[PATCH v5 1/2] core: Factor out flow calculation from get_rps_cpu

[PATCH v5 1/2] core: Factor out flow calculation from get_rps_cpu

[Krishna Kumar]

From: Krishna Kumar <krkumar2@in.ibm.com>

Factor out flow calculation code from get_rps_cpu, since other
functions can use the same code.

Revisions:

v2 (Ben): Separate flow calcuation out and use in select queue.
v3 (Arnd): Don't re-implement MIN.
v4 (Changli): skb->data points to ethernet header in macvtap, and
	make a fast path. Tested macvtap with this patch.
v5 (Changli):
	- Cache skb->rxhash in skb_get_rxhash
	- macvtap may not have pow(2) queues, so change code for
	  queue selection.
    (Arnd):
	- Use first available queue if all fails.


Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com>
---
 include/linux/skbuff.h |    9 +++
 net/core/dev.c         |  106 ++++++++++++++++++++++-----------------
 2 files changed, 71 insertions(+), 44 deletions(-)

diff -ruNp org/include/linux/skbuff.h new/include/linux/skbuff.h
--- org/include/linux/skbuff.h	2010-08-04 20:45:19.000000000 +0530
+++ new/include/linux/skbuff.h	2010-08-04 20:47:02.000000000 +0530
@@ -558,6 +558,15 @@ extern unsigned int   skb_find_text(stru
 				    unsigned int to, struct ts_config *config,
 				    struct ts_state *state);
 
+extern __u32 __skb_get_rxhash(struct sk_buff *skb);
+static inline __u32 skb_get_rxhash(struct sk_buff *skb)
+{
+	if (!skb->rxhash)
+		skb->rxhash = __skb_get_rxhash(skb);
+
+	return skb->rxhash;
+}
+
 #ifdef NET_SKBUFF_DATA_USES_OFFSET
 static inline unsigned char *skb_end_pointer(const struct sk_buff *skb)
 {
diff -ruNp org/net/core/dev.c new/net/core/dev.c
--- org/net/core/dev.c	2010-08-04 20:45:19.000000000 +0530
+++ new/net/core/dev.c	2010-08-04 21:13:37.000000000 +0530
@@ -2259,69 +2259,41 @@ static inline void ____napi_schedule(str
 	__raise_softirq_irqoff(NET_RX_SOFTIRQ);
 }
 
-#ifdef CONFIG_RPS
-
-/* One global table that all flow-based protocols share. */
-struct rps_sock_flow_table *rps_sock_flow_table __read_mostly;
-EXPORT_SYMBOL(rps_sock_flow_table);
-
 /*
- * get_rps_cpu is called from netif_receive_skb and returns the target
- * CPU from the RPS map of the receiving queue for a given skb.
- * rcu_read_lock must be held on entry.
+ * __skb_get_rxhash: calculate a flow hash based on src/dst addresses
+ * and src/dst port numbers. Returns a non-zero hash number on success
+ * and 0 on failure.
  */
-static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
-		       struct rps_dev_flow **rflowp)
+__u32 __skb_get_rxhash(struct sk_buff *skb)
 {
+	int nhoff, hash = 0;
 	struct ipv6hdr *ip6;
 	struct iphdr *ip;
-	struct netdev_rx_queue *rxqueue;
-	struct rps_map *map;
-	struct rps_dev_flow_table *flow_table;
-	struct rps_sock_flow_table *sock_flow_table;
-	int cpu = -1;
 	u8 ip_proto;
-	u16 tcpu;
 	u32 addr1, addr2, ihl;
 	union {
 		u32 v32;
 		u16 v16[2];
 	} ports;
 
-	if (skb_rx_queue_recorded(skb)) {
-		u16 index = skb_get_rx_queue(skb);
-		if (unlikely(index >= dev->num_rx_queues)) {
-			WARN_ONCE(dev->num_rx_queues > 1, "%s received packet "
-				"on queue %u, but number of RX queues is %u\n",
-				dev->name, index, dev->num_rx_queues);
-			goto done;
-		}
-		rxqueue = dev->_rx + index;
-	} else
-		rxqueue = dev->_rx;
-
-	if (!rxqueue->rps_map && !rxqueue->rps_flow_table)
-		goto done;
-
-	if (skb->rxhash)
-		goto got_hash; /* Skip hash computation on packet header */
+	nhoff = skb_network_offset(skb);
 
 	switch (skb->protocol) {
 	case __constant_htons(ETH_P_IP):
-		if (!pskb_may_pull(skb, sizeof(*ip)))
+		if (!pskb_may_pull(skb, sizeof(*ip) + nhoff))
 			goto done;
 
