Message-ID: <20110301223458.GA24403@yandex-team.ru>
Date:	Wed, 2 Mar 2011 01:34:58 +0300
From:	"Oleg V. Ukhno" <olegu@...dex-team.ru>
To:	netdev@...r.kernel.org
Cc:	Jay Vosburgh <fubar@...ibm.com>,
	"David S. Miller" <davem@...emloft.net>
Subject: [PATCH] bonding: added 802.3ad round-robin hashing policy and
 source mac selection mode

Patch introduces two new (related) features to bonding module.
First feature is round-robin hashing policy, which is primarily
intended for use with 802.3ad mode, and puts every next IPv4 and
IPv6 packet into  next availables slave without taling into account
which layer3 and above protocol is used.
Second feature makes possible choosing which MAC-address will be set
in the transmitted packet - when set to src-mac it will force setting
slave's interface real MAC address as source MAC address in every
packet, sent via this slave interface.
Main goal of this patch is to make possible single TCP stream
equally striped for both transmitted and received packets over all
available slaves.
This operating mode is not fully 802.3ad compliant, and will cause
some packet reordering in TCP stream, to some kernel tuning may be
required.
For correct working enabling round-robin hashing policy plus using
real slave's MAC addresses as source MAC addresses in transmitted
packets requires specific switch setting)hashing mode for port-channel
("etherchannel) should be set to src-mac or src-dst-mac to get
correct load-striping on the receiving host's etherchannel.
General requirements for using bonding in this operating mode are:
- even and preferrably equal number of slaves on sending and receiving
hosts;
- equal RTT between sending and receiving hosts on all slaves;
- switch capable of doing etherchannels and using src-mac or src-dst-mac
hashing policy for egress load striping

Signed-off-by: Oleg V. Ukhno <olegu@...dex-team.ru>
---
 
Documentation/networking/bonding.txt |  109 +++++++++++++++++++++++++++++++++++
 drivers/net/bonding/bond_3ad.c       |    2 
 drivers/net/bonding/bond_main.c      |   59 +++++++++++++++++-
 drivers/net/bonding/bond_sysfs.c     |   50 ++++++++++++++++
 drivers/net/bonding/bonding.h        |   11 ++-
 include/linux/if_bonding.h           |    1 
 6 files changed, 223 insertions(+), 9 deletions(-)

diff -uprN -X linux-2.6/Documentation/dontdiff linux-2.6/Documentation/networking/bonding.txt linux-2.6.p/Documentation/networking/bonding.txt
--- linux-2.6/Documentation/networking/bonding.txt	2011-02-08 16:03:01.290281998 +0300
+++ linux-2.6.p/Documentation/networking/bonding.txt	2011-03-01 22:27:56.570282000 +0300
@@ -83,6 +83,7 @@ Table of Contents
 12. Configuring Bonding for Maximum Throughput
 12.1	Maximum Throughput in a Single Switch Topology
 12.1.1		MT Bonding Mode Selection for Single Switch Topology
+12.1.1.1 		Maximizing TCP Throughput for RX/TX for Single Switch Topology using layer2 mechanisms
 12.1.2		MT Link Monitoring for Single Switch Topology
 12.2	Maximum Throughput in a Multiple Switch Topology
 12.2.1		MT Bonding Mode Selection for Multiple Switch Topology
@@ -761,11 +762,62 @@ xmit_hash_policy
 		conversations.  Other implementations of 802.3ad may
 		or may not tolerate this noncompliance.
 
