lkml.org 
[lkml]   [2020]   [Feb]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 3/7] vfio/pci: Introduce VF token
On Thu, 13 Feb 2020 10:23:21 -0700
Alex Williamson <alex.williamson@redhat.com> wrote:

> On Thu, 13 Feb 2020 12:46:54 +0100
> Cornelia Huck <cohuck@redhat.com> wrote:
>
> > On Tue, 11 Feb 2020 16:05:42 -0700
> > Alex Williamson <alex.williamson@redhat.com> wrote:
> >
> > > If we enable SR-IOV on a vfio-pci owned PF, the resulting VFs are not
> > > fully isolated from the PF. The PF can always cause a denial of
> > > service to the VF, if not access data passed through the VF directly.
> > > This is why vfio-pci currently does not bind to PFs with SR-IOV enabled
> > > and does not provide access itself to enabling SR-IOV on a PF. The
> > > IOMMU grouping mechanism might allow us a solution to this lack of
> > > isolation, however the deficiency isn't actually in the DMA path, so
> > > much as the potential cooperation between PF and VF devices. Also,
> > > if we were to force VFs into the same IOMMU group as the PF, we severely
> > > limit the utility of having independent drivers managing PFs and VFs
> > > with vfio.
> > >
> > > Therefore we introduce the concept of a VF token. The token is
> > > implemented as a UUID and represents a shared secret which must be set
> > > by the PF driver and used by the VF drivers in order to access a vfio
> > > device file descriptor for the VF. The ioctl to set the VF token will
> > > be provided in a later commit, this commit implements the underlying
> > > infrastructure. The concept here is to augment the string the user
> > > passes to match a device within a group in order to retrieve access to
> > > the device descriptor. For example, rather than passing only the PCI
> > > device name (ex. "0000:03:00.0") the user would also pass a vf_token
> > > UUID (ex. "2ab74924-c335-45f4-9b16-8569e5b08258"). The device match
> > > string therefore becomes:
> > >
> > > "0000:03:00.0 vf_token=2ab74924-c335-45f4-9b16-8569e5b08258"
> > >
> > > This syntax is expected to be extensible to future options as well, with
> > > the standard being:
> > >
> > > "$DEVICE_NAME $OPTION1=$VALUE1 $OPTION2=$VALUE2"
> >
> > Is this designed to be an AND condition? (I read it as such.)
>
> Not sure I understand, the device name is always required for
> compatibility, then zero or more key/value pairs may be needed
> depending on the vfio bus driver and device requirements. I'm not
> suggesting that the user would pass multiple vf_token= options and the
> implementation here would only parse the first. I'm really only
> suggesting the parsing format we'd use for multiple options, I'm not
> trying to dictate how a bus driver might make use of them. Does that
> make sense, should I change some wording?

Not multiple vf_token= options, but multiple, different options.

E.g. we have something like "$NAME foo=xyz bar=zyx". What is supposed
to happen?

- both the foo= and bar= values have to give a match
- either foo= or bar= have to match
- if foo= doesn't match, try bar=
- foo= and bar= are ignored, if not applicable

(My understanding is that $NAME matching continues to be mandatory?)

What should happen for vf_token= is reasonably clear, but I'm wondering
about further extensions, as you already talk about it.

>
> > >
> > > The device name must be first and option=value pairs are separated by
> > > spaces.
> > >
> > > The vf_token option is only required for VFs where the PF device is
> > > bound to vfio-pci. There is no change for PFs using existing host
> > > drivers.
> > >
> > > Note that in order to protect existing VF users, not only is it required
> > > to set a vf_token on the PF before VFs devices can be accessed, but also
> > > if there are existing VF users, (re)opening the PF device must also
> > > provide the current vf_token as authentication. This is intended to
> > > prevent a VF driver starting with a trusted PF driver and later being
> > > replaced by an unknown driver. A vf_token is not required to open the
> > > PF device when none of the VF devices are in use by vfio-pci drivers.
> > >
> > > Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> > > ---
> > > drivers/vfio/pci/vfio_pci.c | 127 +++++++++++++++++++++++++++++++++++
> > > drivers/vfio/pci/vfio_pci_private.h | 8 ++
> > > 2 files changed, 134 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/vfio/pci/vfio_pci.c b/drivers/vfio/pci/vfio_pci.c
> > > index 2ec6c31d0ab0..26aea9ac4863 100644
> > > --- a/drivers/vfio/pci/vfio_pci.c
> > > +++ b/drivers/vfio/pci/vfio_pci.c
> > > @@ -466,6 +466,35 @@ static void vfio_pci_disable(struct vfio_pci_device *vdev)
> > > vfio_pci_set_power_state(vdev, PCI_D3hot);
> > > }
> > >
> > > +static struct pci_driver vfio_pci_driver;
> > > +
> > > +static void vfio_pci_vf_token_user_add(struct vfio_pci_device *vdev, int val)
> >
> > Suggestion: call this _user_modify(), and have _user_add() and
> > _user_remove() as wrappers. That said, ...
>
> I did consider something along these lines, but it seems overly
> explicit for a helper that's used in two places with only two obvious
> and discrete values. If this were an exposed API, absolutely I'd agree.
>
> > > +{
> > > + struct pci_dev *physfn = pci_physfn(vdev->pdev);
> > > + struct vfio_device *pf_dev;
> > > + struct vfio_pci_device *pf_vdev;
> > > +
> > > + if (!vdev->pdev->is_virtfn)
> > > + return;
> > > +
> > > + pf_dev = vfio_device_get_from_dev(&physfn->dev);
> > > + if (!pf_dev)
> > > + return;
> > > +
> > > + if (pci_dev_driver(physfn) != &vfio_pci_driver) {
> > > + vfio_device_put(pf_dev);
> > > + return;
> > > + }
> > > +
> > > + pf_vdev = vfio_device_data(pf_dev);
> > > +
> > > + mutex_lock(&pf_vdev->vf_token->lock);
> > > + pf_vdev->vf_token->users += val;
> >
> > ...is this expected to always be >= 0? If yes, it looks like a bug if
> > this is called with val==-n for n > users.
>
> I'm not sure if you're inadvertently pointing out the bug in the
> vfio_pci_open() path below where we increment token users before
> vfio_pci_enable() which can fail, or your suggesting a WARN_ON here
> should this go negative. This is a static helper function, so I
> generally don't try to sanitize the inputs like I would for an exposed
> API.

