perf trace cannot work with software events ?

perf trace cannot work with software events ?

[zhao bao]

 Hi,  I want to trace page faults in my program. The result I wanted like this:
 1st page fault at addrA.
 2nd page fault at addrB.
 3rd page fault at addrC
 ...

 Perf seems providing this function, but I can't get what I want.
 I use Perf on Fedora 14(kernel 2.6.35). The following is what I have
done in my machine.
Thanks for help in advance.

[root@localhost lab]# perf record -R -c 1  -d  -e  faults ./hello
hello,world
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.010 MB perf.data (~418 samples) ]

[root@localhost lab]# perf trace
  Fatal: no event_list!

Re: perf trace cannot work with software events ?

[Lin Ming]

On Sat, Mar 26, 2011 at 4:04 PM, zhao bao <baozhao@gmail.com> wrote:
>  Hi,  I want to trace page faults in my program. The result I wanted like this:
>  1st page fault at addrA.
>  2nd page fault at addrB.
>  3rd page fault at addrC
>  ...
>
>  Perf seems providing this function, but I can't get what I want.
>  I use Perf on Fedora 14(kernel 2.6.35). The following is what I have
> done in my machine.
> Thanks for help in advance.
>
> [root@localhost lab]# perf record -R -c 1  -d  -e  faults ./hello

It's used to trace kernel space page faults, not for userspace code.

Lin Ming

> hello,world
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.010 MB perf.data (~418 samples) ]
>
> [root@localhost lab]# perf trace
>  Fatal: no event_list!
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

Fwd: perf trace cannot work with software events ?

[zhao bao]

---------- Forwarded message ----------
From: zhao bao <baozhao@gmail.com>
Date: 2011/3/26
Subject: Re: perf trace cannot work with software events ?
To: Lin Ming <minggr@gmail.com>


>
> It's used to trace kernel space page faults, not for userspace code.
>
> Lin Ming
>

From the ouput of "perf report" command, I think  perf can record the
page faults from userspace code.

[root@localhost lab]# perf record -R -c 1  -d  -e  faults ./hello
hello,world
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.010 MB perf.data (~420 samples) ]
[root@localhost lab]# perf report
# Events: 92  cycles
#
# Overhead  Command      Shared Object  Symbol
# ........  .......  .................  ......
#
   10.87%    hello  ld-2.12.90.so      [.] .L198
    6.52%    hello  ld-2.12.90.so      [.] _dl_load_cache_lookup
    6.52%    hello  ld-2.12.90.so      [.] _dl_lookup_symbol_x
    4.35%    hello  ld-2.12.90.so      [.] dl_main
    3.26%    hello  ld-2.12.90.so      [.] _dl_start
    3.26%    hello  ld-2.12.90.so      [.] _dl_cache_libcmp
    3.26%    hello  ld-2.12.90.so      [.] do_lookup_x
    2.17%    hello  libc-2.12.90.so    [.] _init
    2.17%    hello  libc-2.12.90.so    [.] _IO_file_overflow_internal

Re: Fwd: perf trace cannot work with software events ?

[David Ahern]

On 03/26/11 06:31, zhao bao wrote:
>> It's used to trace kernel space page faults, not for userspace code.
>>
>> Lin Ming
>>
> 
> From the ouput of "perf report" command, I think  perf can record the
> page faults from userspace code.
> 
> [root@localhost lab]# perf record -R -c 1  -d  -e  faults ./hello
> hello,world

'-e faults' is a software event. perf-trace does not handle S/W events.
perf-trace has been renamed to perf-script, and perf-script recently
gained the ability to dump S/W events:

https://lkml.org/lkml/2011/3/10/3

David

> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.010 MB perf.data (~420 samples) ]
> [root@localhost lab]# perf report
> # Events: 92  cycles
> #
> # Overhead  Command      Shared Object  Symbol
> # ........  .......  .................  ......
> #
>    10.87%    hello  ld-2.12.90.so      [.] .L198
>     6.52%    hello  ld-2.12.90.so      [.] _dl_load_cache_lookup
>     6.52%    hello  ld-2.12.90.so      [.] _dl_lookup_symbol_x
>     4.35%    hello  ld-2.12.90.so      [.] dl_main
>     3.26%    hello  ld-2.12.90.so      [.] _dl_start
>     3.26%    hello  ld-2.12.90.so      [.] _dl_cache_libcmp
>     3.26%    hello  ld-2.12.90.so      [.] do_lookup_x
>     2.17%    hello  libc-2.12.90.so    [.] _init
>     2.17%    hello  libc-2.12.90.so    [.] _IO_file_overflow_internal
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Fwd: perf trace cannot work with software events ?

