On Thu, 15 Feb 2007 06:19:10 +0100, Mahesh Siddheshwar <siddheshwar.mahesh at sun.com> wrote:
> Robert Thurlow wrote: >> Glenn Faden wrote: >> >>>> 4) A bug currently prevents a client instance and a server instance >>>> from being safe to use on the same box (apologies, can't quote the >>>> bugid from here). How likely, in your use case, is it that this will >>>> be a problem, i.e. will your boxes be in the position where a zone >>>> needs data shared from another zone as opposed to a separate server? >>> >>> >>> This is a must fix. In TX we want to automount between labeled zones >>> on the same machine. It seems to work with ZFS. Is the deadlock >>> specific to UFS/NFS? >> >> Good question! I don't expect that it is, but perhaps ZFS's use of the >> ARC would insulate it. Maybe Mahesh would know. > > The problem seen in 5065254 and what is seen commonly > in the recent past is mainly due to the interaction between NFS, UFS and > segmap driver**. This scenario, typically, is noticeable only under > heavy load or on systems with a low segmapsize. Since ZFS does not > use the segmap driver, this particular scenario should be averted. > > Currently the loopback mounted configuration is never tested. So I > won't be surprised if we run into other loopings, but with some > effort those should be tractable. > > Mahesh > > ** NFS tries to commit pages, which, on the NFS server > requires UFS to obtain a segmap slot. Before you use the segmap slot, > you need to free/destroy the previous mappings for the segmap slot, which > happens to be a locked NFS page, the lock for which is currently > owned by the commit thread which begun this process. > > There is one more scenario which I have not seen, but > is theoretically possible -- where writing to a UFS file > requires stealing a dirty NFS page which would in > turn require writing to the server, which requires exclusive > locking of the same UFS file. even if you leave this segmap issue aside, it is very likely you encounter different deadlocks, because this is an attempt to stack file systems that are not really stackable file systems and you'd run into issues similar to 4498652 / 4154394 --- frankB
