Re: [PATCH v3] binder: add mutex_lock for mmap and NULL when free

From: Carlos Llamas
Date: Sat Oct 07 2023 - 19:40:02 EST


On Sat, Oct 07, 2023 at 08:28:13PM +0800, Kassey Li wrote:
> - Enforce alloc->mutex in binder_alloc_mmap_handler when add the entry to
> list.
>
> - Assign the freed pages/page_ptr to NULL to catch possible use after free
> with NULL pointer access.
>
> Fixes: 19c987241ca1 ("binder: separate out binder_alloc functions")
> Fixes: f2517eb76f1f ("android: binder: Add global lru shrinker to binder")
> CC: Todd Kjos <tkjos@xxxxxxxxxxx>
> CC: Sherry Yang <sherryy@xxxxxxxxxxx>
> Link: https://lore.kernel.org/all/20231007034046.2352124-1-quic_yingangl@xxxxxxxxxxx/
> Signed-off-by: Kassey Li <quic_yingangl@xxxxxxxxxxx>
> ---
> V1 -> V2: Add Fixes info.
> V2 -> V3: Add this history.
> ---
> drivers/android/binder_alloc.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/android/binder_alloc.c b/drivers/android/binder_alloc.c
> index e3db8297095a..c7d126e04343 100644
> --- a/drivers/android/binder_alloc.c
> +++ b/drivers/android/binder_alloc.c
> @@ -775,6 +775,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
> }
>
> buffer->user_data = alloc->buffer;
> + mutex_lock(&alloc->mutex);

At this stage we are already holding the mm->mmap_lock. If you take the
alloc->mutex here you will deadlock. You should see this warning with
lockdep enabled. Also, you don't need the lock here...

> list_add(&buffer->entry, &alloc->buffers);
> buffer->free = 1;
> binder_insert_free_buffer(alloc, buffer);
> @@ -782,7 +783,7 @@ int binder_alloc_mmap_handler(struct binder_alloc *alloc,
>
> /* Signal binder_alloc is fully initialized */
> binder_alloc_set_vma(alloc, vma);

This barrier will set the alloc->vma pointer. Once set, it signals that
the alloc-> space has been initialized and it is safe to access.

> -
> + mutex_unlock(&alloc->mutex);
> return 0;
>
> err_alloc_buf_struct_failed:
> @@ -856,9 +857,11 @@ void binder_alloc_deferred_release(struct binder_alloc *alloc)
> __func__, alloc->pid, i, page_addr,
> on_lru ? "on lru" : "active");
> __free_page(alloc->pages[i].page_ptr);
> + alloc->pages[i].page_ptr = NULL;
> page_count++;
> }
> kfree(alloc->pages);
> + alloc->pages = NULL;

The process is dying and there aren't any more references to it. It is
pointless to mark the pages NULL. No one is or will use them after.

> }
> mutex_unlock(&alloc->mutex);
> if (alloc->mm)
> --
> 2.25.1
>

I see that you reported a crash on the previous thread here:
https://lore.kernel.org/all/26988068-8c9f-8591-db6e-44c8105af638@xxxxxxxxxxx/
...unfortunately, it doesn't seem to me like setting alloc->pages = NULL
would help fix your issue.

I do agree this sounds like a use-after-free though. Are you able to
reproduce the issue with KASAN enabled? This should tell us where the
actual problem is.

--
Carlos Llamas