ARM/kirkwood: v3.12-rc6: kernel BUG at mm/util.c:390!

[catalin.marinas@arm.com (Catalin Marinas)]

On Sun, Oct 27, 2013 at 02:18:17PM +0000, Simon Baatz wrote:
> On Sun, Oct 27, 2013 at 01:53:44PM +0000, Russell King - ARM Linux wrote:
> > On Sun, Oct 27, 2013 at 09:16:53PM +0800, Ming Lei wrote:
> > > On Sun, Oct 27, 2013 at 8:50 PM, Aaro Koskinen <aaro.koskinen@iki.fi> wrote:
> > > >
> > So...
> > 
> > flush_kernel_dcache_page() is expected to take a struct page pointer.
> > This struct page pointer is part of the kernel's array of struct pages
> > which identifies every single physical page under the control of the
> > kernel.
> > 
> > Arguably, it should not crash if passed a page which has been allocated
> > to the slab cache; as this is not a page cache page,
> > flush_kernel_dcache_page() should merely ignore the call to it and
> > simply return on these.  So this makes total sense:
> 
> In this respect, flush_kernel_dcache_page() is following
> flush_dcache_page(). For example in crypto/scatterwalk.c:
> 
> static void scatterwalk_pagedone(struct scatter_walk *walk, int out,
>                                  unsigned int more)
> {
>         if (out) {
>                 struct page *page;
> 
>                 page = sg_page(walk->sg) + ((walk->offset - 1) >>
> PAGE_SHIFT);
>                 if (!PageSlab(page))
>                         flush_dcache_page(page);
>         }
> ... 
> 
> 
> or in drivers/ata/libata-sff.c:
> 
> ...
>         if (!do_write && !PageSlab(page))
>                 flush_dcache_page(page);
> ...
> 
> 
> (Probably, both cases should have used
> flush_kernel_dcache_page() in the first place). If we say that this
> check belongs in flush_kernel_dcache_page() we should also put it
> into flush_dcache_page(), no?

According to cachetlb.txt, flush_dcache_page() is only called on page
cache pages, so this excludes the PageSlab() check.

For flush_kernel_dcache_page() it says "when the kernel modifies and
user page" and my reading is that this applies to either page cache
or anonymous pages but not slab pages, so I would add such check to the
caller.

-- 
Catalin