Re: [PATCH] accel/qaic: Fix dereferencing freed memory

From: Jeffrey Hugo
Date: Mon Jun 12 2023 - 11:10:25 EST


On 6/12/2023 7:21 AM, Christian König wrote:
Am 12.06.23 um 15:03 schrieb Pranjal Ramajor Asha Kanojiya:


On 6/12/2023 4:52 PM, Christian König wrote:


Am 10.06.23 um 04:12 schrieb Sukrut Bellary:
smatch warning:
    drivers/accel/qaic/qaic_data.c:620 qaic_free_object() error:
        dereferencing freed memory 'obj->import_attach'

obj->import_attach is detached and freed using dma_buf_detach().
But used after free to decrease the dmabuf ref count using
dma_buf_put().

Fixes: ff13be830333 ("accel/qaic: Add datapath")
Signed-off-by: Sukrut Bellary <sukrut.bellary@xxxxxxxxx>
---
  drivers/accel/qaic/qaic_data.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/accel/qaic/qaic_data.c b/drivers/accel/qaic/qaic_data.c
index e42c1f9ffff8..7cba4d680ea8 100644
--- a/drivers/accel/qaic/qaic_data.c
+++ b/drivers/accel/qaic/qaic_data.c
@@ -613,11 +613,13 @@ static int qaic_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struc
  static void qaic_free_object(struct drm_gem_object *obj)
  {
      struct qaic_bo *bo = to_qaic_bo(obj);
+    struct dma_buf *dmabuf;

Maybe move that variable into the if.

      if (obj->import_attach) {
          /* DMABUF/PRIME Path */
+        dmabuf = obj->import_attach->dmabuf;
          dma_buf_detach(obj->import_attach->dmabuf, obj->import_attach);
-        dma_buf_put(obj->import_attach->dmabuf);
+        dma_buf_put(dmabuf);

I strongly assume you are not using the GEM prime helpers for this?

Christian.

Driver uses drm_gem_prime_fd_to_handle() helper function but it also registers for ->gem_prime_import() which is internally called by drm_gem_prime_fd_to_handle(). All the operations done in gem_prime_import() are undone here.

Then why don't you use drm_prime_gem_destroy() which is the cleanup helper for imports created by ->gem_prime_import() ?

That looks pretty much identical to what you do here manually.

I think destroy() wasn't used because we are new to DRM and sometimes confused by the multitude of options. I appreciate the suggestion and will follow up to see if destroy() will work here, or identify what would prevent the use of it.

-Jeff