[PATCH net-next-2.6] pktgen: speedup fragmented skbs

[PATCH net-next-2.6] pktgen: speedup fragmented skbs

[Eric Dumazet]

We spend lot of time clearing pages in pktgen.
(Or not clearing them on ipv6 and leaking kernel memory)

Since we dont modify them, we can use one zeroed page, and get
references on it. This page can use NUMA affinity as well.

Define pktgen_finalize_skb() helper, used both in ipv4 and ipv6

Results using skbs with one frag :

Before patch :

Result: OK: 608980458(c608978520+d1938) nsec, 1000000000
(100byte,1frags)
  1642088pps 1313Mb/sec (1313670400bps) errors: 0

After patch :

Result: OK: 345285014(c345283891+d1123) nsec, 1000000000
(100byte,1frags)
  2896158pps 2316Mb/sec (2316926400bps) errors: 0

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
---
 net/core/pktgen.c |  234 +++++++++++++++++---------------------------
 1 files changed, 93 insertions(+), 141 deletions(-)

diff --git a/net/core/pktgen.c b/net/core/pktgen.c
index a9e7fc4..17c4e16 100644
--- a/net/core/pktgen.c
+++ b/net/core/pktgen.c
@@ -251,6 +251,7 @@ struct pktgen_dev {
 	int max_pkt_size;	/* = ETH_ZLEN; */
 	int pkt_overhead;	/* overhead for MPLS, VLANs, IPSEC etc */
 	int nfrags;
+	struct page *page;
 	u64 delay;		/* nano-seconds */
 
 	__u64 count;		/* Default No packets to send */
@@ -1134,6 +1135,10 @@ static ssize_t pktgen_if_write(struct file *file,
 		if (node_possible(value)) {
 			pkt_dev->node = value;
 			sprintf(pg_result, "OK: node=%d", pkt_dev->node);
+			if (pkt_dev->page) {
+				put_page(pkt_dev->page);
+				pkt_dev->page = 0;
+			}
 		}
 		else
 			sprintf(pg_result, "ERROR: node not possible");
@@ -2605,6 +2610,90 @@ static inline __be16 build_tci(unsigned int id, unsigned int cfi,
 	return htons(id | (cfi << 12) | (prio << 13));
 }
 