Yes, if you let users drop below 0, it's an internal error. Still, I
think it's worth checking, so that we catch those internal errors early
on, so a WARN_ON is probably the right thing to do.

>
> > > + mutex_unlock(&pf_vdev->vf_token->lock);
> > > +
> > > + vfio_device_put(pf_dev);
> > > +}
> > > +
> > > static void vfio_pci_release(void *device_data)
> > > {
> > > struct vfio_pci_device *vdev = device_data;
> > > @@ -475,6 +504,7 @@ static void vfio_pci_release(void *device_data)
> > > if (!(--vdev->refcnt)) {
> > > vfio_spapr_pci_eeh_release(vdev->pdev);
> > > vfio_pci_disable(vdev);
> > > + vfio_pci_vf_token_user_add(vdev, -1);
> > > }
> > >
> > > mutex_unlock(&vdev->reflck->lock);
> > > @@ -493,6 +523,7 @@ static int vfio_pci_open(void *device_data)
> > > mutex_lock(&vdev->reflck->lock);
> > >
> > > if (!vdev->refcnt) {
> > > + vfio_pci_vf_token_user_add(vdev, 1);
> > > ret = vfio_pci_enable(vdev);
> > > if (ret)
> > > goto error;
>
> I think we want to include decrementing token users in the error path
> here.

Yes; good that my comment made you spot it, because I didn't notice :)

>
> > > @@ -1278,11 +1309,86 @@ static void vfio_pci_request(void *device_data, unsigned int count)
> > > mutex_unlock(&vdev->igate);
> > > }
> > >
> > > +#define VF_TOKEN_ARG "vf_token="
> > > +
> > > +/* Called with vdev->vf_token->lock */
> > > +static int vfio_pci_vf_token_match(struct vfio_pci_device *vdev, char *opts)
> > > +{
> > > + char *token;
> > > + uuid_t uuid;
> > > + int ret;
> > > +
> > > + if (!opts)
> > > + return -EINVAL;
> > > +
> > > + token = strstr(opts, VF_TOKEN_ARG);
> > > + if (!token)
> > > + return -EINVAL;
> > > +
> > > + token += strlen(VF_TOKEN_ARG);
> > > +
> > > + ret = uuid_parse(token, &uuid);
> > > + if (ret)
> > > + return ret;
> > > +
> > > + if (!uuid_equal(&uuid, &vdev->vf_token->uuid))
> > > + return -EACCES;
> > > +
> > > + return 0;

Again, I guess my objections below are a matter of taste; especially
because this function does the key=value parsing, and I'm not sure it's
the right place to do so.

