The buffer should only be released by its own complete callback, which only gets called after being successfully queued. Moreover, other uses of `mlx5_fpga_conn_send`, and the related `mlx5_fpga_conn_post_recv` will free after error.

The other part of the patch, that checks for `flow` being NULL may be unnecessary since it looks like the handle is always from an active context. But that's a guess. And it's only unreachable code.

The opinion I have from this is despite other patches being bad ideas, this one doesn't look like it. Because the other patches didn't make it past the mailing list, it demonstrates that the maintainers are doing a good enough job.

You’re right, that wasn’t one of the bad patches: https://lore.kernel.org/lkml/CAK8KejpUVLxmqp026JY7x5GzHU2YJL...