Message-ID: <20230710051818-mutt-send-email-mst@kernel.org>
Date: Mon, 10 Jul 2023 05:40:21 -0400
From: "Michael S. Tsirkin" <mst@...hat.com>
To: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
Cc: virtualization@...ts.linux-foundation.org,
	Jason Wang <jasowang@...hat.com>,
	"David S. Miller" <davem@...emloft.net>,
	Eric Dumazet <edumazet@...gle.com>,
	Jakub Kicinski <kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>,
	Alexei Starovoitov <ast@...nel.org>,
	Daniel Borkmann <daniel@...earbox.net>,
	Jesper Dangaard Brouer <hawk@...nel.org>,
	John Fastabend <john.fastabend@...il.com>, netdev@...r.kernel.org,
	bpf@...r.kernel.org, Christoph Hellwig <hch@...radead.org>
Subject: Re: [PATCH vhost v11 10/10] virtio_net: merge dma operation for one
 page

On Mon, Jul 10, 2023 at 11:42:37AM +0800, Xuan Zhuo wrote:
> Currently, the virtio core will perform a dma operation for each
> operation. Although, the same page may be operated multiple times.
> 
> The driver does the dma operation and manages the dma address based the
> feature premapped of virtio core.
> 
> This way, we can perform only one dma operation for the same page. In
> the case of mtu 1500, this can reduce a lot of dma operations.
> 
> Tested on Aliyun g7.4large machine, in the case of a cpu 100%, pps
> increased from 1893766 to 1901105. An increase of 0.4%.

what kind of dma was there? an IOMMU? which vendors? in which mode
of operation?

> Signed-off-by: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>

This kind of difference is likely in the noise.


> ---
>  drivers/net/virtio_net.c | 283 ++++++++++++++++++++++++++++++++++++---
>  1 file changed, 267 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 486b5849033d..4de845d35bed 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -126,6 +126,27 @@ static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
>  #define VIRTNET_SQ_STATS_LEN	ARRAY_SIZE(virtnet_sq_stats_desc)
>  #define VIRTNET_RQ_STATS_LEN	ARRAY_SIZE(virtnet_rq_stats_desc)
>  
> +/* The bufs on the same page may share this struct. */
> +struct virtnet_rq_dma {
> +	struct virtnet_rq_dma *next;
> +
> +	dma_addr_t addr;
> +
> +	void *buf;
> +	u32 len;
> +
> +	u32 ref;
> +};
> +
> +/* Record the dma and buf. */

I guess I see that. But why?
And these two comments are the extent of the available
documentation, that's not enough I feel.


> +struct virtnet_rq_data {
> +	struct virtnet_rq_data *next;

Is manually reimplementing a linked list the best
we can do?

> +
> +	void *buf;
> +
> +	struct virtnet_rq_dma *dma;
> +};
> +
>  /* Internal representation of a send virtqueue */
>  struct send_queue {
>  	/* Virtqueue associated with this send _queue */
> @@ -175,6 +196,13 @@ struct receive_queue {
>  	char name[16];
>  
>  	struct xdp_rxq_info xdp_rxq;
> +
> +	struct virtnet_rq_data *data_array;
> +	struct virtnet_rq_data *data_free;
> +
> +	struct virtnet_rq_dma *dma_array;
> +	struct virtnet_rq_dma *dma_free;
> +	struct virtnet_rq_dma *last_dma;
>  };
>  
>  /* This structure can contain rss message with maximum settings for indirection table and keysize
> @@ -549,6 +577,176 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>  	return skb;
>  }
>  
> +static void virtnet_rq_unmap(struct receive_queue *rq, struct virtnet_rq_dma *dma)
> +{
> +	struct device *dev;
> +
> +	--dma->ref;
> +
> +	if (dma->ref)
> +		return;
> +

If you don't unmap there is no guarantee valid data will be
there in the buffer.