+	round-robin
+
+		This policy simply puts every next packet into next
+		slave interfaces, providing round-robin load striping
+		for transmitted data. This policy can be enabled with
+		any mode which supports choosing alternate hash policy,
+		but was initially done for 802.3ad mode.
+
+		Main goal for this policy is to stripe TX load without
+		taking into account which layer3 protocol is used, and
+		can be used for single TCP connection load striping. When
+		enabled, it will round-robin packets for IPv4 and IPv6
+		only.
+
+		There is also src_mac_select option, which can be used
+		to configure RX load-striping using switch hashing
+		algorhytms on the receiving side. See detailed description
+		below.
+
+		This algorithm is not 802.3ad compliant.  This hash
+		policy will generally cause TCP packets to be delivered
+		out of order.  Other implementations of 802.3ad may or
+		may not tolerate this noncompliance.
+
 	The default value is layer2.  This option was added in bonding
 	version 2.6.3.  In earlier versions of bonding, this parameter
 	does not exist, and the layer2 policy is the only policy.  The
 	layer2+3 value was added for bonding version 3.2.2.
 
+src_mac_select
+
+	Specifies the source MAC selection method for outgoing packets.
+
+	Possible values are:
+
+	default or 0
+
+		The normal selection policy for the bonding mode is
+		used.  This varies according to the bonding mode.  The
+		balance-xor, balance-rr, 802.3ad and broadcast modes use
+		the MAC address of the bonding master.  The balance-alb
+		and balance-tlb modes use the MAC address of the slave
+		the packet is sent on.
+
+	slave-src or 1
+
+		Sets the source MAC for all outgoing IPv4 and IPv6
+		packets to the MAC address of the slave the packet is
+		sent on.  This is intended to permit fine grained load
+		balancing for 802.3ad and balance-xor modes by changing
+		the slave's MAC addresses.  This is documented in detail
+		in section 12.1.1.1.
+
+	This option was added for bonding version 3.8.0
+
+
 resend_igmp
 
 	Specifies the number of IGMP membership reports to be issued after
@@ -2190,6 +2242,63 @@ balance-alb: This mode is everything tha
 	device driver must support changing the hardware address while
 	the device is open.
 
+12.1.1.1 Maximizing TCP Throughput for RX/TX for Single Switch Topology
+         using layer2 mechanisms
+----------------------------------------------------------------------
+
+	It is also possible to use round-robin packet transmission,
+either in the balance-rr mode or using the round-robin xmit_hash_policy
+setting for balance-xor or 802.3ad modes to evenly stripe traffic across
+the set of slaves.  In conjuction with the src_mac_select option's
+"slave-src" value and a carefully configured network, it is possible to
+achieve high throughput using round robin.
+
+	The network here involves multiple hosts running bonding, all
+connected to a common switch, for example:
+
+  +--------+             +------------+             +--------+
+  |        |eth0    port1|            |port3    eth0|        |
+  | Host A +-------------+  switch    +-------------+ Host B |
+  | bond0  +-------------+            +-------------+ bond0  |
+  |        |eth1    port2|            |port4    eth1|        |
+  +--------+             +------------+             +--------+
+
+	In this configuration, the switch has ports 1 and 2 in one port
+channel group, and ports 3 and 4 in another port channel group.
+These channel groups are set to hash transmitted packets according to
+the source MAC address.
+
+	Host A and Host B each have a bond utilizing a round-robin
+transmit scheme, with src_mac_select set to "slave-src" and each slave's
+MAC address chosen to hash sequentially when run through the switch's
+source MAC address egress hash.
+
+	For clarity, only two slaves / ports are shown in the diagram.
+
+	In this manner, traffic sent from Host A to Host B will
+generally be evenly striped all the way through to Host B.  It is first
+sent round robin by Host A.  The switch will then hash on the source MAC
+address of the packets, whose MAC addresses have been manually selected
+to hash sequentially, thus the switch's egress port channel will tend to
+keep the traffic separated by port similarly to how it arrived (e.g.,
+packets from eth0 arriving on port1 will tend to go out port3, and
+simiarly for eth1, port2 and port4).
+
+	This scheme works best when the number of slaves and switch
+ports match at both ends.  Note that it will work with differing numbers
+of slaves, but the traffic balance may not be optimal, and it may be
+possible for a host with more slaves to overrun a host with fewer
+slaves.
+
+	According to usage experience and test results, balancing multiple
+(8-16) TCP sessions across bonded interface of 2 slaves utilizes 1.9-2.0
+Gbps in both directions, but with 4 slaves in bond - only 3.2-3.5 Gbps.
+When doing single unidirectional data transfer usually full aggregate
+bandwidth utilization is possible.
+
+	This usage scenario requires special switch configuration, as well
+as tuning TCP reordering related sysctl parameters.
+
 12.1.2 MT Link Monitoring for Single Switch Topology
 ----------------------------------------------------
 
