On Tue, Mar 08, 2016 at 12:43:15PM +0100, Heiko Carstens wrote:
> On Tue, Mar 08, 2016 at 09:51:05AM +0100, Christian Borntraeger wrote:
> > On 03/08/2016 01:41 AM, Kees Cook wrote:
> >
> > >> --- a/arch/s390/kernel/vmlinux.lds.S
> > >> +++ b/arch/s390/kernel/vmlinux.lds.S
> > >> @@ -52,6 +52,12 @@ SECTIONS
> > >>
> > >> RW_DATA_SECTION(0x100, PAGE_SIZE, THREAD_SIZE)
> > >>
> > >> + . = ALIGN(PAGE_SIZE)
> >
> >
> > missing ";" ?
> >
> >
> > With that and your fixes, this function claims to mark 0kB and
> > lkdtm can still write. Reason is that _edata is 0xc11008 and start is
> > 0x0c11000.
> >
> > making _edata page aligned as well, does now try to mark one page,
Yes, we do need that fixed.
> > but then
> > we run into the next issue, that
> >
> > static void change_page_attr(unsigned long addr, int numpages,
> > pte_t (*set) (pte_t))
> > {
> > pte_t *ptep;
> > int i;
> >
> > for (i = 0; i < numpages; i++) {
> > ptep = walk_page_table(addr);
> >
> > triggers this
> > if (WARN_ON_ONCE(!ptep))
> > break;
> >
> > because the kernel decided to map this with a large page. So we need
> > to fix this function to then break the large page into a smaller chunk....
>
> Yes... however that's a rather large change. I'll try to come up with a
> patch that has less impact and implement the code that splits the kernel
> mapping later.
> Looking at our vmemmap code makes me realize that this code needs also to
> be improved.
Hm, that's going to be too ugly. So we'll go for the "real" solution, which
modifies the kernel mapping instead.
