"Serge E. Hallyn" <serue@xxxxxxxxxx> writes: > Quoting Trond Myklebust (trond.myklebust@xxxxxxxxxx): >> On Tue, 2009-01-06 at 14:02 -0600, Serge E. Hallyn wrote: >> > Quoting Matt Helsley (matthltc@xxxxxxxxxx): >> > > We can often specify the UTS namespace to use when starting an RPC client. >> > > However sometimes no UTS namespace is available (specifically during > system >> > > shutdown as the last NFS mount in a container is unmounted) so fall >> > > back to the initial UTS namespace. >> > >> > So what happens if we take this patch and do nothing else? >> > >> > The only potential problem situation will be rpc requests >> > made on behalf of a container in which the last task has >> > exited, right? So let's say a container did an nfs mount >> > and then exits, causing an nfs umount request. >> > >> > That umount request will now be sent with the wrong nodename. >> > Does that actually cause problems, will the server use the >> > nodename to try and determine the client sending the request? >> >> The NFSv2/v3 umount rpc call will be sent by the 'umount' program from >> userspace, not the kernel. The problem here is that because lazy mounts >> exist, the lifetime of the RPC client may be longer than that of the > > Right that was what i was referring to. > >> container. In addition, it may be shared among more than 1 container, >> because superblocks can be shared. > > Good point. And in that case what do we care about (even though > apparently we just might not care at all :) - who did the mount, > or who is using it? > > In fact one thing I noticed in Matt's patch 3 was that he copied > in the nodename verbatim, so a future hostname() by the container > wouldn't be reflected, again not sure if that would matter. > >> One thing you need to be aware of here is that inode dirty data >> writebacks may be initiated by completely different processes than the >> one that dirtied the inode. > > Right, but I *was* thinking that we wanted to associate the nodename > on the rpc calls with the hostname of the mounter, not the actor. Maybe > you'll tell me above that that is bogus. > >> IOW: Aside from being extremely ugly, approaches like [PATCH 4/4] which >> rely on being able to determine the container-specific node name at RPC >> generation time are therefore going to return incorrect values. > > So should we use patch 2/4, plus (as someone - was it you? - suggested) > using a DEFAULT instead of init_utsname()->nodename when > current->utsname() == NULL? Is there any reason to believe that the kernel helper threads will ever have a useful namespace value? I don't think so. That implies to me you want to capture the value at mount time, and to pass it in to the rpc_call creation, and only at very specific well defined points where we interact with user space should we examine current->utsname(). At which point there should be no question of current->utsname() is valid as the user space process is alive. Eric -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
