Re: [PATCH v1 3/3] i386/sev: Add KVM_EXIT_SNP_REQ_CERTS support for certificate-fetching

[Michael Roth]

On Thu, Dec 19, 2024 at 01:37:18PM +0000, Daniel P. Berrangé wrote:
> On Thu, Dec 19, 2024 at 07:16:01AM -0600, Michael Roth wrote:
> > On Thu, Dec 19, 2024 at 08:13:44AM +0000, Daniel P. Berrangé wrote:
> > > On Wed, Dec 18, 2024 at 04:29:51PM -0600, Michael Roth wrote:
> > > > On Wed, Dec 18, 2024 at 05:50:52PM +0000, Daniel P. Berrangé wrote:
> > > > > On Wed, Dec 18, 2024 at 09:49:39AM -0600, Michael Roth wrote:
> > > > > > The GHCB specification[1] defines a VMGEXIT-based Guest Request
> > > > > > hypercall to allow an SNP guest to issue encrypted requests 
> > > > > > directly to
> > > > > > SNP firmware to do things like query the attestation report for the
> > > > > > guest. These are generally handled purely in the kernel.
> > > > > > 
> > > > > > In some some cases, it's useful for the host to be able to 
> > > > > > additionally
> > > > > > supply the certificate chain for the signing key that SNP firmware 
> > > > > > uses
> > > > > > to sign these attestation reports. To allow for this, the GHCB
> > > > > > specification defines an Extended Guest Request where this 
> > > > > > certificate
> > > > > > data can be provided in a special format described in the GHCB spec.
> > > > > > This certificate data may be global or guest-specific depending on 
> > > > > > how
> > > > > > the guest was configured. Rather than providing interfaces to manage
> > > > > > these within the kernel, KVM provides a new KVM_EXIT_SNP_REQ_CERTS 
> > > > > > exit
> > > > > > to request the certificate contents from userspace. Implement 
> > > > > > support
> > > > > > for that here.
> > > > > > 
> > > > > > To synchronize delivery of the certificates to the guest in a way 
> > > > > > where
> > > > > > they will not be rendered invalid by updates to SNP firmware or
> > > > > > attestation singing/endorsement keys by management tools outside the
> > > > > > purview of QEMU, it is expected by users of KVM_EXIT_SNP_REQ_CERTS 
> > > > > > to
> > > > > > obtain a shared/read lock on the certificate file prior to 
> > > > > > delivering
> > > > > > them back to KVM. Only after this will the attestation report be
> > > > > > retrieved from firmware and bundled with the certificate data, so 
> > > > > > QEMU
> > > > > > must continue to hold the file lock until KVM confirms that the
> > > > > > attestation report has been retrieved/bundled. This confirmation is 
> > > > > > done
> > > > > > by way of the kvm_immediate_exit callback infrastructure that was
> > > > > > introduced in a previous patch.
> > > > > > 
> > > > > > [1] "Guest Hypervisor Communication Block (GHCB) Standardization",
> > > > > >     https://www.amd.com/en/developer/sev.html
> > > > > > 
> > > > > > Signed-off-by: Michael Roth <michael.roth@amd.com>
> > > > > > ---
> > > > > >  qapi/qom.json                 |  23 +++-
> > > > > >  target/i386/kvm/kvm.c         |  10 ++
> > > > > >  target/i386/sev-sysemu-stub.c |   5 +
> > > > > >  target/i386/sev.c             | 249 
> > > > > > ++++++++++++++++++++++++++++++++++
> > > > > >  target/i386/sev.h             |   2 +
> > > > > >  5 files changed, 288 insertions(+), 1 deletion(-)
> > > > > > 
> > > > > > diff --git a/qapi/qom.json b/qapi/qom.json
> > > > > > index 28ce24cd8d..6eaf0e7721 100644
> > > > > > --- a/qapi/qom.json
> > > > > > +++ b/qapi/qom.json
> > > > > > @@ -1034,6 +1034,25 @@
> > > > > >  #     firmware.  Set this to true to disable the use of VCEK.
> > > > > >  #     (default: false) (since: 9.1)
> > > > > >  #
> > > > > > +# @certs-path: Path to certificate data that can be passed to 
> > > > > > guests via
> > > > > > +#              SNP Extended Guest Requests. File should be in the 
> > > > > > format
> > > > > > +#              described in the GHCB specification. (default: none)
> > > > > > +#              (since: 10.0)
> > > > > 
