Re: [PATCH v2 28/39] timekeeping: Fix a circular include dependency

["Arnd Bergmann" <arnd@arndb.de>]

On Fri, Oct 27, 2023, at 01:54, Kent Overstreet wrote:
> On Fri, Oct 27, 2023 at 01:05:48AM +0200, Thomas Gleixner wrote:
>> On Thu, Oct 26 2023 at 18:33, Suren Baghdasaryan wrote:
>> > On Wed, Oct 25, 2023 at 5:33 PM Thomas Gleixner <tglx@linutronix.de> wrote:
>> >> > This avoids a circular header dependency in an upcoming patch by only
>> >> > making hrtimer.h depend on percpu-defs.h
>> >>
>> >> What's the actual dependency problem?
>> >
>> > Sorry for the delay.
>> > When we instrument per-cpu allocations in [1] we need to include
>> > sched.h in percpu.h to be able to use alloc_tag_save(). sched.h
>> 
>> Including sched.h in percpu.h is fundamentally wrong as sched.h is the
>> initial place of all header recursions.
>> 
>> There is a reason why a lot of funtionalitiy has been split out of
>> sched.h into seperate headers over time in order to avoid that.
>
> Yeah, it's definitely unfortunate. The issue here is that
> alloc_tag_save() needs task_struct - we have to pull that in for
> alloc_tag_save() to be inline, which we really want.
>
> What if we moved task_struct to its own dedicated header? That might be
> good to do anyways...

Yes, I agree that is the best way to handle it. I've prototyped
a more thorough header cleanup with good results (much improved
build speed) in the past, and most of the work to get there is
to seperate out structures like task_struct, mm_struct, net_device,
etc into headers that only depend on the embedded structure
definitions without needing all the inline functions associated
with them.

      Arnd