-		ip = (struct iphdr *) skb->data;
+		ip = (struct iphdr *) skb->data + nhoff;
 		ip_proto = ip->protocol;
 		addr1 = (__force u32) ip->saddr;
 		addr2 = (__force u32) ip->daddr;
 		ihl = ip->ihl;
 		break;
 	case __constant_htons(ETH_P_IPV6):
-		if (!pskb_may_pull(skb, sizeof(*ip6)))
+		if (!pskb_may_pull(skb, sizeof(*ip6) + nhoff))
 			goto done;
 
-		ip6 = (struct ipv6hdr *) skb->data;
+		ip6 = (struct ipv6hdr *) skb->data + nhoff;
 		ip_proto = ip6->nexthdr;
 		addr1 = (__force u32) ip6->saddr.s6_addr32[3];
 		addr2 = (__force u32) ip6->daddr.s6_addr32[3];
@@ -2330,6 +2302,7 @@ static int get_rps_cpu(struct net_device
 	default:
 		goto done;
 	}
+
 	switch (ip_proto) {
 	case IPPROTO_TCP:
 	case IPPROTO_UDP:
@@ -2338,8 +2311,9 @@ static int get_rps_cpu(struct net_device
 	case IPPROTO_AH:
 	case IPPROTO_SCTP:
 	case IPPROTO_UDPLITE:
-		if (pskb_may_pull(skb, (ihl * 4) + 4)) {
-			ports.v32 = * (__force u32 *) (skb->data + (ihl * 4));
+		if (pskb_may_pull(skb, (ihl * 4) + 4 + nhoff)) {
+			ports.v32 = * (__force u32 *) (skb->data + nhoff +
+						       (ihl * 4));
 			if (ports.v16[1] < ports.v16[0])
 				swap(ports.v16[0], ports.v16[1]);
 			break;
@@ -2352,11 +2326,55 @@ static int get_rps_cpu(struct net_device
 	/* get a consistent hash (same value on both flow directions) */
 	if (addr2 < addr1)
 		swap(addr1, addr2);
-	skb->rxhash = jhash_3words(addr1, addr2, ports.v32, hashrnd);
-	if (!skb->rxhash)
-		skb->rxhash = 1;
 
-got_hash:
+	hash = jhash_3words(addr1, addr2, ports.v32, hashrnd);
+	if (!hash)
+		hash = 1;
+
+done:
+	return hash;
+}
+EXPORT_SYMBOL(__skb_get_rxhash);
+
+#ifdef CONFIG_RPS
+
+/* One global table that all flow-based protocols share. */
+struct rps_sock_flow_table *rps_sock_flow_table __read_mostly;
+EXPORT_SYMBOL(rps_sock_flow_table);
+
+/*
+ * get_rps_cpu is called from netif_receive_skb and returns the target
+ * CPU from the RPS map of the receiving queue for a given skb.
+ * rcu_read_lock must be held on entry.
+ */
+static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
+		       struct rps_dev_flow **rflowp)
+{
+	struct netdev_rx_queue *rxqueue;
+	struct rps_map *map;
+	struct rps_dev_flow_table *flow_table;
+	struct rps_sock_flow_table *sock_flow_table;
+	int cpu = -1;
+	u16 tcpu;
+
+	if (skb_rx_queue_recorded(skb)) {
+		u16 index = skb_get_rx_queue(skb);
+		if (unlikely(index >= dev->num_rx_queues)) {
+			WARN_ONCE(dev->num_rx_queues > 1, "%s received packet "
+				"on queue %u, but number of RX queues is %u\n",
+				dev->name, index, dev->num_rx_queues);
+			goto done;
+		}
+		rxqueue = dev->_rx + index;
+	} else
+		rxqueue = dev->_rx;
+
+	if (!rxqueue->rps_map && !rxqueue->rps_flow_table)
+		goto done;
+
+	if (!skb_get_rxhash(skb))
+		goto done;
+
 	flow_table = rcu_dereference(rxqueue->rps_flow_table);
 	sock_flow_table = rcu_dereference(rps_sock_flow_table);
 	if (flow_table && sock_flow_table) {

[PATCH v5 2/2] macvtap: Implement multiqueue for macvtap driver

[Krishna Kumar]

From: Krishna Kumar <krkumar2@in.ibm.com>

Implement multiqueue facility for macvtap driver. The idea is that
a macvtap device can be opened multiple times and the fd's can be
used to register eg, as backend for vhost.

Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com>
---
 drivers/net/macvtap.c      |   99 ++++++++++++++++++++++++++++-------
 include/linux/if_macvlan.h |    9 ++-
 2 files changed, 90 insertions(+), 18 deletions(-)

diff -ruNp org/include/linux/if_macvlan.h new/include/linux/if_macvlan.h
--- org/include/linux/if_macvlan.h	2010-08-04 20:45:19.000000000 +0530
+++ new/include/linux/if_macvlan.h	2010-08-04 20:45:19.000000000 +0530
@@ -40,6 +40,12 @@ struct macvlan_rx_stats {
 	unsigned long		rx_errors;
 };
 
+/*
+ * Maximum times a macvtap device can be opened. This can be used to
+ * configure the number of receive queue, e.g. for multiqueue virtio.
+ */
+#define MAX_MACVTAP_QUEUES	(NR_CPUS < 16 ? NR_CPUS : 16)
+
 struct macvlan_dev {
 	struct net_device	*dev;
 	struct list_head	list;
@@ -50,7 +56,8 @@ struct macvlan_dev {
 	enum macvlan_mode	mode;
 	int (*receive)(struct sk_buff *skb);
 	int (*forward)(struct net_device *dev, struct sk_buff *skb);
-	struct macvtap_queue	*tap;
+	struct macvtap_queue	*taps[MAX_MACVTAP_QUEUES];
+	int			numvtaps;
 };
 
 static inline void macvlan_count_rx(const struct macvlan_dev *vlan,
diff -ruNp org/drivers/net/macvtap.c new/drivers/net/macvtap.c
--- org/drivers/net/macvtap.c	2010-08-04 20:45:19.000000000 +0530
+++ new/drivers/net/macvtap.c	2010-08-04 21:20:49.000000000 +0530
@@ -84,26 +84,45 @@ static const struct proto_ops macvtap_so
 static DEFINE_SPINLOCK(macvtap_lock);
 
 /*
- * Choose the next free queue, for now there is only one
+ * get_slot: return a [unused/occupied] slot in vlan->taps[]:
+ *	- if 'q' is NULL, return the first empty slot;
+ *	- otherwise, return the slot this pointer occupies.
  */
+static int get_slot(struct macvlan_dev *vlan, struct macvtap_queue *q)
+{
+	int i;
+
+	for (i = 0; i < MAX_MACVTAP_QUEUES; i++) {
+		if (rcu_dereference(vlan->taps[i]) == q)
+			return i;
+	}
+
+	/* Should never happen */
+	BUG_ON(1);
+}
+
 static int macvtap_set_queue(struct net_device *dev, struct file *file,
 				struct macvtap_queue *q)
 {
 	struct macvlan_dev *vlan = netdev_priv(dev);
+	int index;
 	int err = -EBUSY;
 
 	spin_lock(&macvtap_lock);
-	if (rcu_dereference(vlan->tap))
+	if (vlan->numvtaps == MAX_MACVTAP_QUEUES)
 		goto out;
 
 	err = 0;
+	index = get_slot(vlan, NULL);
 	rcu_assign_pointer(q->vlan, vlan);
-	rcu_assign_pointer(vlan->tap, q);
+	rcu_assign_pointer(vlan->taps[index], q);
 	sock_hold(&q->sk);
 
 	q->file = file;
 	file->private_data = q;
 
+	vlan->numvtaps++;
+
 out:
 	spin_unlock(&macvtap_lock);
 	return err;
@@ -124,9 +143,12 @@ static void macvtap_put_queue(struct mac
 	spin_lock(&macvtap_lock);
 	vlan = rcu_dereference(q->vlan);
 	if (vlan) {
-		rcu_assign_pointer(vlan->tap, NULL);
+		int index = get_slot(vlan, q);
+
+		rcu_assign_pointer(vlan->taps[index], NULL);
 		rcu_assign_pointer(q->vlan, NULL);
 		sock_put(&q->sk);
+		--vlan->numvtaps;
 	}
 
 	spin_unlock(&macvtap_lock);
@@ -136,39 +158,82 @@ static void macvtap_put_queue(struct mac
 }
 
