Fix some small problems in the live upgrade procedure #385
Conversation
We observe the following deadlock when live upgrade:
- sc.upgradeDaemons
- sc.manager.Lock()
- manager.upgradeNydusDaemon
- m.mu.Lock()
Let's fix it.
Signed-off-by: Huang Jianan <jnhuang@linux.alibaba.com>
By this we can use the new daemon to handle the upcomming mounts. Signed-off-by: Huang Jianan <jnhuang@linux.alibaba.com>
Send states to the new daemon to restore. Signed-off-by: Huang Jianan <jnhuang@linux.alibaba.com>
So we can get the correct state after umount. Signed-off-by: Huang Jianan <jnhuang@linux.alibaba.com>
Codecov ReportBase: 27.99% // Head: 28.03% // Increases project coverage by
Additional details and impacted files@@ Coverage Diff @@
## main #385 +/- ##
==========================================
+ Coverage 27.99% 28.03% +0.03%
==========================================
Files 40 40
Lines 4086 4081 -5
==========================================
Hits 1144 1144
+ Misses 2803 2798 -5
Partials 139 139
Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. ☔ View full report at Codecov. |
| @@ -316,6 +318,11 @@ func (sc *Controller) upgradeNydusDaemon(d *daemon.Daemon, c upgradeRequest) err | |||
| return err | |||
| } | |||
|
|
|||
| su := manager.SupervisorSet.GetSupervisor(d.ID()) | |||
| if err := su.SendStatesTimeout(time.Second * 10); err != nil { | |||
There was a problem hiding this comment.
Why do we have to receive the old nydusd's states before a takeover request to new nydusd ?
There was a problem hiding this comment.
It's a similar procedure to doDaemonFailover function. The snapshotter needs to send the old nydusd's state to new one when takeover.
Here send refers to from snapshotter to nydus, as opposed to sendfd process, which may be a little confusing.
This is to prepare for the upcoming nydusd live upgrade function.