[PATCH] fs: export anon_inode_make_secure_inode() and fix secretmem LSM bypass

[Christian Brauner]

On Fri, Jun 20, 2025 at 08:02:18AM -0700, Sean Christopherson wrote:
> On Thu, Jun 19, 2025, Mike Rapoport wrote:
> > On Thu, Jun 19, 2025 at 02:06:17PM +0200, Christian Brauner wrote:
> > > On Thu, Jun 19, 2025 at 02:01:22PM +0300, Mike Rapoport wrote:
> > > > On Thu, Jun 19, 2025 at 12:38:25PM +0200, Christian Brauner wrote:
> > > > > On Thu, Jun 19, 2025 at 11:13:49AM +0200, Vlastimil Babka wrote:
> > > > > > On 6/19/25 09:31, Shivank Garg wrote:
> > > > > > > Export anon_inode_make_secure_inode() to allow KVM guest_memfd to create
> > > > > > > anonymous inodes with proper security context. This replaces the current
> > > > > > > pattern of calling alloc_anon_inode() followed by
> > > > > > > inode_init_security_anon() for creating security context manually.
> > > > > > > 
> > > > > > > This change also fixes a security regression in secretmem where the
> > > > > > > S_PRIVATE flag was not cleared after alloc_anon_inode(), causing
> > > > > > > LSM/SELinux checks to be bypassed for secretmem file descriptors.
> > > > > > > 
> > > > > > > As guest_memfd currently resides in the KVM module, we need to export this
> > > > > > 
> > > > > > Could we use the new EXPORT_SYMBOL_GPL_FOR_MODULES() thingy to make this
> > > > > > explicit for KVM?
> > > > > 
> > > > > Oh? Enlighten me about that, if you have a second, please. 
> > > > 
> > > > From Documentation/core-api/symbol-namespaces.rst:
> > > > 
> > > > The macro takes a comma separated list of module names, allowing only those
> > > > modules to access this symbol. Simple tail-globs are supported.
> > > > 
> > > > For example::
> > > > 
> > > >   EXPORT_SYMBOL_GPL_FOR_MODULES(preempt_notifier_inc, "kvm,kvm-*")
> > > > 
> > > > will limit usage of this symbol to modules whoes name matches the given
> > > > patterns.
> > > 
> > > Is that still mostly advisory and can still be easily circumenvented?
> 
> Yes and no.  For out-of-tree modules, it's mostly advisory.  Though I can imagine
> if someone tries to report a bug because their module is masquerading as e.g. kvm,
> then they will be told to go away (in far less polite words :-D).
> 
> For in-tree modules, the restriction is much more enforceable.  Renaming a module
> to circumvent a restricted export will raise major red flags, and getting "proper"
> access to a symbol would require an ack from the relevant maintainers.  E.g. for
> many KVM-induced exports, it's not that other module writers are trying to misbehave,
> there simply aren't any guardrails to deter them from using a "dangerous" export.
>  
> The other big benefit I see is documentation, e.g. both for readers/developers to
> understand the intent, and for auditing purposes (I would be shocked if there
> aren't exports that were KVM-induced, but that are no longer necessary).
> 
> And we can utilize the framework to do additional hardening.  E.g. for exports
> that exist solely for KVM, I plan on adding wrappers so that the symbols are
> exproted if and only if KVM is enabled in the kernel .config[*].  Again, that's
> far from perfect, e.g. AFAIK every distro enables KVM, but it should help keep
> everyone honest.
> 
> [*] https://lore.kernel.org/all/ZzJOoFFPjrzYzKir@google.com 
> 
> > The commit message says
> > 
> >    will limit the use of said function to kvm.ko, any other module trying
> >    to use this symbol will refure to load (and get modpost build
> >    failures).
> 
> To Christian's point, the restrictions are trivial to circumvent by out-of-tree
> modules.  E.g. to get access to the above, simply name your module kvm-lol.ko or
> whatever.

Thanks for all the details!
I'm more than happy to switch a bunch of our exports so that we only
allow them for specific modules. But for that we also need
EXPOR_SYMBOL_FOR_MODULES() so we can switch our non-gpl versions.