Re: 2.6.23-rc8 build failure: __you_cannot_kmalloc_that_much in dmi_id_init

[Jean Delvare <khali@linux-fr.org>]

Hi Andrew,

On Thu, 4 Oct 2007 01:48:28 -0700, Andrew Morton wrote:
> On Tue, 2 Oct 2007 23:26:42 +0200 Jean Delvare <khali@linux-fr.org> wrote:
> 
> > On Mon, 1 Oct 2007 23:54:12 +0200, Jean Delvare wrote:
> > > On Mon, 1 Oct 2007 22:54:47 +0200, Jean Delvare wrote:
> > > > 2.6.23-rc8 and 2.6.23-rc8-git4 fail to build on one of my test
> > > > machines, with:
> > > > 
> > > > drivers/built-in.o(.init.text+0x780e): In function `dmi_id_init':
> > > > : undefined reference to `__you_cannot_kmalloc_that_much'
> > > > 
> > > > The code is allocating sizeof(struct device) so it really shouldn't be
> > > > a problem. I have no idea what's wrong. That's on i386, very old
> > > > machine (Pentium 166MMX / Intel TX chipset), with gcc 3.2.3 and
> > > > binutils 2.14.90.0.6. 2.6.22.9 compiles fine on the same system (but it
> > > > doesn't include dmi-id so it's not very surprising).
> > > > 
> > > > .config attached.
> > > 
> > > More information: building the same config on a much more recent system
> > > works fine. This seems to point at a toolchain issue.
> > 
> > More information:
> > * No improvement in 2.6.23-rc9.
> > * Building the same config on a different system with the same
> >   toolchain, fails the same. So it's not just one system acting weirdly,
> >   the bug can be reproduced.
> > * I tried arbitrary values for the kzalloc() in dmi-id.c, the bottom line
> >   is that anything above 64 bytes triggers the bug.
> > * The same kzalloc() in a different driver doesn't trigger the bug.
> > 
> > I'm puzzled, no idea what to try next.
> 
> Yeah, the tricks we play in there do fool some versions of gcc, and you see
> the result.
> 
> Roland came up with this:
> 
> --- a/include/linux/slub_def.h
> +++ b/include/linux/slub_def.h
> @@ -145,7 +145,7 @@ static __always_inline struct kmem_cache *kmalloc_slab(size_t size)
>  	 * testing it here shouldn't be needed.  But some versions of gcc need
>  	 * help.
>  	 */
> -	if (__builtin_constant_p(size) && index < 0) {
> +	if (__builtin_constant_p(index) && index < 0) {
>  		/*
>  		 * Generate a link failure. Would be great if we could
>  		 * do something to stop the compile here.
> 
> 
> 
> Does it fix things for you?

No, it doesn't, but I hardly see how it could. The above change is in
kmalloc_slab() in slub_def.h while my own problem happens in kmalloc()
in slab_def.h.

-- 
Jean Delvare