On Fri, Jul 15, 2022 at 2:09 PM Nathan Chancellor <nathan@xxxxxxxxxx> wrote:
>
> On Fri, Jul 15, 2022 at 01:36:17PM +1000, Dave Airlie wrote:
> > Matthew Auld (1):
> > drm/i915/ttm: fix sg_table construction
>
> This patch breaks i386_defconfig with both GCC and clang:
>
> ld: drivers/gpu/drm/i915/i915_scatterlist.o: in function `i915_rsgt_from_mm_node':
> i915_scatterlist.c:(.text+0x1a7): undefined reference to `__udivdi3'
Yeah, we definitely don't want arbitrary 64x64 divides in the kernel,
and the fact that we don't include libgcc.a once again caught this
horrid issue.
The offending code is
if (sg_alloc_table(st, DIV_ROUND_UP(node->size, segment_pages),
GFP_KERNEL)) {
and I have to say that all of those games with "u64 page_alignment"
that commit aff1e0b09b54 ("drm/i915/ttm: fix sg_table construction")
introduces are absolutely disgusting.
And they are just *pointlessly* disgusting.
Why is that "page_alignment" a "u64"?
And why is it a "size", instead of being a "number of bits"?
The code literally does things like
const u64 max_segment = round_down(UINT_MAX, page_alignment);
which means that
(a) page_alignment must be a power-of-two for this to work
(round_down() only works in powers of two)
(b) the result obviously must fit in an "unsigned int", since it's
rounding down a UINT_MAX!
So (a) makes it doubtful that "page_alignment" should have been a
value (as opposed to mask), and (b) then questions why was that made
an "u64" value when it cannot have a u64 range?
And if max_segments cannot have a 64-bit range, then segment_pages here:
u64 segment_pages = max_segment >> PAGE_SHIFT;
sure cannot.
Fixing those then uncovers other things:
len = min(block_size, max_segment - sg->length);
now complains about mixing types ("max_segment - sg->length" being
u32), because 'block_size' is 64, bit, and that does seem to make some
amount of sense:
block_size = node->size << PAGE_SHIFT;
with the 'node->size' being from drm_mm_node, and that size is a
'u64'. That I *could* see being more than 32 bits on a 64-bit
architecture. Ok.
But then that means that 'len' cannot be a 64-bit value either, and it
should probably have been
u32 len;
..
len = min_t(u64, block_size, max_segment - sg->length);
and that would just have been a lot nicer on 32-bit x86, avoiding a
lot of pointlessly 64-bit things.
That said, even those type simplifications do not fix the fundamental
issue. That "DIV_ROUND_UP()" still ends up being a 64-bit divide,
although now it's at least a "64-by-32" bit divide.
Which needs to be handled by "do_div()" rather than the generic
DIV_ROUND_UP() helper, because sadly, at least gcc still generates a
full __udivdi3() even for the 64-by-32 divides.
Can Intel GPU people please take a look?
Linus
