Vhost used to compare the dma_as against the address_space_memory to detect whether the IOMMU is enabled or not. This might not work well since the virito-bus may call get_dma_as if VIRTIO_F_IOMMU_PLATFORM is set without an actual IOMMU enabled when device is plugged. In the case of PCI where pci_get_address_space() is used, the bus master as is returned. So vhost actually tries to enable device IOTLB even if the IOMMU is not enabled. This will lead a lots of unnecessary transactions between vhost and Qemu and will introduce a huge drop of the performance.
For PCI, an ideal approach is to use pci_device_iommu_address_space() just for get_dma_as. But Qemu may choose to initialize the IOMMU after the virtio-pci which lead a wrong address space is returned during device plugged. So this patch switch to use transport specific way via iommu_enabled() to detect the IOMMU during vhost start. In this case, we are fine since we know the IOMMU is initialized correctly. Signed-off-by: Jason Wang <jasow...@redhat.com> --- hw/virtio/vhost.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c index 7b7bde7657..d2097b7423 100644 --- a/hw/virtio/vhost.c +++ b/hw/virtio/vhost.c @@ -286,7 +286,7 @@ static int vhost_dev_has_iommu(struct vhost_dev *dev) * does not have IOMMU, there's no need to enable this feature * which may cause unnecessary IOTLB miss/update trnasactions. */ - return vdev->dma_as != &address_space_memory && + return virtio_bus_device_iommu_enabled(vdev) && virtio_host_has_feature(vdev, VIRTIO_F_IOMMU_PLATFORM); } -- 2.25.1