On 5/28/2015 11:17 PM, Ouyang, Changchun wrote:
> Vring enqueue need consider the 2 cases:
> 1. Vring descriptors chained together, the first one is for virtio header,
> the rest are for real
> data, virtio driver in Linux usually use this scheme;
> 2. Only one descriptor, virtio header and real data share one single
> descriptor, virtio-net pmd use
> such scheme;
>
> So does vring dequeue, it should not assume vring descriptor is chained or
> not chained, virtio in
> different Linux version has different behavior, e.g. fedora 20 use chained
> vring descriptor, while
> fedora 21 use one single vring descriptor for tx.
>
> Changes in v2
> - drop the uncompleted packet
> - refine code logic
>
> Signed-off-by: Changchun Ouyang <changchun.ouyang at intel.com>
> ---
> lib/librte_vhost/vhost_rxtx.c | 65
> +++++++++++++++++++++++++++++++++----------
> 1 file changed, 50 insertions(+), 15 deletions(-)
>
> diff --git a/lib/librte_vhost/vhost_rxtx.c b/lib/librte_vhost/vhost_rxtx.c
> index 4809d32..06ae2df 100644
> --- a/lib/librte_vhost/vhost_rxtx.c
> +++ b/lib/librte_vhost/vhost_rxtx.c
> @@ -59,7 +59,7 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
> struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, 0, 0, 0, 0, 0}, 0};
> uint64_t buff_addr = 0;
> uint64_t buff_hdr_addr = 0;
> - uint32_t head[MAX_PKT_BURST], packet_len = 0;
> + uint32_t head[MAX_PKT_BURST];
> uint32_t head_idx, packet_success = 0;
> uint16_t avail_idx, res_cur_idx;
> uint16_t res_base_idx, res_end_idx;
> @@ -113,6 +113,10 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
> rte_prefetch0(&vq->desc[head[packet_success]]);
>
> while (res_cur_idx != res_end_idx) {
> + uint32_t offset = 0;
> + uint32_t data_len, len_to_cpy;
> + uint8_t hdr = 0, uncompleted_pkt = 0;
> +
> /* Get descriptor from available ring */
> desc = &vq->desc[head[packet_success]];
>
> @@ -125,7 +129,6 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
>
> /* Copy virtio_hdr to packet and increment buffer address */
> buff_hdr_addr = buff_addr;
> - packet_len = rte_pktmbuf_data_len(buff) + vq->vhost_hlen;
>
> /*
> * If the descriptors are chained the header and data are
> @@ -136,28 +139,55 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
> desc = &vq->desc[desc->next];
> /* Buffer address translation. */
> buff_addr = gpa_to_vva(dev, desc->addr);
> - desc->len = rte_pktmbuf_data_len(buff);
Do we got confirm from virtio SPEC that it is OK to only update used->len?
> } else {
> buff_addr += vq->vhost_hlen;
> - desc->len = packet_len;
> + hdr = 1;
> }
>
> + data_len = rte_pktmbuf_data_len(buff);
> + len_to_cpy = RTE_MIN(data_len,
> + hdr ? desc->len - vq->vhost_hlen : desc->len);
> + while (len_to_cpy > 0) {
> + /* Copy mbuf data to buffer */
> + rte_memcpy((void *)(uintptr_t)buff_addr,
> + (const void *)(rte_pktmbuf_mtod(buff, const
> char *) + offset),
> + len_to_cpy);
> + PRINT_PACKET(dev, (uintptr_t)buff_addr,
> + len_to_cpy, 0);
> +
> + offset += len_to_cpy;
> +
> + if (offset == data_len)
> + break;
Ok, i see scatter gather case handling is in patch 5.
> +
> + if (desc->flags & VRING_DESC_F_NEXT) {
> + desc = &vq->desc[desc->next];
> + buff_addr = gpa_to_vva(dev, desc->addr);
> + len_to_cpy = RTE_MIN(data_len - offset,
> desc->len);
> + } else {
> + /* Room in vring buffer is not enough */
> + uncompleted_pkt = 1;
> + break;
> + }
> + };
> +
> /* Update used ring with desc information */
> vq->used->ring[res_cur_idx & (vq->size - 1)].id =
> head[packet_success];
> - vq->used->ring[res_cur_idx & (vq->size - 1)].len = packet_len;
>
> - /* Copy mbuf data to buffer */
> - /* FIXME for sg mbuf and the case that desc couldn't hold the
> mbuf data */
> - rte_memcpy((void *)(uintptr_t)buff_addr,
> - rte_pktmbuf_mtod(buff, const void *),
> - rte_pktmbuf_data_len(buff));
> - PRINT_PACKET(dev, (uintptr_t)buff_addr,
> - rte_pktmbuf_data_len(buff), 0);
> + /* Drop the packet if it is uncompleted */
> + if (unlikely(uncompleted_pkt == 1))
> + vq->used->ring[res_cur_idx & (vq->size - 1)].len = 0;
Here things become complicated with the previous lockless reserve.
What is the consequence when guest sees zero in used->len? At least, do
we check with virtio-net implementation?
> + else
> + vq->used->ring[res_cur_idx & (vq->size - 1)].len =
> + offset + vq->vhost_hlen;
Two questions here,
1. add virtio header len?
2. Why not use packet_len rather than offset?
>
> res_cur_idx++;
> packet_success++;
>
> + if (unlikely(uncompleted_pkt == 1))
> + continue;
> +
> rte_memcpy((void *)(uintptr_t)buff_hdr_addr,
> (const void *)&virtio_hdr, vq->vhost_hlen);
>
> @@ -589,7 +619,14 @@ rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t
> queue_id,
> desc = &vq->desc[head[entry_success]];
>
> /* Discard first buffer as it is the virtio header */
> - desc = &vq->desc[desc->next];
> + if (desc->flags & VRING_DESC_F_NEXT) {
> + desc = &vq->desc[desc->next];
> + vb_offset = 0;
> + vb_avail = desc->len;
> + } else {
> + vb_offset = vq->vhost_hlen;
> + vb_avail = desc->len - vb_offset;
> + }
>
> /* Buffer address translation. */
> vb_addr = gpa_to_vva(dev, desc->addr);
> @@ -608,8 +645,6 @@ rte_vhost_dequeue_burst(struct virtio_net *dev, uint16_t
> queue_id,
> vq->used->ring[used_idx].id = head[entry_success];
> vq->used->ring[used_idx].len = 0;
>
> - vb_offset = 0;
> - vb_avail = desc->len;
> /* Allocate an mbuf and populate the structure. */
> m = rte_pktmbuf_alloc(mbuf_pool);
> if (unlikely(m == NULL)) {