[zhao bao]

2011/3/26 David Ahern <dsahern@gmail.com>:
> On 03/26/11 06:31, zhao bao wrote:
>>> It's used to trace kernel space page faults, not for userspace code.
>>>
>>> Lin Ming
>>>
>>
>> From the ouput of "perf report" command, I think  perf can record the
>> page faults from userspace code.
>>
>> [root@localhost lab]# perf record -R -c 1  -d  -e  faults ./hello
>> hello,world
>
> '-e faults' is a software event. perf-trace does not handle S/W events.
> perf-trace has been renamed to perf-script, and perf-script recently
> gained the ability to dump S/W events:
>
> https://lkml.org/lkml/2011/3/10/3
>
> David
>
Thanks. How can I apply your patch?  May I need upgrade to 2.6.38?

Re: Fwd: perf trace cannot work with software events ?

[David Ahern]

On 03/26/11 08:00, zhao bao wrote:
> 2011/3/26 David Ahern <dsahern@gmail.com>:
>> On 03/26/11 06:31, zhao bao wrote:
>>>> It's used to trace kernel space page faults, not for userspace code.
>>>>
>>>> Lin Ming
>>>>
>>>
>>> From the ouput of "perf report" command, I think  perf can record the
>>> page faults from userspace code.
>>>
>>> [root@localhost lab]# perf record -R -c 1  -d  -e  faults ./hello
>>> hello,world
>>
>> '-e faults' is a software event. perf-trace does not handle S/W events.
>> perf-trace has been renamed to perf-script, and perf-script recently
>> gained the ability to dump S/W events:
>>
>> https://lkml.org/lkml/2011/3/10/3
>>
>> David
>>
> Thanks. How can I apply your patch?  May I need upgrade to 2.6.38?

You can compile latest perf userpsace and it should work with files from
older kernels.

David

Re: Fwd: perf trace cannot work with software events ?

[zhao bao]

> You can compile latest perf userpsace and it should work with files from
> older kernels.
>
> David
>

Follow your guide, I git the latest kernel and compiled the perf tools
, then perf script works.
But I found the outputed address is  instruction address, not the
address caused  the page  fault. Can I get these addresses with perf
tool?
[root@localhost lab]# /root/bin/perf  script
           hello 30242 [000] 48324.304652: page-faults:
c05b6d35 clear_user ([kernel.kallsyms])
           hello 30242 [000] 48324.304684: page-faults:
c05b6d35 clear_user ([kernel.kallsyms])
           hello 30242 [000] 48324.304701: page-faults:
c05b68a0 __copy_to_user_ll ([kernel.kallsyms])
           hello 30242 [000] 48324.304746: page-faults:
870 _start (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304750: page-faults:
4920 _dl_start (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304753: page-faults:
17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304755: page-faults:
4939 _dl_start (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304765: page-faults:
a790 _dl_setup_hash (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304768: page-faults:
14480 _dl_sysdep_start (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304771: page-faults:
14560 _dl_sysdep_start (/lib/ld-2.13.so)
           hello 30242 [000] 48324.304774: page-faults:
159e0 brk (/lib/ld-2.13.so)

Re: Fwd: perf trace cannot work with software events ?

