Hello Osama,

On Thu, 12 Nov 2026 20:16:50 +0100
Osama Abdelkader <[email protected]> wrote:

> panthor_heap_grow() allocates a new chunk with panthor_alloc_heap_chunk(),
> which adds the chunk at the head of heap->chunks under heap->lock, and
> then drops the lock. panthor_heap_grow() then assumes the head of the
> list is the chunk it just added, and reads its GPU VA through
> list_first_entry() without holding heap->lock.
> 
> panthor_heap_grow() and panthor_heap_return_chunk() only take pool->lock
> in read mode, and tiler OOM events of different groups are processed in
> parallel on the unbound heap_alloc_wq. If two groups use the same heap
> context, they can modify the chunk list concurrently:
> 
>   group A OOM work                     group B OOM work
>   alloc C_A, list_add()
>                                        alloc C_B, list_add()
>   list_first_entry() -> C_B
>                                        group scheduled out:
>                                        panthor_heap_return_chunk(C_B)
>                                          list_del(), destroy BO, kfree()
>   C_B->bo                              <- use-after-free
> 
> Even without the free, both groups end up being handed the same chunk,
> and the chunk allocated by group A is never linked to the heap context.
> 
> Fix this by having panthor_alloc_heap_chunk() return the GPU VA of the
> chunk it allocated, retrieved before the chunk is published in the list,
> instead of looking it up afterwards.
> 
> Fixes: 9cca48fa4f89 ("drm/panthor: Add the heap logical block")
> Cc: [email protected]
> Assisted-by: Claude:claude-opus-5.5
> Signed-off-by: Osama Abdelkader <[email protected]>
> ---
>  drivers/gpu/drm/panthor/panthor_heap.c | 23 +++++++++++++++--------
>  1 file changed, 15 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/panthor/panthor_heap.c 
> b/drivers/gpu/drm/panthor/panthor_heap.c
> index 99311abdf1e9..d80f514366fd 100644
> --- a/drivers/gpu/drm/panthor/panthor_heap.c
> +++ b/drivers/gpu/drm/panthor/panthor_heap.c
> @@ -139,7 +139,8 @@ static void panthor_free_heap_chunk(struct 
> panthor_heap_pool *pool,
>  
>  static int panthor_alloc_heap_chunk(struct panthor_heap_pool *pool,
>                                   struct panthor_heap *heap,
> -                                 bool initial_chunk)
> +                                 bool initial_chunk,
> +                                 u64 *chunk_gpu_va)
>  {
>       struct panthor_heap_chunk *chunk;
>       struct panthor_heap_chunk_header *hdr;
> @@ -181,6 +182,15 @@ static int panthor_alloc_heap_chunk(struct 
> panthor_heap_pool *pool,
>  
>       panthor_kernel_bo_vunmap(chunk->bo);
>  
> +     /*
> +      * Retrieve the GPU VA before the chunk becomes visible in the heap
> +      * chunk list: once heap->lock is released, a concurrent grow or
> +      * return_chunk on the same heap can modify the list, and the chunk
> +      * can be freed.
> +      */
> +     if (chunk_gpu_va)
> +             *chunk_gpu_va = panthor_kernel_bo_gpuva(chunk->bo);

Let's make this unconditional and force the user to pass a non-NULL
chunk_gpu_va.

> +
>       mutex_lock(&heap->lock);
>       list_add(&chunk->node, &heap->chunks);
>       heap->chunk_count++;
> @@ -216,7 +226,7 @@ static int panthor_alloc_heap_chunks(struct 
> panthor_heap_pool *pool,

Let's have panthor_alloc_heap_chunks() retrieve the first chunk GPU VA
to simplify panthor_heap_create(). The first_chunk_gpu_va pointer can
be passed directly to panthor_alloc_heap_chunk().

The rest looks good to me.

Thanks,

Boris

Reply via email to