[FELIX-5887] Cycles in DS depending on bundle order

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: scr-2.0.14, scr-2.1.0
• Component/s: Declarative Services (SCR)
• Environment: MacOS + Windows + Linux
• Assignee: Unassigned
• Reporter: Marko Herchet
• Created: 13/Jul/18 09:47
• Updated: 22/Feb/19 15:37

Description

Following test case:

top.jar: Top @Reference volatile List<Bottom>
bottom.jar: Bottom @Reference Top

In SCR 2.0.2:
top.jar, bottom.jar -> OK, see trace
new Top()
Top.activate()
new Bottom()
Bottom.top=<top>
Bottom.activate()
Top.bottom +=<bottom>

bottom.jar, top.jar -> FAILS but recovers
[osgi.enroute.examples.concurrency.cycle.bottom.Bottom(1)]
Circular reference detected, getService returning null
new Top()
Top.activate()
new Bottom()
Bottom.top=<top>
Bottom.activate()
Top.bottom +=<bottom>

In SCR 2.0.4 and later

top.jar,bottom.jar -> OK

new Top()
Top.activate()
new Bottom()
Bottom.top=<top>
Bottom.activate()
Top.bottom +=<bottom>

bottom.jar, top.jar -> FAILS and DOES NOT RECOVER!
new Bottom()
new Top()
Top.activate()
new Bottom()

The Bottom component never gets added to Top. This is a very serious problem for reliable systems. I see this behaviour on 2.0.4, 2.0.6, and 2.0.8

----------
I've traced the cycle problem in 2.0.2. What happens is:
1) The Bottom configuration is started. There is no Top, so it is not satisfied
2) The Top configuration is started. It is satisfied since it can live with 0 Bottoms.
3) The top configuration registers Top
4) The bottom configuration sees a Top and initiates Bottom
5) The bottom configuration gets Top (which is under construction)
6) The top configuration looks at its dependencies
7) The top configuration sees there is a Bottom now and tries to get it
8) The bottom factory is called again on the same thread and blows up

I am not sure what happens in 2.0.4 and later but it does look scary that Bottom is now never added to the Top.

---- Fix
I tried to do the activation in a background thread that but that failed. I think the cycle should be detected and and de Top should not get its dynamic dependencies after the activate method returned. But, just a guess, this is complicated stuff.

---- Workaround
The workaround is to not let Top register as a service but manually register it at the end of the activate method. Since the service is then already initialized and registered, the Bottom will not get activated until the Top is done.

There is a bnd workspace that consistently shows this output.
https://github.com/pkriens/felix.scr.cycle

I've added 4 log files. 2.0.4 and 2.0.8 seem to act very similar

Issue Links

• is a clone of: FELIX-5780 Cycles in DS depending on bundle order