On 11/18/20 12:04 PM, Vladimir Sementsov-Ogievskiy wrote:
Hi all!
The problem
Assume we have mirror job with nbd target node with enabled reconnect.
Connection failed. So, all current requests to nbd node are waiting for
nbd driver to reconnect. And they will wait for reconnect-delay time
specified in nbd blockdev options. This timeout may be long enough, for
example, we in Virtuozzo use 300 seconds by default.
So, if at this moment user tries to cancel the job, job will wait for
its in-flight requests to finish up to 300 seconds. From the user point
of view, cancelling the job takes a long time. Bad.
Solution
Let's just cancel "waiting for reconnect in in-flight request coroutines"
on mirror (and backup) cancel. Welcome the series below.
Because of my question on 4/11, I did not queue the entire series yet.
But 6/11 was trivial enough to queue now.