On Tue, 14 Nov 2023 11:26:42 +0800, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> On Thu, Nov 9, 2023 at 7:06 PM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, 9 Nov 2023 14:37:38 +0800, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > On Tue, Nov 7, 2023 at 11:12 AM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > If the xsk is enabling, the xsk tx will share the send queue.
> > > > But the xsk requires that the send queue use the premapped mode.
> > > > So the send queue must support premapped mode.
> > > >
> > > > Signed-off-by: Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx>
> > > > ---
> > > > drivers/net/virtio/main.c | 163 ++++++++++++++++++++++++++++----
> > > > drivers/net/virtio/virtio_net.h | 16 ++++
> > > > 2 files changed, 163 insertions(+), 16 deletions(-)
> > > >
> > > > diff --git a/drivers/net/virtio/main.c b/drivers/net/virtio/main.c
> > > > index 16e75c08639e..f052db459156 100644
> > > > --- a/drivers/net/virtio/main.c
> > > > +++ b/drivers/net/virtio/main.c
> > > > @@ -46,6 +46,7 @@ module_param(napi_tx, bool, 0644);
> > > > #define VIRTIO_XDP_REDIR BIT(1)
> > > >
> > > > #define VIRTIO_XDP_FLAG BIT(0)
> > > > +#define VIRTIO_XMIT_DATA_MASK (VIRTIO_XDP_FLAG)
> > > >
> > > > #define VIRTNET_DRIVER_VERSION "1.0.0"
> > > >
> > > > @@ -167,6 +168,29 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
> > > > return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
> > > > }
> > > >
> > > > +static inline void *virtnet_sq_unmap(struct virtnet_sq *sq, void *data)
> > > > +{
> > > > + struct virtnet_sq_dma *next, *head;
> > > > +
> > > > + head = (void *)((unsigned long)data & ~VIRTIO_XMIT_DATA_MASK);
> > > > +
> > > > + data = head->data;
> > > > +
> > > > + while (head) {
> > > > + virtqueue_dma_unmap_single_attrs(sq->vq, head->addr, head->len,
> > > > + DMA_TO_DEVICE, 0);
> > > > +
> > > > + next = head->next;
> > > > +
> > > > + head->next = sq->dmainfo.free;
> > > > + sq->dmainfo.free = head;
> > > > +
> > > > + head = next;
> > > > + }
> > > > +
> > > > + return data;
> > > > +}
> > > > +
> > > > static void __free_old_xmit(struct virtnet_sq *sq, bool in_napi,
> > > > u64 *bytes, u64 *packets)
> > > > {
> > > > @@ -175,14 +199,24 @@ static void __free_old_xmit(struct virtnet_sq *sq, bool in_napi,
> > > >
> > > > while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
> > > > if (!is_xdp_frame(ptr)) {
> > > > - struct sk_buff *skb = ptr;
> > > > + struct sk_buff *skb;
> > > > +
> > > > + if (sq->do_dma)
> > > > + ptr = virtnet_sq_unmap(sq, ptr);
> > > > +
> > > > + skb = ptr;
> > > >
> > > > pr_debug("Sent skb %p\n", skb);
> > > >
> > > > *bytes += skb->len;
> > > > napi_consume_skb(skb, in_napi);
> > > > } else {
> > > > - struct xdp_frame *frame = ptr_to_xdp(ptr);
> > > > + struct xdp_frame *frame;
> > > > +
> > > > + if (sq->do_dma)
> > > > + ptr = virtnet_sq_unmap(sq, ptr);
> > > > +
> > > > + frame = ptr_to_xdp(ptr);
> > > >
> > > > *bytes += xdp_get_frame_len(frame);
> > > > xdp_return_frame(frame);
> > > > @@ -567,22 +601,104 @@ static void *virtnet_rq_alloc(struct virtnet_rq *rq, u32 size, gfp_t gfp)
> > > > return buf;
> > > > }
> > > >
> > > > -static void virtnet_rq_set_premapped(struct virtnet_info *vi)
> > > > +static int virtnet_sq_set_premapped(struct virtnet_sq *sq)
> > > > {
> > > > - int i;
> > > > + struct virtnet_sq_dma *d;
> > > > + int err, size, i;
> > > >
> > > > - /* disable for big mode */
> > > > - if (!vi->mergeable_rx_bufs && vi->big_packets)
> > > > - return;
> > > > + size = virtqueue_get_vring_size(sq->vq);
> > > > +
> > > > + size += MAX_SKB_FRAGS + 2;
> > >
> > > Btw, the dmainfo seems per sg? If I'm correct, how can vq_size +
> > > MAX_SKB_FRAGS + 2 work?
> >
> >
> > We may alloc dmainfo items when the vq is full. So I prepare more dmainfo items.