 /*
- * Since we only support one queue, just dereference the pointer.
+ * Select a queue based on the rxq of the device on which this packet
+ * arrived. If the incoming device is not mq, calculate a flow hash
+ * to select a queue. If all fails, find the first available queue.
+ * Cache vlan->numvtaps since it can become zero during the execution
+ * of this function.
  */
 static struct macvtap_queue *macvtap_get_queue(struct net_device *dev,
 					       struct sk_buff *skb)
 {
 	struct macvlan_dev *vlan = netdev_priv(dev);
+	struct macvtap_queue *tap = NULL;
+	int numvtaps = vlan->numvtaps;
+	__u32 rxq;
+
+	if (!numvtaps)
+		goto out;
+
+	if (likely(skb_rx_queue_recorded(skb))) {
+		rxq = skb_get_rx_queue(skb);
+
+		while (unlikely(rxq >= numvtaps))
+			rxq -= numvtaps;
+
+		tap = rcu_dereference(vlan->taps[rxq]);
+		if (tap)
+			goto out;
+	}
+
+	/* Check if we can use flow to select a queue */
+	rxq = skb_get_rxhash(skb);
+	if (rxq) {
+		tap = rcu_dereference(vlan->taps[rxq % numvtaps]);
+		if (tap)
+			goto out;
+	}
 
-	return rcu_dereference(vlan->tap);
+	/* Everything failed - find first available queue */
+	for (rxq = 0; rxq < MAX_MACVTAP_QUEUES; rxq++) {
+		tap = rcu_dereference(vlan->taps[rxq]);
+		if (tap)
+			break;
+	}
+
+out:
+	return tap;
 }
 
 /*
  * The net_device is going away, give up the reference
- * that it holds on the queue (all the queues one day)
- * and safely set the pointer from the queues to NULL.
+ * that it holds on all queues and safely set the pointer
+ * from the queues to NULL.
  */
 static void macvtap_del_queues(struct net_device *dev)
 {
 	struct macvlan_dev *vlan = netdev_priv(dev);
-	struct macvtap_queue *q;
+	struct macvtap_queue *q, *qlist[MAX_MACVTAP_QUEUES];
+	int i, j = 0;
 
+	/* macvtap_put_queue can free some slots, so go through all slots */
 	spin_lock(&macvtap_lock);
-	q = rcu_dereference(vlan->tap);
-	if (!q) {
-		spin_unlock(&macvtap_lock);
-		return;
+	for (i = 0; i < MAX_MACVTAP_QUEUES && vlan->numvtaps; i++) {
+		q = rcu_dereference(vlan->taps[i]);
+		if (q) {
+			qlist[j++] = q;
+			rcu_assign_pointer(vlan->taps[i], NULL);
+			rcu_assign_pointer(q->vlan, NULL);
+			vlan->numvtaps--;
+		}
 	}
-
-	rcu_assign_pointer(vlan->tap, NULL);
-	rcu_assign_pointer(q->vlan, NULL);
+	BUG_ON(vlan->numvtaps != 0);
 	spin_unlock(&macvtap_lock);
 
 	synchronize_rcu();
-	sock_put(&q->sk);
+
+	for (--j; j >= 0; j--)
+		sock_put(&qlist[j]->sk);
 }
 
 /*

Re: [PATCH v5 1/2] core: Factor out flow calculation from get_rps_cpu

[David Miller]

From: Krishna Kumar <krkumar2@in.ibm.com>
Date: Wed, 04 Aug 2010 21:45:52 +0530

> Factor out flow calculation code from get_rps_cpu, since other
> functions can use the same code.

Applied.

Re: [PATCH v5 2/2] macvtap: Implement multiqueue for macvtap driver

[David Miller]

From: Krishna Kumar <krkumar2@in.ibm.com>
Date: Wed, 04 Aug 2010 21:45:59 +0530

> Implement multiqueue facility for macvtap driver. The idea is that
> a macvtap device can be opened multiple times and the fd's can be
> used to register eg, as backend for vhost.

Applied.