diff -uprN -X linux-2.6/Documentation/dontdiff linux-2.6/drivers/net/bonding/bond_3ad.c linux-2.6.p/drivers/net/bonding/bond_3ad.c
--- linux-2.6/drivers/net/bonding/bond_3ad.c	2011-02-16 00:59:18.710282002 +0300
+++ linux-2.6.p/drivers/net/bonding/bond_3ad.c	2011-03-01 22:57:17.530282004 +0300
@@ -2419,7 +2419,7 @@ int bond_3ad_xmit_xor(struct sk_buff *sk
 		goto out;
 	}
 
-	slave_agg_no = bond->xmit_hash_policy(skb, slaves_in_agg);
+	slave_agg_no = bond->xmit_hash_policy(skb, slaves_in_agg, dev);
 
 	bond_for_each_slave(bond, slave, i) {
 		struct aggregator *agg = SLAVE_AD_INFO(slave).port.aggregator;
diff -uprN -X linux-2.6/Documentation/dontdiff linux-2.6/drivers/net/bonding/bonding.h linux-2.6.p/drivers/net/bonding/bonding.h
--- linux-2.6/drivers/net/bonding/bonding.h	2011-02-16 00:59:18.720282002 +0300
+++ linux-2.6.p/drivers/net/bonding/bonding.h	2011-03-01 22:53:16.160281999 +0300
@@ -23,8 +23,8 @@
 #include "bond_3ad.h"
 #include "bond_alb.h"
 
-#define DRV_VERSION	"3.7.0"
-#define DRV_RELDATE	"June 2, 2010"
+#define DRV_VERSION	"3.8.0"
+#define DRV_RELDATE	"March 3, 2011"
 #define DRV_NAME	"bonding"
 #define DRV_DESCRIPTION	"Ethernet Channel Bonding Driver"
 
@@ -162,6 +162,7 @@ struct bond_params {
 	int tx_queues;
 	int all_slaves_active;
 	int resend_igmp;
+	int src_mac_select;
 };
 
 struct bond_parm_tbl {
@@ -235,7 +236,7 @@ struct bonding {
 #endif /* CONFIG_PROC_FS */
 	struct   list_head bond_list;
 	struct   netdev_hw_addr_list mc_list;
-	int      (*xmit_hash_policy)(struct sk_buff *, int);
+	int      (*xmit_hash_policy)(struct sk_buff *, int, struct net_device *);
 	__be32   master_ip;
 	u16      flags;
 	u16      rr_tx_counter;
@@ -308,6 +309,9 @@ static inline bool bond_is_lb(const stru
 #define BOND_ARP_VALIDATE_ALL		(BOND_ARP_VALIDATE_ACTIVE | \
 					 BOND_ARP_VALIDATE_BACKUP)
 
+#define BOND_SRC_MAC_DEFAULT		0
+#define BOND_SRC_MAC_SLAVE		1
+
 static inline int slave_do_arp_validate(struct bonding *bond,
 					struct slave *slave)
 {
@@ -402,6 +406,7 @@ extern const struct bond_parm_tbl arp_va
 extern const struct bond_parm_tbl fail_over_mac_tbl[];
 extern const struct bond_parm_tbl pri_reselect_tbl[];
 extern struct bond_parm_tbl ad_select_tbl[];
+extern const struct bond_parm_tbl src_mac_select_tbl[];
 
