Re: lib/klist.c: bit 0 in pointer can't be used as flag

[Jesper Nilsson <Jesper.Nilsson@axis.com>]

On Wed, Jan 14, 2009 at 07:17:05PM +0100, Greg KH wrote:
> On Wed, Jan 14, 2009 at 04:12:21PM +0100, Jesper Nilsson wrote:
> > On Wed, Jan 14, 2009 at 11:45:28AM +0100, David Miller wrote:
> > > From: Jesper Nilsson <Jesper.Nilsson@axis.com>
> > > Date: Wed, 14 Jan 2009 11:36:24 +0100
> > > > On Wed, Jan 14, 2009 at 11:21:40AM +0100, David Miller wrote:
> > > > > From: Jesper Nilsson <Jesper.Nilsson@axis.com>
> > > > > Date: Wed, 14 Jan 2009 11:19:08 +0100
> > > > > > The following patch works for CRIS, and is less intrusive than
> > > > > > my earlier patch. If this the way to go I can resubmit a proper patch.
> > > > > 
> > > > > Out of curiosity, is there a way to get gcc to output code such
> > > > > that data objects are aligned more naturally?  Some option or
> > > > > similar?
> > > > 
> > > > There's flags for alignment of objects, but no flags for changing
> > > > structure layout or size, which is probably what we run into here.
> > > 
> > > Really?
> > 
> > Yes, after some more research, I found that the bug is indeed triggered by
> > the struct klist being aligned to odd bytes inside other structs.
> > 
> > Since CRIS uses packed structs, having a char * inside the kobjects
> > shifts the alignment for all data after any struct that contains
> > a kobject (and indeed inside the kobject itself)
> > 
> > In this case it was the class pointer of the struct device
> > that had a klist in the private pointer (struct class_private),
> > after a struct kset (which contains a kobject).
> 
> So, in order to make things smaller, your "use an external flag" patch
> would probably be better than forcing the alignment on a 4 byte boundry,
> right?

Unfortunately not, the GCC flags for CRIS only act on objects (address
of the first member in the struct), the structs themselves are
still subject to alignment on odd bytes. So the patch for aligning
all klists on 16 or 32 bits is still needed for CRIS.
It should however be a no-op for all architectures that do struct padding.

> thanks,
> greg k-h

/^JN - Jesper Nilsson
-- 
               Jesper Nilsson -- jesper.nilsson@axis.com