Re: Value of __*{init,exit} anotations?

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Thu, 2008-01-31 at 00:32 +0200, Adrian Bunk wrote:
> On Wed, Jan 30, 2008 at 03:41:35PM -0600, James Bottomley wrote:
> > On Wed, 2008-01-30 at 22:20 +0100, Sam Ravnborg wrote:
> > > On Wed, Jan 30, 2008 at 03:00:16PM -0600, James Bottomley wrote:
> >...
> > > > __init is possibly justifiable with a few hundred k savings on boot.
> > > > __devinit and the rest are surely killable on the grounds they provide
> > > > little benefit for all the pain they cause.
> > > For the embedded people a few kb here and there is worth it.
> > > 
> > > > all __exit seems to do is set us up for unreferenced pointers in
> > > > discarded sections, so could we kill that too?
> > > Again - savings when we build-in the drivers.
> > > And without the checks we see 'funny' linker errors on the architectues
> > > that can continue to add the .exit.text in /DISCARD/
> > 
> > Perhaps you have different figures, but my standard kernel linking ones
> > tell me that the discard sections only save tens of k (not hundreds that
> > the init ones save), so I really do think they have no real benefit and
> > land us huge problems of pointer references into discarded sections.
> > 
> > I don't deny we can invest large amounts of work to fix our current
> > issues and build large scriptable checks to ensure we keep it fixed ...
> > I'm just asking if, at the end of the day, it's really worth it.
> 
> Some people consider it worth it for their memory restricted systems
> and would like to drive the annotations even further. [1]
> 
> My experience while fixing section bugs during the last years is that 
> the __dev{init,exit}* are actually the main question since they are both 
> the majority of annotations and the ones that bring benefits only 
> in a case that has become very exotic (CONFIG_HOTPLUG=n).
> 
> All the other annotations either both bring value for everyone
> (plain __init* and __exit*) or are nothing normal drivers would
> use (__cpu* and _mem*).
> 
> People at linux-arch (Cc'ed) might be better at explaining how often 
> CONFIG_HOTPLUG gets used in real-life systems and how big the savings 
> are there.
> 
> That might be a good basis for deciding whether it's worth it.

I'll certainly buy this.  Perhaps killing everything other than __init
and __exit (meaning discardable whether the system is hotplug, suspend
or whatever) might get rid of 90% of the problem while still preserving
90% of the benefits.  I think a lot of the issues do come from confusion
over whether it should be __init, __devinint etc .

We can argue later over the benefit of __exit ...

James