Skip to contentSkip to conversation

net: Implement fast TX queue selection

PATCH Applied Krishna Kumar 2 participants 6 messages Oct 20, 2009 – Oct 21, 2009

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

v4 cover net: Implement fast TX queue selection

From: Krishna Kumar <[email protected]>

Notes:
------
1. Eric suggested:
- To use u16 for txq#, but I am using an "int" for now as that avoids one unnecessary subtraction during tx.
- An improvement of caching the txq at connection establishment time so as to use rxq# = txq# (TBD later).
- Drivers can call sk_tx_queue_set() to set the txq if they are going to call skb_tx_hash() internally.
2. v3 & v4 patch stress tested with 1000 netperfs, reboot's, etc.

Changelog [from v3]:
--------------------
1. Changed the order of patches so that the patch setting the txq is moved to the end. This results in bisect-safe patches.
2. Fixed a build failure.

Changelog [from v2]:
--------------------
1. Changed names of functions setting, getting and returning the txq#; and added a new one to reset the txq#.
2. Free sk doesn't need to reset txq#.

Changelog [from v1]:
--------------------
1. Changed IPv6 code to call __sk_dst_reset() directly.
2. Removed the patch re-arranging ("encapsulating") __sk_dst_reset()

Multiqueue cards on routers/firewalls set skb->queue_mapping on input which helps in faster xmit. Implement fast queue selection for locally generated packets also, by saving the txq# for connected sockets (in dev_pick_tx) and use it in subsequent iterations. Locally generated packets for a connection will xmit on the same txq, but routing & firewall loads should not be affected by this patch. Tests shows the distribution across txq's for 1-4 netperf sessions is similar to existing code.

Testing & results:
------------------

1. Cycles/Iter (C/I) used by dev_pick_tx:

         (B -> Billion,   M -> Million)
   |--------------|------------------------|------------------------|
   |              |          ORG           |          NEW           |
   |  Test        |--------|---------|-----|--------|---------|-----|
   |              | Cycles |  Iters  | C/I | Cycles | Iters   | C/I |
   |--------------|--------|---------|-----|--------|---------|-----|
   | [TCP_STREAM, | 3.98 B | 12.47 M | 320 | 1.95 B | 12.92 M | 152 |
   |  UDP_STREAM, |        |         |     |        |         |     |         
   |  TCP_RR,     |        |         |     |        |         |     |        
   |  UDP_RR]     |        |         |     |        |         |     |        
   |--------------|--------|---------|-----|--------|---------|-----|        
   | [TCP_STREAM, | 8.92 B | 29.66 M | 300 | 3.82 B | 38.88 M | 98  |        
   |  TCP_RR,     |        |         |     |        |         |     |         
   |  UDP_RR]     |        |         |     |        |         |     |         
   |--------------|--------|---------|-----|--------|---------|-----|

2. Stress test (over 48 hours) : 1000 netperfs running combination of TCP_STREAM/RR, UDP_STREAM/RR (v4/6, NODELAY/~NODELAY for all tests), with some ssh sessions, reboots, modprobe -r driver, etc.

3. Performance test (10 hours): Single 10 hour netperf run of TCP_STREAM/RR, TCP_STREAM + NO_DELAY and UDP_RR. Results show an improvement in both performance and cpu utilization.

Tested on a 4-processor AMD Opteron 2.8 GHz system with 1GB memory, 10G Chelsio card. Each BW number is the sum of 3 iterations of individual tests using 512, 16K, 64K & 128K I/O sizes, in Mb/s:

------------------------ TCP Tests -----------------------
#procs Org BW New BW (%) Org SD New SD (%)
------------------------------------------------------------

1       77777.7    81011.0 (4.15)    42.3     40.2 (-5.11)
4       91599.2    91878.8 (.30)    955.9    919.3 (-3.83)
6       89533.3    91792.2 (2.52)  2262.0   2143.0 (-5.25)
8       87507.5    89161.9 (1.89)  4363.4   4073.6 (-6.64)
10      85152.4    85607.8 (.53)   6890.4   6851.2 (-.56)

------------------------------------------------------------

------------------------- TCP NO_DELAY Tests ---------------
#procs Org BW New BW (%) Org SD New SD (%)
------------------------------------------------------------

1       57001.9    57888.0 (1.55)     67.7      70.2 (3.75)
4       69555.1    69957.4 (.57)     823.0     834.3 (1.36)
6       71359.3    71918.7 (.78)    1740.8    1724.5 (-.93)
8       72577.6    72496.1 (-.11)   2955.4    2937.7 (-.59)
10      70829.6    71444.2 (.86)    4826.1    4673.4 (-3.16)

------------------------------------------------------------

----------------------- Request Response Tests --------------------
#procs Org TPS New TPS (%) Org SD New SD (%)
(1-10)
-------------------------------------------------------------------

TCP     1019245.9   1042626.4 (2.29) 16352.9   16459.8 (.65)
UDP     934598.64   942956.9  (.89)  11607.3   11593.2 (-.12)

-------------------------------------------------------------------

Thanks,

- KK

---

#1 lore raw

PATCH v4 1/4 net: Introduce sk_tx_queue_mapping

From: Krishna Kumar <[email protected]>

Introduce sk_tx_queue_mapping; and functions that set, test and get this value. Reset sk_tx_queue_mapping to -1 whenever the dst cache is set/reset, and in socket alloc. Setting txq to -1 and using valid txq=<0 to n-1> allows the tx path to use the value of sk_tx_queue_mapping directly instead of subtracting 1 on every tx.

---
 include/net/sock.h |   26 ++++++++++++++++++++++++++
 net/core/sock.c    |    5 ++++-
 2 files changed, 30 insertions(+), 1 deletion(-)
include/net/sock.h +26 -0
diff -ruNp org/include/net/sock.h new/include/net/sock.h
@@ -107,6 +107,7 @@ struct net;
  *	@skc_node: main hash linkage for various protocol lookup tables
  *	@skc_nulls_node: main hash linkage for UDP/UDP-Lite protocol
  *	@skc_refcnt: reference count
+ *	@skc_tx_queue_mapping: tx queue number for this connection
  *	@skc_hash: hash value used with various protocol lookup tables
  *	@skc_family: network address family
  *	@skc_state: Connection state
@@ -128,6 +129,7 @@ struct sock_common {
 		struct hlist_nulls_node skc_nulls_node;
 	};
 	atomic_t		skc_refcnt;
+	int			skc_tx_queue_mapping;
 
 	unsigned int		skc_hash;
 	unsigned short		skc_family;
@@ -215,6 +217,7 @@ struct sock {
 #define sk_node			__sk_common.skc_node
 #define sk_nulls_node		__sk_common.skc_nulls_node
 #define sk_refcnt		__sk_common.skc_refcnt
+#define sk_tx_queue_mapping	__sk_common.skc_tx_queue_mapping
 
 #define sk_copy_start		__sk_common.skc_hash
 #define sk_hash			__sk_common.skc_hash
@@ -1094,8 +1097,29 @@ static inline void sock_put(struct sock 
 extern int sk_receive_skb(struct sock *sk, struct sk_buff *skb,
 			  const int nested);
 
+static inline void sk_tx_queue_set(struct sock *sk, int tx_queue)
+{
+	sk->sk_tx_queue_mapping = tx_queue;
+}
+
+static inline void sk_tx_queue_clear(struct sock *sk)
+{
+	sk->sk_tx_queue_mapping = -1;
+}
+
+static inline int sk_tx_queue_get(const struct sock *sk)
+{
+	return sk->sk_tx_queue_mapping;
+}
+
+static inline bool sk_tx_queue_recorded(const struct sock *sk)
+{
+	return (sk && sk->sk_tx_queue_mapping >= 0);
+}
+
 static inline void sk_set_socket(struct sock *sk, struct socket *sock)
 {
+	sk_tx_queue_clear(sk);
 	sk->sk_socket = sock;
 }
 
@@ -1152,6 +1176,7 @@ __sk_dst_set(struct sock *sk, struct dst
 {
 	struct dst_entry *old_dst;
 
+	sk_tx_queue_clear(sk);
 	old_dst = sk->sk_dst_cache;
 	sk->sk_dst_cache = dst;
 	dst_release(old_dst);
@@ -1170,6 +1195,7 @@ __sk_dst_reset(struct sock *sk)
 {
 	struct dst_entry *old_dst;
 
+	sk_tx_queue_clear(sk);
 	old_dst = sk->sk_dst_cache;
 	sk->sk_dst_cache = NULL;
 	dst_release(old_dst);
net/core/sock.c +4 -1
diff -ruNp org/net/core/sock.c new/net/core/sock.c
@@ -357,6 +357,7 @@ struct dst_entry *__sk_dst_check(struct 
 	struct dst_entry *dst = sk->sk_dst_cache;
 
 	if (dst && dst->obsolete && dst->ops->check(dst, cookie) == NULL) {
+		sk_tx_queue_clear(sk);
 		sk->sk_dst_cache = NULL;
 		dst_release(dst);
 		return NULL;
@@ -953,7 +954,8 @@ static void sock_copy(struct sock *nsk, 
 	void *sptr = nsk->sk_security;
 #endif
 	BUILD_BUG_ON(offsetof(struct sock, sk_copy_start) !=
-		     sizeof(osk->sk_node) + sizeof(osk->sk_refcnt));
+		     sizeof(osk->sk_node) + sizeof(osk->sk_refcnt) +
+		     sizeof(osk->sk_tx_queue_mapping));
 	memcpy(&nsk->sk_copy_start, &osk->sk_copy_start,
 	       osk->sk_prot->obj_size - offsetof(struct sock, sk_copy_start));
 #ifdef CONFIG_SECURITY_NETWORK
@@ -997,6 +999,7 @@ static struct sock *sk_prot_alloc(struct
 
 		if (!try_module_get(prot->owner))
 			goto out_free_sec;
+		sk_tx_queue_clear(sk);
 	}
 
 	return sk;
#2 lore raw

PATCH v4 2/4 net: IPv6 changes

In reply to Krishna Kumar

From: Krishna Kumar <[email protected]>

IPv6: Reset sk_tx_queue_mapping when dst_cache is reset. Use existing macro to do the work.

---
 net/ipv6/inet6_connection_sock.c |    3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
net/ipv6/inet6_connection_sock.c +1 -2
diff -ruNp org/net/ipv6/inet6_connection_sock.c new/net/ipv6/inet6_connection_sock.c
@@ -168,8 +168,7 @@ struct dst_entry *__inet6_csk_dst_check(
 	if (dst) {
 		struct rt6_info *rt = (struct rt6_info *)dst;
 		if (rt->rt6i_flow_cache_genid != atomic_read(&flow_cache_genid)) {
-			sk->sk_dst_cache = NULL;
-			dst_release(dst);
+			__sk_dst_reset(sk);
 			dst = NULL;
 		}
 	}
#3 lore raw

PATCH v4 3/4 net: Fix for dst_negative_advice

In reply to Krishna Kumar

From: Krishna Kumar <[email protected]>

dst_negative_advice() should check for changed dst and reset sk_tx_queue_mapping accordingly. Pass sock to the callers of dst_negative_advice.

(sk_reset_txq is defined just for use by dst_negative_advice. The only way I could find to get around this is to move dst_negative_() from dst.h to dst.c, include sock.h in dst.c, etc)

---
 include/net/dst.h      |   12 ++++++++++--
 net/core/sock.c        |    6 ++++++
 net/dccp/timer.c       |    4 ++--
 net/decnet/af_decnet.c |    2 +-
 net/ipv4/tcp_timer.c   |    4 ++--
 5 files changed, 21 insertions(+), 7 deletions(-)
include/net/dst.h +10 -2
diff -ruNp org/include/net/dst.h new/include/net/dst.h
@@ -222,11 +222,19 @@ static inline void dst_confirm(struct ds
 		neigh_confirm(dst->neighbour);
 }
 
-static inline void dst_negative_advice(struct dst_entry **dst_p)
+static inline void dst_negative_advice(struct dst_entry **dst_p,
+				       struct sock *sk)
 {
 	struct dst_entry * dst = *dst_p;
-	if (dst && dst->ops->negative_advice)
+	if (dst && dst->ops->negative_advice) {
 		*dst_p = dst->ops->negative_advice(dst);
+
+		if (dst != *dst_p) {
+			extern void sk_reset_txq(struct sock *sk);
+
+			sk_reset_txq(sk);
+		}
+	}
 }
 
 static inline void dst_link_failure(struct sk_buff *skb)
net/core/sock.c +6 -0
diff -ruNp org/net/core/sock.c new/net/core/sock.c
@@ -352,6 +352,12 @@ discard_and_relse:
 }
 EXPORT_SYMBOL(sk_receive_skb);
 
+void sk_reset_txq(struct sock *sk)
+{
+	sk_tx_queue_clear(sk);
+}
+EXPORT_SYMBOL(sk_reset_txq);
+
 struct dst_entry *__sk_dst_check(struct sock *sk, u32 cookie)
 {
 	struct dst_entry *dst = sk->sk_dst_cache;
net/dccp/timer.c +2 -2
diff -ruNp org/net/dccp/timer.c new/net/dccp/timer.c
@@ -38,7 +38,7 @@ static int dccp_write_timeout(struct soc
 
 	if (sk->sk_state == DCCP_REQUESTING || sk->sk_state == DCCP_PARTOPEN) {
 		if (icsk->icsk_retransmits != 0)
-			dst_negative_advice(&sk->sk_dst_cache);
+			dst_negative_advice(&sk->sk_dst_cache, sk);
 		retry_until = icsk->icsk_syn_retries ?
 			    : sysctl_dccp_request_retries;
 	} else {
@@ -63,7 +63,7 @@ static int dccp_write_timeout(struct soc
 			   Golden words :-).
 		   */
 
-			dst_negative_advice(&sk->sk_dst_cache);
+			dst_negative_advice(&sk->sk_dst_cache, sk);
 		}
 
 		retry_until = sysctl_dccp_retries2;
net/decnet/af_decnet.c +1 -1
diff -ruNp org/net/decnet/af_decnet.c new/net/decnet/af_decnet.c
@@ -1955,7 +1955,7 @@ static int dn_sendmsg(struct kiocb *iocb
 	}
 
 	if ((flags & MSG_TRYHARD) && sk->sk_dst_cache)
-		dst_negative_advice(&sk->sk_dst_cache);
+		dst_negative_advice(&sk->sk_dst_cache, sk);
 
 	mss = scp->segsize_rem;
 	fctype = scp->services_rem & NSP_FC_MASK;
net/ipv4/tcp_timer.c +2 -2
diff -ruNp org/net/ipv4/tcp_timer.c new/net/ipv4/tcp_timer.c
@@ -141,14 +141,14 @@ static int tcp_write_timeout(struct sock
 
 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 		if (icsk->icsk_retransmits)
-			dst_negative_advice(&sk->sk_dst_cache);
+			dst_negative_advice(&sk->sk_dst_cache, sk);
 		retry_until = icsk->icsk_syn_retries ? : sysctl_tcp_syn_retries;
 	} else {
 		if (retransmits_timed_out(sk, sysctl_tcp_retries1)) {
 			/* Black hole detection */
 			tcp_mtu_probing(icsk, sk);
 
-			dst_negative_advice(&sk->sk_dst_cache);
+			dst_negative_advice(&sk->sk_dst_cache, sk);
 		}
 
 		retry_until = sysctl_tcp_retries2;
#4 lore raw

PATCH v4 4/4 net: Use sk_tx_queue_mapping for connected sockets

In reply to Krishna Kumar

From: Krishna Kumar <[email protected]>

For connected sockets, the first run of dev_pick_tx saves the calculated txq in sk_tx_queue_mapping. This is not saved if either the device has a queue select or the socket is not connected. Next iterations of dev_pick_tx uses the cached value of sk_tx_queue_mapping.

---
 net/core/dev.c |   24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)
net/core/dev.c +18 -6
diff -ruNp org/net/core/dev.c new/net/core/dev.c
@@ -1791,13 +1791,25 @@ EXPORT_SYMBOL(skb_tx_hash);
 static struct netdev_queue *dev_pick_tx(struct net_device *dev,
 					struct sk_buff *skb)
 {
-	const struct net_device_ops *ops = dev->netdev_ops;
-	u16 queue_index = 0;
+	u16 queue_index;
+	struct sock *sk = skb->sk;
+
+	if (sk_tx_queue_recorded(sk)) {
+		queue_index = sk_tx_queue_get(sk);
+	} else {
+		const struct net_device_ops *ops = dev->netdev_ops;
 
-	if (ops->ndo_select_queue)
-		queue_index = ops->ndo_select_queue(dev, skb);
-	else if (dev->real_num_tx_queues > 1)
-		queue_index = skb_tx_hash(dev, skb);
+		if (ops->ndo_select_queue) {
+			queue_index = ops->ndo_select_queue(dev, skb);
+		} else {
+			queue_index = 0;
+			if (dev->real_num_tx_queues > 1)
+				queue_index = skb_tx_hash(dev, skb);
+
+			if (sk && sk->sk_dst_cache)
+				sk_tx_queue_set(sk, queue_index);
+		}
+	}
 
 	skb_set_queue_mapping(skb, queue_index);
 	return netdev_get_tx_queue(dev, queue_index);