Re: Upcoming: Notifications, FS notifications and fsinfo()

[Lennart Poettering <mzxreary@0pointer.de>]

On Di, 31.03.20 10:56, Miklos Szeredi (miklos@szeredi.hu) wrote:

> On Tue, Mar 31, 2020 at 10:34 AM Karel Zak <kzak@redhat.com> wrote:
> >
> > On Tue, Mar 31, 2020 at 07:11:11AM +0200, Miklos Szeredi wrote:
> > > On Mon, Mar 30, 2020 at 11:17 PM Christian Brauner
> > > <christian.brauner@ubuntu.com> wrote:
> > >
> > > > Fwiw, putting down my kernel hat and speaking as someone who maintains
> > > > two container runtimes and various other low-level bits and pieces in
> > > > userspace who'd make heavy use of this stuff I would prefer the fd-based
> > > > fsinfo() approach especially in the light of across namespace
> > > > operations, querying all properties of a mount atomically all-at-once,
> > >
> > > fsinfo(2) doesn't meet the atomically all-at-once requirement.
> >
> > I guess your /proc based idea have exactly the same problem...
>
> Yes, that's exactly what I wanted to demonstrate: there's no
> fundamental difference between the two API's in this respect.
>
> > I see two possible ways:
> >
> > - after open("/mnt", O_PATH) create copy-on-write object in kernel to
> >   represent mount node -- kernel will able to modify it, but userspace
> >   will get unchanged data from the FD until to close()
> >
> > - improve fsinfo() to provide set (list) of the attributes by one call
>
> I think we are approaching this from the wrong end.   Let's just
> ignore all of the proposed interfaces for now and only concentrate on
> what this will be used for.
>
> Start with a set of use cases by all interested parties.  E.g.
>
>  - systemd wants to keep track attached mounts in a namespace, as well
> as new detached mounts created by fsmount()
>
>  - systemd need to keep information (such as parent, children, mount
> flags, fs options, etc) up to date on any change of topology or
> attributes.

- We also have code that recursively remounts r/o or unmounts some
  directory tree (with filters), which is currently nasty to do since
  the relationships between dirs are not always clear from
  /proc/self/mountinfo alone, in particular not in an even remotely
  atomic fashion, or when stuff is overmounted.

- We also have code that needs to check if /dev/ is plain tmpfs or
  devtmpfs. We cannot use statfs for that, since in both cases
  TMPFS_MAGIC is reported, hence we currently parse
  /proc/self/mountinfo for that to find the fstype string there, which
  is different for both cases.

Lennart

--
Lennart Poettering, Berlin