Virtualization - Nov 2018 - [PATCH net-next 1/2] vhost_net: mitigate page reference counting during page frag refill

We do a get_page() which involves a atomic operation. This patch tries
to mitigate a per packet atomic operation by maintaining a reference
bias which is initially USHRT_MAX. Each time a page is got, instead of
calling get_page() we decrease the bias and when we find it's time to
use a new page we will decrease the bias at one time through
__page_cache_drain_cache().

Testpmd(virtio_user + vhost_net) + XDP_DROP on TAP shows about 1.6%
improvement.

Before: 4.63Mpps
After:  4.71Mpps

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/vhost/net.c | 54 ++++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 51 insertions(+), 3 deletions(-)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index ab11b2bee273..d919284f103b 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c
@@ -141,6 +141,10 @@ struct vhost_net {
 	unsigned tx_zcopy_err;
 	/* Flush in progress. Protected by tx vq lock. */
 	bool tx_flush;
+	/* Private page frag */
+	struct page_frag page_frag;
+	/* Refcount bias of page frag */
+	int refcnt_bias;
 };
 
 static unsigned vhost_net_zcopy_mask __read_mostly;
@@ -637,14 +641,53 @@ static bool tx_can_batch(struct vhost_virtqueue *vq,
size_t total_len)
 	       !vhost_vq_avail_empty(vq->dev, vq);
 }
 
+#define SKB_FRAG_PAGE_ORDER     get_order(32768)
+
+static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
+				       struct page_frag *pfrag, gfp_t gfp)
+{
+	if (pfrag->page) {
+		if (pfrag->offset + sz <= pfrag->size)
+			return true;
+		__page_frag_cache_drain(pfrag->page, net->refcnt_bias);
+	}
+
+	pfrag->offset = 0;
+	net->refcnt_bias = 0;
+	if (SKB_FRAG_PAGE_ORDER) {
+		/* Avoid direct reclaim but allow kswapd to wake */
+		pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
+					  __GFP_COMP | __GFP_NOWARN |
+					  __GFP_NORETRY,
+					  SKB_FRAG_PAGE_ORDER);
+		if (likely(pfrag->page)) {
+			pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
+			goto done;
+		}
+	}
+	pfrag->page = alloc_page(gfp);
+	if (likely(pfrag->page)) {
+		pfrag->size = PAGE_SIZE;
+		goto done;
+	}
+	return false;
+
+done:
+	net->refcnt_bias = USHRT_MAX;
+	page_ref_add(pfrag->page, USHRT_MAX - 1);
+	return true;
+}
+
 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
 
 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
 			       struct iov_iter *from)
 {
 	struct vhost_virtqueue *vq = &nvq->vq;
+	struct vhost_net *net = container_of(vq->dev, struct vhost_net,
+					     dev);
 	struct socket *sock = vq->private_data;
-	struct page_frag *alloc_frag = &current->task_frag;
+	struct page_frag *alloc_frag = &net->page_frag;
 	struct virtio_net_hdr *gso;
 	struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
 	struct tun_xdp_hdr *hdr;
@@ -665,7 +708,8 @@ static int vhost_net_build_xdp(struct vhost_net_virtqueue
*nvq,
 
 	buflen += SKB_DATA_ALIGN(len + pad);
 	alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
-	if (unlikely(!skb_page_frag_refill(buflen, alloc_frag, GFP_KERNEL)))
+	if (unlikely(!vhost_net_page_frag_refill(net, buflen,
+						 alloc_frag, GFP_KERNEL)))
 		return -ENOMEM;
 
 	buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
@@ -703,7 +747,7 @@ static int vhost_net_build_xdp(struct vhost_net_virtqueue
*nvq,
 	xdp->data_end = xdp->data + len;
 	hdr->buflen = buflen;
 
-	get_page(alloc_frag->page);
+	--net->refcnt_bias;
 	alloc_frag->offset += buflen;
 
 	++nvq->batched_xdp;
@@ -1292,6 +1336,8 @@ static int vhost_net_open(struct inode *inode, struct file
*f)
 	vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
 
 	f->private_data = n;
+	n->page_frag.page = NULL;
+	n->refcnt_bias = 0;
 
 	return 0;
 }
@@ -1366,6 +1412,8 @@ static int vhost_net_release(struct inode *inode, struct
file *f)
 	kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
 	kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
 	kfree(n->dev.vqs);
+	if (n->page_frag.page)
+		__page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
 	kvfree(n);
 	return 0;
 }
-- 
2.17.1

Thanks to the batched XDP buffs through msg_control. Instead of
calling put_page() for each page which involves a atomic operation,
let's batch them by record the last page that needs to be freed and
its refcnt count and free them in a batch.

Testpmd(virtio-user + vhost_net) + XDP_DROP shows 3.8% improvement.

Before: 4.71Mpps
After : 4.89Mpps

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/net/tun.c | 29 ++++++++++++++++++++++++++---
 1 file changed, 26 insertions(+), 3 deletions(-)

diff --git a/drivers/net/tun.c b/drivers/net/tun.c
index a65779c6d72f..e90a7923a5f6 100644
--- a/drivers/net/tun.c
+++ b/drivers/net/tun.c
@@ -188,6 +188,11 @@ struct tun_file {
 	struct xdp_rxq_info xdp_rxq;
 };
 
+struct tun_page {
+	struct page *page;
+	int count;
+};
+
 struct tun_flow_entry {
 	struct hlist_node hash_link;
 	struct rcu_head rcu;
@@ -2377,9 +2382,16 @@ static void tun_sock_write_space(struct sock *sk)
 	kill_fasync(&tfile->fasync, SIGIO, POLL_OUT);
 }
 
+static void tun_put_page(struct tun_page *tpage)
+{
+	if (tpage->page)
+		__page_frag_cache_drain(tpage->page, tpage->count);
+}
+
 static int tun_xdp_one(struct tun_struct *tun,
 		       struct tun_file *tfile,
-		       struct xdp_buff *xdp, int *flush)
+		       struct xdp_buff *xdp, int *flush,
+		       struct tun_page *tpage)
 {
 	struct tun_xdp_hdr *hdr = xdp->data_hard_start;
 	struct virtio_net_hdr *gso = &hdr->gso;
@@ -2390,6 +2402,7 @@ static int tun_xdp_one(struct tun_struct *tun,
 	int buflen = hdr->buflen;
 	int err = 0;
 	bool skb_xdp = false;
+	struct page *page;
 
 	xdp_prog = rcu_dereference(tun->xdp_prog);
 	if (xdp_prog) {
@@ -2416,7 +2429,14 @@ static int tun_xdp_one(struct tun_struct *tun,
 		case XDP_PASS:
 			break;
 		default:
-			put_page(virt_to_head_page(xdp->data));
+			page = virt_to_head_page(xdp->data);
+			if (tpage->page == page) {
+				++tpage->count;
+			} else {
+				tun_put_page(tpage);
+				tpage->page = page;
+				tpage->count = 1;
+			}
 			return 0;
 		}
 	}
@@ -2480,6 +2500,7 @@ static int tun_sendmsg(struct socket *sock, struct msghdr
*m, size_t total_len)
 		return -EBADFD;
 
 	if (ctl && (ctl->type == TUN_MSG_PTR)) {
+		struct tun_page tpage = {0};
 		int n = ctl->num;
 		int flush = 0;
 
@@ -2488,7 +2509,7 @@ static int tun_sendmsg(struct socket *sock, struct msghdr
*m, size_t total_len)
 
 		for (i = 0; i < n; i++) {
 			xdp = &((struct xdp_buff *)ctl->ptr)[i];
-			tun_xdp_one(tun, tfile, xdp, &flush);
+			tun_xdp_one(tun, tfile, xdp, &flush, &tpage);
 		}
 
 		if (flush)
@@ -2497,6 +2518,8 @@ static int tun_sendmsg(struct socket *sock, struct msghdr
*m, size_t total_len)
 		rcu_read_unlock();
 		local_bh_enable();
 
+		tun_put_page(&tpage);
+
 		ret = total_len;
 		goto out;
 	}
-- 
2.17.1

From: Jason Wang <jasowang at redhat.com>
Date: Thu, 15 Nov 2018 17:43:09 +0800
> We do a get_page() which involves a atomic operation. This patch tries
> to mitigate a per packet atomic operation by maintaining a reference
> bias which is initially USHRT_MAX. Each time a page is got, instead of
> calling get_page() we decrease the bias and when we find it's time to
> use a new page we will decrease the bias at one time through
> __page_cache_drain_cache().
> 
> Testpmd(virtio_user + vhost_net) + XDP_DROP on TAP shows about 1.6%
> improvement.
> 
> Before: 4.63Mpps
> After:  4.71Mpps
> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>
Applied.

From: Jason Wang <jasowang at redhat.com>
Date: Thu, 15 Nov 2018 17:43:10 +0800
> Thanks to the batched XDP buffs through msg_control. Instead of
> calling put_page() for each page which involves a atomic operation,
> let's batch them by record the last page that needs to be freed and
> its refcnt count and free them in a batch.
> 
> Testpmd(virtio-user + vhost_net) + XDP_DROP shows 3.8% improvement.
> 
> Before: 4.71Mpps
> After : 4.89Mpps
> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>
Applied.