Re: [PATCH 09/10] MCDE: Add build files and bus

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Tue, Nov 30, 2010 at 03:05:33PM -0800, Greg KH wrote:
> On Tue, Nov 30, 2010 at 10:05:50PM +0000, Russell King - ARM Linux wrote:
> > On Tue, Nov 30, 2010 at 10:48:34AM -0800, Greg KH wrote:
> > > On Tue, Nov 30, 2010 at 06:40:49PM +0000, Russell King - ARM Linux wrote:
> > > > There's lots of static devices, not only platform devices, in the ARM
> > > > tree.  It's going to be a hell of a lot of work to fix this all up
> > > > properly.
> > > 
> > > I agree, it's been abused for many years this way :(
> > 
> > I don't agree that it is abuse - it was something explicitly allowed by
> > the original device model design by Patrick, with the condition that
> > such a device was never unregistered.  That's exactly the way we treat
> > these devices.
> 
> I understand Pat allowed this, I just don't agree that it's the correct
> thing to do :)
> 
> -mm had a patch for a long time that would throw up warnings if you ever
> did this for x86 so that arch should be clean of this issue by now.
> 
> > What I'm slightly concerned about is that this is going to needlessly
> > bloat the kernel - we're going to have to find some other way to store
> > this information, and create devices from that - which means additional
> > code to do the creation, and data structures for it to create these from.
> > There will be additional wastage from kmalloc as kmalloc doesn't allocate
> > just the size you ask for, but normally a power of two which will contain
> > the size.
> > 
> > That could potentially mean that as the device structure is 216 bytes,
> > kmalloc will use the 256 byte allocation size, which means a wastage of
> > 40 bytes per device structure.  On top of that goes the size of
> > resources with the allocation slop on top for that, and then there's
> > another allocation for the platform data.
> > 
> > Has anyone considered these implications before making this choice?
> 
> Yes, I have, which is one reason I haven't done this type of change yet.
> I need to figure out a way to not drasticly increase the size and still
> make it easy and simple for the platform and driver write their code.
> 
> It's a work in progress, but wherever possible, I encourage people to
> not make 'struct device' static.

Right, so saying to ARM developers that they can't submit code which
adds new static device structures is rather problematical then, and
effectively brings a section of kernel development to a complete
standstill - it means no support for additional ARM platforms until
this issue is resolved.  (This "condition" was mentioned by Arnd
earlier in this thread, and was put in such a way that it was now
a hard and fast rule.)

I feel it would be better to allow the current situation to continue.
If we start telling people that they can't use statically declared
devices without first having an alternative, we'll end up with people
inventing their own individual - and different - solutions to this
problem, which could actually make the problem harder to resolve in
the longer term.