+static void pktgen_finalize_skb(struct pktgen_dev *pkt_dev, struct sk_buff *skb,
+				int datalen)
+{
+	struct timeval timestamp;
+	struct pktgen_hdr *pgh;
+
+	pgh = (struct pktgen_hdr *)skb_put(skb, sizeof(*pgh));
+	datalen -= sizeof(*pgh);
+
+	if (pkt_dev->nfrags <= 0) {
+		pgh = (struct pktgen_hdr *)skb_put(skb, datalen);
+		memset(pgh + 1, 0, datalen);
+	} else {
+		int frags = pkt_dev->nfrags;
+		int i, len;
+
+
+		if (frags > MAX_SKB_FRAGS)
+			frags = MAX_SKB_FRAGS;
+		len = datalen - frags * PAGE_SIZE;
+		if (len > 0) {
+			memset(skb_put(skb, len), 0, len);
+			datalen = frags * PAGE_SIZE;
+		}
+
+		i = 0;
+		while (datalen > 0) {
+			if (unlikely(!pkt_dev->page)) {
+				int node = numa_node_id();
+
+				if (pkt_dev->node >= 0 && (pkt_dev->flags & F_NODE))
+					node = pkt_dev->node;
+				pkt_dev->page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
+				if (!pkt_dev->page)
+					break;
+			}
+			skb_shinfo(skb)->frags[i].page = pkt_dev->page;
+			get_page(pkt_dev->page);
+			skb_shinfo(skb)->frags[i].page_offset = 0;
+			skb_shinfo(skb)->frags[i].size =
+			    (datalen < PAGE_SIZE ? datalen : PAGE_SIZE);
+			datalen -= skb_shinfo(skb)->frags[i].size;
+			skb->len += skb_shinfo(skb)->frags[i].size;
+			skb->data_len += skb_shinfo(skb)->frags[i].size;
+			i++;
+			skb_shinfo(skb)->nr_frags = i;
+		}
+
+		while (i < frags) {
+			int rem;
+
+			if (i == 0)
+				break;
+
+			rem = skb_shinfo(skb)->frags[i - 1].size / 2;
+			if (rem == 0)
+				break;
+
+			skb_shinfo(skb)->frags[i - 1].size -= rem;
+
+			skb_shinfo(skb)->frags[i] =
+			    skb_shinfo(skb)->frags[i - 1];
+			get_page(skb_shinfo(skb)->frags[i].page);
+			skb_shinfo(skb)->frags[i].page =
+			    skb_shinfo(skb)->frags[i - 1].page;
+			skb_shinfo(skb)->frags[i].page_offset +=
+			    skb_shinfo(skb)->frags[i - 1].size;
+			skb_shinfo(skb)->frags[i].size = rem;
+			i++;
+			skb_shinfo(skb)->nr_frags = i;
+		}
+	}
+
+	/* Stamp the time, and sequence number,
+	 * convert them to network byte order
+	 */
+	pgh->pgh_magic = htonl(PKTGEN_MAGIC);
+	pgh->seq_num = htonl(pkt_dev->seq_num);
+
+	do_gettimeofday(&timestamp);
+	pgh->tv_sec = htonl(timestamp.tv_sec);
+	pgh->tv_usec = htonl(timestamp.tv_usec);
+}
+
 static struct sk_buff *fill_packet_ipv4(struct net_device *odev,
 					struct pktgen_dev *pkt_dev)
 {
@@ -2613,7 +2702,6 @@ static struct sk_buff *fill_packet_ipv4(struct net_device *odev,
 	struct udphdr *udph;
 	int datalen, iplen;
 	struct iphdr *iph;
-	struct pktgen_hdr *pgh = NULL;
 	__be16 protocol = htons(ETH_P_IP);
 	__be32 *mpls;
 	__be16 *vlan_tci = NULL;                 /* Encapsulates priority and VLAN ID */
@@ -2729,76 +2817,7 @@ static struct sk_buff *fill_packet_ipv4(struct net_device *odev,
 			   pkt_dev->pkt_overhead);
 	skb->dev = odev;
 	skb->pkt_type = PACKET_HOST;
-
-	if (pkt_dev->nfrags <= 0) {
-		pgh = (struct pktgen_hdr *)skb_put(skb, datalen);
-		memset(pgh + 1, 0, datalen - sizeof(struct pktgen_hdr));
-	} else {
-		int frags = pkt_dev->nfrags;
-		int i, len;
-
-		pgh = (struct pktgen_hdr *)(((char *)(udph)) + 8);
-
-		if (frags > MAX_SKB_FRAGS)
-			frags = MAX_SKB_FRAGS;
-		if (datalen > frags * PAGE_SIZE) {
-			len = datalen - frags * PAGE_SIZE;
-			memset(skb_put(skb, len), 0, len);
-			datalen = frags * PAGE_SIZE;
-		}
-
-		i = 0;
-		while (datalen > 0) {
-			struct page *page = alloc_pages(GFP_KERNEL | __GFP_ZERO, 0);
-			skb_shinfo(skb)->frags[i].page = page;
-			skb_shinfo(skb)->frags[i].page_offset = 0;
-			skb_shinfo(skb)->frags[i].size =
-			    (datalen < PAGE_SIZE ? datalen : PAGE_SIZE);
-			datalen -= skb_shinfo(skb)->frags[i].size;
-			skb->len += skb_shinfo(skb)->frags[i].size;
-			skb->data_len += skb_shinfo(skb)->frags[i].size;
-			i++;
-			skb_shinfo(skb)->nr_frags = i;
-		}
-
-		while (i < frags) {
-			int rem;
-
-			if (i == 0)
-				break;
-
-			rem = skb_shinfo(skb)->frags[i - 1].size / 2;
-			if (rem == 0)
-				break;
-
-			skb_shinfo(skb)->frags[i - 1].size -= rem;
-
-			skb_shinfo(skb)->frags[i] =
-			    skb_shinfo(skb)->frags[i - 1];
-			get_page(skb_shinfo(skb)->frags[i].page);
-			skb_shinfo(skb)->frags[i].page =
-			    skb_shinfo(skb)->frags[i - 1].page;
-			skb_shinfo(skb)->frags[i].page_offset +=
-			    skb_shinfo(skb)->frags[i - 1].size;
-			skb_shinfo(skb)->frags[i].size = rem;
-			i++;
-			skb_shinfo(skb)->nr_frags = i;
-		}
-	}
-
-	/* Stamp the time, and sequence number,
-	 * convert them to network byte order
-	 */
-	if (pgh) {
-		struct timeval timestamp;
-
-		pgh->pgh_magic = htonl(PKTGEN_MAGIC);
-		pgh->seq_num = htonl(pkt_dev->seq_num);
-
-		do_gettimeofday(&timestamp);
-		pgh->tv_sec = htonl(timestamp.tv_sec);
-		pgh->tv_usec = htonl(timestamp.tv_usec);
-	}
+	pktgen_finalize_skb(pkt_dev, skb, datalen);
 