> +	dev = virtqueue_dma_dev(rq->vq);
> +
> +	dma_unmap_page(dev, dma->addr, dma->len, DMA_FROM_DEVICE);





> +
> +	dma->next = rq->dma_free;
> +	rq->dma_free = dma;
> +}
> +
> +static void *virtnet_rq_recycle_data(struct receive_queue *rq,
> +				     struct virtnet_rq_data *data)
> +{
> +	void *buf;
> +
> +	buf = data->buf;
> +
> +	data->next = rq->data_free;
> +	rq->data_free = data;
> +
> +	return buf;
> +}
> +
> +static struct virtnet_rq_data *virtnet_rq_get_data(struct receive_queue *rq,
> +						   void *buf,
> +						   struct virtnet_rq_dma *dma)
> +{
> +	struct virtnet_rq_data *data;
> +
> +	data = rq->data_free;
> +	rq->data_free = data->next;
> +
> +	data->buf = buf;
> +	data->dma = dma;
> +
> +	return data;
> +}
> +
> +static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
> +{
> +	struct virtnet_rq_data *data;
> +	void *buf;
> +
> +	buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
> +	if (!buf || !rq->data_array)
> +		return buf;
> +
> +	data = buf;
> +
> +	virtnet_rq_unmap(rq, data->dma);
> +
> +	return virtnet_rq_recycle_data(rq, data);
> +}
> +
> +static void *virtnet_rq_detach_unused_buf(struct receive_queue *rq)
> +{
> +	struct virtnet_rq_data *data;
> +	void *buf;
> +
> +	buf = virtqueue_detach_unused_buf(rq->vq);
> +	if (!buf || !rq->data_array)
> +		return buf;
> +
> +	data = buf;
> +
> +	virtnet_rq_unmap(rq, data->dma);
> +
> +	return virtnet_rq_recycle_data(rq, data);
> +}
> +
> +static int virtnet_rq_map_sg(struct receive_queue *rq, void *buf, u32 len)
> +{
> +	struct virtnet_rq_dma *dma = rq->last_dma;
> +	struct device *dev;
> +	u32 off, map_len;
> +	dma_addr_t addr;
> +	void *end;
> +
> +	if (likely(dma) && buf >= dma->buf && (buf + len <= dma->buf + dma->len)) {
> +		++dma->ref;
> +		addr = dma->addr + (buf - dma->buf);
> +		goto ok;
> +	}

So this is the meat of the proposed optimization. I guess that
if the last buffer we allocated happens to be in the same page
as this one then they can both be mapped for DMA together.
Why last one specifically? Whether next one happens to
be close depends on luck. If you want to try optimizing this
the right thing to do is likely by using a page pool.
There's actually work upstream on page pool, look it up.

> +
> +	end = buf + len - 1;
> +	off = offset_in_page(end);
> +	map_len = len + PAGE_SIZE - off;
> +
> +	dev = virtqueue_dma_dev(rq->vq);
> +
> +	addr = dma_map_page_attrs(dev, virt_to_page(buf), offset_in_page(buf),
> +				  map_len, DMA_FROM_DEVICE, 0);
> +	if (addr == DMA_MAPPING_ERROR)
> +		return -ENOMEM;
> +
> +	dma = rq->dma_free;
> +	rq->dma_free = dma->next;
> +
> +	dma->ref = 1;
> +	dma->buf = buf;
> +	dma->addr = addr;
> +	dma->len = map_len;
> +
> +	rq->last_dma = dma;
> +
> +ok:
> +	sg_init_table(rq->sg, 1);
> +	rq->sg[0].dma_address = addr;
> +	rq->sg[0].length = len;
> +
> +	return 0;
> +}
> +
> +static int virtnet_rq_merge_map_init(struct virtnet_info *vi)
> +{
> +	struct receive_queue *rq;
> +	int i, err, j, num;
> +
> +	/* disable for big mode */
> +	if (!vi->mergeable_rx_bufs && vi->big_packets)
> +		return 0;
> +
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
> +		err = virtqueue_set_premapped(vi->rq[i].vq);
> +		if (err)
> +			continue;
> +
> +		rq = &vi->rq[i];
> +
> +		num = virtqueue_get_vring_size(rq->vq);
> +
> +		rq->data_array = kmalloc_array(num, sizeof(*rq->data_array), GFP_KERNEL);
> +		if (!rq->data_array)
> +			goto err;
> +
> +		rq->dma_array = kmalloc_array(num, sizeof(*rq->dma_array), GFP_KERNEL);
> +		if (!rq->dma_array)
> +			goto err;
> +
> +		for (j = 0; j < num; ++j) {
> +			rq->data_array[j].next = rq->data_free;
> +			rq->data_free = &rq->data_array[j];
> +
> +			rq->dma_array[j].next = rq->dma_free;
> +			rq->dma_free = &rq->dma_array[j];
> +		}
> +	}
> +
> +	return 0;
> +
> +err:
> +	for (i = 0; i < vi->max_queue_pairs; i++) {
> +		struct receive_queue *rq;
> +
> +		rq = &vi->rq[i];
> +
> +		kfree(rq->dma_array);
> +		kfree(rq->data_array);
> +	}
> +
> +	return -ENOMEM;
> +}
> +
>  static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
>  {
>  	unsigned int len;
> @@ -835,7 +1033,7 @@ static struct page *xdp_linearize_page(struct receive_queue *rq,
>  		void *buf;
>  		int off;
>  
> -		buf = virtqueue_get_buf(rq->vq, &buflen);
> +		buf = virtnet_rq_get_buf(rq, &buflen, NULL);
>  		if (unlikely(!buf))
>  			goto err_buf;
>  
> @@ -1126,7 +1324,7 @@ static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
>  		return -EINVAL;
>  
>  	while (--*num_buf > 0) {
> -		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
> +		buf = virtnet_rq_get_buf(rq, &len, &ctx);
>  		if (unlikely(!buf)) {
>  			pr_debug("%s: rx error: %d buffers out of %d missing\n",
>  				 dev->name, *num_buf,
> @@ -1351,7 +1549,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  	while (--num_buf) {
>  		int num_skb_frags;
>  
> -		buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
> +		buf = virtnet_rq_get_buf(rq, &len, &ctx);
>  		if (unlikely(!buf)) {
>  			pr_debug("%s: rx error: %d buffers out of %d missing\n",
>  				 dev->name, num_buf,
> @@ -1414,7 +1612,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  err_skb:
>  	put_page(page);
>  	while (num_buf-- > 1) {
> -		buf = virtqueue_get_buf(rq->vq, &len);
> +		buf = virtnet_rq_get_buf(rq, &len, NULL);
>  		if (unlikely(!buf)) {
>  			pr_debug("%s: rx error: %d buffers missing\n",
>  				 dev->name, num_buf);
> @@ -1529,6 +1727,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
>  	unsigned int xdp_headroom = virtnet_get_headroom(vi);
>  	void *ctx = (void *)(unsigned long)xdp_headroom;
>  	int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
> +	struct virtnet_rq_data *data;
>  	int err;
>  
>  	len = SKB_DATA_ALIGN(len) +
> @@ -1539,11 +1738,34 @@ static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
>  	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
>  	get_page(alloc_frag->page);
>  	alloc_frag->offset += len;
> -	sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
