On 28/09/2026 19:44, Bill Wendling wrote:
On Mon, Sep 28, 2026 at 5:20 AM Matthew Auld <[email protected]> wrote:
On 25/09/2026 21:44, Bill Wendling wrote:
In 'struct gpu_buddy', the 'roots' field points to an array of pointers
to 'struct gpu_buddy_block'. The number of allocated roots is tracked
by the 'n_roots' field.
Annotate the 'roots' pointer with the '__counted_by_ptr' attribute
referencing 'n_roots' to enable compile-time and runtime bounds-checking
via KASAN and '__builtin_dynamic_object_size'.
In 'gpu_buddy_init', 'mm->n_roots' is initialized first, and 'mm->roots' is
subsequently allocated with 'kmalloc_objs(struct gpu_buddy_block *,
mm->n_roots)'.
Since the bounds associated with 'roots' ('n_roots') are fully set prior
to any array allocation or access and remain invariant, this annotation
will not cause runtime panics or false-positive bounds checks.
Cc: [email protected]
Assisted-by: LLM
Signed-off-by: Bill Wendling <[email protected]>
Reviewed-by: Matthew Auld <[email protected]>
Would it make sense to also do this for *_scoreboard and free_tree, if
we add a new field n_orders?
Sure! There's already a 'max_order'. I assume that 'n_orders' would be
more dynamic?
Yeah, I think just mm.n_orders = max_order + 1. I assume the
__counted_by_ptr() only works with a field?
-bw
---
include/linux/gpu_buddy.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/gpu_buddy.h b/include/linux/gpu_buddy.h
index 2c36124bb696..d7249e500ab6 100644
--- a/include/linux/gpu_buddy.h
+++ b/include/linux/gpu_buddy.h
@@ -172,7 +172,7 @@ struct gpu_buddy {
* a power of two, with each root being the largest power-of-two
* that fits in the remaining space.
*/
- struct gpu_buddy_block **roots;
+ struct gpu_buddy_block **roots __counted_by_ptr(n_roots);
/*
* Per-order free block scoreboard: free_scoreboard[order] holds the
* number of blocks of that order currently in the free state.
