Re: [Devel] [PATCH 2/6] nfsd: swap fs root in NFSd kthreads

["J. Bruce Fields" <bfields@fieldses.org>]

On Tue, Dec 11, 2012 at 07:07:00PM +0400, Stanislav Kinsbursky wrote:
> 11.12.2012 18:56, J. Bruce Fields пишет:
> >On Tue, Dec 11, 2012 at 06:12:40PM +0400, Stanislav Kinsbursky wrote:
> >>UID: 9899
> >>
> >>11.12.2012 18:00, Stanislav Kinsbursky пишет:
> >>>11.12.2012 00:28, J. Bruce Fields пишет:
> >>>>On Thu, Dec 06, 2012 at 06:34:47PM +0300, Stanislav Kinsbursky wrote:
> >>>>>NFSd does lookup. Lookup is done starting from current->fs->root.
> >>>>>NFSd is a kthread, cloned by kthreadd, and thus have global (but luckely
> >>>>>unshared) root.
> >>>>>So we have to swap root to those, which process, started NFSd, has. Because
> >>>>>that process can be in a container with it's own root.
> >>>>
> >>>>This doesn't sound right to me.
> >>>>
> >>>>Which lookups exactly do you see being done relative to
> >>>>current->fs->root ?
> >>>>
> >>>
> >>>Ok, you are right. I was mistaken here.
> >>>This is not a exactly lookup, but d_path() problem in svc_export_request().
> >>>I.e. without root swapping, d_path() will give not local export path (like "/export")
> >>>but something like this "/root/containers_root/export".
> >>>
> >>
> >>We, actually, can do it less in less aggressive way.
> >>I.e. instead root swap and current svc_export_request() implementation:
> >>
> >>void svc_export_request(...)
> >>{
> >>	<snip>
> >>         pth = d_path(&exp->ex_path, *bpp, *blen);
> >>	<snip>
> >>}
> >>
> >>we can do something like this:
> >>
> >>void svc_export_request(...)
> >>{
> >>	struct nfsd_net *nn = ...
> >>	<snip>
> >>	spin_lock(&dcache_lock);
> >>         pth = __d_path(&exp->ex_path, &nn->root, *bpp, *blen);
> >>	spin_unlock(&dcache_lock);
> >>	<snip>
> >>}
> >
> >That looks simpler, but I still don't understand why we need it.
> >
> >I'm confused about how d_path works; I would have thought that
> >filesystem namespaces would have their own vfsmount trees and hence that
> >the (vfsmount, dentry) would be enough to specify the path.  Is the root
> >argument for the case of chroot?  Do we care about that?
> >
> 
> It works very simple: just traverse the tree from specified dentry up to current->fs->root.dentry.
> Having container in some fully separated mount point is great, of course. But:
> 1) this is a limitation we really want to avoid. I.e. container can be chrooted into some path like "/root/containers_root/" as in example above.
> 2) NFSd kthread works in init root environment. But we anyway want to get proper path string in container root, but not in kthreads root.
> 
> >Also, svc_export_request is called from mountd's read of
> >/proc/net/rpc/nfsd.export/channel.  If mountd's root is wrong, then
> >nothing's going to work anyway.
> >
> 
> I don't really understand, how  mountd's root can be wrong. I.e.
> its' always right as I see it. NFSd kthreads have to swap/use
> relative path/whatever to communicate with proper mountd.
> Or I'm missing something?

Ugh, I see the problem: I thought svc_export_request was called at the
time mountd does the read, but instead its done at the time nfsd does
the upcall.

I suspect that's wrong, and we really want this done in the context of
the mountd process when it does the read call.  If d_path is called
there then we have no problem.

--b.