Re: [PATCH for 7.2? V2] vhost: fix vq dirty bitmap syncing when vIOMMU is enabled

[Jason Wang]

On Wed, Nov 30, 2022 at 12:08 AM Eric Auger <eric.auger@redhat.com> wrote:
>
> Hi Michael,
>
> On 11/29/22 16:44, Michael S. Tsirkin wrote:
> > On Tue, Nov 29, 2022 at 10:52:29AM +0100, Eric Auger wrote:
> >> Hi Jason,
> >>
> >> On 11/29/22 05:02, Jason Wang wrote:
> >>> When vIOMMU is enabled, the vq->used_phys is actually the IOVA not
> >>> GPA. So we need to translate it to GPA before the syncing otherwise we
> >>> may hit the following crash since IOVA could be out of the scope of
> >>> the GPA log size. This could be noted when using virtio-IOMMU with
> >>> vhost using 1G memory.
> >>>
> >>> Fixes: c471ad0e9bd46 ("vhost_net: device IOTLB support")
> >>> Cc: qemu-stable@nongnu.org
> >>> Tested-by: Lei Yang <leiyang@redhat.com>
> >>> Reported-by: Yalan Zhang <yalzhang@redhat.com>
> >>> Signed-off-by: Jason Wang <jasowang@redhat.com>
> >>> ---
> >>> Changes since V1:
> >>> - Fix the address calculation when used ring is not page aligned
> >>> - Fix the length for each round of dirty bitmap syncing
> >>> - Use LOG_GUEST_ERROR to log wrong used adddress
> >>> - Various other tweaks
> >>> ---
> >>>  hw/virtio/vhost.c | 76 ++++++++++++++++++++++++++++++++++-------------
> >>>  1 file changed, 56 insertions(+), 20 deletions(-)
> >>>
> >>> diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
> >>> index d1c4c20b8c..0cd5f25fcb 100644
> >>> --- a/hw/virtio/vhost.c
> >>> +++ b/hw/virtio/vhost.c
> >>> @@ -20,6 +20,7 @@
> >>>  #include "qemu/range.h"
> >>>  #include "qemu/error-report.h"
> >>>  #include "qemu/memfd.h"
> >>> +#include "qemu/log.h"
> >>>  #include "standard-headers/linux/vhost_types.h"
> >>>  #include "hw/virtio/virtio-bus.h"
> >>>  #include "hw/virtio/virtio-access.h"
> >>> @@ -106,6 +107,24 @@ static void vhost_dev_sync_region(struct vhost_dev 
> >>> *dev,
> >>>      }
> >>>  }
> >>>
> >>> +static bool vhost_dev_has_iommu(struct vhost_dev *dev)
> >>> +{
> >>> +    VirtIODevice *vdev = dev->vdev;
> >>> +
> >>> +    /*
> >>> +     * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
> >>> +     * incremental memory mapping API via IOTLB API. For platform that
> >>> +     * does not have IOMMU, there's no need to enable this feature
> >>> +     * which may cause unnecessary IOTLB miss/update transactions.
> >>> +     */
> >>> +    if (vdev) {
> >>> +        return virtio_bus_device_iommu_enabled(vdev) &&
> >>> +            virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
> >>> +    } else {
> >>> +        return false;
> >>> +    }
> >>> +}
> >>> +
> >>>  static int vhost_sync_dirty_bitmap(struct vhost_dev *dev,
> >>>                                     MemoryRegionSection *section,
> >>>                                     hwaddr first,
> >>> @@ -137,8 +156,43 @@ static int vhost_sync_dirty_bitmap(struct vhost_dev 
> >>> *dev,
> >>>              continue;
> >>>          }
> >>>
> >>> -        vhost_dev_sync_region(dev, section, start_addr, end_addr, 
> >>> vq->used_phys,
> >>> -                              range_get_last(vq->used_phys, 
> >>> vq->used_size));
> >>> +        if (vhost_dev_has_iommu(dev)) {
> >>> +            IOMMUTLBEntry iotlb;
> >>> +            hwaddr used_phys = vq->used_phys, used_size = vq->used_size;
> >>> +            hwaddr phys, s;
> >>> +
> >>> +            while (used_size) {
> >>> +                rcu_read_lock();
> >>> +                iotlb = address_space_get_iotlb_entry(dev->vdev->dma_as,
> >>> +                                                      used_phys,
> >>> +                                                      true, 
> >>> MEMTXATTRS_UNSPECIFIED);
> >>> +                rcu_read_unlock();
> >>> +
> >>> +                if (!iotlb.target_as) {
> >>> +                    qemu_log_mask(LOG_GUEST_ERROR, "translation "
> >>> +                                  "failure for used_phys %"PRIx64"\n", 
> >>> used_phys);
> >> looks weird to see translation of "used_phys" whereas it is an iova. At
> >> least I would reword the msg

