Hi Eitan, On 09:11 Thu 04 May , Eitan Zahavi wrote: > > So how would the partition manager signal to the LID manager that some > transaction were made and needed to be waited on?
It does not. > Normally a manager reports back the need for waiting in its return code > that is assigned as the next signal to the state machine. > > So if this signal is not provided from the partition manager (which by > your patch returns void) how would the state machine actually wait for > it if the LID manager returns OSM_SIGNAL_DONE ? Then this will wait at the next "wait point" (link_manager iirc). But then if if we will continue there w/out waiting too, and so on... I will need to check, not sure that there is "unconditional wait point" somewhere. Anyway, I see your point now. > Although this can be fixed by forwarding the result of the partition > manager to affect the LID manager next signal calculation Right, or just simple conditional wait for outstanding transactions if any SMPs were sent and not yet responded. Actually then we may group even more independent resweeper components for parallel execution in simpler way. > I would rather > not do it. Keeping the state machine as "linear" as possible have great > merits in avoiding extra complexity and bugs. This may only simplify the state machine. Other obvious advantage of this is performance. Sasha. _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
