On Thu, Aug 02, 2018 at 07:21:22PM +0200, Maxime Coquelin wrote:
> IOTLB entries contain the host virtual address of the guest
> pages. When receiving a new VHOST_USER_SET_MEM_TABLE request,
> the previous regions get unmapped, so the IOTLB entries, if any,
> will be invalid. It does cause the vhost-user process to
> segfault.
>
> This patch introduces a new function to flush the IOTLB cache,
> and call it as soon as the backend handles a VHOST_USER_SET_MEM
> request.
>
> Fixes: 69c90e98f483 ("vhost: enable IOMMU support")
> Cc: sta...@dpdk.org
>
> Signed-off-by: Maxime Coquelin <maxime.coque...@redhat.com>
> ---
> Changes since v1:
> - Fix indentation (Stephen)
> - Fix double iotlb-lock lock
>
> lib/librte_vhost/iotlb.c | 10 ++++++++--
> lib/librte_vhost/iotlb.h | 2 +-
> lib/librte_vhost/vhost_user.c | 5 +++++
> 3 files changed, 14 insertions(+), 3 deletions(-)
>
> diff --git a/lib/librte_vhost/iotlb.c b/lib/librte_vhost/iotlb.c
> index c11ebcaac..c6354fef7 100644
> --- a/lib/librte_vhost/iotlb.c
> +++ b/lib/librte_vhost/iotlb.c
> @@ -303,6 +303,13 @@ vhost_user_iotlb_cache_find(struct vhost_virtqueue *vq,
> uint64_t iova,
> return vva;
> }
>
> +void
> +vhost_user_iotlb_flush_all(struct vhost_virtqueue *vq)
> +{
> + vhost_user_iotlb_cache_remove_all(vq);
> + vhost_user_iotlb_pending_remove_all(vq);
> +}
> +
> int
> vhost_user_iotlb_init(struct virtio_net *dev, int vq_index)
> {
> @@ -315,8 +322,7 @@ vhost_user_iotlb_init(struct virtio_net *dev, int
> vq_index)
> * The cache has already been initialized,
> * just drop all cached and pending entries.
> */
> - vhost_user_iotlb_cache_remove_all(vq);
> - vhost_user_iotlb_pending_remove_all(vq);
> + vhost_user_iotlb_flush_all(vq);
> }
>
> #ifdef RTE_LIBRTE_VHOST_NUMA
> diff --git a/lib/librte_vhost/iotlb.h b/lib/librte_vhost/iotlb.h
> index e7083e37b..60b9e4c57 100644
> --- a/lib/librte_vhost/iotlb.h
> +++ b/lib/librte_vhost/iotlb.h
> @@ -73,7 +73,7 @@ void vhost_user_iotlb_pending_insert(struct vhost_virtqueue
> *vq, uint64_t iova,
> uint8_t perm);
> void vhost_user_iotlb_pending_remove(struct vhost_virtqueue *vq, uint64_t
> iova,
> uint64_t size, uint8_t perm);
> -
> +void vhost_user_iotlb_flush_all(struct vhost_virtqueue *vq);
> int vhost_user_iotlb_init(struct virtio_net *dev, int vq_index);
>
> #endif /* _VHOST_IOTLB_H_ */
> diff --git a/lib/librte_vhost/vhost_user.c b/lib/librte_vhost/vhost_user.c
> index dc53ff712..a2d4c9ffc 100644
> --- a/lib/librte_vhost/vhost_user.c
> +++ b/lib/librte_vhost/vhost_user.c
> @@ -813,6 +813,11 @@ vhost_user_set_mem_table(struct virtio_net **pdev,
> struct VhostUserMsg *pmsg)
> dev->mem = NULL;
> }
>
> + /* Flush IOTLB cache as previous HVAs are now invalid */
> + if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
> + for (i = 0; i < dev->nr_vring; i++)
> + vhost_user_iotlb_flush_all(dev->virtqueue[i]);
Why is the pending list also flushed?
Thanks,
Tiwei
