mount / unmount and mountcheckdirs()

[Konstantin Belousov]

On Fri, Sep 15, 2017 at 05:06:10PM +0100, Edward Napierala wrote:
> 2017-09-15 12:45 GMT+01:00 Konstantin Belousov <kostikbel at gmail.com>:
> 
> > On Fri, Sep 15, 2017 at 12:31:05PM +0100, Edward Napierala wrote:
> > > 2017-09-15 12:00 GMT+01:00 Konstantin Belousov <kostikbel at gmail.com>:
> > >
> > > > On Fri, Sep 15, 2017 at 11:49:17AM +0100, Edward Napierala wrote:
> > > > > 2017-09-15 11:30 GMT+01:00 Konstantin Belousov <kostikbel at gmail.com
> > >:
> > > > >
> > > > > > On Fri, Sep 15, 2017 at 11:08:31AM +0100, Edward Napierala wrote:
> > > > > > > 2017-09-15 10:20 GMT+01:00 Konstantin Belousov <
> > kostikbel at gmail.com
> > > > >:
> > > > > > > > I believe that the current autofs does not allow a process to
> > get
> > > > into
> > > > > > > > this situation at all.
> > > > > > > >
> > > > > > >
> > > > > > > It does.  For example:
> > > > > > >
> > > > > > > [trasz at v2:~]% cd /media/md0
> > > > > > > [trasz at v2:/media/md0]% mount
> > > > > > > /dev/ada0s1a on / (ufs, local, noatime, journaled soft-updates)
> > > > > > > devfs on /dev (devfs, local, multilabel)
> > > > > > > map -hosts on /net (autofs)
> > > > > > > map -media on /media (autofs)
> > > > > > > [trasz at v2:/media/md0]% ls
> > > > > > > [trasz at v2:/media/md0]% mount
> > > > > > > /dev/ada0s1a on / (ufs, local, noatime, journaled soft-updates)
> > > > > > > devfs on /dev (devfs, local, multilabel)
> > > > > > > map -hosts on /net (autofs)
> > > > > > > map -media on /media (autofs)
> > > > > > > /dev/md0 on /media/md0 (ufs, local, noatime, nosuid, automounted)
> > > > > > >
> > > > > > > Getting rid of mountcheckdirs() in the unmount path should be
> > fine, I
> > > > > > think.
> > > > > > How the example proves that mountcheckdirs() can be removed ?
> > > > >
> > > > >
> > > > > In the unmount path?  It doesn't; I just don't think autofs would be
> > > > > affected,
> > > > > since having a current working directory in a mountpoint prevents the
> > > > > unmount
> > > > > (unless it's forced).
> > > > >
> > > > > In the mount path?  See below.
> > > > >
> > > > >
> > > > > > How can
> > > > > > we see which directory content was printed by ls, the covered or
> > > > mounted ?
> > > > > >
> > > > >
> > > > > Well, it would be pretty useless if you got the covered directory.
> > > > This is exactly what is being discussed.
> > > >
> > > > > But ok, here's
> > > > > a better example, one that actually shows that:
> > > > >
> > > > > [trasz at v2:/media/md0]% mount
> > > > > /dev/ada0s1a on / (ufs, local, noatime, journaled soft-updates)
> > > > > devfs on /dev (devfs, local, multilabel)
> > > > > map -hosts on /net (autofs)
> > > > > map -media on /media (autofs)
> > > > > [trasz at v2:/media/md0]% ls -al
> > > > > total 9
> > > > > drwxr-xr-x  3 root  wheel     512 Sep 10 10:16 .
> > > > > drwxr-xr-x  3 root  wheel     512 Sep  8 11:42 ..
> > > > > drwxrwxr-x  2 root  operator  512 Sep 10 10:16 .snap
> > > > > [trasz at v2:/media/md0]% mount
> > > > > /dev/ada0s1a on / (ufs, local, noatime, journaled soft-updates)
> > > > > devfs on /dev (devfs, local, multilabel)
> > > > > map -hosts on /net (autofs)
> > > > > map -media on /media (autofs)
> > > > > /dev/md0 on /media/md0 (ufs, local, noatime, nosuid, automounted)
> > > > >
> > > > > Notice the .snap/ directory; it's an empty UFS filesystem.
> > > > And the .snap is from the /dev/md0 volume, right ?
> > > >
> > >
> > > Yes.
> > >
> > >
> > > > If yes, then this demonstration indeed proves that autofs behaves as
> > > > intended on mount.
> > > >
> > >
> > > Exactly.  And, from what I understand, that would break if
> > mountcheckdirs()
> > > got removed from the mount path.
> >
> > Even for autofs ?  Don't autofs postpone lookup() completion until the
> > mount occurs ?
> >
> 
> It does, and that case might work - I'm not sure, but it might.  But what
> happens afterwards, after the syscall that triggered the mount completes?
> You would still have a shell with cwd in that dir, now covered by another
> filesystem.  Autofs doesn't do anything to "lift" the process to the newly
> mounted filesystem root.
The change of the current directory should have already triggered the action,
because it was preceeded by lookup.  At least this is my understanding of
the flow.