Currently, IOVAs are assumed to be GPA when the dirty page logging is performed.
Add support for guest IOVA_VA mode by translating addresses to GPA before doing the dirty page logging. Also, the guest_pages tracking mechanism used for zero_copy will not work either if IOVA_VA is enabled. In fact, enabling dequeue_zero_copy and IOMMU is dangerous altogether because IOTLB invalidation of guest pages can happen while the buffers are being used. So, ensure this cannot happen by preventing both features to be activated simultaneously. Further information: https://bugs.dpdk.org/show_bug.cgi?id=337 Fixes: 69c90e98f483 ("vhost: enable IOMMU support") Cc: maxime.coque...@redhat.com Cc: sta...@dpdk.org --- v5 changes: Rebase on top of dpdk-next-virtio v4 changes: Address review comments v3 changes: Address style warnings v2 changes: Also translate buffer addresses ---- Adrian Moreno (3): vhost: translate incoming log address to gpa vhost: convert buffer addresses to GPA for logging vhost: prevent zero copy mode if iommu is on lib/librte_vhost/rte_vhost.h | 2 +- lib/librte_vhost/socket.c | 8 ++++++ lib/librte_vhost/vdpa.c | 3 ++- lib/librte_vhost/vhost.c | 41 ++++++++++++++++++++++++++++ lib/librte_vhost/vhost.h | 51 +++++++++++++++++++++++++++++++++++ lib/librte_vhost/vhost_user.c | 43 ++++++++++++++++++++++++++++- lib/librte_vhost/virtio_net.c | 12 +++++---- 7 files changed, 152 insertions(+), 8 deletions(-) -- 2.21.0
