Solver manager schedules incorrectly
Created by: subhacom
In certain cases, after automatic solver setup by Cell (i.e., the solver manager), ticks do not point to their intended targets. This leads to segfaults, since the RecvFunc is correct, but target element is something else entirely.
File: TESTS/bugs/solve_sched.g
Reported by: niraj_
Activity
Created by: subhacom
Logged In: YES user_id=1708347 Originator: YES
The problem is more general: the same one that Upi encountered in setting up the KineticManager. Creation and scheduling of the solver elements, and descheduling of solved elements leads to the Tick conn vector and (in this case,) the MsgSrc being changed.
Can't think of any quick hacks. A good idea, from Upi's notes, is to: "- queue reinit operations. I need some kind of queueing operations anyway for other kinds of runtime creation of objects."
Implementation ideas, again from Upi's notes:
- Add a 'inQ' flag to ProcInfo. This indicates if the command is coming through the usual messages or through a queueue.
- Any operation (process or reinit) that needs orderly rescheduling, must trap this flag and push itself onto the queue. Ideally through a return msg, (but there are issues of hardcoding funcs to lessen memory use.)
- The queue is operated after all ticks are done (otherwise too many checks). Just a matter of rerunning the function with a different flag.
Sounds good, but this still leaves the unlikely but finite possibility of the queue iterators themselves getting invalidated.
Original comment by: niraj_
Created by: subhacom
Logged In: YES user_id=1708347 Originator: YES
This issue now does not arise with the new form of ConnTainer iteration. 'Send' will continue to step through a list of message targets correctly, even when the list is modified from any of the message calls. (Confirmed in revision 517).
Original comment by: niraj_
