Re: [PATCH 2/8] loopfs: implement loopfs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Sun, Apr 12, 2020 at 02:03:00PM +0200, Christian Brauner wrote:
> On Sun, Apr 12, 2020 at 12:38:54PM +0200, David Rheinsberg wrote:
> > Hey
> > 
> > On Thu, Apr 9, 2020 at 10:27 AM Christian Brauner
> > <christian.brauner@xxxxxxxxxx> wrote:
> > > On Thu, Apr 09, 2020 at 07:39:18AM +0200, David Rheinsberg wrote:
> > > > With loopfs in place, any process can create its own user_ns, mount
> > > > their private loopfs and create as many loop-devices as they want.
> > > > Hence, this limit does not serve as an effective global
> > > > resource-control. Secondly, anyone with access to `loop-control` can
> > > > now create loop instances until this limit is hit, thus causing anyone
> > > > else to be unable to create more. This effectively prevents you from
> > > > sharing a loopfs between non-trusting parties. I am unsure where that
> > > > limit would actually be used?
> > >
> > > Restricting it globally indeed wasn't the intended use-case for it. This
> > > was more so that you can specify an instance limit, bind-mount that
> > > instance into several places and sufficiently locked down users cannot
> > > exceed the instance limit.
> > 
> > But then these users can each exhaust the limit individually. As such,
> > you cannot share this instance across users that have no
> > trust-relationship. Fine with me, but I still don't understand in
> 
> Well, you can't really share anything across clients with the same
> privilege level if one of them is untrusted.
> 
> > which scenario the limit would be useful. Anyone can create a user-ns,
> > create a new loopfs mount, and just happily create more loop-devices.
> > So what is so special that you want to restrict the devices on a
> > _single_ mount instance?
> 
> To share that instance across namespaces. You can e.g. create the
> mount instance in one mount namespace owned by userns1, create a second
> user namespace usern2 with the same mapping which is blocked from
> creating additional user namespaces either by seccomp or by
> /proc/sys/user/max_user_namespaces or lsms what have you. Because it
> doesn't own the mount namespace the loopfs mount it is in it can't
> remount it and can't exceed the local limit.
> 
> > 
> > > I don't think we'd be getting much out of a global limit per se I think
> > > the initial namespace being able to reserve a bunch of devices
> > > they can always rely on being able create when they need them is more
> > > interesting. This is similat to what devpts implements with the
> > > "reserved" mount option and what I initially proposed for binderfs. For
> > > the latter it was deemed unnecessary by others so I dropped it from
> > > loopfs too.
> > 
> > The `reserve` of devpts has a fixed 2-tier system: A global limit, and
> > a init-ns reserve. This does nothing to protect one container from
> > another.
> 
> What I was getting at is that what matters first and foremost is
> protecting init userns.
> 
> > 
> > Furthermore, how do you intend to limit user-space from creating an
> > unbound amount of loop devices? Unless I am mistaken, with your
> > proposal *any* process can create a new loopfs with a basically
> > unlimited amount of loop-devices, thus easily triggering unbound
> > kernel allocations. I think this needs to be accounted. The classic
> > way is to put a per-uid limit into `struct user_struct` (done by
> > pipes, mlock, epoll, mq, etc.). An alternative is `struct ucount`,
> > which allows hierarchical management (inotify uses that, as an
> > example).
> 
> Yeah, I know. We can certainly do this.
> 
> > 
> > > I also expect most users to pre-create devices in the initial namespace
> > > instance they need (e.g. similar to what binderfs does or what loop
> > > devices currently have). Does that make sense to you?
> > 
> > Our use-case is to get programmatic access to loop-devices, so we can
> > build customer images on request (especially to create XFS images,
> > since mkfs.xfs cannot write them, IIRC). We would be perfectly happy
> > with a kernel-interface that takes a file-descriptor to a regular file
> > and returns us a file-descriptor to a newly created block device
> > (which is automatically destroyed when the last file-descriptor to it
> > is closed). This would be ideal *to us*, since it would do automatic
> > cleanup on crashes.
> > 
> > We don't need any representation of the loop-device in the
> > file-system, as long as we can somehow mount it (either by passing the
> > bdev-FD to the new mount-api, or by using /proc/self/fd/ as
> > mount-source).
> 
> We want the ability to have a filesystem representation as it will allow
> us to handle a host of legacy workloads cleanly e.g. that users can just
> call mount -o loop /bla whenever they have opted into syscall
> interception for a particular filesystem. In addition, we can cover your
> use case completely was well I think. Both with the old and new mount api.
> 
> > 
> > With your proposed loop-fs we could achieve something close to it:
> > Mount a private loopfs, create a loop-device, and rely on automatic
> > cleanup when the mount-namespace is destroyed.
> 
> With loopfs you can do this with the old or new mount api and you don't
> need to have loopfs mounted for that at all. Here's a sample program
> that works right now with the old mount api:

That also led me to discover a bug I need to fix, so thanks!
Christian



[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux