Bluetooth: hci_core: lookup hci_conn on RX path on protocol side

Bluetooth: hci_core: lookup hci_conn on RX path on protocol side

The hdev lock/lookup/unlock/use pattern in the packet RX path doesn't
ensure hci_conn* is not concurrently modified/deleted. This locking
appears to be leftover from before conn_hash started using RCU
commit bf4c63252490b ("Bluetooth: convert conn hash to RCU")
and not clear if it had purpose since then.

Currently, there are code paths that delete hci_conn* from elsewhere
than the ordered hdev->workqueue where the RX work runs in. E.g.
commit 5af1f84ed13a ("Bluetooth: hci_sync: Fix UAF on hci_abort_conn_sync")
introduced some of these, and there probably were a few others before
it.  It's better to do the locking so that even if these run
concurrently no UAF is possible.

Move the lookup of hci_conn and associated socket-specific conn to
protocol recv handlers, and do them within a single critical section
to cover hci_conn* usage and lookup.

syzkaller has reported a crash that appears to be this issue:

    [Task hdev->workqueue]          [Task 2]
                                    hci_disconnect_all_sync
    l2cap_recv_acldata(hcon)
                                      hci_conn_get(hcon)
                                      hci_abort_conn_sync(hcon)
                                        hci_dev_lock
      hci_dev_lock
                                        hci_conn_del(hcon)
      v-------------------------------- hci_dev_unlock
                                      hci_conn_put(hcon)
      conn = hcon->l2cap_data (UAF)

Fixes: 5af1f84ed13a ("Bluetooth: hci_sync: Fix UAF on hci_abort_conn_sync")
Reported-by: syzbot+d32d77220b92eddd89ad@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=d32d77220b92eddd89ad
Signed-off-by: Pauli Virtanen <pav@iki.fi>
Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>