On Sat, Jul 16, 2022 at 2:35 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> 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.
Hmm. The "DIV_ROUND_UP()" issue could be solved by just making the
rule be that the max_segment size is always a power of two.
Then you don't need the (expensive!) DIV_ROUND_UP(), and can just use
the regular "round_up()" that works on powers-of-two.
And the simplest way to do that is to just make "max_segments" be 2GB.
The whole "round_down(UINT_MAX, page_alignment)" seems entirely
pointless. Do you really want segments that are some odd number just
under the 4GB mark, and force expensive divides?
For consistency, I used the same value in
i915_rsgt_from_buddy_resource(). I have no idea if that makes sense.
Anyway, the attached patch is COMPLETELY UNTESTED. But it at least
seems to compile. Maybe.
Linus
drivers/gpu/drm/i915/i915_scatterlist.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_scatterlist.c b/drivers/gpu/drm/i915/i915_scatterlist.c
index f63b50b71e10..830dcd833d4e 100644
--- a/drivers/gpu/drm/i915/i915_scatterlist.c
+++ b/drivers/gpu/drm/i915/i915_scatterlist.c
@@ -81,7 +81,9 @@ struct i915_refct_sgt *i915_rsgt_from_mm_node(const struct drm_mm_node *node,
u64 region_start,
u64 page_alignment)
{
- const u64 max_segment = round_down(UINT_MAX, page_alignment);
+ // Make sure max_segment (and thus segment_pages) is
+ // a power of two that fits in 32 bits.
+ const u64 max_segment = 1ul << 31;
u64 segment_pages = max_segment >> PAGE_SHIFT;
u64 block_size, offset, prev_end;
struct i915_refct_sgt *rsgt;
@@ -96,7 +98,7 @@ struct i915_refct_sgt *i915_rsgt_from_mm_node(const struct drm_mm_node *node,
i915_refct_sgt_init(rsgt, node->size << PAGE_SHIFT);
st = &rsgt->table;
- if (sg_alloc_table(st, DIV_ROUND_UP(node->size, segment_pages),
+ if (sg_alloc_table(st, round_up(node->size, segment_pages),
GFP_KERNEL)) {
i915_refct_sgt_put(rsgt);
return ERR_PTR(-ENOMEM);
@@ -159,7 +161,7 @@ struct i915_refct_sgt *i915_rsgt_from_buddy_resource(struct ttm_resource *res,
{
struct i915_ttm_buddy_resource *bman_res = to_ttm_buddy_resource(res);
const u64 size = res->num_pages << PAGE_SHIFT;
- const u64 max_segment = round_down(UINT_MAX, page_alignment);
+ const u64 max_segment = 1u << 31;
struct drm_buddy *mm = bman_res->mm;
struct list_head *blocks = &bman_res->blocks;
struct drm_buddy_block *block;
