Re: [RFC PATCH v1 0/6] proc: Add allowlist for procfs files

[Alexey Gladkov]

On Thu, Jan 26, 2023 at 11:16:07AM +0100, Christian Brauner wrote:
> On Wed, Jan 25, 2023 at 03:36:28PM -0800, Andrew Morton wrote:
> > On Wed, 25 Jan 2023 16:28:47 +0100 Alexey Gladkov <legion@xxxxxxxxxx> wrote:
> > 
> > > The patch expands subset= option. If the proc is mounted with the
> > > subset=allowlist option, the /proc/allowlist file will appear. This file
> > > contains the filenames and directories that are allowed for this
> > > mountpoint. By default, /proc/allowlist contains only its own name.
> > > Changing the allowlist is possible as long as it is present in the
> > > allowlist itself.
> > > 
> > > This allowlist is applied in lookup/readdir so files that will create
> > > modules after mounting will not be visible.
> > > 
> > > Compared to the previous patches [1][2], I switched to a special virtual
> > > file from listing filenames in the mount options.
> > > 
> > 
> > Changlog doesn't explain why you think Linux needs this feature.  The
> > [2/6] changelog hints that containers might be involved.  IOW, please
> > fully describe the requirement and use-case(s).
> > 
> > Also, please describe why /proc/allowlist is made available via a mount
> > option, rather than being permanently present.
> > 
> > And why add to subset=, instead of a separate mount option.
> > 
> > Does /proc/allowlist work in subdirectories?  Like, permit presence of
> > /proc/sys/vm/compact_memory?
> > 
> > I think the whole thing is misnamed, really.  "allowlist" implies
> > access permissions.  Some of the test here uses "visibility" and other
> > places use "presence", which are better.  "presentlist" and
> > /proc/presentlist might be better.  But why not simply /proc/contents?
> 
> Currently, a lot of container runtimes - even if they mount a new procfs
> instance - overmount various procfs files and directories to ensure that
> they're hidden from the container workload. (The motivations for this
> are mixed and usually it's only needed for containers that run with the
> same privilege level as the host.)
> 
> The consequence of overmounting is that we need to refuse mounting
> procfs again somewhere else otherwise the procfs instance might reveal
> files and directories that were supposed to be hidden.
> 
> So this patchset moves the ability to hide entries into the kernel
> through an allowlist. This way you can hide files and directories while
> being able to mount procfs again because it will inherit the same
> allowlist.
> 
> I get the motivation. The question is whether this belongs into the
> kernel at all. I'm unfortunately not convinced.
> 
> This adds a lot of string parsing to procfs and I think we would also
> need to decide what a reasonable maximum limit for such allowlists would
> be.> The data structure likely shouldn't be a linked list but at least an
> rbtree especially if the size isn't limited.

There is a limit. So far I've limited the file size to 128k. I think this
is a reasonable limit.

> But fundamentally I think it moves something that should be and
> currently is a userspace policy into the kernel which I think is wrong.

We don't have mechanisms to implement this userspace policy. overmount is
not a solution but plugging holes in the absence of other ways to control
the visibility of files in procfs.

> Sure you can't predict what files show up in procfs over time but then
> subset=pid is already your friend - even if not as fine-grained.
> 
> If this where another simple subset style mount option that allowlists a
> bunch of well-known global proc files then sure. But making this
> dynamically configurable from userspace doesn't make sense to me. I
> mean, users could write /gobble/dy/gook into /proc/allowlist or use it
> to stash secrets or hashes or whatever as we have no way of figuring out
> whether the entry they allowlist does or will actually ever exist.

BTW I only allow printable data to be written to the file.

We can make this file write-only and then writing any extraneous data
there will not make sense.

> In general, such flexibility belongs into userspace imho.
> 
> Frankly, if that is really required it would almost make more sense to
> be able to attach a new bpf program type to procfs that would allow to
> filter procfs entries. Then the filter could be done purely in
> userspace. If signed bpf lands one could then even ship signed programs
> that are attachable by userns root.

I'll ask the podman developers how much more comfortable they would be
using bpf to control file visibility in procfs. thanks for the idea.

-- 
Rgrds, legion