> > > > > Can we document the required format here explicitly, rather than 
> > > > > expecting
> > > > > users to go searching for specs which are often practically impossible
> > > > > to find, and even harder to read & interpret ?
> > > > 
> > > > It'll be difficult to summarize in a way that will be self-reliant,
> > > > since knowing the certificate format is not sufficient to make sure
> > > > it coincides with the endorsement key being used by firmware. So I can't
> > > > promise to completely reduce reliance on external specs, but at least
> > > > give a better of the format and where those external specs will come
> > > > into play in filling out the data.
> > > > 
> > > > If it needs to be at least somewhat self-sufficient then that might
> > > > warrant a separate document in 
> > > > docs/system/i386/amd-memory-encryption.rst
> > > > or somewhere thereabouts that summarizes the whole attestation flow and
> > > > how certificates tie into that.
> > > > 
> > > > Any preferences?
> > > > 
> > > > > 
> > > > > > diff --git a/target/i386/sev.c b/target/i386/sev.c
> > > > > > index 1a4eb1ada6..2c41bdbccf 100644
> > > > > > --- a/target/i386/sev.c
> > > > > > +++ b/target/i386/sev.c
> > > > > > @@ -157,6 +157,9 @@ struct SevSnpGuestState {
> > > > > >      char *id_auth_base64;
> > > > > >      uint8_t *id_auth;
> > > > > >      char *host_data;
> > > > > > +    char *certs_path;
> > > > > > +    int certs_fd;
> > > > > > +    uint32_t certs_timeout;
> > > > > >  
> > > > > >      struct kvm_sev_snp_launch_start kvm_start_conf;
> > > > > >      struct kvm_sev_snp_launch_finish kvm_finish_conf;
> > > > > > @@ -1355,6 +1358,215 @@ sev_snp_launch_finish(SevCommonState 
> > > > > > *sev_common)
> > > > > >      }
> > > > > >  }
> > > > > >  
> > > > > > +static int open_certs_locked(SevSnpGuestState *sev_snp_guest)
> > > > > > +{
> > > > > > +    int fd, ret;
> > > > > > +
> > > > > > +    if (sev_snp_guest->certs_fd != -1) {
> > > > > > +        return 0;
> > > > > > +    }
> > > > > > +
> > > > > > +    fd = qemu_open(sev_snp_guest->certs_path, O_RDONLY, NULL);
> > > > > > +    if (fd == -1) {
> > > > > > +        error_report("Unable to open certificate blob at path %s, 
> > > > > > ret %d",
> > > > > > +                     sev_snp_guest->certs_path, fd);
> > > > > > +        return fd;
> > > > > > +    }
> > > > > > +
> > > > > > +    ret = qemu_lock_fd(fd, 0, 0, false);
> > > > > > +    if (ret == -EAGAIN || ret == -EACCES) {
> > > > > > +        ret = -EAGAIN;
> > > > > > +        goto out_close;
> > > > > > +    } else if (ret) {
> > > > > > +        goto out_close;
> > > > > > +    }
> > > > > 
> > > > > This locking scheme is likely unsafe. Consider this sequence
> > > > > 
> > > > >   * QEMU runs qemu_open(path)
> > > > >   * External mgmt app runs unlink(path)
> > > > >   * External mgmt app runs open(path)
> > > > >   * External mgmt app runs lock(fd)
> > > > >   * QEMU runs qemu_lock_fd(fd)
> > > > > 
> > > > > QEMU has successfully acquired a lock on an FD that corresponds to a
> > > > > deleted file, not the current existing file.
> > > > > 
> > > > > Avoiding this problem requires either that the external mgmt app 
> > > > > agrees
> > > > > to *NEVER* unlink() the files under any circumstance, or for QEMU to
> > > > > run its open + lock logic in a loop, checking 'stat' and 'fstat' 
> > > > > before
> > > > > opening and after locking, in order to detect a replaced file from its
> > > > > changed inode.
> > > > > 
> > > > > I'm not inclined to rely on mgmt apps never unlink()ing as that's to
> > > > > easy to mess up IMHO.
> > > > 
> > > > Yah I went into more detail in my response to Markus, but long story
> > > > short is that we are assuming mgmt is cooperative in this case, and
> > > > so as you mentioned, it would never unlink files while SNP guests are
> > > > running, but instead take an exclusive lock on them and update them in
> > > > place with the understanding that doing anything otherwise would open
