This looks like a sensible way to do this.
Yet something to improve:
On Tue, Jun 18, 2024 at 04:44:56PM +0200, Jiri Pirko wrote:
> From: Jiri Pirko <j...@nvidia.com>
>
> Add support for Byte Queue Limits (BQL).
>
> Tested on qemu emulated virtio_net device with 1, 2 and 4 queues.
> Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is
> running in background. Netperf TCP_RR results:
>
> NOBQL FQC 1q: 159.56 159.33 158.50 154.31 agv: 157.925
> NOBQL FQC 2q: 184.64 184.96 174.73 174.15 agv: 179.62
> NOBQL FQC 4q: 994.46 441.96 416.50 499.56 agv: 588.12
> NOBQL PFF 1q: 148.68 148.92 145.95 149.48 agv: 148.2575
> NOBQL PFF 2q: 171.86 171.20 170.42 169.42 agv: 170.725
> NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99 agv: 2159.7875
> BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57 agv: 1282.4375
> BQL FQC 2q: 768.30 817.72 864.43 974.40 agv: 856.2125
> BQL FQC 4q: 945.66 942.68 878.51 822.82 agv: 897.4175
> BQL PFF 1q: 149.69 151.49 149.40 147.47 agv: 149.5125
> BQL PFF 2q: 2059.32 798.74 1844.12 381.80 agv: 1270.995
> BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16 agv: 6119.1875
>
> Signed-off-by: Jiri Pirko <j...@nvidia.com>
> ---
> v2->v3:
> - fixed the switch from/to orphan mode while skbs are yet to be
> completed by using the second least significant bit in virtqueue
> token pointer to indicate skb is orphan. Don't account orphan
> skbs in completion.
> - reorganized parallel skb/xdp free stats accounting to napi/others.
> - fixed kick condition check in orphan mode
> v1->v2:
> - moved netdev_tx_completed_queue() call into __free_old_xmit(),
> propagate use_napi flag to __free_old_xmit() and only call
> netdev_tx_completed_queue() in case it is true
> - added forgotten call to netdev_tx_reset_queue()
> - fixed stats for xdp packets
> - fixed bql accounting when __free_old_xmit() is called from xdp path
> - handle the !use_napi case in start_xmit() kick section
> ---
> drivers/net/virtio_net.c | 81 ++++++++++++++++++++++++++++------------
> 1 file changed, 57 insertions(+), 24 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 61a57d134544..9f9b86874173 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -47,7 +47,8 @@ module_param(napi_tx, bool, 0644);
> #define VIRTIO_XDP_TX BIT(0)
> #define VIRTIO_XDP_REDIR BIT(1)
>
> -#define VIRTIO_XDP_FLAG BIT(0)
> +#define VIRTIO_XDP_FLAG BIT(0)
> +#define VIRTIO_ORPHAN_FLAG BIT(1)
>
> /* RX packet size EWMA. The average packet size is used to determine the
> packet
> * buffer size when refilling RX rings. As the entire RX ring may be refilled
> @@ -85,6 +86,8 @@ struct virtnet_stat_desc {
> struct virtnet_sq_free_stats {
> u64 packets;
> u64 bytes;
> + u64 napi_packets;
> + u64 napi_bytes;
> };
>
> struct virtnet_sq_stats {
> @@ -506,29 +509,50 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
> return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
> }
>
> -static void __free_old_xmit(struct send_queue *sq, bool in_napi,
> - struct virtnet_sq_free_stats *stats)
> +static bool is_orphan_skb(void *ptr)
> +{
> + return (unsigned long)ptr & VIRTIO_ORPHAN_FLAG;
> +}
> +
> +static void *skb_to_ptr(struct sk_buff *skb, bool orphan)
> +{
> + return (void *)((unsigned long)skb | (orphan ? VIRTIO_ORPHAN_FLAG : 0));
> +}
> +
> +static struct sk_buff *ptr_to_skb(void *ptr)
> +{
> + return (struct sk_buff *)((unsigned long)ptr & ~VIRTIO_ORPHAN_FLAG);
> +}
> +
> +static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
> + bool in_napi, struct virtnet_sq_free_stats *stats)
> {
> unsigned int len;
> void *ptr;
>
> while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
> - ++stats->packets;
> -
> if (!is_xdp_frame(ptr)) {
> - struct sk_buff *skb = ptr;
> + struct sk_buff *skb = ptr_to_skb(ptr);
>
> pr_debug("Sent skb %p\n", skb);
>
> - stats->bytes += skb->len;
> + if (is_orphan_skb(ptr)) {
> + stats->packets++;
> + stats->bytes += skb->len;
> + } else {
> + stats->napi_packets++;
> + stats->napi_bytes += skb->len;
> + }
> napi_consume_skb(skb, in_napi);
> } else {
> struct xdp_frame *frame = ptr_to_xdp(ptr);
>
> + stats->packets++;
> stats->bytes += xdp_get_frame_len(frame);
> xdp_return_frame(frame);
> }
> }
> + netdev_tx_completed_queue(txq, stats->napi_packets, stats->napi_bytes);
Are you sure it's right? You are completing larger and larger
number of bytes and packets each time.
For example as won't this eventually trigger this inside dql_completed:
BUG_ON(count > num_queued - dql->num_completed);
?
If I am right the perf testing has to be redone with this fixed ...
> }
>
> /* Converting between virtqueue no. and kernel tx/rx queue no.