> > > +}
> > > +
> > > static int vfio_pci_match(void *device_data, char *buf)
> > > {
> > > struct vfio_pci_device *vdev = device_data;
> > > + char *opts;
> > > +
> > > + opts = strchr(buf, ' ');
> > > + if (opts) {
> > > + *opts = 0;
> > > + opts++;
> > > + }
> > > +
> > > + if (strcmp(pci_name(vdev->pdev), buf))
> > > + return 0; /* No match */
> >
> > Up to here, this function is fine; but below, it gets a bit hard to
> > follow...
> >
> > > +
> > > + if (vdev->pdev->is_virtfn) {
> > > + struct pci_dev *physfn = pci_physfn(vdev->pdev);
> > > + struct vfio_device *pf_dev;
> > > + int ret = 0;
> > > +
> > > + pf_dev = vfio_device_get_from_dev(&physfn->dev);
> > > + if (pf_dev) {
> > > + if (pci_dev_driver(physfn) == &vfio_pci_driver) {
> > > + struct vfio_pci_device *pf_vdev =
> > > + vfio_device_data(pf_dev);
> > > +
> > > + mutex_lock(&pf_vdev->vf_token->lock);
> > > + ret = vfio_pci_vf_token_match(pf_vdev, opts);
> > > + mutex_unlock(&pf_vdev->vf_token->lock);
> > > + }
> > > +
> > > + vfio_device_put(pf_dev);
> > > +
> > > + if (ret)
> > > + return -EACCES;
> > > + }
> > > + }
> >
> > If we are a VF, and the PF is bound to vfio, pass whatever stuff other
> > than the device name that was passed in to an opaque match function.
>
> vfio_pci_match() is an opaque match function relative to vfio.c, but
> there's nothing opaque here. We have a VF where the associated PF is
> bound to vfio-pci, therefore we require that the additional options
> include a vf_token matching the PF and we go looking to verify that.
>
> > > - return !strcmp(pci_name(vdev->pdev), buf);
> > > + if (vdev->vf_token) {
> > > + int ret = 0;
> > > +
> > > + mutex_lock(&vdev->vf_token->lock);
> > > +
> > > + if (vdev->vf_token->users)
> > > + ret = vfio_pci_vf_token_match(vdev, opts);
> > > +
> > > + mutex_unlock(&vdev->vf_token->lock);
> > > +
> > > + if (ret)
> > > + return -EACCES;
> > > + }
> >
> > If we have a VF token with users, pass whatever stuff other than the
> > device name that was passed in to an opaque match function.
>
> This description strays further off course a bit. If we have a
> vf_token then we are a PF and clearly bound to vfio-pci. If there are
> existing VF users then we require the user to provide a vf_token
> matching the one currently on the device.

Maybe my wording is just a bit off...

>
> > What about the following instead:
> >
> > - parse the passed-in string into device name and key/value pairs
> > - maybe reject anything with an unknown key
>
> This is definitely something that we should decided whether or not we
> want to do it. I think an argument for it is that a user can pick
> arbitrary key=value options that would be ignored with this
> implementation, but later might match a key that gets defined and then
> we break them. Misuse of the API on the part of the user, but maybe
> better to just prevent it ahead of time.

Yes, it's probably good to do this now.

>
> > - check the device name
> > - if we're a VF with the PF bound to vfio, require a VF token to be
> > specified and pass it to a token match function
> > - if we have a VF token with users, require a VF token to be specified
> > and pass it to a token match function
>
> This is essentially what we do above. Maybe we just disagree about
> whether we're calling an "opaque match function" versus a "token match
> function".

Maybe I should have said "function parsing a string, which might
contain a lot of unrelated stuff" vs "function explicitly handling a
vf_token value".

>
> >
> > My main gripes with the current code are:
> > - key=value parsing is done in the match function for vf_token
> > - it looks hard to extend the list of supported key/value pairs
> > - I don't see a good way to find out (as the user) _why_ the VF isn't
> > matching
>
> If we want to reject unknown options, then yes, the parsing should be
> done ahead. I don't see that it's hard to extend though, each new
> requirement can follow the same methodology to check for it in the
> remaining options string.

If you pre-parse into option/value pairs, you see quite easily if you
managed to obtain a required option, if an option has been specified
more than once, or if an unknown option has been specified. If a new
option is introduced, you just need to handle whatever has been parsed
already. Extending is probably not exactly hard, but pre-parsing likely
makes it easier, as you don't have implicit assumptions.

>
> The last point is the one I brought up in the cover letter and where
> I'm also not happy with the opaque error condition, but I have no
> thoughts on how to resolve it. Either we block the user from getting
> the device file descriptor, and they're left scratching their heads
> why, or we give them access to the file descriptor AND we need to
> impose barriers to access mechanisms (ex. block read/write/mmap, limit
> ioctls) AND we need to use VFIO_DEVICE_FEATURE and VFIO_DEVICE_GET_INFO
> as a mechanism for the user to figure out that the device requires
> "something" to get full access. With the latter, I'm concerned whether
> existing userspace code will fail in predictable ways and that it
> snowballs into an ugly API. For instance, if we add a flag to device
> info to indicate it's locked, existing code won't know about that flag,
> so we have to cripple device info to report no regions and no irqs to
> make that code fail. Then a user needs to know which feature to probe
> for to figure out how the device is locked, then once they do we make
> device info report real values? It's maybe a little more deterministic
> than blocking access to the file descriptor, but I'm not sure it's
> worth it. We could do a log-once if the match fails for token, but we
> need to be careful not to provide an obvious point where the user could
> spam the logs. I've also considered if we could write an error back
> into the user's buffer, but the ioctl isn't designed that way and we
> don't know if we'd break how the user consumes that buffer later.

Some extended reporting mechanism is likely to become unwieldy,
especially when we realize we missed something. A simple log message
that a vf_token is required (pointing to a more verbose description, if
possible) looks best (obviously rate limited or only printed once).
Just enough to give the user a hint so that they are not left
completely baffled.

>
> Ideas greatly welcomed in this space. Thanks for the review!
>
> Alex

\
 
 \ /
  Last update: 2020-02-13 19:35    [W:0.075 / U:99.916 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site