> > > > a race window where guests might get stale certificates.
> > > 
> > > If there's an expectation & requirement that no SNP guests are running,
> > > then IMHO this whole thing is just over-engineered. Just remove all this
> > > locking code entirely, and document that none of this must be changed
> > > while QEMU is running - which is a common requirement for a great many
> > > things on the host.
> > 
> > VCEK endorsement keys can change as a result of SNP firmware updates,
> > which can be done while SNP guests are running and are often done in such
> > a way to patch bugs/security holes. VLEK endorsement keys can similarly be
> > updated on a live host. Both these sorts of interactions cannot be made
> > compatible with bundling certificates with attestation reports without some
> > orchestration in place to keep them atomic relative to the endorsement
> > key being used by firmware to sign attestation reports. Every CSP
> > implementing this will need to solve it in some way, and I'm sure some
> > will handle all this completely differently. But it will make
> > interoperative management/tooling a mess, and having a reference
> > implementation based around something common will make it easier to
> > steer CSPs to that common solution and give management tools authors
> > *some* reference approach to target rather than expecting to retrofit
> > some custom solution on top.
> > 
> > With these patches, you can update firmware and endorsement keys while
> > SNP guests are running, but it requires write locks on any active
> > certificates as defined here and in the kernel, and doing certificate
> > file updates in place while that write lock is still held. I don't really
> > think that's over-engineered. I think it's surprisingly simple given the
> > potential complexity of the above-mentioned requirements.
> > 
> > But yes if management tries to unlink certificates while SNP guests are
> > running, all bets are off. But at that point they are not cooperating
> > with kernel/QEMU, and we don't need to care about that. And if they
> > really do need to blow away certificates for a complete re-install
> > or data-wipe or whatever, at that point they'd just need to ensure
> > they stop all their SNP guests first.
> 
> IMHO we msut consider unlink() to be a valid thing, because the right
> way for apps to perform crash safe atomic updates of existing files,
> is to use rename() from a temporary file, and the rename() in has an
> implicit unlink as part of its operation. ie apps would be doing:
> 
>    fd = open("foo.tmp")
>    write(fd, ...)
>    fsync(fd)
>    close(fd)
>    rename("foo.tmp", "foo")

If we still want to allow for this rather than enforcing in-place
update, one alternative would be to just allow a separate lock file
to be specified rather than locking the certificate file itself. That
would provide a bit more flexibility.

I can update the QEMU implementation to take -certs-lock-file in
addition to -certs-file so they can be specified separately. And if
-certs-lock-file is not specified then QEMU will just assume
management handles things different or has agreed to not do endorsement
key updates while SNP guests are running.

I think we'd considered something like that originally but the thinking
was that locking the certs themselves was more organic in terms of an
"obvious"/natural solution. But it does end up being a bit more
inflexible WRT how libraries/etc. might manage file updates underneath
the covers, so maybe a lock file is the better approach after all.

-Mike

> 
> That final rename operation will have the same effect on the locks
> as unlink(). To cope with this anything doing locking has to run
> in a loop comparing the inode either side of acquiring the lock
> 
> TLDR: if we're going to do locking in QEMU, it needs to be done
> robustly.
> 
> With regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
>