 #ifdef CONFIG_XFRM
 	if (!process_ipsec(pkt_dev, skb, protocol))
@@ -2980,7 +2999,6 @@ static struct sk_buff *fill_packet_ipv6(struct net_device *odev,
 	struct udphdr *udph;
 	int datalen;
 	struct ipv6hdr *iph;
-	struct pktgen_hdr *pgh = NULL;
 	__be16 protocol = htons(ETH_P_IPV6);
 	__be32 *mpls;
 	__be16 *vlan_tci = NULL;                 /* Encapsulates priority and VLAN ID */
@@ -3083,75 +3101,7 @@ static struct sk_buff *fill_packet_ipv6(struct net_device *odev,
 	skb->dev = odev;
 	skb->pkt_type = PACKET_HOST;
 
-	if (pkt_dev->nfrags <= 0)
-		pgh = (struct pktgen_hdr *)skb_put(skb, datalen);
-	else {
-		int frags = pkt_dev->nfrags;
-		int i;
-
-		pgh = (struct pktgen_hdr *)(((char *)(udph)) + 8);
-
-		if (frags > MAX_SKB_FRAGS)
-			frags = MAX_SKB_FRAGS;
-		if (datalen > frags * PAGE_SIZE) {
-			skb_put(skb, datalen - frags * PAGE_SIZE);
-			datalen = frags * PAGE_SIZE;
-		}
-
-		i = 0;
-		while (datalen > 0) {
-			struct page *page = alloc_pages(GFP_KERNEL, 0);
-			skb_shinfo(skb)->frags[i].page = page;
-			skb_shinfo(skb)->frags[i].page_offset = 0;
-			skb_shinfo(skb)->frags[i].size =
-			    (datalen < PAGE_SIZE ? datalen : PAGE_SIZE);
-			datalen -= skb_shinfo(skb)->frags[i].size;
-			skb->len += skb_shinfo(skb)->frags[i].size;
-			skb->data_len += skb_shinfo(skb)->frags[i].size;
-			i++;
-			skb_shinfo(skb)->nr_frags = i;
-		}
-
-		while (i < frags) {
-			int rem;
-
-			if (i == 0)
-				break;
-
-			rem = skb_shinfo(skb)->frags[i - 1].size / 2;
-			if (rem == 0)
-				break;
-
-			skb_shinfo(skb)->frags[i - 1].size -= rem;
-
-			skb_shinfo(skb)->frags[i] =
-			    skb_shinfo(skb)->frags[i - 1];
-			get_page(skb_shinfo(skb)->frags[i].page);
-			skb_shinfo(skb)->frags[i].page =
-			    skb_shinfo(skb)->frags[i - 1].page;
-			skb_shinfo(skb)->frags[i].page_offset +=
-			    skb_shinfo(skb)->frags[i - 1].size;
-			skb_shinfo(skb)->frags[i].size = rem;
-			i++;
-			skb_shinfo(skb)->nr_frags = i;
-		}
-	}
-
-	/* Stamp the time, and sequence number,
-	 * convert them to network byte order
-	 * should we update cloned packets too ?
-	 */
-	if (pgh) {
-		struct timeval timestamp;
-
-		pgh->pgh_magic = htonl(PKTGEN_MAGIC);
-		pgh->seq_num = htonl(pkt_dev->seq_num);
-
-		do_gettimeofday(&timestamp);
-		pgh->tv_sec = htonl(timestamp.tv_sec);
-		pgh->tv_usec = htonl(timestamp.tv_usec);
-	}
-	/* pkt_dev->seq_num++; FF: you really mean this? */
+	pktgen_finalize_skb(pkt_dev, skb, datalen);
 
