Re: [PATCH net-next v2 16/21] virtio_net: xsk: rx: introduce add_recvbuf_xsk()

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Fri, Nov 10, 2023 at 10:38:04AM +0800, Xuan Zhuo wrote:
> On Thu, 9 Nov 2023 17:26:33 +0100, Maciej Fijalkowski <maciej.fijalkowski@intel.com> wrote:
> > On Thu, Nov 09, 2023 at 07:11:46PM +0800, Xuan Zhuo wrote:
> > > On Thu, 9 Nov 2023 03:12:27 -0500, "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > > On Tue, Nov 07, 2023 at 11:12:22AM +0800, Xuan Zhuo wrote:
> > > > > Implement the logic of filling rq with XSK buffers.
> > > > >
> > > > > Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
> > > > > ---
> > > > >  drivers/net/virtio/main.c       |  4 ++-
> > > > >  drivers/net/virtio/virtio_net.h |  5 ++++
> > > > >  drivers/net/virtio/xsk.c        | 49 ++++++++++++++++++++++++++++++++-
> > > > >  drivers/net/virtio/xsk.h        |  2 ++
> > > > >  4 files changed, 58 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/drivers/net/virtio/main.c b/drivers/net/virtio/main.c
> > > > > index 6210a6e37396..15943a22e17d 100644
> > > > > --- a/drivers/net/virtio/main.c
> > > > > +++ b/drivers/net/virtio/main.c
> > > > > @@ -1798,7 +1798,9 @@ static bool try_fill_recv(struct virtnet_info *vi, struct virtnet_rq *rq,
> > > > >  	bool oom;
> > > > >
> > > > >  	do {
> > > > > -		if (vi->mergeable_rx_bufs)
> > > > > +		if (rq->xsk.pool)
> > > > > +			err = virtnet_add_recvbuf_xsk(vi, rq, rq->xsk.pool, gfp);
> > > > > +		else if (vi->mergeable_rx_bufs)
> > > > >  			err = add_recvbuf_mergeable(vi, rq, gfp);
> > > > >  		else if (vi->big_packets)
> > > > >  			err = add_recvbuf_big(vi, rq, gfp);
> > > >
> > > > I'm not sure I understand. How does this handle mergeable flag still being set?
> > >
> > >
> > > You has the same question as Jason.
> > >
> > > So I think maybe I should put the handle into the
> > > add_recvbuf_mergeable and add_recvbuf_small.
> > >
> > > Let me think about this.
> > >
> > >
> > > >
> > > >
> > > > > diff --git a/drivers/net/virtio/virtio_net.h b/drivers/net/virtio/virtio_net.h
> > > > > index a13d6d301fdb..1242785e311e 100644
> > > > > --- a/drivers/net/virtio/virtio_net.h
> > > > > +++ b/drivers/net/virtio/virtio_net.h
> > > > > @@ -140,6 +140,11 @@ struct virtnet_rq {
> > > > >
> > > > >  		/* xdp rxq used by xsk */
> > > > >  		struct xdp_rxq_info xdp_rxq;
> > > > > +
> > > > > +		struct xdp_buff **xsk_buffs;
> > > > > +		u32 nxt_idx;
> > > > > +		u32 num;
> > > > > +		u32 size;
> > > > >  	} xsk;
> > > > >  };
> > > > >
> > > > > diff --git a/drivers/net/virtio/xsk.c b/drivers/net/virtio/xsk.c
> > > > > index ea5804ddd44e..e737c3353212 100644
> > > > > --- a/drivers/net/virtio/xsk.c
> > > > > +++ b/drivers/net/virtio/xsk.c
> > > > > @@ -38,6 +38,41 @@ static void virtnet_xsk_check_queue(struct virtnet_sq *sq)
> > > > >  		netif_stop_subqueue(dev, qnum);
> > > > >  }
> > > > >
> > > > > +int virtnet_add_recvbuf_xsk(struct virtnet_info *vi, struct virtnet_rq *rq,
> > > > > +			    struct xsk_buff_pool *pool, gfp_t gfp)
> > > > > +{
> > > > > +	struct xdp_buff **xsk_buffs;
> > > > > +	dma_addr_t addr;
> > > > > +	u32 len, i;
> > > > > +	int err = 0;
> > > > > +
> > > > > +	xsk_buffs = rq->xsk.xsk_buffs;
> > > > > +
> > > > > +	if (rq->xsk.nxt_idx >= rq->xsk.num) {
> > > > > +		rq->xsk.num = xsk_buff_alloc_batch(pool, xsk_buffs, rq->xsk.size);
> > > > > +		if (!rq->xsk.num)
> > > > > +			return -ENOMEM;
> > > > > +		rq->xsk.nxt_idx = 0;
> > > > > +	}
> > > >
> > > > Another manually rolled linked list implementation.
> > > > Please, don't.
> > >
> > >
> > > The array is for speedup.
> > >
> > > xsk_buff_alloc_batch will return many xsk_buff that will be more efficient than
> > > the xsk_buff_alloc.
> >
> > But your sg list just contains a single entry?
> > I think that you have to walk through the xsk_buffs array, retrieve dma
> > addrs from there and have sg list sized to the value
> > xsk_buff_alloc_batch() returned.
> >
> > I don't think your logic based on nxt_idx is needed. Please take a look
> > how other drivers use xsk_buff_alloc_batch().
> >
> > I don't see callsites of virtnet_add_recvbuf_xsk() though.
> 
> 
> virtnet_add_recvbuf_xsk is called by the above try_fill_recv()
> And the loop is in there.