> @@ -955,21 +979,22 @@ static void virtnet_rq_unmap_free_buf(struct virtqueue
> *vq, void *buf)
> virtnet_rq_free_buf(vi, rq, buf);
> }
>
> -static void free_old_xmit(struct send_queue *sq, bool in_napi)
> +static void free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
> + bool in_napi)
> {
> struct virtnet_sq_free_stats stats = {0};
>
> - __free_old_xmit(sq, in_napi, &stats);
> + __free_old_xmit(sq, txq, in_napi, &stats);
>
> /* Avoid overhead when no packets have been processed
> * happens when called speculatively from start_xmit.
> */
> - if (!stats.packets)
> + if (!stats.packets && !stats.napi_packets)
> return;
>
> u64_stats_update_begin(&sq->stats.syncp);
> - u64_stats_add(&sq->stats.bytes, stats.bytes);
> - u64_stats_add(&sq->stats.packets, stats.packets);
> + u64_stats_add(&sq->stats.bytes, stats.bytes + stats.napi_bytes);
> + u64_stats_add(&sq->stats.packets, stats.packets + stats.napi_packets);
> u64_stats_update_end(&sq->stats.syncp);
> }
>
> @@ -1003,7 +1028,9 @@ static void check_sq_full_and_disable(struct
> virtnet_info *vi,
> * early means 16 slots are typically wasted.
> */
> if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
> - netif_stop_subqueue(dev, qnum);
> + struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
> +
> + netif_tx_stop_queue(txq);
> u64_stats_update_begin(&sq->stats.syncp);
> u64_stats_inc(&sq->stats.stop);
> u64_stats_update_end(&sq->stats.syncp);
> @@ -1012,7 +1039,7 @@ static void check_sq_full_and_disable(struct
> virtnet_info *vi,
> virtqueue_napi_schedule(&sq->napi, sq->vq);
> } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
> /* More just got used, free them then recheck. */
> - free_old_xmit(sq, false);
> + free_old_xmit(sq, txq, false);
> if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
> netif_start_subqueue(dev, qnum);
> u64_stats_update_begin(&sq->stats.syncp);
> @@ -1138,7 +1165,8 @@ static int virtnet_xdp_xmit(struct net_device *dev,
> }
>
> /* Free up any pending old buffers before queueing new ones. */
> - __free_old_xmit(sq, false, &stats);
> + __free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq),
> + false, &stats);
>
> for (i = 0; i < n; i++) {
> struct xdp_frame *xdpf = frames[i];
> @@ -2313,7 +2341,7 @@ static void virtnet_poll_cleantx(struct receive_queue
> *rq)
>
> do {
> virtqueue_disable_cb(sq->vq);
> - free_old_xmit(sq, true);
> + free_old_xmit(sq, txq, true);
> } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
>
> if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
> @@ -2412,6 +2440,7 @@ static int virtnet_enable_queue_pair(struct
> virtnet_info *vi, int qp_index)
> goto err_xdp_reg_mem_model;
>
> virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
> + netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, qp_index));
> virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
>
> return 0;
> @@ -2471,7 +2500,7 @@ static int virtnet_poll_tx(struct napi_struct *napi,
> int budget)
> txq = netdev_get_tx_queue(vi->dev, index);
> __netif_tx_lock(txq, raw_smp_processor_id());
> virtqueue_disable_cb(sq->vq);
> - free_old_xmit(sq, true);
> + free_old_xmit(sq, txq, true);
>
> if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
> if (netif_tx_queue_stopped(txq)) {
> @@ -2505,7 +2534,7 @@ static int virtnet_poll_tx(struct napi_struct *napi,
> int budget)
> return 0;
> }
>
> -static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
> +static int xmit_skb(struct send_queue *sq, struct sk_buff *skb, bool orphan)
> {
> struct virtio_net_hdr_mrg_rxbuf *hdr;
> const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
> @@ -2549,7 +2578,8 @@ static int xmit_skb(struct send_queue *sq, struct
> sk_buff *skb)
> return num_sg;
> num_sg++;
> }
> - return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
> + return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg,
> + skb_to_ptr(skb, orphan), GFP_ATOMIC);
> }
>
> static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> @@ -2559,24 +2589,25 @@ static netdev_tx_t start_xmit(struct sk_buff *skb,
> struct net_device *dev)
> struct send_queue *sq = &vi->sq[qnum];
> int err;
> struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
> - bool kick = !netdev_xmit_more();
> + bool xmit_more = netdev_xmit_more();
> bool use_napi = sq->napi.weight;
> + bool kick;
>
> /* Free up any pending old buffers before queueing new ones. */
> do {
> if (use_napi)
> virtqueue_disable_cb(sq->vq);
>
> - free_old_xmit(sq, false);
> + free_old_xmit(sq, txq, false);
>
> - } while (use_napi && kick &&
> + } while (use_napi && !xmit_more &&
> unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
>
> /* timestamp packet in software */
> skb_tx_timestamp(skb);
>
> /* Try to transmit */
> - err = xmit_skb(sq, skb);
> + err = xmit_skb(sq, skb, !use_napi);
>
> /* This should not happen! */
> if (unlikely(err)) {
> @@ -2598,7 +2629,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb,
> struct net_device *dev)
>
> check_sq_full_and_disable(vi, dev, sq);
>
> - if (kick || netif_xmit_stopped(txq)) {
> + kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
> + !xmit_more || netif_xmit_stopped(txq);
> + if (kick) {
> if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
> {
> u64_stats_update_begin(&sq->stats.syncp);
> u64_stats_inc(&sq->stats.kicks);
> --
> 2.45.1
