Re: [perfmon2] How to reset sampling buffer before it overflows?

[<xu...@ic...>]

> 2010/11/22  <xu...@ic...>:
>> hi,
>>
>> I am writing my own app using the libpfm 3.10 API (interface v2.x) on a
>> X86 system.
>>
>> I setup a sampling context for the target-process to sample every 1
>> million insts retired. Traditionally, I should wait for the buffer
>> overflow message and process with the sampling buffer and then restart.
>> But I use another different mechanism (I will explain why, later) that I
>> setup another timer to count for a while (1s) and then stop the
>> target-process with PTRACE_ATTACH. I set the buffer big enough to ensure
>> it would not overflow before the timer expires. After reading the
>> sampling
>> entries, I try to reset the buffer using pfm_restart(), but it failed
>> with
>> error EBUSY (hdr_count field in the buffer header didn't reset).
>>
>> So, how to reset sampling buffer before it overflows?
>>
> Are you doing this with the kernel code in the GIT tree or a released
> patch?
>
I am using the official kernel 2.6.29 and the kernel patch from
http://sourceforge.net/projects/perfmon2/files/

> I looked at the code in __pfm_restart() and you cannot call pfm_restart()
> if you did not get a notification (can_restart> 0), unfortunately. I think
> this
> is because pfm_restart() does more than just reset the buffer. It also
> reloads
> the sampling periods.
>
Yes, I checked the kernel code too. So can I just add another syscall
which only re-writes the hdr_count (=0) and hdr_cur_off (=sizeof(*header))
field to reset the buffer? Will it be correct for my simple case where the
buffer never overflows?

Sorry I have little knowledge about the perfmon kernel.
Thanks for you help.

>> I do this because actually I am monitoring multiple processes that run
>> concurrently, I would like a inst-based sampling for each process.
>> However, due to their IPC difference, their sampling buffers won't
>> overflow simultaneously. That would involves CPU fragment if I wait for
>> all the buffers to overflow in my single-threaded program.
>>
>> Thanks very much!
>>
>>
>>
>>
>
>