[David Ahern]

On 03/26/11 10:13, zhao bao wrote:
> Follow your guide, I git the latest kernel and compiled the perf tools
> , then perf script works.
> But I found the outputed address is  instruction address, not the
> address caused  the page  fault. Can I get these addresses with perf
> tool?
> [root@localhost lab]# /root/bin/perf  script
>            hello 30242 [000] 48324.304652: page-faults:
> c05b6d35 clear_user ([kernel.kallsyms])
>            hello 30242 [000] 48324.304684: page-faults:
> c05b6d35 clear_user ([kernel.kallsyms])
>            hello 30242 [000] 48324.304701: page-faults:
> c05b68a0 __copy_to_user_ll ([kernel.kallsyms])
>            hello 30242 [000] 48324.304746: page-faults:
> 870 _start (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304750: page-faults:
> 4920 _dl_start (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304753: page-faults:
> 17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304755: page-faults:
> 4939 _dl_start (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304765: page-faults:
> a790 _dl_setup_hash (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304768: page-faults:
> 14480 _dl_sysdep_start (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304771: page-faults:
> 14560 _dl_sysdep_start (/lib/ld-2.13.so)
>            hello 30242 [000] 48324.304774: page-faults:
> 159e0 brk (/lib/ld-2.13.so)

Without callchains it shows the IP when the event triggered. If you want
more detail try adding -g to perf-record to get callchains.

Re: Fwd: perf trace cannot work with software events ?

[zhao bao]

2011/3/27 David Ahern <dsahern@gmail.com>:
>
>
> On 03/26/11 10:13, zhao bao wrote:
>> Follow your guide, I git the latest kernel and compiled the perf tools
>> , then perf script works.
>> But I found the outputed address is  instruction address, not the
>> address caused  the page  fault. Can I get these addresses with perf
>> tool?
>> [root@localhost lab]# /root/bin/perf  script
>>            hello 30242 [000] 48324.304652: page-faults:
>> c05b6d35 clear_user ([kernel.kallsyms])
>>            hello 30242 [000] 48324.304684: page-faults:
>> c05b6d35 clear_user ([kernel.kallsyms])
>>            hello 30242 [000] 48324.304701: page-faults:
>> c05b68a0 __copy_to_user_ll ([kernel.kallsyms])
>>            hello 30242 [000] 48324.304746: page-faults:
>> 870 _start (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304750: page-faults:
>> 4920 _dl_start (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304753: page-faults:
>> 17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304755: page-faults:
>> 4939 _dl_start (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304765: page-faults:
>> a790 _dl_setup_hash (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304768: page-faults:
>> 14480 _dl_sysdep_start (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304771: page-faults:
>> 14560 _dl_sysdep_start (/lib/ld-2.13.so)
>>            hello 30242 [000] 48324.304774: page-faults:
>> 159e0 brk (/lib/ld-2.13.so)
>
> Without callchains it shows the IP when the event triggered. If you want
> more detail try adding -g to perf-record to get callchains.
>
Thanks for your kind reply.I am sorry  for misleading you.

The address I want is the value stored  in X86 CR2 register when page
fault  occurs, Does perf provide these  information?

Another question.
I find  a new utility called 'trace' mentioned in
http://lwn.net/Articles/415728/  which provides I want (trace report
-p). But I git the tip tree,  and make -j install in   tools/perf
directory,  I couldn't find trace utility. Where is the trace utility?

Re: Fwd: perf trace cannot work with software events ?

[David Ahern]

On 03/27/11 01:44, zhao bao wrote:
> The address I want is the value stored  in X86 CR2 register when page
> fault  occurs, Does perf provide these  information?

Add the -d option to perf-record, and try the following change on
perf-script:

diff --git a/tools/perf/builtin-script.c b/tools/perf/builtin-script.c
index ac574ea..85376b7 100644
--- a/tools/perf/builtin-script.c
+++ b/tools/perf/builtin-script.c
@@ -32,6 +32,7 @@ enum perf_output_field {
    PERF_OUTPUT_EVNAME          = 1U << 5,
    PERF_OUTPUT_TRACE           = 1U << 6,
    PERF_OUTPUT_SYM             = 1U << 7,
+   PERF_OUTPUT_ADDR            = 1U << 8,
 };

 struct output_option {
@@ -46,6 +47,7 @@ struct output_option {
    {.str = "event", .field = PERF_OUTPUT_EVNAME},
    {.str = "trace", .field = PERF_OUTPUT_TRACE},
    {.str = "sym",   .field = PERF_OUTPUT_SYM},
+   {.str = "addr",   .field = PERF_OUTPUT_ADDR},
 };

