse_events__scanner (str=str@entry=0x7fffffff8480 "event=0x2e,umask=0x4f\n", data=data@entry=0x7fffffff83f0, 
>     start_token=start_token@entry=259) at util/parse-events.c:1081
> #3  0x000000000048ca52 in parse_events_terms (terms=terms@entry=0x18e4a88, str=str@entry=0x7fffffff8480 "event=0x2e,umask=0x4f\n")
>     at util/parse-events.c:1104
> #4  0x00000000004b2497 in __perf_pmu__new_alias (list=list@entry=0x7fffffffa700, dir=dir@entry=0x7fffffff96b0 "/sys/bus/event_source/devices/cpu/events", 
>     name=name@entry=0x18d864b "cache-references", desc=desc@entry=0x0, val=val@entry=0x7fffffff8480 "event=0x2e,umask=0x4f\n") at util/pmu.c:224
> #5  0x00000000004b25c0 in perf_pmu__new_alias (list=list@entry=0x7fffffffa700, dir=dir@entry=0x7fffffff96b0 "/sys/bus/event_source/devices/cpu/events", 
>     name=name@entry=0x18d864b "cache-references", file=file@entry=0x18d67e0) at util/pmu.c:258
> #6  0x00000000004b2788 in pmu_aliases_parse (dir=dir@entry=0x7fffffff96b0 "/sys/bus/event_source/devices/cpu/events", head=head@entry=0x7fffffffa700)
>     at util/pmu.c:313
> #7  0x00000000004b2864 in pmu_aliases (name=name@entry=0x5b9c47 "cpu", head=head@entry=0x7fffffffa700) at util/pmu.c:340
> #8  0x00000000004b328a in pmu_lookup (name=<optimized out>, name@entry=0x5b9c47 "cpu") at util/pmu.c:477
> #9  0x00000000004b33de in perf_pmu__find (name=name@entry=0x5b9c47 "cpu") at util/pmu.c:541
> #10 0x000000000048a445 in perf_pmu__parse_init () at util/parse-events.c:1008
> #11 0x000000000048b86d in perf_pmu__parse_check (name=name@entry=0x18d6c40 "file.c") at util/parse-events.c:1054
> #12 0x00000000004ad4e2 in pmu_str_check (scanner=scanner@entry=0x18d6f20) at util/parse-events.l:74

this is support for the event term shortcut, like using

  perf record -e mem-loads ...
    
instead of:
    
  perf record -e cpu/mem-loads/

it calls 'perf_pmu__find("cpu") which triggers 'cpu' pmu 
lookup 'pmu_lookup("cpu")'

and that one triggers alias parsing above as well  ;-)

> #13 0x00000000004af78b in parse_events_lex (yylval_param=yylval_param@entry=0x7fffffffa900, yylloc_param=yylloc_param@entry=0x7fffffffa920, 
>     yyscanner=yyscanner@entry=0x18d6f20) at util/parse-events.l:267
> #14 0x00000000004b042c in parse_events_parse (_data=_data@entry=0x7fffffffbec0, scanner=0x18d6f20) at util/parse-events-bison.c:1570
> #15 0x000000000048a5f6 in parse_events__scanner (str=str@entry=0x7fffffffe778 "file.c", data=data@entry=0x7fffffffbec0, start_token=start_token@entry=258)
>     at util/parse-events.c:1086
> #16 0x000000000048b955 in parse_events (evlist=0x18d7b90, str=str@entry=0x7fffffffe778 "file.c", err=err@entry=0x7fffffffbf10) at util/parse-events.c:1126

events parsing starts here ^^^

> #17 0x000000000048b9f7 in parse_events_option (opt=<optimized out>, str=0x7fffffffe778 "file.c", unset=<optimized out>) at util/parse-events.c:1224
> #18 0x0000000000487f30 in get_value (p=p@entry=0x7fffffffc0b0, opt=0x831c00 <__record_options>, flags=flags@entry=1) at util/parse-options.c:151
> #19 0x000000000048823e in parse_short_opt (p=p@entry=0x7fffffffc0b0, options=<optimized out>, options@entry=0x831c00 <__record_options>)
>     at util/parse-options.c:230
> #20 0x0000000000488d6e in parse_options_step (ctx=ctx@entry=0x7fffffffc0b0, options=options@entry=0x831c00 <__record_options>, 
>     usagestr=usagestr@entry=0x599850 <__record_usage>) at util/parse-options.c:401
> #21 0x000000000048918c in parse_options_subcommand (argc=argc@entry=5, argv=argv@entry=0x7fffffffe520, options=0x831c00 <__record_options>, 
>     subcommands=subcommands@entry=0x0, usagestr=0x599850 <__record_usage>, flags=flags@entry=2) at util/parse-options.c:518
> #22 0x00000000004892da in parse_options (argc=argc@entry=5, argv=argv@entry=0x7fffffffe520, options=<optimized out>, usagestr=<optimized out>, 
>     flags=flags@entry=2) at util/parse-options.c:553
> #23 0x000000000042ad70 in cmd_record (argc=5, argv=0x7fffffffe520, prefix=<optimized out>) at builtin-record.c:1102
> #24 0x0000000000474253 in run_builtin (p=p@entry=0x83c7f0 <commands+144>, argc=argc@entry=5, argv=argv@entry=0x7fffffffe520) at perf.c:370
> #25 0x000000000047443f in handle_internal_command (argc=5, argv=0x7fffffffe520) at perf.c:429
> #26 0x00000000004744b0 in run_argv (argcp=argcp@entry=0x7fffffffe38c, argv=argv@entry=0x7fffffffe380) at perf.c:473
> #27 0x0000000000474726 in main (argc=5, argv=0x7fffffffe520) at perf.c:588
> (gdb) 
> 
> /me trying to figure out this problem, but also how to avoid doing all this
> seemingly unnecessary work, i.e. reading everything on the sysfs pmu files when
> trying to figure out if this is a bpf filter event (-e foo.o).

it's deep, but it's initialized only once.. after it's initialized,
there's no other parsing

jirka