> -		    vi->hdr_len + GOOD_PACKET_LEN);
> -	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
> +
> +	if (rq->data_array) {
> +		err = virtnet_rq_map_sg(rq, buf + VIRTNET_RX_PAD + xdp_headroom,
> +					vi->hdr_len + GOOD_PACKET_LEN);
> +		if (err)
> +			goto map_err;
> +
> +		data = virtnet_rq_get_data(rq, buf, rq->last_dma);
> +	} else {
> +		sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
> +			    vi->hdr_len + GOOD_PACKET_LEN);
> +		data = (void *)buf;
> +	}
> +
> +	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, data, ctx, gfp);
>  	if (err < 0)
> -		put_page(virt_to_head_page(buf));
> +		goto add_err;
> +
> +	return err;
> +
> +add_err:
> +	if (rq->data_array) {
> +		virtnet_rq_unmap(rq, data->dma);
> +		virtnet_rq_recycle_data(rq, data);
> +	}
> +
> +map_err:
> +	put_page(virt_to_head_page(buf));
>  	return err;
>  }
>  
> @@ -1620,6 +1842,7 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi,
>  	unsigned int headroom = virtnet_get_headroom(vi);
>  	unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
>  	unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
> +	struct virtnet_rq_data *data;
>  	char *buf;
>  	void *ctx;
>  	int err;
> @@ -1650,12 +1873,32 @@ static int add_recvbuf_mergeable(struct virtnet_info *vi,
>  		alloc_frag->offset += hole;
>  	}
>  
> -	sg_init_one(rq->sg, buf, len);
> +	if (rq->data_array) {
> +		err = virtnet_rq_map_sg(rq, buf, len);
> +		if (err)
> +			goto map_err;
> +
> +		data = virtnet_rq_get_data(rq, buf, rq->last_dma);
> +	} else {
> +		sg_init_one(rq->sg, buf, len);
> +		data = (void *)buf;
> +	}
> +
>  	ctx = mergeable_len_to_ctx(len + room, headroom);
> -	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
> +	err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, data, ctx, gfp);
>  	if (err < 0)
> -		put_page(virt_to_head_page(buf));
> +		goto add_err;
> +
> +	return 0;
> +
> +add_err:
> +	if (rq->data_array) {
> +		virtnet_rq_unmap(rq, data->dma);
> +		virtnet_rq_recycle_data(rq, data);
> +	}
>  
> +map_err:
> +	put_page(virt_to_head_page(buf));
>  	return err;
>  }
>  
> @@ -1775,13 +2018,13 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
>  		void *ctx;
>  
>  		while (stats.packets < budget &&
> -		       (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
> +		       (buf = virtnet_rq_get_buf(rq, &len, &ctx))) {
>  			receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
>  			stats.packets++;
>  		}
>  	} else {
>  		while (stats.packets < budget &&
> -		       (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
> +		       (buf = virtnet_rq_get_buf(rq, &len, NULL)) != NULL) {
>  			receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
>  			stats.packets++;
>  		}
> @@ -3514,6 +3757,9 @@ 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->rq[i].data_array);
> +		kfree(vi->rq[i].dma_array);
>  	}
>  
>  	/* We called __netif_napi_del(),
> @@ -3591,9 +3837,10 @@ static void free_unused_bufs(struct virtnet_info *vi)
>  	}
>  
>  	for (i = 0; i < vi->max_queue_pairs; i++) {
> -		struct virtqueue *vq = vi->rq[i].vq;
> -		while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
> -			virtnet_rq_free_unused_buf(vq, buf);
> +		struct receive_queue *rq = &vi->rq[i];
> +
> +		while ((buf = virtnet_rq_detach_unused_buf(rq)) != NULL)
> +			virtnet_rq_free_unused_buf(rq->vq, buf);
>  		cond_resched();
>  	}
>  }
> @@ -3767,6 +4014,10 @@ static int init_vqs(struct virtnet_info *vi)
>  	if (ret)
>  		goto err_free;
>  
> +	ret = virtnet_rq_merge_map_init(vi);
> +	if (ret)
> +		goto err_free;
> +
>  	cpus_read_lock();
>  	virtnet_set_affinity(vi);
>  	cpus_read_unlock();
> -- 
> 2.32.0.3.g01195cf9f

Powered by blists - more mailing lists