 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
 void bond_send_unsolicited_na(struct bonding *bond);
diff -uprN -X linux-2.6/Documentation/dontdiff linux-2.6/drivers/net/bonding/bond_main.c linux-2.6.p/drivers/net/bonding/bond_main.c
--- linux-2.6/drivers/net/bonding/bond_main.c	2011-02-16 00:59:18.720282002 +0300
+++ linux-2.6.p/drivers/net/bonding/bond_main.c	2011-03-01 23:00:24.770282003 +0300
@@ -111,6 +111,7 @@ static char *fail_over_mac;
 static int all_slaves_active = 0;
 static struct bond_params bonding_defaults;
 static int resend_igmp = BOND_DEFAULT_RESEND_IGMP;
+static char *src_mac_select;
 
 module_param(max_bonds, int, 0);
 MODULE_PARM_DESC(max_bonds, "Max number of bonded devices");
@@ -152,7 +153,7 @@ module_param(ad_select, charp, 0);
 MODULE_PARM_DESC(ad_select, "803.ad aggregation selection logic: stable (0, default), bandwidth (1), count (2)");
 module_param(xmit_hash_policy, charp, 0);
 MODULE_PARM_DESC(xmit_hash_policy, "XOR hashing method: 0 for layer 2 (default)"
-				   ", 1 for layer 3+4");
+				   ", 1 for layer 3+4, 3 for round-robin");
 module_param(arp_interval, int, 0);
 MODULE_PARM_DESC(arp_interval, "arp interval in milliseconds");
 module_param_array(arp_ip_target, charp, NULL, 0);
@@ -167,6 +168,9 @@ MODULE_PARM_DESC(all_slaves_active, "Kee
 				     "0 for never (default), 1 for always.");
 module_param(resend_igmp, int, 0);
 MODULE_PARM_DESC(resend_igmp, "Number of IGMP membership reports to send on link failure");
+module_param(src_mac_select, charp, 0);
+MODULE_PARM_DESC(src_mac_select, "Source MAC selection mode: 0  or default (default),"
+				 "1 or slave-src to use slave's MAC as packet's src MAC");
 
 /*----------------------------- Global variables ----------------------------*/
 
@@ -206,6 +210,7 @@ const struct bond_parm_tbl xmit_hashtype
 {	"layer2",		BOND_XMIT_POLICY_LAYER2},
 {	"layer3+4",		BOND_XMIT_POLICY_LAYER34},
 {	"layer2+3",		BOND_XMIT_POLICY_LAYER23},
+{	"round-robin",		BOND_XMIT_POLICY_RR},
 {	NULL,			-1},
 };
 
@@ -238,6 +243,12 @@ struct bond_parm_tbl ad_select_tbl[] = {
 {	NULL,		-1},
 };
 
+const struct bond_parm_tbl src_mac_select_tbl[] = {
+{	"default",	BOND_SRC_MAC_DEFAULT},
+{	"slave-src",	BOND_SRC_MAC_SLAVE},
+{	NULL,		-1},
+};
+
 /*-------------------------- Forward declarations ---------------------------*/
 
 static void bond_send_gratuitous_arp(struct bonding *bond);
@@ -422,6 +433,7 @@ struct vlan_entry *bond_next_vlan(struct
 int bond_dev_queue_xmit(struct bonding *bond, struct sk_buff *skb,
 			struct net_device *slave_dev)
 {
+	struct ethhdr *eth_data;
 	skb->dev = slave_dev;
 	skb->priority = 1;
 #ifdef CONFIG_NET_POLL_CONTROLLER
@@ -433,6 +445,15 @@ int bond_dev_queue_xmit(struct bonding *
 		slave_dev->priv_flags &= ~IFF_IN_NETPOLL;
 	} else
 #endif
+		if (bond->params.src_mac_select == BOND_SRC_MAC_SLAVE &&
+		   (skb->protocol == htons(ETH_P_IP) ||
+		   skb->protocol == htons(ETH_P_IPV6))) {
+			skb_reset_mac_header(skb);
+			eth_data = eth_hdr(skb);
+			memcpy(eth_data->h_source, slave_dev->perm_addr,
+				ETH_ALEN);
+		}
+
 		dev_queue_xmit(skb);
 
 	return 0;
@@ -3261,6 +3282,10 @@ static void bond_info_show_master(struct
 			bond->params.xmit_policy);
 	}
 
