The core mistake was that the webconn doesn't really go out of scope
once the connection disconnects. It is kept in the webhub connIndex
to be reconnected if the user connects again. This was the new
behavior as part of reliable websockets.
Therefore, it was a mistake to return the session to the pool
once the connection drops. Because the connection would still
recieve events from the web_hub.
And once you release the session, another login might acquire the
session and set some props, while the web_hub might still try
to send events to it, which will cause a read of the map prop.
The following test case illustrates such a race. It is very
hard to trigger it organically, hence I artificially wrote
the code.
The right fix is to release the session only when the connection
is stale and gets deleted from the conn index. The PR has been
load tested in `-race` mode just for extra sanity check.
```go
func TestHubSessionRace(t *testing.T) {
th := Setup(t).InitBasic()
defer th.TearDown()
s := httptest.NewServer(dummyWebsocketHandler(t))
defer s.Close()
th.Server.HubStart()
wc1 := registerDummyWebConn(t, th.App, s.Listener.Addr(), th.BasicUser.Id)
defer wc1.Close()
var wg sync.WaitGroup
wg.Add(2)
go func() {
defer wg.Done()
token := wc1.GetSessionToken()
// Return to pool after *WebConn.Pump finishes
wc1.App.Srv().userService.ReturnSessionToPool(wc1.GetSession())
// A new HTTP requests acquires a session which gets it from the pool
sess, _ := wc1.App.GetSession(token)
// Login happens which sets some session properties
sess.AddProp(model.SessionPropPlatform, "chrome")
}()
go func() {
defer wg.Done()
// Called from *WebConn.shouldSendEvent
t.Log("session: ", wc1.GetSession().Props[model.SessionPropIsGuest] == "true")
}()
wg.Wait()
}
```
https://mattermost.atlassian.net/browse/MM-46604
```release-note
NONE
```
c5d7ac0876