As discussed: http://lore.kernel.org/all/CACGkMEvq0No8QGC46U4mGsMtuD44fD_cfLcPaVmJ3rHYqRZxYg@xxxxxxxxxxxxxx If virtio is in pre-mapped mode, the driver should manage the DMA info itself. However, due to the indirect feature, the virtio-net driver may need to maintain an excessive amount of DMA info. Therefore, we only enable pre-mapped mode for the send queue (SQ) when af-xdp is bound to it. We have set a limit on the amount of DMA info we manage. If the kernel stack or xdp tx/redirect attempts to use more descriptors than this limit, virtnet_add_outbuf() will return an -ENOMEM error. However, AF-XDP can continue working. The last patch aims to demonstrate the logic of enabling pre-mapped mode when af-xdp is bound to the SQ, and we can discuss it further. Although the APIs are not used, this patch might belong in the final patch set (virtio-net supports AF_XDP). But I hope the first four patches can be merged first to reduce the number of patches in the final set. Thanks. Xuan Zhuo (5): virtio_ring: introduce vring_need_unmap_buffer virtio_ring: introduce dma map api for page virtio_ring: introduce virtqueue_dma_map_sg_attrs virtio_ring: virtqueue_set_dma_premapped() support to disable virtio_net: sq support premapped mode drivers/net/virtio_net.c | 212 ++++++++++++++++++++++++++++++++++- drivers/virtio/virtio_ring.c | 118 ++++++++++++++++--- include/linux/virtio.h | 12 +- 3 files changed, 319 insertions(+), 23 deletions(-) -- 2.32.0.3.g01195cf9f