+	seq_printf(seq, "Source MAC select is: %s (%d)\n",
+		src_mac_select_tbl[bond->params.src_mac_select].modename,
+		bond->params.src_mac_select);
+
 	if (USES_PRIMARY(bond->params.mode)) {
 		seq_printf(seq, "Primary Slave: %s",
 			   (bond->primary_slave) ?
@@ -3717,7 +3742,8 @@ void bond_unregister_arp(struct bonding
  * Hash for the output device based upon layer 2 and layer 3 data. If
  * the packet is not IP mimic bond_xmit_hash_policy_l2()
  */
-static int bond_xmit_hash_policy_l23(struct sk_buff *skb, int count)
+static int bond_xmit_hash_policy_l23(struct sk_buff *skb, int count,
+					 struct net_device *bond_dev)
 {
 	struct ethhdr *data = (struct ethhdr *)skb->data;
 	struct iphdr *iph = ip_hdr(skb);
@@ -3735,7 +3761,8 @@ static int bond_xmit_hash_policy_l23(str
  * the packet is a frag or not TCP or UDP, just use layer 3 data.  If it is
  * altogether not IP, mimic bond_xmit_hash_policy_l2()
  */
-static int bond_xmit_hash_policy_l34(struct sk_buff *skb, int count)
+static int bond_xmit_hash_policy_l34(struct sk_buff *skb, int count,
+					 struct net_device *bond_dev)
 {
 	struct ethhdr *data = (struct ethhdr *)skb->data;
 	struct iphdr *iph = ip_hdr(skb);
@@ -3759,13 +3786,31 @@ static int bond_xmit_hash_policy_l34(str
 /*
  * Hash for the output device based upon layer 2 data
  */
-static int bond_xmit_hash_policy_l2(struct sk_buff *skb, int count)
+static int bond_xmit_hash_policy_l2(struct sk_buff *skb, int count,
+					 struct net_device *bond_dev)
 {
 	struct ethhdr *data = (struct ethhdr *)skb->data;
 
 	return (data->h_dest[5] ^ data->h_source[5]) % count;
 }
 
+/*
+ * Round-robin over all active slaves(one packet per slave) for IP and IPv6,
+ * otherwise mimic bond_xmit_hash_policy_l2() for IP IGMP traffic
+ */
+static int bond_xmit_hash_policy_rr(struct sk_buff *skb, int count,
+					 struct net_device *bond_dev)
+{
+	struct ethhdr *data = (struct ethhdr *)skb->data;
+	struct iphdr *iph = ip_hdr(skb);
+	struct bonding *bond = netdev_priv(bond_dev);
+	if ((iph->protocol == IPPROTO_IGMP) &&
+		(skb->protocol == htons(ETH_P_IP))) {
+			return (data->h_dest[5] ^ data->h_source[5]) % count;
+	}
+	return bond->rr_tx_counter++ % count;
+}
+
 /*-------------------------- Device entry points ----------------------------*/
 
 static int bond_open(struct net_device *bond_dev)
@@ -4395,7 +4440,8 @@ static int bond_xmit_xor(struct sk_buff
 	if (!BOND_IS_OK(bond))
 		goto out;
 
