[PATCH 3/3] virt: Add sev_secret module to expose confidential computing secrets

[Ard Biesheuvel]

On Fri, 13 Aug 2021 at 15:05, Andrew Scull <ascull at google.com> wrote:
>
> On Mon, Aug 09, 2021 at 07:01:57PM +0000, Dov Murik wrote:
> > The new sev_secret module exposes the confidential computing (coco)
> > secret area via securityfs interface.
> >
> > When the module is loaded (and securityfs is mounted, typically under
> > /sys/kernel/security), a "coco/sev_secret" directory is created in
> > securityfs.  In it, a file is created for each secret entry.  The name
> > of each such file is the GUID of the secret entry, and its content is
> > the secret data.
> >
> > This allows applications running in a confidential computing setting to
> > read secrets provided by the guest owner via a secure secret injection
> > mechanism (such as AMD SEV's LAUNCH_SECRET command).
> >
> > Removing (unlinking) files in the "coco/sev_secret" directory will zero
> > out the secret in memory, and remove the filesystem entry.  If the
> > module is removed and loaded again, that secret will not appear in the
> > filesystem.
>
> We've also been looking into a similar secret mechanism recently in the
> context of Android and protected KVM [1]. Our secrets would come from a
> different source, likely described as a reserved-memory node in the DT,
> but would need to be exposed to userspace in the same way as the SEV
> secrets. Originally I tried using a character device, but this approach
> with securityfs feels neater to me.
>

Agreed. I particularly like how deleting the file wipes the secret from memory.

> We're also looking to pass secrets from the bootloader to Linux, outside
> of any virtualization or confidential compute context (at least a far as
> I have understood the meaning of the term). Again, this feels like it
> would be exposed to userspace in the same way.
>

Indeed.

> It would be good to be able to share the parts that would be common. I
> expect that would mean the operations for a secret file and for a
> directory of secrets at a minimum. But it might also influence the paths
> in securityfs; I see, looking back, that the "coco" directory was added
> since the RFC but would a generalized "secret" subsystem make sense? Or
> would it be preferable for each case to define their own path?
>

I think we should avoid 'secret', to be honest. Even if protected KVM
is not riding the SEV/TDX wave, I think confidential computing is
still an accurate description of its semantics.

> [1] -- https://lwn.net/Articles/836693/
>
> > +static int sev_secret_unlink(struct inode *dir, struct dentry *dentry)
> > +{
> > +     struct sev_secret *s = sev_secret_get();
> > +     struct inode *inode = d_inode(dentry);
> > +     struct secret_entry *e = (struct secret_entry *)inode->i_private;
> > +     int i;
> > +
> > +     if (e) {
> > +             /* Zero out the secret data */
> > +             memzero_explicit(e->data, secret_entry_data_len(e));
>
> Would there be a benefit in flushing these zeros?
>

Do you mean cache clean+invalidate? Better to be precise here.


> > +             e->guid = NULL_GUID;
> > +     }
> > +
> > +     inode->i_private = NULL;
> > +
> > +     for (i = 0; i < SEV_SECRET_NUM_FILES; i++)
> > +             if (s->fs_files[i] == dentry)
> > +                     s->fs_files[i] = NULL;
> > +
> > +     /*
> > +      * securityfs_remove tries to lock the directory's inode, but we reach
> > +      * the unlink callback when it's already locked
> > +      */
> > +     inode_unlock(dir);
> > +     securityfs_remove(dentry);
> > +     inode_lock(dir);
> > +
> > +     return 0;
> > +}