Re: [BUG][Resend] 2.6.12-rc3-mm3: Kernel BUG at "mm/slab.c":1219 [update]

[Andrew Morton <akpm@osdl.org>]

"Rafael J. Wysocki" <rjw@sisk.pl> wrote:
>
> > Clearly init_bio() is not passing in a NULL `name' parameter.  Maybe the
>  > backtrace is screwed due to dopey gcc autoinlining and the bad caller is
>  > really biovec_init_slabs().  Try removing the
>  > __cacheline_aligned_mostly_readonly from the declaration of bvec_slabs[].
> 
>  Heh, it boots without the __cacheline_aligned_mostly_readonly (ie the BUG is
>  only triggered if the __cacheline_aligned_mostly_readonly is present in the
>  declaration of bvec_slabs[]).  I've double-checked it.  Interesting ... ;-)

oops.

+#ifdef CONFIG_X86
+#define __cacheline_aligned_mostly_readonly		\
+  __attribute__((__aligned__(SMP_CACHE_BYTES),		\
+		 __section__(".data.mostly_readonly")))
+#else
+#define __cacheline_aligned_mostly_readonly __cacheline_aligned
+#endif

So on x86_64 we're putting __cacheline_aligned_mostly_readonly stuff into a
section which is not implemented anywhere.  I'm rather surprised that the
kernel linked at all.  But I'm more surprised that it mysteriously oopsed.

Oh well, I'll change that to CONFIG_X86 && !CONFIG_X86_64, thanks.

Or maybe Christoph wants to rustle up the x86_64 patch?  We really should
patch all architectures if we're going to do this thing.