On Wed, Nov 18, 2015 at 05:33:43PM +0300, Andrey Ryabinin wrote:

On 11/16/2015 07:51 PM, Catalin Marinas wrote:

quoted

On Mon, Nov 16, 2015 at 06:34:27PM +0300, Andrey Ryabinin wrote:

quoted

On 11/16/2015 02:16 PM, Suzuki K. Poulose wrote:

quoted

arch/arm64/mm/kasan_init.c:95:2: note: in expansion of macro ?BUILD_BUG_ON?
  BUILD_BUG_ON(!IS_ALIGNED(KASAN_SHADOW_END, PGDIR_SIZE));

The problem is that the PGDIR_SIZE is (1UL << 47) with 16K+48bit, which makes
the KASAN_SHADOW_END unaligned(which is aligned to (1UL << (48 - 3)) ). Is the
alignment really needed ? Thoughts on how best we could fix this ?

Yes, it's really needed, because some code relies on this (e.g.
clear_pgs() and kasan_init()). But it should be possible to get rid of
this requirement.

I don't think clear_pgds() and kasan_init() are the only problems. IIUC,
kasan_populate_zero_shadow() also assumes that KASan shadow covers
multiple pgds. You need some kind of recursive writing which avoids
populating an entry which is not empty (like kasan_early_pud_populate).

I think kasan_populate_zero_shadow() should be fine. We call pgd_populate() only
if address range covers the entire pgd:

		if (IS_ALIGNED(addr, PGDIR_SIZE) && end - addr >= PGDIR_SIZE) {
....
			pgd_populate(&init_mm, pgd, kasan_zero_pud);
....

and otherwise we check for pgd_none(*pgd):
		if (pgd_none(*pgd)) {
			pgd_populate(&init_mm, pgd,
				early_alloc(PAGE_SIZE, NUMA_NO_NODE));
		}

OK, I missed the fact that zero_pud_populate() handles the pmd/pte
population with kasan_zero_*.

So if it's only tmp_pg_dir, as you said already, you can add a tmp_pud
for the case where KASAN_SHADOW_SIZE is smaller than PGDIR_SIZE and
change clear_pgds() to erase the puds.

-- 
Catalin