Re: [PATCH] VFS autmounter support v2

[David Howells]

> > Then have a daemon that can take a request to mount and then reply with
> > the mount parameters, allowing the trap fs to obtain a vfsmount via
> > do_kern_mount(). I would make the trap fs supply the daemon with an fd
> > attached to the trap rootdir to act as a token representing the request
> > (and controlling its lifetime).
> > 
> 
> You would have to go this route.  I think Al's opinion in this is that 
> your original proposal allows arbitrary dentry's in the tree to act as 
> traps.  As such, there is no way for a derived namespace to manipulate 
> that trap at all.

> By implying that the trap is installed via mount says you are now proposing
> that every trap is represented by its own superblock.

This is more or less what Al suggested, except that he suggested "traps" are
special vfsmounts that don't have superblocks, dentries, and inodes.
 
> You're new proposal is exactly what I have been working on, autofs 
> direct mountpoints using the less intrusive follow_link magic Anvin has 
> mentioned on a previous thread both here and on autofs@vger.
> 
> The one problem with this solution is the following breaks:
> 
> # installtrap /foo host:/export/foo
> <userspace daemon listens for requests to mount>
> # newnssh
> newnssh # cd /foo
> 
> Oops, the daemon started from the initial namespace doesn't have access 
> to the namespace in my second shell.

Yes. You can see that happening now with autofs and amd. However it does work
with my suggestion because the "automounter" code just returns a namespace
independent vfsmount, which the VFS can then bind into the appropriate
namespace and the appropriate place.

> The most reasonable way I can see to cope with this is to allow 
> CAP_SYS_ADMIN processes the ability to change namespaces.  Eg, the 
> daemon can be told which pid triggered the trap on /foo, 
> open(/proc/<pid>/mounts) and perform a ioctl(IOC_USENAMESPACE) on it.
> 
> What do you guys think?

I think a better way is for the kernel to pass the daemon a file descriptor
attached to the mount point. This would then act as a token representing the
request, and as such it automatically includes the mount point info (struct
file has vfsmount/dentry pointers) and can also be used to manage the lifetime
of the request.

I'd then make there be either a "mount" ioctl/fcntl on that fd that uses the
info stored on the struct file, or perhaps provide an "fmount" syscall (but
you have to be careful - otherwise someone can use an arbitrary
cross-namespace fd to mangle some other namespace).

David