Re: [PATCH] perf symbol: Fix kernel symbol address display

[Thomas Richter]

On 4/24/20 8:23 PM, Arnaldo Carvalho de Melo wrote:
> Em Fri, Apr 24, 2020 at 08:01:22PM +0200, Thomas Richter escreveu:
>> On 4/24/20 5:06 PM, Arnaldo Carvalho de Melo wrote:
>>> Em Fri, Apr 24, 2020 at 02:37:01PM +0200, Thomas Richter escreveu:
>>>> On 4/20/20 10:46 PM, Arnaldo Carvalho de Melo wrote:
>>>>> Em Wed, Apr 15, 2020 at 09:07:44AM +0200, Thomas Richter escreveu:
>>>>>> Running commands
>>>>>>
>>>>>>    ./perf record -e rb0000 -- find .
>>>>>>    ./perf report -v
>>>>>
>>>>> Or when pressing 'V' in the TUI.
>>>>>  
>>>>>> reveals symbol names and its addresses. There is a mismatch between
>>>>>
>>>>> Yeah, an address that at some point was put there to help with debugging
>>>>> the symbol resolution, IIRC how it looked like when looking at
>>>>>
>>>>>   readelf -sW vmlinux
>>>>>
>>>>> Or any other DSO, for instance, for a glibc symbol here:
>>>>>
>>>>> Using 'perf report -s pid,dso,sym' then pressing 'V':
>>>>>
>>>>>    1.55%    20325:perf  /usr/lib64/libc-2.30.so   0x161825   B [.] __strlen_avx2
>>>>>
>>>>> [acme@five perf]$ readelf -sW /usr/lib64/libc-2.30.so | grep strlen_avx2
>>>>>  24371: 0000000000161810   414 FUNC    LOCAL  DEFAULT   15 __strlen_avx2
>>>>> [acme@five perf]$ 
>>>>>
>>>>> Can you check if doing in /lib/modules/.../build/vmlinux produces what
>>>>> appears when 'V' is in place?
>>>>>
>>>>> And perhaps we can also show the DSO offset and the rip as it gets laid
>>>>> out in memory in the end? So we have all the informations?
>>>>>
>>>>> - Arnaldo
>>>>>
>>>>
>>>> Arnaldo,
>>>>
>>>> having verified this also works in the TUI mode using 'V', are you going to pick
>>>> this patch?
>>>
>>> That would be a change in behaviour, the original intent was to show the
>>> value one would get from the ELF symbol table, isn't that the case?
>>>
>>
>> I think there is a misunderstanding.
>>
>> Your example above refers to libc, which shows the addresses
>> correctly (on x86 and s390). I was refering to the kernel dso.
>> The issue only shows up on the kernel DSO (vmlinux and kallsyms)
>>
>> When I use the -v option (or 'V' in TUI) I get this
>>
>>   3.55%  find /lib/modules/.../build/vmlinux  0xf11ec  v [k] check_chain_key
>>                                               ^^^^^^^
>> This address is not correct as compared to readelf on vmlinux file:
>>
>> [root@m35lp76 linux]# readelf -sW vmlinux| fgrep check_chain_key
>>  20698: 00000000001f0c70   486 FUNC    LOCAL  DEFAULT    1 check_chain_key
>> [root@m35lp76 linux]# 
>>
>> The problem are the mapping functions with are applied for the kernel DSO.
>> In dso__process_kernel_symbol() the mapping is adjusted (around line 903)
>> and  function hist_entry_iter_add() saves the modified address in the
>> struct hist_entry->ip. This address is used from now on and printed.
>>
>> Let me know and I will double check this and rework the commit message
>> as it was not good enough.
> 
> Yeah, this may be me not understanding something right, I saw this was a
> corner case, that only happens in debug mode (ok, better provide good
> info here...), so I haven't devoted more than cursory attention to this
> and may be making a bad judgement, so I thought that since this doesn't
> seems urgent, better leave it to when I manage to have solid time to
> devote to this or someone else looks at this.
> 
> Sorry, but these have been difficult time for everybody, and I fear, for
> me, I have to prioritise all the time.
> 
> - Arnaldo
>  

Yes, it is not easy these days.
I just wanted to make sure this patch is not lost. I already have been
asked when this will be fixed.

-- 
Thomas Richter, Dept 3252, IBM s390 Linux Development, Boeblingen, Germany
--
Vorsitzender des Aufsichtsrats: Matthias Hartmann
GeschÃftsfÃhrung: Dirk Wittkopp
Sitz der Gesellschaft: BÃblingen / Registergericht: Amtsgericht Stuttgart, HRB 243294