On Tue, Sep 28, 2010 at 11:10:42AM -0600, Alex Williamson wrote:
> On Tue, 2010-09-28 at 16:40 +0200, Michael S. Tsirkin wrote:
> > On Tue, Sep 28, 2010 at 08:26:21AM -0600, Alex Williamson wrote:
> > > On Tue, 2010-09-28 at 12:09 +0200, Michael S. Tsirkin wrote:
> > > > On Mon, Sep 27, 2010 at 02:46:32PM -0600, Alex Williamson wrote:
> > > > > On Wed, 2010-09-22 at 14:18 -0700, Tom Lyon wrote:
> > > > > > +ssize_t vfio_mem_readwrite(
> > > > > > + int write,
> > > > > > + struct vfio_dev *vdev,
> > > > > > + char __user *buf,
> > > > > > + size_t count,
> > > > > > + loff_t *ppos)
> > > > > > +{
> > > > > > + struct pci_dev *pdev = vdev->pdev;
> > > > > > + resource_size_t end;
> > > > > > + void __iomem *io;
> > > > > > + loff_t pos;
> > > > > > + int pci_space;
> > > > > > +
> > > > > > + pci_space = vfio_offset_to_pci_space(*ppos);
> > > > > > + pos = vfio_offset_to_pci_offset(*ppos);
> > > > > > +
> > > > > > + if (!pci_resource_start(pdev, pci_space))
> > > > > > + return -EINVAL;
> > > > > > + end = pci_resource_len(pdev, pci_space);
> > > > > > + if (vdev->barmap[pci_space] == NULL)
> > > > > > + vdev->barmap[pci_space] = pci_iomap(pdev, pci_space, 0);
> > > > > > + io = vdev->barmap[pci_space];
> > > > > > +
> > > > >
> > > > > So we do a pci_iomap, but never do corresponding pci_iounmap. This also
> > > > > only works for the first 6 BARs since the ROM BAR needs pci_map_rom.
> > > >
> > > >
> > > > An issue with ROM is that I think it can not be enabled together
> > > > with BARs. This is why pci_read_rom/pci_write_rom do what
> > > > they do.
> > >
> > > I don't think that's an issue. pci_read/write_rom doesn't do anything
> > > about disabling BARs when the ROM is read.
> >
> > Oh yes it does: it calls pci_map rom which calls pci_enable_rom.
> > And this enables ROM space decoder, which, PCI spec says,
> > can disable BAR decoder in device.
> >
> > See it now?
>
> Ok, I see the comment there now, thanks.
>
> > > The reason that it maps and
> > > unmaps around the read is to be resource friendly.
> >
> > Not only that. you can not use both ROM and BARs
> > at the same time: if you enable ROM you must not
> > access BARs.
>
> So we either need to shadow the ROM if we want to support an mmap to it
> or map/unmap around reads.
> > > For vfio, since
> > > we've assigned a device to vfio and it's open, I think it's valid to
> > > assign all the resources and sit on them. Otherwise I don't think we
> > > can guarantee that a read that worked last time will still get resources
> > > and work this time.
> >
> > The issue is that BARs are not accessible while ROM is enabled. It
> > might be a security hole to try to do that from an app, and I do know
> > there is hardware where it will not work, and PCI spec says as much:
> >
> > 6.2.5.2:
> > In order to minimize the number of address decoders needed, a device may
> > share a decoder between the Expansion ROM Base Address register and
> > other Base Address registers.47 When expansion ROM decode is enabled,
> > the decoder is used for accesses to the expansion ROM and device
> > independent software must not access the device through any other Base
> > Address registers.
> >
> > Thus I think applications must shadow the ROM and kernel must prevent
> > ROM access after mapping BARs. We could shadow in kernel as well but this
> > would be expensive in memory as ROMs are sometimes quite large.
>
> Shadowing in the VFIO kernel driver seems too resource intensive,
> probably best to disable mmap for the ROM and only allow read write,
> letting the app shadow if it wants. Preventing BAR access is
> troublesome since we support mmaps. I don't like the idea of overly
> restrictive ordering, but should we shutdown ROM access after the domain
> is set(?)
We probably must disable it on BAR mmap. Further existing sysfs
lacks locking so I think it's racy wrt multiple readers, etc.
> > > > > I
> > > > > wonder if we should be doing all the BAR mapping at open and unmap at
> > > > > close so that we can fail if the device can't get basic resources.
> > > >
> > > > I belive we should do this on ioctl so that e.g. hotunplug
> > > > can reset the device clean it up. Unused device should also not
> > > > consume resources.
> > >
> > > If it's assigned to vfio and opened, then it's not unused. A hotunplug
> > > would involve a close of the vfio device, which could then reset the
> > > device. I think we probably are still missing a device reset, but I'm
> > > not sure if it needs an ioctl or if a reset on open would be sufficient.
> > >
> > > Alex
> >
> > You are only looking at qemu, but look at the whole stack:
> >
> > First, I think we want reset on close, so the device is in sane
> > state when it's returned to the OS.
>
> Yep, I'd agree with that.
>
> > Second, we typically will want libvirt to open the device and pass it to qemu,
> > qemu might start using it much later, and it might keep it around,
> > unused, for as long as it wants. So we do want a way to reset without close.
>
> I'm fine with that, but we also need a reset on open to prevent leaking
> host state to guest/userspace.
Can't hurt ...
> > Finally, we will want a way for libvirt to verify that qemu has
> > closed the device, so it's safe to pass it to another guest.
>
> This gets tricky since we'd really like a revoke to insure there isn't a
> mmap still lingering. What do you propose?
revoke isn't practical I am afraid :) We can have an ioctl making the fd
exclusive, so that it fails until all fds for the old one go away. Or
maybe just make open exclusive in this way (might be problematic as open
does not have NONBLOCK flag).
