On Wed, 29 Mar 2023, David Gow <davidgow@xxxxxxxxxx> wrote: > The drm buddy allocator tests were broken on 32-bit systems, as > rounddown_pow_of_two() takes a long, and the buddy allocator handles > 64-bit sizes even on 32-bit systems. > > This can be reproduced with the drm_buddy_allocator KUnit tests on i386: > ./tools/testing/kunit/kunit.py run --arch i386 \ > --kunitconfig ./drivers/gpu/drm/tests drm_buddy > > (It results in kernel BUG_ON() when too many blocks are created, due to > the block size being too small.) > > This was independently uncovered (and fixed) by Luís Mendes, whose patch > added a new u64 variant of rounddown_pow_of_two(). This version instead > recalculates the size based on the order. > > Reported-by: Luís Mendes <luis.p.mendes@xxxxxxxxx> > Link: https://lore.kernel.org/lkml/CAEzXK1oghXAB_KpKpm=-CviDQbNaH0qfgYTSSjZgvvyj4U78AA@xxxxxxxxxxxxxx/T/ > Signed-off-by: David Gow <davidgow@xxxxxxxxxx> > --- > drivers/gpu/drm/drm_buddy.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/drivers/gpu/drm/drm_buddy.c b/drivers/gpu/drm/drm_buddy.c > index 3d1f50f481cf..7098f125b54a 100644 > --- a/drivers/gpu/drm/drm_buddy.c > +++ b/drivers/gpu/drm/drm_buddy.c > @@ -146,8 +146,8 @@ int drm_buddy_init(struct drm_buddy *mm, u64 size, u64 chunk_size) > unsigned int order; > u64 root_size; > > - root_size = rounddown_pow_of_two(size); > - order = ilog2(root_size) - ilog2(chunk_size); > + order = ilog2(size) - ilog2(chunk_size); > + root_size = chunk_size << order; Just noticed near the beginning of the function there's also: if (!is_power_of_2(chunk_size)) return -EINVAL; which is also wrong for 32-bit. BR, Jani. > > root = drm_block_alloc(mm, NULL, order, offset); > if (!root) -- Jani Nikula, Intel Open Source Graphics Center
