Re: [PATCH 2/5] perf_counter: Export various perf helpers for external users

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, Sep 10, 2009 at 09:28:30PM +1000, Paul Mackerras wrote:
> Frederic Weisbecker writes:
> 
> > Export various perf helpers that initialize, destroy, attach and detach
> > a perf counter for future external users like the hardware breakpoint api.
> 
> You are exporting things that are quite deep into the guts of the
> perf_counter code, which makes me think that what you're exporting
> isn't the right abstraction.  At the least, your commit message needs
> to talk about what these external users want to do and why they need
> to reach so deeply into the perf_counter internals.




Yeah those are quite deep in perf internals but I don't have
much the choice. I can't (I shouldn't) call sys_perf_open()
directly so I need to do about the same things that are done
from this syscall, without the file handling.

But I agree this needs more comments and explanations in the
changelog. I'll add these for the v2.


 
> > The allocation and initialization of a perf counter have been split up
> > so that an external user can first allocate and then prefill the
> > counter before initialize it properly.
> 
> Once again, this sounds like the wrong abstraction to me.  Can you
> explain why first allocating and then prefilling it before
> initializing it properly is necessary or desirable?
> 
> Paul.


You're right this is wrong. And this was even meant to be
reverted in further incremental patches.

The reason is that struct perf_counter embeds a struct hw_breakpoint.
And I need it to be inserted right between the counter allocation
and its initialization for it to take effect while calling
bp_perf_counter_init()

Ingo noticed the 1:1 relationship between struct hw_breakpoint
and struct perf_counter and suggested to remove the former and create
core breakpoint field inside perf attributes.

I like the idea, that would shrink the code and also bring a
unified way to set the breakpoints parameters in the counter
(for both syscall and in-kernel uses).

And then we won't need anymore this alloc/init split, which means
this change would have been reverted in further patches.

But I guess I should do that right in the beginning (in this
patchset) instead of later. So I will do that in the v2.

Thanks.