Let me tweak this in the next version.

> >>> +                    return -EINVAL;
> >>> +                }
> >>> +
> >>> +                phys = iotlb.translated_addr + (used_phys & 
> >>> iotlb.addr_mask);
> >> you may use a local variable storing this offset =
> >>
> >> used_phys & iotlb.addr_mask

Ok.

> >>
> >>> +
> >>> +                /* Distance from start of used ring until last byte of
> >>> +                   IOMMU page */
> >> you can avoid checkpatch warnings here
> >>> +                s = iotlb.addr_mask - (used_phys & iotlb.addr_mask);
> >>> +                /* Size of used ring, or of the part of it until end
> >>> +                   of IOMMU page */
> >> and here

Will fix.

> >>
> >> I would suggest to rewrite this into
> >> s =iotlb.addr_mask - (used_phys & iotlb.addr_mask) + 1
> >> s = MIN(s, used_size);
> > This does not work - if iotlb.addr_mask - (used_phys & iotlb.addr_mask)
> > is all-ones then + 1 gives you 0 and MIN gives you 0.
> > Theoretical but worth being safe here IMHO.
> Ah OK, I should have read your previous discussion more thoroughly ...
> Maybe just add a short comment then to justify the gym below and avoid
> tempting sbdy else to rewrite it in a more common but wrong way.

That's fine.

Thanks

>
> Thanks
>
> Eric
> >
> >
> >>> +                s = MIN(s, used_size - 1) + 1;
> >>> +
> >>> +                vhost_dev_sync_region(dev, section, start_addr, 
> >>> end_addr, phys,
> >>> +                                      range_get_last(phys, s));
> >>> +                used_size -= s;
> >>> +                used_phys += s;
> >>> +            }
> >>> +        } else {
> >>> +            vhost_dev_sync_region(dev, section, start_addr,
> >>> +                                  end_addr, vq->used_phys,
> >>> +                                  range_get_last(vq->used_phys, 
> >>> vq->used_size));
> >>> +        }
> >>>      }
> >>>      return 0;
> >>>  }
> >>> @@ -306,24 +360,6 @@ static inline void vhost_dev_log_resize(struct 
> >>> vhost_dev *dev, uint64_t size)
> >>>      dev->log_size = size;
> >>>  }
> >>>
> >>> -static bool vhost_dev_has_iommu(struct vhost_dev *dev)
> >>> -{
> >>> -    VirtIODevice *vdev = dev->vdev;
> >>> -
> >>> -    /*
> >>> -     * For vhost, VIRTIO_F_IOMMU_PLATFORM means the backend support
> >>> -     * incremental memory mapping API via IOTLB API. For platform that
> >>> -     * does not have IOMMU, there's no need to enable this feature
> >>> -     * which may cause unnecessary IOTLB miss/update transactions.
> >>> -     */
> >>> -    if (vdev) {
> >>> -        return virtio_bus_device_iommu_enabled(vdev) &&
> >>> -            virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM);
> >>> -    } else {
> >>> -        return false;
> >>> -    }
> >>> -}
> >>> -
> >>>  static void *vhost_memory_map(struct vhost_dev *dev, hwaddr addr,
> >>>                                hwaddr *plen, bool is_write)
> >>>  {
> >> Besides,
> >>
> >> Tested-by: Eric Auger <eric.auger@redhat.com>
> >>
> >> Eric
>