 /* default set to maintain compatibility with current format */
@@ -158,6 +160,8 @@ static void print_sample_start(struct perf_sample
*sample,

        printf("%s: ", evname ? evname : "(unknown)");
    }
+   if (PRINT_FIELD(ADDR))
+       printf("0x%" PRIx64, sample->addr);
 }

 static void process_event(union perf_event *event __unused,
@@ -829,7 +833,7 @@ static const struct option options[] = {
    OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
            "Look for files with symbols relative to this directory"),
    OPT_CALLBACK('f', "fields", NULL, "str",
-            "comma separated output fields prepend with 'type:'. Valid
types: hw,sw,trace. Fields: comm,tid,pid,time,cpu,event,trace,sym",
+            "comma separated output fields prepend with 'type:'. Valid
types: hw,sw,trace. Fields: comm,tid,pid,time,cpu,event,trace,sym,addr",
             parse_output_fields),

    OPT_END()


> 
> Another question.
> I find  a new utility called 'trace' mentioned in
> http://lwn.net/Articles/415728/  which provides I want (trace report
> -p). But I git the tip tree,  and make -j install in   tools/perf
> directory,  I couldn't find trace utility. Where is the trace utility?

Re: Fwd: perf trace cannot work with software events ?

[zhao bao]

2011/3/28 David Ahern <dsahern@gmail.com>:
>
>
> On 03/27/11 01:44, zhao bao wrote:
>> The address I want is the value stored  in X86 CR2 register when page
>> fault  occurs, Does perf provide these  information?
>
> Add the -d option to perf-record, and try the following change on
> perf-script:
>
> diff --git a/tools/perf/builtin-script.c b/tools/perf/builtin-script.c
> index ac574ea..85376b7 100644
> --- a/tools/perf/builtin-script.c
> +++ b/tools/perf/builtin-script.c
> @@ -32,6 +32,7 @@ enum perf_output_field {
>    PERF_OUTPUT_EVNAME          = 1U << 5,
>    PERF_OUTPUT_TRACE           = 1U << 6,
>    PERF_OUTPUT_SYM             = 1U << 7,
> +   PERF_OUTPUT_ADDR            = 1U << 8,
>  };
>
>  struct output_option {
> @@ -46,6 +47,7 @@ struct output_option {
>    {.str = "event", .field = PERF_OUTPUT_EVNAME},
>    {.str = "trace", .field = PERF_OUTPUT_TRACE},
>    {.str = "sym",   .field = PERF_OUTPUT_SYM},
> +   {.str = "addr",   .field = PERF_OUTPUT_ADDR},
>  };
>
>  /* default set to maintain compatibility with current format */
> @@ -158,6 +160,8 @@ static void print_sample_start(struct perf_sample
> *sample,
>
>        printf("%s: ", evname ? evname : "(unknown)");
>    }
> +   if (PRINT_FIELD(ADDR))
> +       printf("0x%" PRIx64, sample->addr);
>  }
>
>  static void process_event(union perf_event *event __unused,
> @@ -829,7 +833,7 @@ static const struct option options[] = {
>    OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
>            "Look for files with symbols relative to this directory"),
>    OPT_CALLBACK('f', "fields", NULL, "str",
> -            "comma separated output fields prepend with 'type:'. Valid
> types: hw,sw,trace. Fields: comm,tid,pid,time,cpu,event,trace,sym",
> +            "comma separated output fields prepend with 'type:'. Valid
> types: hw,sw,trace. Fields: comm,tid,pid,time,cpu,event,trace,sym,addr",
>             parse_output_fields),
>
>    OPT_END()
>
>
>>
I applyed this patch in the builtin-script.c , recompile and install
it. But I found it still doesn't  show any CR2 address info. The patch
seems to list a addr column, but perf script still work as before.

I tried two ways.
(1) In fedora 14,  new "/root/bin/perf record -d" cannot work because
of  "-d" option, so I use old perf record (perf version
2.6.35.6-45.fc14.i686), then use new "perf script" , result is still
the same as before.

[root@localhost lab]# perf --version
perf version 2.6.35.6-45.fc14.i686

[root@localhost lab]# /root/bin/perf --version
perf version 2.6.38.2972.gf88210.dirty

[root@localhost lab]# /root/bin/perf record  -c 1 -e faults  -R -d  ./hello

  Error: sys_perf_event_open() syscall returned with 22 (Invalid
argument).  /bin/dmesg may provide additional information.