-	slave_no = bond->xmit_hash_policy(skb, bond->slave_cnt);
+	slave_no = bond->xmit_hash_policy(skb, bond->slave_cnt,
+					bond_dev);
 
 	bond_for_each_slave(bond, slave, i) {
 		slave_no--;
@@ -4492,6 +4538,9 @@ static void bond_set_xmit_hash_policy(st
 	case BOND_XMIT_POLICY_LAYER34:
 		bond->xmit_hash_policy = bond_xmit_hash_policy_l34;
 		break;
+	case BOND_XMIT_POLICY_RR:
+		bond->xmit_hash_policy = bond_xmit_hash_policy_rr;
+		break;
 	case BOND_XMIT_POLICY_LAYER2:
 	default:
 		bond->xmit_hash_policy = bond_xmit_hash_policy_l2;
diff -uprN -X linux-2.6/Documentation/dontdiff linux-2.6/drivers/net/bonding/bond_sysfs.c linux-2.6.p/drivers/net/bonding/bond_sysfs.c
--- linux-2.6/drivers/net/bonding/bond_sysfs.c	2011-02-08 16:03:02.950282003 +0300
+++ linux-2.6.p/drivers/net/bonding/bond_sysfs.c	2011-02-16 02:05:58.650281999 +0300
@@ -1643,6 +1643,55 @@ out:
 static DEVICE_ATTR(resend_igmp, S_IRUGO | S_IWUSR,
 		   bonding_show_resend_igmp, bonding_store_resend_igmp);
 
+/*
+ * Show and set the bonding src_mac_select param.
+ */
+
+static ssize_t bonding_show_src_mac_select(struct device *d,
+					struct device_attribute *attr,
+					char *buf)
+{
+	struct bonding *bond = to_bond(d);
+
+	return sprintf(buf, "%s %d\n",
+			src_mac_select_tbl[bond->params.src_mac_select].modename,
+			bond->params.src_mac_select);
+}
+
+static ssize_t bonding_store_src_mac_select(struct device *d,
+					struct device_attribute *attr,
+					const char *buf, size_t count)
+{
+	int new_value, ret = count;
+	struct bonding *bond = to_bond(d);
+
+	if (bond->dev->flags & IFF_UP) {
+		pr_err("%s: Interface is up. Unable to update src mac select policy.\n",
+			bond->dev->name);
+		ret = -EPERM;
+		goto out;
+	}
+
+	new_value = bond_parse_parm(buf, src_mac_select_tbl);
+	if (new_value < 0)  {
+		pr_err("%s: Ignoring invalid src mac select policy value %.*s.\n",
+			bond->dev->name,
+			(int)strlen(buf) - 1, buf);
+		ret = -EINVAL;
+		goto out;
+	} else {
+		bond->params.src_mac_select = new_value;
+		pr_info("%s: setting src mac select policy to %s (%d).\n",
+			bond->dev->name,
+			src_mac_select_tbl[new_value].modename, new_value);
+	}
+out:
+	return ret;
+}
+
+static DEVICE_ATTR(src_mac_select, S_IRUGO | S_IWUSR,
+		bonding_show_src_mac_select, bonding_store_src_mac_select);
+
 static struct attribute *per_bond_attrs[] = {
 	&dev_attr_slaves.attr,
 	&dev_attr_mode.attr,
@@ -1671,6 +1720,7 @@ static struct attribute *per_bond_attrs[
 	&dev_attr_queue_id.attr,
 	&dev_attr_all_slaves_active.attr,
 	&dev_attr_resend_igmp.attr,
+	&dev_attr_src_mac_select.attr,
 	NULL,
 };
 
diff -uprN -X linux-2.6/Documentation/dontdiff linux-2.6/include/linux/if_bonding.h linux-2.6.p/include/linux/if_bonding.h
--- linux-2.6/include/linux/if_bonding.h	2011-02-16 00:59:18.720282002 +0300
+++ linux-2.6.p/include/linux/if_bonding.h	2011-03-01 23:00:56.610282019 +0300
@@ -91,6 +91,7 @@
 #define BOND_XMIT_POLICY_LAYER2		0 /* layer 2 (MAC only), default */
 #define BOND_XMIT_POLICY_LAYER34	1 /* layer 3+4 (IP ^ (TCP || UDP)) */
 #define BOND_XMIT_POLICY_LAYER23	2 /* layer 2+3 (IP ^ MAC) */
+#define BOND_XMIT_POLICY_RR		3 /* round-robin mode */
 
 typedef struct ifbond {
 	__s32 bond_mode;
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists