[RFC PATCH net-next 1/3] ppp: convert ppp->rlock to rwlock to improve RX

[RFC PATCH net-next 1/3] ppp: convert ppp->rlock to rwlock to improve RX

[Qingfang Deng]

The rlock spinlock in struct ppp protects the receive path, but it is
frequently used in a read-mostly manner. Converting it to an rwlock_t
allows multiple CPU to enter the receive path and improve its
performance.

Signed-off-by: Qingfang Deng <dqfext@gmail.com>
---
 drivers/net/ppp/ppp_generic.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ppp/ppp_generic.c b/drivers/net/ppp/ppp_generic.c
index 6220866258fc..15e270e9bf36 100644
--- a/drivers/net/ppp/ppp_generic.c
+++ b/drivers/net/ppp/ppp_generic.c
@@ -118,7 +118,7 @@ struct ppp {
 	struct file	*owner;		/* file that owns this unit 48 */
 	struct list_head channels;	/* list of attached channels 4c */
 	int		n_channels;	/* how many channels are attached 54 */
-	spinlock_t	rlock;		/* lock for receive side 58 */
+	rwlock_t	rlock;		/* lock for receive side 58 */
 	spinlock_t	wlock;		/* lock for transmit side 5c */
 	int __percpu	*xmit_recursion; /* xmit recursion detect */
 	int		mru;		/* max receive unit 60 */
@@ -372,12 +372,14 @@ static const int npindex_to_ethertype[NUM_NP] = {
  */
 #define ppp_xmit_lock(ppp)	spin_lock_bh(&(ppp)->wlock)
 #define ppp_xmit_unlock(ppp)	spin_unlock_bh(&(ppp)->wlock)
-#define ppp_recv_lock(ppp)	spin_lock_bh(&(ppp)->rlock)
-#define ppp_recv_unlock(ppp)	spin_unlock_bh(&(ppp)->rlock)
+#define ppp_recv_lock(ppp)	write_lock_bh(&(ppp)->rlock)
+#define ppp_recv_unlock(ppp)	write_unlock_bh(&(ppp)->rlock)
 #define ppp_lock(ppp)		do { ppp_xmit_lock(ppp); \
 				     ppp_recv_lock(ppp); } while (0)
 #define ppp_unlock(ppp)		do { ppp_recv_unlock(ppp); \
 				     ppp_xmit_unlock(ppp); } while (0)
+#define ppp_recv_read_lock(ppp)		read_lock_bh(&(ppp)->rlock)
+#define ppp_recv_read_unlock(ppp)	read_unlock_bh(&(ppp)->rlock)
 
 /*
  * /dev/ppp device routines.
@@ -1252,7 +1254,7 @@ static int ppp_dev_configure(struct net *src_net, struct net_device *dev,
 	for (indx = 0; indx < NUM_NP; ++indx)
 		ppp->npmode[indx] = NPMODE_PASS;
 	INIT_LIST_HEAD(&ppp->channels);
-	spin_lock_init(&ppp->rlock);
+	rwlock_init(&ppp->rlock);
 	spin_lock_init(&ppp->wlock);
 
 	ppp->xmit_recursion = alloc_percpu(int);
@@ -2210,12 +2212,12 @@ struct ppp_mp_skb_parm {
 static inline void
 ppp_do_recv(struct ppp *ppp, struct sk_buff *skb, struct channel *pch)
 {
-	ppp_recv_lock(ppp);
+	ppp_recv_read_lock(ppp);
 	if (!ppp->closing)
 		ppp_receive_frame(ppp, skb, pch);
 	else
 		kfree_skb(skb);
-	ppp_recv_unlock(ppp);
+	ppp_recv_read_unlock(ppp);
 }
 
 /**
-- 
2.43.0

[RFC PATCH net-next 2/3] pppoe: call ppp_input directly when PPPOX_BOUND

[Qingfang Deng]

bh_lock_sock() in sk_receive_skb() is causing unnecessary lock
contensions. When PPP is connected, call ppp_input directly.

Signed-off-by: Qingfang Deng <dqfext@gmail.com>
---
 drivers/net/ppp/pppoe.c | 16 ++++++++++++----
 1 file changed, 12 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ppp/pppoe.c b/drivers/net/ppp/pppoe.c
index 2ea4f4890d23..26f86c9730bb 100644
--- a/drivers/net/ppp/pppoe.c
+++ b/drivers/net/ppp/pppoe.c
@@ -372,9 +372,6 @@ static int pppoe_rcv_core(struct sock *sk, struct sk_buff *skb)
 	 * can't change.
 	 */
 
-	if (skb->pkt_type == PACKET_OTHERHOST)
-		goto abort_kfree;
-
 	if (sk->sk_state & PPPOX_BOUND) {
 		ppp_input(&po->chan, skb);
 	} else if (sk->sk_state & PPPOX_RELAY) {
@@ -416,8 +413,12 @@ static int pppoe_rcv(struct sk_buff *skb, struct net_device *dev,
 	struct pppoe_hdr *ph;
 	struct pppox_sock *po;
 	struct pppoe_net *pn;
+	struct sock *sk;
 	int len;
 
+	if (skb->pkt_type == PACKET_OTHERHOST)
+		goto drop;
+
 	skb = skb_share_check(skb, GFP_ATOMIC);
 	if (!skb)
 		goto out;
@@ -448,7 +449,14 @@ static int pppoe_rcv(struct sk_buff *skb, struct net_device *dev,
 	if (!po)
 		goto drop;
 
-	return sk_receive_skb(sk_pppox(po), skb, 0);
+	sk = sk_pppox(po);
+	if (sk->sk_state & PPPOX_BOUND) {
+		ppp_input(&po->chan, skb);
+		sock_put(sk);
+		return NET_RX_SUCCESS;
+	}
+
+	return sk_receive_skb(sk, skb, 0);
 
 drop:
 	kfree_skb(skb);
-- 
2.43.0

[RFC PATCH net-next 3/3] ppp: synchronize netstats update

[Qingfang Deng]

As the receive path can be run concurrently, synchronize netstats update
by using percpu vars for normal stats and atomic ops for error stats.

Signed-off-by: Qingfang Deng <dqfext@gmail.com>
---
 drivers/net/ppp/ppp_generic.c | 74 +++++++++++++++--------------------
 1 file changed, 32 insertions(+), 42 deletions(-)

diff --git a/drivers/net/ppp/ppp_generic.c b/drivers/net/ppp/ppp_generic.c
index 15e270e9bf36..b133f370a258 100644
--- a/drivers/net/ppp/ppp_generic.c
+++ b/drivers/net/ppp/ppp_generic.c
@@ -95,18 +95,6 @@ struct ppp_file {
 #define PF_TO_PPP(pf)		PF_TO_X(pf, struct ppp)
 #define PF_TO_CHANNEL(pf)	PF_TO_X(pf, struct channel)
 
-/*
- * Data structure to hold primary network stats for which
- * we want to use 64 bit storage.  Other network stats
- * are stored in dev->stats of the ppp strucute.
- */
-struct ppp_link_stats {
-	u64 rx_packets;
-	u64 tx_packets;
-	u64 rx_bytes;
-	u64 tx_bytes;
-};
-
 /*
  * Data structure describing one ppp unit.
  * A ppp unit corresponds to a ppp network interface device
@@ -150,7 +138,6 @@ struct ppp {
 	struct bpf_prog *active_filter; /* filter for pkts to reset idle */
 #endif /* CONFIG_PPP_FILTER */
 	struct net	*ppp_net;	/* the net we belong to */
-	struct ppp_link_stats stats64;	/* 64 bit network stats */
 };
 
 /*
@@ -1484,7 +1471,7 @@ ppp_start_xmit(struct sk_buff *skb, struct net_device *dev)
 
  outf:
 	kfree_skb(skb);
-	++dev->stats.tx_dropped;
+	DEV_STATS_INC(dev, tx_dropped);
 	return NETDEV_TX_OK;
 }
 
@@ -1534,29 +1521,22 @@ ppp_net_siocdevprivate(struct net_device *dev, struct ifreq *ifr,
 static void
 ppp_get_stats64(struct net_device *dev, struct rtnl_link_stats64 *stats64)
 {
-	struct ppp *ppp = netdev_priv(dev);
-
-	ppp_recv_lock(ppp);
-	stats64->rx_packets = ppp->stats64.rx_packets;
-	stats64->rx_bytes   = ppp->stats64.rx_bytes;
-	ppp_recv_unlock(ppp);
-
-	ppp_xmit_lock(ppp);
-	stats64->tx_packets = ppp->stats64.tx_packets;
-	stats64->tx_bytes   = ppp->stats64.tx_bytes;
-	ppp_xmit_unlock(ppp);
-
 	stats64->rx_errors        = dev->stats.rx_errors;
 	stats64->tx_errors        = dev->stats.tx_errors;
 	stats64->rx_dropped       = dev->stats.rx_dropped;
 	stats64->tx_dropped       = dev->stats.tx_dropped;
 	stats64->rx_length_errors = dev->stats.rx_length_errors;
+	dev_fetch_sw_netstats(stats64, dev->tstats);
 }
 
 static int ppp_dev_init(struct net_device *dev)
 {
 	struct ppp *ppp;
 
+	dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
+	if (!dev->tstats)
+		return -ENOMEM;
+
 	netdev_lockdep_set_classes(dev);
 
 	ppp = netdev_priv(dev);
@@ -1586,6 +1566,7 @@ static void ppp_dev_uninit(struct net_device *dev)
 
 	ppp->file.dead = 1;
 	wake_up_interruptible(&ppp->file.rwait);
+	free_percpu(dev->tstats);
 }
 
 static void ppp_dev_priv_destructor(struct net_device *dev)
@@ -1791,8 +1772,7 @@ ppp_send_frame(struct ppp *ppp, struct sk_buff *skb)
 #endif /* CONFIG_PPP_FILTER */
 	}
 
-	++ppp->stats64.tx_packets;
-	ppp->stats64.tx_bytes += skb->len - PPP_PROTO_LEN;
+	dev_sw_netstats_tx_add(ppp->dev, 1, skb->len - PPP_PROTO_LEN);
 
 	switch (proto) {
 	case PPP_IP:
@@ -1868,7 +1848,7 @@ ppp_send_frame(struct ppp *ppp, struct sk_buff *skb)
 
  drop:
 	kfree_skb(skb);
-	++ppp->dev->stats.tx_errors;
+	DEV_STATS_INC(ppp->dev, tx_errors);
 }
 
 /*
@@ -2151,7 +2131,7 @@ static int ppp_mp_explode(struct ppp *ppp, struct sk_buff *skb)
 	spin_unlock(&pch->downl);
 	if (ppp->debug & 1)
 		netdev_err(ppp->dev, "PPP: no memory (fragment)\n");
-	++ppp->dev->stats.tx_errors;
+	DEV_STATS_INC(ppp->dev, tx_errors);
 	++ppp->nxseq;
 	return 1;	/* abandon the frame */
 }
@@ -2313,7 +2293,7 @@ ppp_input(struct ppp_channel *chan, struct sk_buff *skb)
 	if (!ppp_decompress_proto(skb)) {
 		kfree_skb(skb);
 		if (pch->ppp) {
-			++pch->ppp->dev->stats.rx_length_errors;
+			DEV_STATS_INC(pch->ppp->dev, rx_length_errors);
 			ppp_receive_error(pch->ppp);
 		}
 		goto done;
@@ -2384,7 +2364,7 @@ ppp_receive_frame(struct ppp *ppp, struct sk_buff *skb, struct channel *pch)
 static void
 ppp_receive_error(struct ppp *ppp)
 {
-	++ppp->dev->stats.rx_errors;
+	DEV_STATS_INC(ppp->dev, rx_errors);
 	if (ppp->vj)
 		slhc_toss(ppp->vj);
 }
@@ -2469,8 +2449,7 @@ ppp_receive_nonmp_frame(struct ppp *ppp, struct sk_buff *skb)
 		break;
 	}
 
-	++ppp->stats64.rx_packets;
-	ppp->stats64.rx_bytes += skb->len - 2;
+	dev_sw_netstats_rx_add(ppp->dev, skb->len - PPP_PROTO_LEN);
 
 	npi = proto_to_npindex(proto);
 	if (npi < 0) {
@@ -2653,7 +2632,7 @@ ppp_receive_mp_frame(struct ppp *ppp, struct sk_buff *skb, struct channel *pch)
 	 */
 	if (seq_before(seq, ppp->nextseq)) {
 		kfree_skb(skb);
-		++ppp->dev->stats.rx_dropped;
+		DEV_STATS_INC(ppp->dev, rx_dropped);
 		ppp_receive_error(ppp);
 		return;
 	}
@@ -2689,7 +2668,7 @@ ppp_receive_mp_frame(struct ppp *ppp, struct sk_buff *skb, struct channel *pch)
 		if (pskb_may_pull(skb, 2))
 			ppp_receive_nonmp_frame(ppp, skb);
 		else {
-			++ppp->dev->stats.rx_length_errors;
+			DEV_STATS_INC(ppp->dev, rx_length_errors);
 			kfree_skb(skb);
 			ppp_receive_error(ppp);
 		}
@@ -2795,7 +2774,7 @@ ppp_mp_reconstruct(struct ppp *ppp)
 		if (lost == 0 && (PPP_MP_CB(p)->BEbits & E) &&
 		    (PPP_MP_CB(head)->BEbits & B)) {
 			if (len > ppp->mrru + 2) {
-				++ppp->dev->stats.rx_length_errors;
+				DEV_STATS_INC(ppp->dev, rx_length_errors);
 				netdev_printk(KERN_DEBUG, ppp->dev,
 					      "PPP: reconstructed packet"
 					      " is too long (%d)\n", len);
@@ -2850,7 +2829,7 @@ ppp_mp_reconstruct(struct ppp *ppp)
 					      "  missed pkts %u..%u\n",
 					      ppp->nextseq,
 					      PPP_MP_CB(head)->sequence-1);
-			++ppp->dev->stats.rx_dropped;
+			DEV_STATS_INC(ppp->dev, rx_dropped);
 			ppp_receive_error(ppp);
 		}
 
@@ -3299,14 +3278,25 @@ static void
 ppp_get_stats(struct ppp *ppp, struct ppp_stats *st)
 {
 	struct slcompress *vj = ppp->vj;
+	int cpu;
 
 	memset(st, 0, sizeof(*st));
-	st->p.ppp_ipackets = ppp->stats64.rx_packets;
+	for_each_possible_cpu(cpu) {
+		struct pcpu_sw_netstats *p = per_cpu_ptr(ppp->dev->tstats, cpu);
+		u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
+
+		rx_packets = u64_stats_read(&p->rx_packets);
+		rx_bytes = u64_stats_read(&p->rx_bytes);
+		tx_packets = u64_stats_read(&p->tx_packets);
+		tx_bytes = u64_stats_read(&p->tx_bytes);
+
+		st->p.ppp_ipackets += rx_packets;
+		st->p.ppp_ibytes += rx_bytes;
+		st->p.ppp_opackets += tx_packets;
+		st->p.ppp_obytes += tx_bytes;
+	}
 	st->p.ppp_ierrors = ppp->dev->stats.rx_errors;
-	st->p.ppp_ibytes = ppp->stats64.rx_bytes;
-	st->p.ppp_opackets = ppp->stats64.tx_packets;
 	st->p.ppp_oerrors = ppp->dev->stats.tx_errors;
-	st->p.ppp_obytes = ppp->stats64.tx_bytes;
 	if (!vj)
 		return;
 	st->vj.vjs_packets = vj->sls_o_compressed + vj->sls_o_uncompressed;
-- 
2.43.0