  Fatal: No CONFIG_PERF_EVENTS=y kernel support configured?

[root@localhost lab]# perf record  -c 1 -e faults  -R -d  ./hello
hello,world
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.010 MB perf.data (~415 samples) ]

[root@localhost lab]# /root/bin/perf script
           hello  2924 [000]  1422.290866: page-faults:
c05b6d35 clear_user ([kernel.kallsyms])
           hello  2924 [000]  1422.290894: page-faults:
c05b6d35 clear_user ([kernel.kallsyms])
           hello  2924 [000]  1422.290910: page-faults:
c05b68a0 __copy_to_user_ll ([kernel.kallsyms])
           hello  2924 [000]  1422.290946: page-faults:
870 _start (/lib/ld-2.13.so)

(2) I compiled the 2.6.38.rc6 kernel(I downloaded a month ago) ,
reboot machien. Now new "/root/bin/perf record -d" can work, but  new
"perf script" shows no change.

[root@localhost lab]#  /root/bin/perf record  -c 1 -e faults  -R -d  ./hello
hello,world
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.013 MB perf.data (~557 samples) ]
[root@localhost lab]#  /root/bin/perf script
           hello  2229 [000]  1339.388865: page-faults:
c05ccbb3 clear_user ([kernel.kallsyms])
           hello  2229 [000]  1339.388898: page-faults:
c05ccbb3 clear_user ([kernel.kallsyms])
           hello  2229 [000]  1339.388915: page-faults:
c05cc70c __copy_to_user_ll ([kernel.kallsyms])
           hello  2229 [000]  1339.388945: page-faults:
870 _start (/lib/ld-2.13.so)
           hello  2229 [000]  1339.388950: page-faults:
4920 _dl_start (/lib/ld-2.13.so)
           hello  2229 [000]  1339.388954: page-faults:
17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)

Re: Fwd: perf trace cannot work with software events ?

[baozhao]

zhao bao <baozhao <at> gmail.com> writes:


> I applyed this patch in the builtin-script.c , recompile and install
> it. But I found it still doesn't  show any CR2 address info. The patch
> seems to list a addr column, but perf script still work as before.

 Hi, I think I find the problem.
In builtin-script.c,   array output_fields should also be
changed,PERF_OUTPUT_ADDR should added to  [PERF_TYPE_SOFTWARE],  am I
right?

static u64 output_fields[PERF_TYPE_MAX] = {
       ....
      [PERF_TYPE_SOFTWARE] = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | \
                              PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | \
                              PERF_OUTPUT_EVNAME | PERF_OUTPUT_SYM | \
                              PERF_OUTPUT_ADDR,


Here is the output:
[root@localhost lab]#  /root/bin/perf script
          hello  2229 [000]  1339.388865: page-faults: 0x804963c
  c05ccbb3 clear_user ([kernel.kallsyms])
          hello  2229 [000]  1339.388898: page-faults: 0x670834
 c05ccbb3 clear_user ([kernel.kallsyms])
          hello  2229 [000]  1339.388915: page-faults: 0xbf905d6b
   c05cc70c __copy_to_user_ll ([kernel.kallsyms])
          hello  2229 [000]  1339.388945: page-faults: 0x652870
      870 _start (/lib/ld-2.13.so)
          hello  2229 [000]  1339.388950: page-faults: 0x656920
     4920 _dl_start (/lib/ld-2.13.so)
          hello  2229 [000]  1339.388954: page-faults: 0x669acb
    17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)
          hello  2229 [000]  1339.388958: page-faults: 0x66fe54
     4939 _dl_start (/lib/ld-2.13.so)

Re: Fwd: perf trace cannot work with software events ?

[David Ahern]

On 03/29/11 10:55, baozhao wrote:
> zhao bao <baozhao <at> gmail.com> writes:
> 
> 
>> I applyed this patch in the builtin-script.c , recompile and install
>> it. But I found it still doesn't  show any CR2 address info. The patch
>> seems to list a addr column, but perf script still work as before.
> 
>  Hi, I think I find the problem.
> In builtin-script.c,   array output_fields should also be
> changed,PERF_OUTPUT_ADDR should added to  [PERF_TYPE_SOFTWARE],  am I
> right?

