Hi Ajit, > But even in the case VxFS cfs, the other node is not allowed to > mount if the first node has not mounted the filesystem with share > mount option, so I was just wondering if ufs has any such mechanism.
Not that I was aware of, no. Again, I agree that such a mechanism could be helpful to avoid errors in corner cases, but it would add more complexity, so I think there should a strong reason before an implementation of such a mechanism was considered. >> If the application needs "raw" device access, then, yes, it should >> cooperate. If it doesn't or when in doubt, you can use the OHAC/Sun Cluster >> framework to guarantee that the application will only ever run on one node >> at a time (for instance using Failover_mode HARD as a last resort, see >> r_properties(5)). > > as far as I could understand about the different Failover_mode > options, it tells what is behavior if any of the methods (prestart, > start, stop or poststop) fail. It does say any thing about the device > access in fail-over cluster. Correct, what I was referring to is that the cluster can guarantee that the application will only ever run once by making sure that it gets stopped somehow: Either by a successful stop routine (and PMF reporting all processes to have exited) or, as a last resort, by panicking the node. So this mechanism should implicitly solve issues with potential concurrent access. > Looks like its a expected behavior that offline devices are accessible > same as online devices in even fail-over cluster. Regarding design questions, I would rather refer you to someone from the core team, but my understanding is that implementation of global, uniform access to devices (from all nodes) as been a very early design decision in Sun Cluster 3 history and has been promoted as an important feature: http://docs.sun.com/app/docs/doc/820-2554/cacheafd?l=en&a=view And, yes, the behavior you are seeing is exactly what is to be expected in Sun Cluster. Maybe your question is related to the fact that OHAC/Sun Cluster does not have a notion of a Scalable or HA Cluster, but those notions apply to resource groups. Nils