> > > > > I
> > > > > believe we should also be calling pci_request_regions in here somewhere.
> > > > > Perhaps something like this:
> > > > >
> > > > > diff --git a/drivers/vfio/vfio_main.c b/drivers/vfio/vfio_main.c
> > > > > index a18e39a..d3886d9 100644
> > > > > --- a/drivers/vfio/vfio_main.c
> > > > > +++ b/drivers/vfio/vfio_main.c
> > > > > @@ -85,6 +85,53 @@ static inline int overlap(int a1, int b1, int a2, int b2)
> > > > > return !(b2 <= a1 || b1 <= a2);
> > > > > }
> > > > >
> > > > > +static int vfio_setup_pci(struct vfio_dev *vdev)
> > > > > +{
> > > > > + int ret, bar;
> > > > > +
> > > > > + ret = pci_enable_device(vdev->pdev);
> > > > > + if (ret)
> > > > > + return ret;
> > > > > +
> > > > > + ret = pci_request_regions(vdev->pdev, "VFIO");
> > > > > + if (ret) {
> > > > > + pci_disable_device(vdev->pdev);
> > > > > + return ret;
> > > > > + }
> > > > > +
> > > > > + for (bar = PCI_STD_RESOURCES; bar <= PCI_ROM_RESOURCE; bar++) {
> > > > > + if (!pci_resource_len(vdev->pdev, bar))
> > > > > + continue;
> > > > > + if (bar != PCI_ROM_RESOURCE) {
> > > > > + if (!pci_resource_start(vdev->pdev, bar))
> > > > > + continue;
> > > > > + vdev->barmap[bar] = pci_iomap(vdev->pdev, bar, 0);
> > > > > + } else {
> > > > > + size_t size;
> > > > > + vdev->barmap[bar] = pci_map_rom(vdev->pdev, &size);
> > > > > + }
> > > > > + }
> > > > > + return ret;
> > > > > +}
> > > > > +
> > > > > +static void vfio_disable_pci(struct vfio_dev *vdev)
> > > > > +{
> > > > > + int bar;
> > > > > +
> > > > > + for (bar = PCI_STD_RESOURCES; bar <= PCI_ROM_RESOURCE; bar++) {
> > > > > + if (!vdev->barmap[bar])
> > > > > + continue;
> > > > > + if (bar != PCI_ROM_RESOURCE)
> > > > > + pci_iounmap(vdev->pdev, vdev->barmap[bar]);
> > > > > + else
> > > > > + pci_unmap_rom(vdev->pdev, vdev->barmap[bar]);
> > > > > + vdev->barmap[bar] = NULL;
> > > > > + }
> > > > > +
> > > > > + pci_release_regions(vdev->pdev);
> > > > > + pci_disable_device(vdev->pdev);
> > > > > +}
> > > > > +
> > > > > static int vfio_open(struct inode *inode, struct file *filep)
> > > > > {
> > > > > struct vfio_dev *vdev;
> > > > > @@ -110,7 +157,7 @@ static int vfio_open(struct inode *inode, struct file *filep)
> > > > > INIT_LIST_HEAD(&listener->dm_list);
> > > > > filep->private_data = listener;
> > > > > if (vdev->listeners == 0)
> > > > > - ret = pci_enable_device(vdev->pdev);
> > > > > + ret = vfio_setup_pci(vdev);
> > > > > if (ret == 0)
> > > > > vdev->listeners++;
> > > > > mutex_unlock(&vdev->lgate);
> > > > > @@ -151,7 +198,7 @@ static int vfio_release(struct inode *inode, struct file *filep)
> > > > > vdev->vconfig = NULL;
> > > > > kfree(vdev->pci_config_map);
> > > > > vdev->pci_config_map = NULL;
> > > > > - pci_disable_device(vdev->pdev);
> > > > > + vfio_disable_pci(vdev);
> > > > > vfio_domain_unset(vdev);
> > > > > wake_up(&vdev->dev_idle_q);
> > > > > }
> > > > > diff --git a/drivers/vfio/vfio_rdwr.c b/drivers/vfio/vfio_rdwr.c
> > > > > index 1fd50a6..7705b45 100644
> > > > > --- a/drivers/vfio/vfio_rdwr.c
> > > > > +++ b/drivers/vfio/vfio_rdwr.c
> > > > > @@ -64,7 +64,7 @@ ssize_t vfio_io_readwrite(
> > > > > if (pos + count > end)
> > > > > return -EINVAL;
> > > > > if (vdev->barmap[pci_space] == NULL)
> > > > > - vdev->barmap[pci_space] = pci_iomap(pdev, pci_space, 0);
> > > > > + return -EINVAL;
> > > > > io = vdev->barmap[pci_space];
> > > > >
> > > > > while (count > 0) {
> > > > > @@ -137,7 +137,12 @@ ssize_t vfio_mem_readwrite(
> > > > > return -EINVAL;
> > > > > end = pci_resource_len(pdev, pci_space);
> > > > > if (vdev->barmap[pci_space] == NULL)
> > > > > - vdev->barmap[pci_space] = pci_iomap(pdev, pci_space, 0);
> > > > > + return -EINVAL;
> > > > > + if (pci_space == PCI_ROM_RESOURCE) {
> > > > > + u32 rom = *(u32 *)(vdev->vconfig + PCI_ROM_ADDRESS);
> > > > > + if (!(rom & PCI_ROM_ADDRESS_ENABLE))
> > > > > + return -EINVAL;
> > > > > + }
> > > > > io = vdev->barmap[pci_space];
> > > > >
> > > > > if (pos > end)
> > > > >
> > >
> > >
>
>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