> >
> >
> > >
> > > > +
> > > > + sq->dmainfo.head = kcalloc(size, sizeof(*sq->dmainfo.head), GFP_KERNEL);
> > > > + if (!sq->dmainfo.head)
> > > > + return -ENOMEM;
> > > > +
> > > > + err = virtqueue_set_dma_premapped(sq->vq);
> > > > + if (err) {
> > > > + kfree(sq->dmainfo.head);
> > > > + return err;
> > > > + }
> > >
> > > Allocating after set_dma_premapped() seems easier.
> >
> > Yes. But, we donot has the way to disable premapped mode.
> >
> > That is my mistake. :)
> >
> >
> > >
> > > Btw, is there a benchmark of TX PPS just for this patch to demonstrate
> > > the impact of the performance?
> >
> > We will have that.
> >
> >
> > >
> > > > +
> > > > + sq->dmainfo.free = NULL;
> > > > +
> > > > + sq->do_dma = true;
> > > > +
> > > > + for (i = 0; i < size; ++i) {
> > > > + d = &sq->dmainfo.head[i];
> > > > +
> > > > + d->next = sq->dmainfo.free;
> > > > + sq->dmainfo.free = d;
> > > > + }
> > > > +
> > > > + return 0;
> > > > +}
> > > > +
> > > > +static void virtnet_set_premapped(struct virtnet_info *vi)
> > > > +{
> > > > + int i;
> > > >
> > > > for (i = 0; i < vi->max_queue_pairs; i++) {
> > > > - if (virtqueue_set_dma_premapped(vi->rq[i].vq))
> > > > - continue;
> > > > + virtnet_sq_set_premapped(&vi->sq[i]);
> > > >
> > > > - vi->rq[i].do_dma = true;
> > > > + /* disable for big mode */
> > > > + if (vi->mergeable_rx_bufs || !vi->big_packets) {
> > > > + if (!virtqueue_set_dma_premapped(vi->rq[i].vq))
> > > > + vi->rq[i].do_dma = true;
> > >
> > > How about sticking a virtnet_rq_set_premapped() and calling it here?
> > >
> > > It seems more clean.
> >
> > OK.
> >
> >
> > >
> > > Btw, the big mode support for pre mapping is still worthwhile
> > > regardless whether or not XDP is supported. It has a page pool so we
> > > can avoid redundant DMA map/unmap there.
> >
> > Yes.
> >
> > I post other patch set to do this.
> >
> >
> > >
> > > > + }
> > > > }
> > > > }
> > > >
> > > > +static struct virtnet_sq_dma *virtnet_sq_map_sg(struct virtnet_sq *sq, int nents, void *data)
> > > > +{
> > > > + struct virtnet_sq_dma *d, *head;
> > > > + struct scatterlist *sg;
> > > > + int i;
> > > > +
> > > > + head = NULL;
> > > > +
> > > > + for_each_sg(sq->sg, sg, nents, i) {
> > > > + sg->dma_address = virtqueue_dma_map_single_attrs(sq->vq, sg_virt(sg),
> > > > + sg->length,
> > > > + DMA_TO_DEVICE, 0);
> > > > + if (virtqueue_dma_mapping_error(sq->vq, sg->dma_address))
> > > > + goto err;
> > > > +
> > > > + d = sq->dmainfo.free;
> > > > + sq->dmainfo.free = d->next;
> > > > +
> > > > + d->addr = sg->dma_address;
> > > > + d->len = sg->length;
> > > > +
> > > > + d->next = head;
> > > > + head = d;
> > > > + }
> > > > +
> > > > + head->data = data;
> > > > +
> > > > + return (void *)((unsigned long)head | ((unsigned long)data & VIRTIO_XMIT_DATA_MASK));
> > >
> > > So head contains a pointer to data, any reason we still need to pack a
> > > data pointer here?
> >
> > Maybe you are right. We can skip this.
