On Thu, Mar 12, 2020 at 10:07:57AM +0000, David Howells wrote:

David Howells [off-list ref] wrote:

quoted

Is ext4_dio_read_iter() broken?  It calls:

	file_accessed(iocb->ki_filp);

at the end of the function - but surely iocb should be expected to have been
freed when iocb->ki_complete() was called?

The iocb is refcounted and only completed when the refcount hits zero,
and an extra reference is held until the submission has completed.
Take a look at iocb_put().

I think it's actually worse than that.  You also can't call
inode_unlock_shared(inode) because you no longer own a ref on the inode since
->ki_complete() is expected to call fput() on iocb->ki_filp.

the file reference also hold an inode reference.

Yes, you own a shared lock on it, but unless somewhere along the
fput-dput-iput chain the inode lock is taken exclusively, the inode can be
freed whilst you're still holding the lock.

Oh - and ext4_dax_read_iter() is also similarly broken.

In addition to that DAX never executes asynchronously.

And xfs_file_dio_aio_read() appears to be broken as it touches the inode after
calling iomap_dio_rw() to unlock it.

Same as above.