Re: Trouble profiling kernel module

[<job...@ao...>]

Sent from my iPhone 3G

On Nov 18, 2009, at 4:00 PM, Maynard Johnson <may...@us...>  
wrote:

> job...@ao... wrote:
>> On Nov 18, 2009, at 10:08 AM, Maynard Johnson <may...@us...>
>> wrote:
>>
>>> job...@ao... wrote:
>>>> On Nov 17, 2009, at 9:23 AM, job...@ao... wrote:
>>>>
>>>>> By looking at EIP values, I can see that samples are being  
>>>>> collected
>>>>> for my test module.  In the function get_file in daemon/ 
>>>>> opd_sfile.c
>>>>> which is called after issuing the opcontrol --shutdown command, I
>>>>> can see trans->pc values that correspond to the EIP values of
>>>>> samples collected for my test module but no new sample file is
>>>>> created for these samples.
>>>>>
>>>>> Since I can see the samples are there for my test module, why  
>>>>> isn't
>>>>> a sample file created for the test module samples?
>>>> When get_file is called for the samples for my test module, the  
>>>> file
>>>> it gets is /var/lib/oprofile/samples//current/{kern}/no-vmlinux/
>>>> {dep}/
>>>> {kern}/no-vmlinux/TIMER.0.0.all.all.all, which has a nonzero count.
>>>> Thus, no sample file is created for my test module.
>>>>
>>>> Why would samples for my test module result in getting the sample
>>>> file
>>>> for no-vmlinux?
>>> Back on Nov 10, Jose Santos correctly told you that using the --no-
>>> vmlinux option will not work if you want to see samples for kernel
>>> modules.  When you use that option, oprofile throws all kernel
>>> samples (whether from base kernel or kernel modules) into the no-
>>> vmlinux sample file.
>>>
>>> -Maynard
>>
>> Thank you for the clarification.  Here are the commands I am now  
>> using
>> after the system boots up:
>>
>> Eclipse # rm -rf /var/lib/oprofile/
>> Eclipse # mkdir /var/lib/oprofile/
>> Eclipse # mount -t ramfs non /var/lib/oprofile/
>> Eclipse # rm /root/.oprofile/daemonrc
>> Eclipse # rm /var/log/messages
>> Eclipse # insmod oprofile.ko
>> Eclipse # insmod helloworld.ko
>> Eclipse # /usr/bin/opcontrol --setup --kernel-
>> range=0xC0000000,0xC9500000 --vml
>> inux=/root/vmlinux.PRO_MOD
>> Eclipse # /usr/bin/opcontrol --start
>> Using 2.6+ OProfile kernel interface.
>> Reading module info.
>> Using log file /var/lib/oprofile/samples/oprofiled.log
>> Daemon started.
>> Profiler running.
>> Eclipse # /usr/bin/opcontrol --shutdown
>> Stopping profiling.
>> Killing daemon.
>> Eclipse # /usr/bin/opreport --image-path=/root
>> CPU: CPU with timer interrupt, speed 0 MHz (estimated)
>> Profiling through timer interrupt
>>           TIMER:0|
>>   samples|      %|
>> ------------------
>>      9373 93.0600 vmlinux.PRO_MOD
>>       587  5.8280 libc-2.6.so
>>        39  0.3872 oprofiled
>>        27  0.2681EclipseMSW.elf.NOV8
>>        23  0.2284 busybox-1.15.2.mtab
>>        19  0.1886 ld-2.6.so
>>         3  0.0298 snmpd
>>         1  0.0099 gateway
>>
>>
>>
>> Using printf's, I can see in get_file that for the first sample (and
>> all subsequent samples) with a trans->pc value that corresponds to
>> helloworld (e.g., 0xC907C0DC, helloworld start=0xC907C000
>> end=0xC907C5a0), odb_open_count(file) returns 1.  Since  
>> odb_open_count
>
> hmmm . . . from the addresses you state above for where your kernel  
> module resides, it seems you are incorrectly specifying the --kernel- 
> range value.  That value should *only* be for the base kernel.  On  
> most platforms, you don't need to specify this value at all --  
> opcontrol will automagically figure it out.
>
> -Maynard

You are correct.  The end of the .text section should be 0xC02633B0.   
I now see my sample module in the report.

Thank you.

>
>> (file) returns 1, no new sample file is created for helloworld.ko.
>> The value of  odb->data->filename in odb_open_count is /var/lib/
>> oprofile/samples//current/{root}//root/vmlinux.PRO_MOD/{dep}/{root}//
>> root/vmlinux.PRO_MOD/TIMER.0.0.all.all.all.  Thus, the samples for
>> helloworld.ko are put into the sample file for vmlinux.PRO_MOD.
>>
>> Should odb_open_count(file) return 1 when odb_open_count(file) is
>> called for the first time for a sample from helloworld.ko?  Shouldn't
>> odb->data be NULL the first time a sample is processed for
>> helloworld.ko so that a sample file for helloworld.ko is created?
>>
>>
>>>>> On Nov 12, 2009, at 3:42 PM, job...@ao... wrote:
>>>>>
>>>>>> Could the fact that I am following the instructions below be an
>>>>>> issue as far as sampling inserted kernel modules?  Of course,
>>>>>> samples for other than the kernel modules seem to be stored just
>>>>>> fine.
>>>>>> If running from an image in the internal jffs2 file system, you
>>>>>> will need to mount another file system to store the samples.  
>>>>>> Below
>>>>>> are examples of the commands to set that up. You will need to
>>>>>> modify for local conditions.
>>>>>> mkdir /var/lib/oprofile
>>>>>> mount -t ramfs none /var/lib/oprofile
>>>>>>
>>>>>> =
>>>>> =
>>
>