Re: [PATCH perf/core v3 3/3] [BUGFIX] perf probe: Show usage even if the last event is skipped

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

On 2015/06/16 23:46, Arnaldo Carvalho de Melo wrote:
> Em Tue, Jun 16, 2015 at 08:50:57PM +0900, Masami Hiramatsu escreveu:
>> When the last part of converted events are blacklisted or out-of-text,
>> those are skipped and perf probe doesn't show usage examples.
>> This fixes it to show the example even if the last part of event list
>> is skipped.
>>
>> E.g. without this patch, events are added, but suddenly end;
> 
> End what? Stop being added?

"End without the last message", I meant.

> I.e. not all eligible events are added? From
> your description the problem seems to be that that last message: "You
> can now use it..."  is not presented, but here, without this patch, it
> is:

I see, actually, this happens only if the skipped symbols (
vfs_caches_init_early or vfs_caches_init) are placed at the end of the
matched symbol list. On Ubuntu 15.04 kernel, it doesn't have
vfs_load_quota_inode etc. and the vfs_caches_init is the last part of
the matched list. Since it is hard to reproduce, I've added a Note on
the end of description :)

----

>>
>> Note that this can be reproduced ONLY IF the vfs_caches_init*
>> is the last part of matched symbol list. I've checked this happens
>> on "3.19.0-generic #18-Ubuntu" kernel binary.
>>

----

To reproduce this bug, you need to find a good symbol matching pattern
which (1) matches both of valid function in .text and invalid function
in .inittext (2) invalid one must be on the end of matched function list.

I fortunately hit such pattern and found this bug, but it depends on
the kernel binary.

[...]


> I.e. the only problem I found was this:
> 
> [root@zoo ~]# time perf probe -l > /dev/null
> 
> real	0m15.408s
> user	0m14.892s
> sys	0m0.534s
> [root@zoo ~]# 
> [root@zoo ~]# perf stat perf probe -l > /dev/null
> 
>  Performance counter stats for 'perf probe -l':
> 
>       15256.588897      task-clock (msec)         #    1.001 CPUs utilized          
>                116      context-switches          #    0.008 K/sec                  
>                  4      cpu-migrations            #    0.000 K/sec                  
>            230,720      page-faults               #    0.015 M/sec                  
>     47,830,405,530      cycles                    #    3.135 GHz                    
>     43,974,134,505      stalled-cycles-frontend   #   91.94% frontend cycles idle   
>    <not supported>      stalled-cycles-backend   
>     11,540,587,038      instructions              #    0.24  insns per cycle        
>                                                   #    3.81  stalled cycles per insn
>      2,807,769,592      branches                  #  184.037 M/sec                  
>         20,087,075      branch-misses             #    0.72% of all branches        
> 
>       15.240796324 seconds time elapsed
> 
> [root@zoo ~]#
> 
> Can you check why it takes so long and check the need for this patch?

It is because perf probe -l is not optimized to show a lot of probes yet.
It initializes and loads debuginfo for each probe. I guess we can reuse
debuginfo among them. let me try...

Thank you,


-- 
Masami HIRAMATSU
Linux Technology Research Center, System Productivity Research Dept.
Center for Technology Innovation - Systems Engineering
Hitachi, Ltd., Research & Development Group
E-mail: masami.hiramatsu.pt@hitachi.com