Re: [PATCH 10/11] x86: perf_event_intel_bts.c: use arch_initcall to hook in enabling

[Paul Gortmaker <paul.gortmaker@windriver.com>]

On 15-06-01 10:14 AM, Peter Zijlstra wrote:
> On Mon, 2015-06-01 at 10:11 -0400, Paul Gortmaker wrote:
>> On 15-06-01 03:04 AM, Peter Zijlstra wrote:
>>> On Sun, May 31, 2015 at 08:54:11PM -0400, Paul Gortmaker wrote:
>>>> This was using module_init, but there is no way this code can
>>>> be modular.  
>>>
>>> No, I think you could actually make it modular if you really wanted to.
>>
>> By "really wanted to" -- do you mean future changes that someone might do?
> 
> Yeah, I think we read "no way this code can be modular" differently.
> 
> I'm reading it like; it can easily be, all it requires is this little
> patch, where you're reading it like Kconfig changes.

OK, so how about I update both logs with something like this, to
clarify I meant "no way this code can be modular currently, given
the existing Kconfig situation":

----------------------8<---------------------------

This was using module_init, but the current Kconfig situation is
as follows:

In arch/x86/kernel/cpu/Makefile:

  obj-$(CONFIG_CPU_SUP_INTEL)    += perf_event_intel_pt.o perf_event_intel_bts.o

and in arch/x86/Kconfig.cpu:

  config CPU_SUP_INTEL
        default y
        bool "Support Intel processors" if PROCESSOR_SELECT

So currently, the end user can not build this code into a module.
If in the future, there is desire for this to be modular, then
it can be changed to include <linux/module.h> and use module_init.

But currently, in the non-modular case, a module_init becomes a
device_initcall.  But this really isn't a device, so we should
choose a more appropriate initcall bucket to put it in.

The obvious choice here seems to be arch_initcall, but that does
make it earlier than it was currently through device_initcall.
As long as perf_pmu_register() is functional, we should be OK.

 
----------------------8<----------------------------------

Paul.

> 
>