> >
> >
> > >
> > >
> > > > +err:
> > > > + virtnet_sq_unmap(sq, head);
> > > > + return NULL;
> > > > +}
> > > > +
> > > > +static int virtnet_add_outbuf(struct virtnet_sq *sq, u32 num, void *data)
> > > > +{
> > > > + int ret;
> > > > +
> > > > + if (sq->do_dma) {
> > > > + data = virtnet_sq_map_sg(sq, num, data);
> > > > + if (!data)
> > > > + return -ENOMEM;
> > > > + }
> > > > +
> > > > + ret = virtqueue_add_outbuf(sq->vq, sq->sg, num, data, GFP_ATOMIC);
> > > > + if (ret && sq->do_dma)
> > > > + virtnet_sq_unmap(sq, data);
> > > > +
> > > > + return ret;
> > > > +}
> > > > +
> > > > static void free_old_xmit(struct virtnet_sq *sq, bool in_napi)
> > > > {
> > > > u64 bytes, packets = 0;
> > > > @@ -686,8 +802,7 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
> > > > skb_frag_size(frag), skb_frag_off(frag));
> > > > }
> > > >
> > > > - err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
> > > > - xdp_to_ptr(xdpf), GFP_ATOMIC);
> > > > + err = virtnet_add_outbuf(sq, nr_frags + 1, xdp_to_ptr(xdpf));
> > > > if (unlikely(err))
> > > > return -ENOSPC; /* Caller handle free/refcnt */
> > > >
> > > > @@ -2126,7 +2241,8 @@ static int xmit_skb(struct virtnet_sq *sq, struct sk_buff *skb)
> > > > return num_sg;
> > > > num_sg++;
> > > > }
> > > > - return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
> > > > +
> > > > + return virtnet_add_outbuf(sq, num_sg, skb);
> > > > }
> > > >
> > > > static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> > > > @@ -3818,6 +3934,8 @@ static void virtnet_free_queues(struct virtnet_info *vi)
> > > > for (i = 0; i < vi->max_queue_pairs; i++) {
> > > > __netif_napi_del(&vi->rq[i].napi);
> > > > __netif_napi_del(&vi->sq[i].napi);
> > > > +
> > > > + kfree(vi->sq[i].dmainfo.head);
> > > > }
> > > >
> > > > /* We called __netif_napi_del(),
> > > > @@ -3866,10 +3984,23 @@ static void free_receive_page_frags(struct virtnet_info *vi)
> > > >
> > > > static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
> > > > {
> > > > - if (!is_xdp_frame(buf))
> > > > + struct virtnet_info *vi = vq->vdev->priv;
> > > > + struct virtnet_sq *sq;
> > > > + int i = vq2rxq(vq);
> > > > +
> > > > + sq = &vi->sq[i];
> > > > +
> > > > + if (!is_xdp_frame(buf)) {
> > > > + if (sq->do_dma)
> > > > + buf = virtnet_sq_unmap(sq, buf);
> > > > +
> > > > dev_kfree_skb(buf);
> > > > - else
> > > > + } else {
> > > > + if (sq->do_dma)
> > > > + buf = virtnet_sq_unmap(sq, buf);
> > > > +
> > > > xdp_return_frame(ptr_to_xdp(buf));
> > > > + }
> > > > }
> > > >
> > > > static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf)
> > > > @@ -4075,7 +4206,7 @@ static int init_vqs(struct virtnet_info *vi)
> > > > if (ret)
> > > > goto err_free;
> > > >
> > > > - virtnet_rq_set_premapped(vi);
> > > > + virtnet_set_premapped(vi);
> > > >
> > > > cpus_read_lock();
> > > > virtnet_set_affinity(vi);
> > > > diff --git a/drivers/net/virtio/virtio_net.h b/drivers/net/virtio/virtio_net.h
> > > > index d814341d9f97..ce806afb6d64 100644
> > > > --- a/drivers/net/virtio/virtio_net.h
> > > > +++ b/drivers/net/virtio/virtio_net.h
> > > > @@ -48,6 +48,18 @@ struct virtnet_rq_dma {
> > > > u16 need_sync;
> > > > };
> > > >
> > > > +struct virtnet_sq_dma {
> > > > + struct virtnet_sq_dma *next;
> > > > + dma_addr_t addr;
> > > > + u32 len;
> > > > + void *data;
> > >
> > > I think we need to seek a way to reuse what has been stored by virtio
> > > core. It should be much more efficient.
> >
> >
> > Yes.
> >
> > But that is for net-next branch.
> >
> > Can we do that as a fix after that is merged to 6.8?
>
> We still have time. I would like to do it from the start.
I want to finish the job including new AF_XDP ZC feature.
Because that this must wait the merge window.
Base on that, the optimizing work can be done everytime.
If we work from the new virtio prepare, that can be merged to 6.8.
And the AF_XDP zc must wait 6.9. right?
Thanks
>
> Thanks
>
> >
> > Thanks.
> >
> >
> > >
> > > Thanks
> > >
> > > > +};
> > > > +
> > > > +struct virtnet_sq_dma_head {
> > > > + struct virtnet_sq_dma *free;
> > > > + struct virtnet_sq_dma *head;
> > > > +};
> > > > +
> > > > /* Internal representation of a send virtqueue */
> > > > struct virtnet_sq {
> > > > /* Virtqueue associated with this virtnet_sq */
> > > > @@ -67,6 +79,10 @@ struct virtnet_sq {
> > > >
> > > > /* Record whether sq is in reset state. */
> > > > bool reset;
> > > > +
> > > > + bool do_dma;
> > > > +
> > > > + struct virtnet_sq_dma_head dmainfo;
> > > > };
> > > >
> > > > /* Internal representation of a receive virtqueue */
> > > > --
> > > > 2.32.0.3.g01195cf9f
> > > >
> > >
> > >
> >
>