The defaults do need to better detect when it exists. Until then you can
add addr to the --fields option.


> 
> static u64 output_fields[PERF_TYPE_MAX] = {
>        ....
>       [PERF_TYPE_SOFTWARE] = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | \
>                               PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | \
>                               PERF_OUTPUT_EVNAME | PERF_OUTPUT_SYM | \
>                               PERF_OUTPUT_ADDR,
> 
> 
> Here is the output:
> [root@localhost lab]#  /root/bin/perf script
>           hello  2229 [000]  1339.388865: page-faults: 0x804963c
>   c05ccbb3 clear_user ([kernel.kallsyms])
>           hello  2229 [000]  1339.388898: page-faults: 0x670834
>  c05ccbb3 clear_user ([kernel.kallsyms])
>           hello  2229 [000]  1339.388915: page-faults: 0xbf905d6b
>    c05cc70c __copy_to_user_ll ([kernel.kallsyms])
>           hello  2229 [000]  1339.388945: page-faults: 0x652870
>       870 _start (/lib/ld-2.13.so)
>           hello  2229 [000]  1339.388950: page-faults: 0x656920
>      4920 _dl_start (/lib/ld-2.13.so)
>           hello  2229 [000]  1339.388954: page-faults: 0x669acb
>     17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)
>           hello  2229 [000]  1339.388958: page-faults: 0x66fe54
>      4939 _dl_start (/lib/ld-2.13.so)

And per another thread we can look at converting the sample addresses to
symbols/dsos.

David


> 
> 
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

perf trace

[Kumar Ranjit-B04060]

Hi David
I need to generate instruction trace for Freescale e500mc core. Please tell me whether I can use perf for that. 
If yes then please tell me some steps to do that.
Regards
Ranjit

-----Original Message-----
From: linux-perf-users-owner@vger.kernel.org [mailto:linux-perf-users-owner@vger.kernel.org] On Behalf Of David Ahern
Sent: Wednesday, March 30, 2011 1:36 AM
To: baozhao
Cc: linux-perf-users@vger.kernel.org
Subject: Re: Fwd: perf trace cannot work with software events ?



On 03/29/11 10:55, baozhao wrote:
> zhao bao <baozhao <at> gmail.com> writes:
> 
> 
>> I applyed this patch in the builtin-script.c , recompile and install 
>> it. But I found it still doesn't  show any CR2 address info. The 
>> patch seems to list a addr column, but perf script still work as before.
> 
>  Hi, I think I find the problem.
> In builtin-script.c,   array output_fields should also be
> changed,PERF_OUTPUT_ADDR should added to  [PERF_TYPE_SOFTWARE],  am I 
> right?

The defaults do need to better detect when it exists. Until then you can add addr to the --fields option.


> 
> static u64 output_fields[PERF_TYPE_MAX] = {
>        ....
>       [PERF_TYPE_SOFTWARE] = PERF_OUTPUT_COMM | PERF_OUTPUT_TID | \
>                               PERF_OUTPUT_CPU | PERF_OUTPUT_TIME | \
>                               PERF_OUTPUT_EVNAME | PERF_OUTPUT_SYM | \
>                               PERF_OUTPUT_ADDR,
> 
> 
> Here is the output:
> [root@localhost lab]#  /root/bin/perf script
>           hello  2229 [000]  1339.388865: page-faults: 0x804963c
>   c05ccbb3 clear_user ([kernel.kallsyms])
>           hello  2229 [000]  1339.388898: page-faults: 0x670834
>  c05ccbb3 clear_user ([kernel.kallsyms])
>           hello  2229 [000]  1339.388915: page-faults: 0xbf905d6b
>    c05cc70c __copy_to_user_ll ([kernel.kallsyms])
>           hello  2229 [000]  1339.388945: page-faults: 0x652870
>       870 _start (/lib/ld-2.13.so)
>           hello  2229 [000]  1339.388950: page-faults: 0x656920
>      4920 _dl_start (/lib/ld-2.13.so)
>           hello  2229 [000]  1339.388954: page-faults: 0x669acb
>     17acb __i686.get_pc_thunk.bx (/lib/ld-2.13.so)
>           hello  2229 [000]  1339.388958: page-faults: 0x66fe54
>      4939 _dl_start (/lib/ld-2.13.so)

And per another thread we can look at converting the sample addresses to symbols/dsos.

