Thank you Tommaso. Please see my comments inline.
I understand it does not make much sense. Nevertheless, a broken link is exactly what I want to simulate (e.g., to evaluate the TCP performance when a radio link failure occurs for a limited period of time).
Yes, when the link failure occurs nothing should change (i.e., at the IP level) and the application should continue generating packets that, however, should not be physically sent through the link and, consequently, should be buffered. Therefore, I expect that, as long as a TCP connection is concerned, the RTO will expire and, as soon as the full link rate will be recovered, the packets will start flowing again through the link in slow start.
Now the question is: how to simulate the link breakout?