Re: [PATCH 00/17] VFS: Filesystem information and notifications [ver #17]

[Miklos Szeredi <miklos@szeredi.hu>]

On Tue, Mar 3, 2020 at 6:28 AM Ian Kent <raven@themaw.net> wrote:
>
> On Mon, 2020-03-02 at 10:09 +0100, Miklos Szeredi wrote:
> > On Fri, Feb 28, 2020 at 5:36 PM David Howells <dhowells@redhat.com>
> > wrote:
> > > sysfs also has some other disadvantages for this:
> > >
> > >  (1) There's a potential chicken-and-egg problem in that you have
> > > to create a
> > >      bunch of files and dirs in sysfs for every created mount and
> > > superblock
> > >      (possibly excluding special ones like the socket mount) - but
> > > this
> > >      includes sysfs itself.  This might work - provided you create
> > > sysfs
> > >      first.
> >
> > Sysfs architecture looks something like this (I hope Greg will
> > correct
> > me if I'm wrong):
> >
> > device driver -> kobj tree <- sysfs tree
> >
> > The kobj tree is created by the device driver, and the dentry tree is
> > created on demand from the kobj tree.   Lifetime of kobjs is bound to
> > both the sysfs objects and the device but not the other way round.
> > I.e. device can go away while the sysfs object is still being
> > referenced, and sysfs can be freely mounted and unmounted
> > independently of device initialization.
> >
> > So there's no ordering requirement between sysfs mounts and other
> > mounts.   I might be wrong on the details, since mounts are created
> > very early in the boot process...
> >
> > >  (2) sysfs is memory intensive.  The directory structure has to be
> > > backed by
> > >      dentries and inodes that linger as long as the referenced
> > > object does
> > >      (procfs is more efficient in this regard for files that aren't
> > > being
> > >      accessed)
> >
> > See above: I don't think dentries and inodes are pinned, only kobjs
> > and their associated cruft.  Which may be too heavy, depending on the
> > details of the kobj tree.
> >
> > >  (3) It gives people extra, indirect ways to pin mount objects and
> > >      superblocks.
> >
> > See above.
> >
> > > For the moment, fsinfo() gives you three ways of referring to a
> > > filesystem
> > > object:
> > >
> > >  (a) Directly by path.
> >
> > A path is always representable by an O_PATH descriptor.
> >
> > >  (b) By path associated with an fd.
> >
> > See my proposal about linking from /proc/$PID/fdmount/$FD ->
> > /sys/devices/virtual/mounts/$MOUNT_ID.
> >
> > >  (c) By mount ID (perm checked by working back up the tree).
> >
> > Check that perm on lookup of /sys/devices/virtual/mounts/$MOUNT_ID.
> > The proc symlink would bypass the lookup check by directly jumping to
> > the mountinfo dir.
> >
> > > but will need to add:
> > >
> > >  (d) By fscontext fd (which is hard to find in sysfs).  Indeed, the
> > > superblock
> > >      may not even exist yet.
> >
> > Proc symlink would work for that too.
>
> There's mounts enumeration too, ordering is required to identify the
> top (or bottom depending on terminology) with more than one mount on
> a mount point.
>
> >
> > If sysfs is too heavy, this could be proc or a completely new
> > filesystem.  The implementation is much less relevant at this stage
> > of
> > the discussion than the interface.
>
> Ha, proc with the seq file interface, that's already proved to not
> work properly and looks difficult to fix.

I'm doing a patch.   Let's see how it fares in the face of all these
preconceptions.

Thanks,
Miklos