David


> 
> 
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe 
> linux-perf-users" in the body of a message to 
> majordomo@vger.kernel.org More majordomo info at  
> http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in the body of a message to majordomo@vger.kernel.org More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: perf trace

[David Ahern]

On 04/17/11 21:43, Kumar Ranjit-B04060 wrote:
> Hi David
> I need to generate instruction trace for Freescale e500mc core. Please tell me whether I can use perf for that. 
> If yes then please tell me some steps to do that.
> Regards
> Ranjit

I have no idea if that processor is supported by perf for H/W counters.

What do you mean by "instruction trace"?

David

RE: perf trace

[Kumar Ranjit-B04060]

It is supported by perf for H/W counter as i can use perf stat for various H/W events. I want to see actual instruction flow for a particular application. Like we have Kernel Function Trace for the kernel.
Regards
Ranjit

________________________________________
From: linux-perf-users-owner@vger.kernel.org [linux-perf-users-owner@vger.kernel.org] on behalf of David Ahern [dsahern@gmail.com]
Sent: Monday, April 18, 2011 11:38 AM
To: Kumar Ranjit-B04060
Cc: linux-perf-users@vger.kernel.org; baozhao
Subject: Re: perf trace

On 04/17/11 21:43, Kumar Ranjit-B04060 wrote:
> Hi David
> I need to generate instruction trace for Freescale e500mc core. Please tell me whether I can use perf for that.
> If yes then please tell me some steps to do that.
> Regards
> Ranjit

I have no idea if that processor is supported by perf for H/W counters.

What do you mean by "instruction trace"?

David

Re: perf trace

[David Ahern]

On 04/18/11 00:38, Kumar Ranjit-B04060 wrote:
> It is supported by perf for H/W counter as i can use perf stat for various H/W events. I want to see actual instruction flow for a particular application. Like we have Kernel Function Trace for the kernel.

meaning periodic samples of the counter? perf-record does not generate
periodic reads of H/W counters.

David

RE: perf trace

[Kumar Ranjit-B04060]

David


 > meaning periodic samples of the counter? perf-record does not generate periodic reads of H/W counters.

Just curious, How does then the profiling works? Perf report shows % no of sample point corresponding a particular event. So it must be sampling the PC periodically.
Regards
Ranjit

Re: perf trace

[David Ahern]

On 04/18/11 11:46, Kumar Ranjit-B04060 wrote:
> David
> 
> 
>  > meaning periodic samples of the counter? perf-record does not generate periodic reads of H/W counters.
> 
> Just curious, How does then the profiling works? Perf report shows % no of sample point corresponding a particular event. So it must be sampling the PC periodically.
> Regards
> Ranjit

With H/W counters the values of the counters are not read when samples
are created.

David

Re: Fwd: perf trace cannot work with software events ?

[zhao bao]

2011/3/28 David Ahern <dsahern@gmail.com>:
>
>
> On 03/27/11 01:44, zhao bao wrote:
>> The address I want is the value stored  in X86 CR2 register when page
>> fault  occurs, Does perf provide these  information?
>
> Add the -d option to perf-record, and try the following change on
> perf-script:
>

Many thanks to you! I hope your patch would be merged into mainline.

Re: Fwd: perf trace cannot work with software events ?

[Lin Ming]

On Sat, Mar 26, 2011 at 10:00 PM, zhao bao <baozhao@gmail.com> wrote:
> 2011/3/26 David Ahern <dsahern@gmail.com>:
>> On 03/26/11 06:31, zhao bao wrote:
>>>> It's used to trace kernel space page faults, not for userspace code.
>>>>
>>>> Lin Ming
>>>>
>>>
>>> From the ouput of "perf report" command, I think  perf can record the
>>> page faults from userspace code.

Right, I misunderstood your original question.