 	return skb;
 }
@@ -3884,6 +3834,8 @@ static int pktgen_remove_device(struct pktgen_thread *t,
 	free_SAs(pkt_dev);
 #endif
 	vfree(pkt_dev->flows);
+	if (pkt_dev->page)
+		put_page(pkt_dev->page);
 	kfree(pkt_dev);
 	return 0;
 }

Re: [PATCH net-next-2.6] pktgen: speedup fragmented skbs

[Ben Greear]

On 01/25/2011 09:13 AM, Eric Dumazet wrote:
> We spend lot of time clearing pages in pktgen.
> (Or not clearing them on ipv6 and leaking kernel memory)
>
> Since we dont modify them, we can use one zeroed page, and get
> references on it. This page can use NUMA affinity as well.
>
> Define pktgen_finalize_skb() helper, used both in ipv4 and ipv6

Some devices, like vlans, can change the skb, but perhaps they
will not actually mess with the paged data?

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com

Re: [PATCH net-next-2.6] pktgen: speedup fragmented skbs

[Eric Dumazet]

Le mardi 25 janvier 2011 à 09:22 -0800, Ben Greear a écrit :
> On 01/25/2011 09:13 AM, Eric Dumazet wrote:
> > We spend lot of time clearing pages in pktgen.
> > (Or not clearing them on ipv6 and leaking kernel memory)
> >
> > Since we dont modify them, we can use one zeroed page, and get
> > references on it. This page can use NUMA affinity as well.
> >
> > Define pktgen_finalize_skb() helper, used both in ipv4 and ipv6
> 
> Some devices, like vlans, can change the skb, but perhaps they
> will not actually mess with the paged data?

Yes, its absolutely forbidden to write on paged data.

If necessary, a COW must be done.

Re: [PATCH net-next-2.6] pktgen: speedup fragmented skbs

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Tue, 25 Jan 2011 18:13:55 +0100

> We spend lot of time clearing pages in pktgen.
> (Or not clearing them on ipv6 and leaking kernel memory)
> 
> Since we dont modify them, we can use one zeroed page, and get
> references on it. This page can use NUMA affinity as well.
> 
> Define pktgen_finalize_skb() helper, used both in ipv4 and ipv6
> 
> Results using skbs with one frag :
> 
> Before patch :
> 
> Result: OK: 608980458(c608978520+d1938) nsec, 1000000000
> (100byte,1frags)
>   1642088pps 1313Mb/sec (1313670400bps) errors: 0
> 
> After patch :
> 
> Result: OK: 345285014(c345283891+d1123) nsec, 1000000000
> (100byte,1frags)
>   2896158pps 2316Mb/sec (2316926400bps) errors: 0
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>

Applied, although I changed "->page = 0;" to "->page = NULL;"

Thanks!

Re: [PATCH net-next-2.6] pktgen: speedup fragmented skbs

[Eric Dumazet]

Le mardi 25 janvier 2011 à 13:26 -0800, David Miller a écrit :

> Applied, although I changed "->page = 0;" to "->page = NULL;"
> 

Ok, thanks !