Ah sorry I was looking for another patch to call it as it used to be in
v1.

> 
> Jason want to reuse the loop of the try_fill_recv().
> So in this function I just consume one item.
> 
> The nxt_idx is used to cross the try_fill_recv.
> 
> If we drop the nxt_idx. This patch will like this:
> 
> diff --git a/drivers/net/virtio/main.c b/drivers/net/virtio/main.c
> index 6210a6e37396..88bff83ad0d8 100644
> --- a/drivers/net/virtio/main.c
> +++ b/drivers/net/virtio/main.c
> @@ -1797,6 +1797,15 @@ static bool try_fill_recv(struct virtnet_info *vi, struct virtnet_rq *rq,
>  	int err;
>  	bool oom;
> 
> +	if (rq->xsk.pool) {
> +		err = virtnet_add_recvbuf_xsk(vi, rq, rq->xsk.pool, gfp);
> +		oom = err == -ENOMEM;
> +		if (err > 0)
> +			goto kick;
> +
> +		return err;
> +	}
> +
>  	do {
>  		if (vi->mergeable_rx_bufs)
>  			err = add_recvbuf_mergeable(vi, rq, gfp);
> @@ -1809,6 +1818,7 @@ static bool try_fill_recv(struct virtnet_info *vi, struct virtnet_rq *rq,
>  		if (err)
>  			break;
>  	} while (rq->vq->num_free);
> +kick:
>  	if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
>  		unsigned long flags;
> 
> diff --git a/drivers/net/virtio/virtio_net.h b/drivers/net/virtio/virtio_net.h
> index a13d6d301fdb..184866014a19 100644
> --- a/drivers/net/virtio/virtio_net.h
> +++ b/drivers/net/virtio/virtio_net.h
> @@ -140,6 +140,8 @@ struct virtnet_rq {
> 
>  		/* xdp rxq used by xsk */
>  		struct xdp_rxq_info xdp_rxq;
> +
> +		struct xdp_buff **xsk_buffs;
>  	} xsk;
>  };
> 
> diff --git a/drivers/net/virtio/xsk.c b/drivers/net/virtio/xsk.c
> index ea5804ddd44e..73c9323bffd3 100644
> --- a/drivers/net/virtio/xsk.c
> +++ b/drivers/net/virtio/xsk.c
> @@ -38,6 +38,46 @@ static void virtnet_xsk_check_queue(struct virtnet_sq *sq)
>  		netif_stop_subqueue(dev, qnum);
>  }
> 
> +int virtnet_add_recvbuf_xsk(struct virtnet_info *vi, struct virtnet_rq *rq,
> +			    struct xsk_buff_pool *pool, gfp_t gfp)
> +{
> +	struct xdp_buff **xsk_buffs;
> +	dma_addr_t addr;
> +	u32 len, i;
> +	int err = 0;
> +	int num;
> +
> +	xsk_buffs = rq->xsk.xsk_buffs;
> +
> +	num = xsk_buff_alloc_batch(pool, xsk_buffs, rq->vq->num_free);
> +	if (!num)
> +		return -ENOMEM;
> +
> +	for (i = 0; i < num; ++i) {
> +		/* use the part of XDP_PACKET_HEADROOM as the virtnet hdr space */
> +		addr = xsk_buff_xdp_get_dma(xsk_buffs[i]) - vi->hdr_len;
> +		len = xsk_pool_get_rx_frame_size(pool) + vi->hdr_len;

len can be pulled out of loop...

