This is indeed a regression.
Perhaps we should also revert: 12a0b6622107 ("nvme: don't hold nvmf_transports_rwsem for more than transport lookups")
Which inherently caused this by removing the serialization of .create_ctrl()...
no, I believe the patch on the semaphore is correct. Otherwise - things can be blocked a long time.. a minute (1 cmd timeout) or even multiple minutes in the case where a command failure in core layers effectively gets ignored and thus doesn't cause the error path in the transport. There can be multiple /dev/nvme-fabrics commands stacked up that can make the delays look much longer to the last guy.
as far as creation vs teardown... yeah, not fun, but there are other ways to deal with it. FC: I got rid of the separate create/reconnect threads a while ago thus the return-control-while-reconnecting behavior, so I've had to deal with it. It's one area it'd be nice to see some convergence in implementation again between transports.
Doesn't fc have a bug there? in create_ctrl after flushing the connect_work, what is telling it if delete is running in with it (or that it already ran...)