Revert "nvmet-tcp: Don't free SQ on authentication success"

In an attempt to fix REPLACETLSPSK we stopped freeing the secrets on
successful connections. This resulted in memory leaks in the kernel, so
let's revert the commit. A improved fix is being developed to just avoid
clearing the tls_key variable.

This reverts commit 2e6eb6b.

Signed-off-by: Alistair Francis <[email protected]>

Author: alistair23

drivers/nvme/target/fabrics-cmd-auth.c

@@ -395,10 +395,9 @@ void nvmet_execute_auth_send(struct nvmet_req *req)
 		goto complete;
 	}
 	/* Final states, clear up variables */
-	if (req->sq->dhchap_step == NVME_AUTH_DHCHAP_MESSAGE_FAILURE2) {
-		nvmet_auth_sq_free(req->sq);
+	nvmet_auth_sq_free(req->sq);
+	if (req->sq->dhchap_step == NVME_AUTH_DHCHAP_MESSAGE_FAILURE2)
 		nvmet_ctrl_fatal_error(ctrl);
-	}
 
 complete:
 	nvmet_req_complete(req, status);
@@ -574,7 +573,9 @@ void nvmet_execute_auth_receive(struct nvmet_req *req)
 	status = nvmet_copy_to_sgl(req, 0, d, al);
 	kfree(d);
 done:
-	if (req->sq->dhchap_step == NVME_AUTH_DHCHAP_MESSAGE_FAILURE1) {
+	if (req->sq->dhchap_step == NVME_AUTH_DHCHAP_MESSAGE_SUCCESS2)
+		nvmet_auth_sq_free(req->sq);
+	else if (req->sq->dhchap_step == NVME_AUTH_DHCHAP_MESSAGE_FAILURE1) {
 		nvmet_auth_sq_free(req->sq);
 		nvmet_ctrl_fatal_error(ctrl);
 	}