I got it. Thanks for your explaination.

It is strange that minibuffer-message-timeout is defined in C.
But the timeout is processed in lisp (in the function minibuffer-message).

在 2019年11月8日 +0800 AM6:58，Juri Linkov <juri@linkov.net>，写道：
The prompt is replaced with the message from autorevert. And after
about 2~3 seconds (not fixed), the prompt comes back. What controls
the delay(2~3 seconds)? It is not the value of
minibuffer-message-timeout, which is 0.6. Is this expected?

The problem is that when auto-revert-handler calls minibuffer-message,
the current buffer is not the minibuffer, because functions that call
auto-revert-handler (auto-revert-notify-handler, auto-revert--end-lockout,
or auto-revert-buffers) change the current buffer using with-current-buffer.

When the current buffer is not the minibuffer then minibuffer-message
just calls (message "%s" message) and then does
(sit-for (or minibuffer-message-timeout 1000000))

And it behaves differently with the following test code:

(progn (run-with-idle-timer 3 nil
(lambda ()
(minibuffer-message "Reverting buffer `%s'." (buffer-name))))
(call-interactively 'find-file))

This works well. The prompt is NOT replaced. The message is appended
to the end of the prompt and disappears after 0.6 second.

To work well like this, minibuffer-message should be called outside
of with-current-buffer code block. Yesterday I fixed Man-bgproc-sentinel
in bug#19064 to call minibuffer-message outside of with-current-buffer.

But auto-revert functions require complete rewrite. I don't see
how this could be fixed with a simple change.