Re: [PATCH] vfio: pci: fix oops in case of vfio_msi_set_vector_signal failure

[Alex Williamson]

On Mon, 2016-02-01 at 18:27 +0100, Eric Auger wrote:
> Hi Alex,
> On 01/29/2016 10:41 PM, Alex Williamson wrote:
> > On Fri, 2016-01-29 at 14:43 +0000, Eric Auger wrote:
> > > In case vfio_msi_set_vector_signal fails we tear down everything.
> > > In the tear down loop we compare int j against unsigned start. Given
> > > the arithmetic conversion I think it is converted into an unsigned and
> > > becomes 0xffffffff, leading to the loop being entered again and things
> > > turn bad when accessing vdev->msix[vector].vector. So let's use int
> > > parameters instead.
> > > Â
> > > Signed-off-by: Eric Auger <eric.auger@xxxxxxxxxx>
> > > ---
> > > Âdrivers/vfio/pci/vfio_pci_intrs.c | 4 ++--
> > > Â1 file changed, 2 insertions(+), 2 deletions(-)
> > > Â
> > > diff --git a/drivers/vfio/pci/vfio_pci_intrs.c b/drivers/vfio/pci/vfio_pci_intrs.c
> > > index 3b3ba15..510c48d 100644
> > > --- a/drivers/vfio/pci/vfio_pci_intrs.c
> > > +++ b/drivers/vfio/pci/vfio_pci_intrs.c
> > > @@ -374,8 +374,8 @@ static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
> > > Â	return 0;
> > > Â}
> > > Â
> > > -static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
> > > -			ÂÂÂÂÂÂunsigned count, int32_t *fds, bool msix)
> > > +static int vfio_msi_set_block(struct vfio_pci_device *vdev, int start,
> > > +			ÂÂÂÂÂÂint count, int32_t *fds, bool msix)
> > > Â{
> > > Â	int i, j, ret = 0;
> > > Â
> >Â
> > Nice find, I don't think that's the only bug there though.ÂÂIf @start is
> > -1 (UINT32_MAX) and @count is 1, then @j gets set to -1 in the setup and
> > we hit the same index dereference problem.ÂÂWhat if we did this instead:
> >Â
> > diff --git a/drivers/vfio/pci/vfio_pci_intrs.c b/drivers/vfio/pci/vfio_pci_intrs.c
> > index 3b3ba15..2ae84ad 100644
> > --- a/drivers/vfio/pci/vfio_pci_intrs.c
> > +++ b/drivers/vfio/pci/vfio_pci_intrs.c
> > @@ -309,14 +309,14 @@ static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
> > Â				ÂÂÂÂÂÂint vector, int fd, bool msix)
> > Â{
> > Â	struct pci_dev *pdev = vdev->pdev;
> > -	int irq = msix ? vdev->msix[vector].vector : pdev->irq + vector;
> > -	char *name = msix ? "vfio-msix" : "vfio-msi";
> > Â	struct eventfd_ctx *trigger;
> > -	int ret;
> > +	int irq, ret;
> > Â
> > -	if (vector >= vdev->num_ctx)
> > +	if (vector < 0 || vector >= vdev->num_ctx)
> > Â		return -EINVAL;
> > Â
> > +	irq = msix ? vdev->msix[vector].vector : pdev->irq + vector;
> > +
> > Â	if (vdev->ctx[vector].trigger) {
> > Â		free_irq(irq, vdev->ctx[vector].trigger);
> > Â		irq_bypass_unregister_producer(&vdev->ctx[vector].producer);
> > @@ -328,8 +328,9 @@ static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
> > Â	if (fd < 0)
> > Â		return 0;
> > Â
> > -	vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "%s[%d](%s)",
> > -					ÂÂÂname, vector, pci_name(pdev));
> > +	vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "vfio-msi%s[%d](%s)",
> > +					ÂÂÂmsix ? "x" : "", vector,
> > +					ÂÂÂpci_name(pdev));
> > Â	if (!vdev->ctx[vector].name)
> > Â		return -ENOMEM;
> > Â
> > @@ -379,7 +380,7 @@ static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
> > Â{
> > Â	int i, j, ret = 0;
> > Â
> > -	if (start + count > vdev->num_ctx)
> > +	if (start >= vdev->num_ctx || start + count > vdev->num_ctx)
> > Â		return -EINVAL;
> > Â
> > Â	for (i = 0, j = start; i < count && !ret; i++, j++) {
> > @@ -388,7 +389,7 @@ static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
> > Â	}
> > Â
> > Â	if (ret) {
> > -		for (--j; j >= start; j--)
> > +		for (--j; j >= 0 && j >= start; j--)
> > Â			vfio_msi_set_vector_signal(vdev, j, -1, msix);
> > Â	}
> > Â
> >Â
> > So we fix the problem with vfio_msi_set_vector_signal() dereferencing
> > the array before it validates the index (even though it shouldn't be
> > able to get there anymore), and then we do a better job of verifying
> > start and count (comparing to num_ctx will use unsigned even though
> > num_ctx itself is signed) and finally explicitly test the <0 case, which
> > I suppose we could also do by casting start at that point (we know it's
> > within the bounds of a signed integer given the previous tests).
>Â
> Yes it looks OK to me.
>Â
> I guess you submit? I will test it.

Yep, I'll post a real patch.ÂÂThanks,

Alex