> +
> +		sg_init_table(rq->sg, 1);
> +		sg_fill_dma(rq->sg, addr, len);

... but when I first commented I did not understand why you were not
passing dma from xsk_buff_pool like this:

	sg_init_table(rq->sg, num);
	len = xsk_pool_get_rx_frame_size(pool) + vi->hdr_len;

	for (i = 0; i < num; ++i) {
		/* use the part of XDP_PACKET_HEADROOM as the virtnet hdr space */
		addr = xsk_buff_xdp_get_dma(xsk_buffs[i]) - vi->hdr_len;
		/* TODO: extend scatterlist size in receive_queue */
		sg_fill_dma(&rq->sg[i], addr, len);
	}

	err = virtqueue_add_inbuf(rq->vq, rq->sg, num, xsk_buffs, gfp);

and now I see that the problem is with 'data' argument above (or xsk_buffs
in this particular example).

Why do you need to pass xdp_buff to virtio_ring? You already have the
rq->xsk.xsk_buffs which you can use on rx side.

Can someone shed some light on it?

> +
> +		err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, xsk_buffs[i], gfp);
> +		if (err)
> +			goto err;
> +	}
> +
> +	return num;
> +
> +err:
> +	if (i)
> +		err = i;
> +
> +	for (; i < num; ++i)
> +		xsk_buff_free(xsk_buffs[i]);
> +
> +	return err;
> +}
> +
>  static int virtnet_xsk_xmit_one(struct virtnet_sq *sq,
>  				struct xsk_buff_pool *pool,
>  				struct xdp_desc *desc)
> @@ -213,7 +253,7 @@ static int virtnet_xsk_pool_enable(struct net_device *dev,
>  	struct virtnet_sq *sq;
>  	struct device *dma_dev;
>  	dma_addr_t hdr_dma;
> -	int err;
> +	int err, size;
> 
>  	/* In big_packets mode, xdp cannot work, so there is no need to
>  	 * initialize xsk of rq.
> @@ -249,6 +289,12 @@ static int virtnet_xsk_pool_enable(struct net_device *dev,
>  	if (!dma_dev)
>  		return -EPERM;
> 
> +	size = virtqueue_get_vring_size(rq->vq);
> +
> +	rq->xsk.xsk_buffs = kcalloc(size, sizeof(*rq->xsk.xsk_buffs), GFP_KERNEL);
> +	if (!rq->xsk.xsk_buffs)
> +		return -ENOMEM;
> +
>  	hdr_dma = dma_map_single(dma_dev, &xsk_hdr, vi->hdr_len, DMA_TO_DEVICE);
>  	if (dma_mapping_error(dma_dev, hdr_dma))
>  		return -ENOMEM;
> @@ -307,6 +353,8 @@ static int virtnet_xsk_pool_disable(struct net_device *dev, u16 qid)
> 
>  	dma_unmap_single(dma_dev, sq->xsk.hdr_dma_address, vi->hdr_len, DMA_TO_DEVICE);
> 
> +	kfree(rq->xsk.xsk_buffs);
> +
>  	return err1 | err2;
>  }
> 
> diff --git a/drivers/net/virtio/xsk.h b/drivers/net/virtio/xsk.h
> index 7ebc9bda7aee..bef41a3f954e 100644
> --- a/drivers/net/virtio/xsk.h
> +++ b/drivers/net/virtio/xsk.h
> @@ -23,4 +23,6 @@ int virtnet_xsk_pool_setup(struct net_device *dev, struct netdev_bpf *xdp);
>  bool virtnet_xsk_xmit(struct virtnet_sq *sq, struct xsk_buff_pool *pool,
>  		      int budget);
>  int virtnet_xsk_wakeup(struct net_device *dev, u32 qid, u32 flag);
> +int virtnet_add_recvbuf_xsk(struct virtnet_info *vi, struct virtnet_rq *rq,
> +			    struct xsk_buff_pool *pool, gfp_t gfp);
>  #endif
> 
>