On 2021-12-14 14:44, Greg KH wrote:
On Tue, Dec 14, 2021 at 01:35:40PM +0100, Jens Wiklander wrote:
Since the tee subsystem does not keep a strong reference to its idle shared memory buffers, it races with other threads that try to destroy a shared memory through a close of its dma-buf fd or by unmapping the memory.
In tee_shm_get_from_id() when a lookup in teedev->idr has been successful, it is possible that the tee_shm is in the dma-buf teardown path, but that path is blocked by the teedev mutex. Since we don't have an API to tell if the tee_shm is in the dma-buf teardown path or not we must find another way of detecting this condition.
Fix this by doing the reference counting directly on the tee_shm using a new refcount_t refcount field. dma-buf is replaced by using anon_inode_getfd() instead, this separates the life-cycle of the underlying file from the tee_shm. tee_shm_put() is updated to hold the mutex when decreasing the refcount to 0 and then remove the tee_shm from teedev->idr before releasing the mutex. This means that the tee_shm can never be found unless it has a refcount larger than 0.
So you are dropping dma-buf support entirely? And anon_inode_getfd() works instead? Why do more people not do this as well?
Indeed, thinking about it, does it really makes sense to do mmap() on an anon_inode_getfd() fd ? It is a singleton inode used there so don't we breach some contract with the linux mm ? The dma-buf code for creating the file object is more complex, it creates a unique inode for each object.
I am by no means claiming to understand inodes' interaction with mmap, just sharing a concern that popped up in my head.
- Lars