Skip to contentSkip to conversation

Add support to get xmit slave

PATCH Superseded Maor Gottlieb 4 participants 21 messages Apr 23, 2020 – Apr 25, 2020

Subsystems (9)

v5 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

v5 cover Add support to get xmit slave

Hi Dave,

This series is a combination of netdev and RDMA, so in order to avoid conflicts, we would like to ask you to route this series through mlx5-next shared branch. It is based on v5.7-rc1 tag.

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

The following series adds support to get the LAG master xmit slave by introducing new .ndo - ndo_get_xmit_slave. Every LAG module can implement it and it first implemented in the bond driver. This is follow-up to the RFC discussion [1].

The main motivation for doing this is for drivers that offload part of the LAG functionality. For example, Mellanox Connect-X hardware implements RoCE LAG which selects the TX affinity when the resources are created and port is remapped when it goes down.

The first part of this patchset introduces the new .ndo and add the support to the bonding module.

The second part adds support to get the RoCE LAG xmit slave by building skb of the RoCE packet based on the AH attributes and call to the new .ndo.

The third part change the mlx5 driver driver to set the QP's affinity port according to the slave which found by the .ndo.

Thanks

[1]
https://lore.kernel.org/netdev/20200126132126.9981-1-maorg@xxxxxxxxxxxx/

Change log:
v5: patch 1 - Remove rcu lock.
patch 10 - Refactor patch that group the AH attributes in struct. patch 11 - call the ndo while holding the rcu and initialize xmit_slave. patch 12 - Store the xmit slave in rdma_ah_init_attr and qp_attr.

v4: 1. Rename master_get_xmit_slave to netdev_get_xmit_slave and move the implementation to dev.c
2. Remove unnecessary check of NULL pointer.
3. Fix typo.
v3: 1. Move master_get_xmit_slave to netdevice.h and change the flags arg.
to bool.
2. Split helper functions commit to multiple commits for each bond
mode.
3. Extract refcotring changes to seperate commits.
v2: The first patch wasn't sent in v1.
v1:
https://lore.kernel.org/netdev/ac373456-b838-29cf-645f-b1ea1a93e3b0@xxxxxxxxx/T/#t

Maor Gottlieb (16):
net/core: Introduce netdev_get_xmit_slave
bonding: Export skip slave logic to function
bonding: Rename slave_arr to usable_slaves bonding/alb: Add helper functions to get the xmit slave
bonding: Add helper function to get the xmit slave based on hash
bonding: Add helper function to get the xmit slave in rr mode
bonding: Add function to get the xmit slave in active-backup mode
bonding: Add array of all slaves
bonding: Implement ndo_get_xmit_slave
RDMA: Group create AH arguments in struct
RDMA/core: Add LAG functionality
RDMA/core: Get xmit slave for LAG
net/mlx5: Change lag mutex lock to spin lock
net/mlx5: Add support to get lag physical port RDMA/mlx5: Refactor affinity related code
RDMA/mlx5: Set lag tx affinity according to slave

 drivers/infiniband/core/Makefile              |   2 +-
 drivers/infiniband/core/lag.c                 | 141 +++++++++
 drivers/infiniband/core/verbs.c               |  59 ++--
 drivers/infiniband/hw/bnxt_re/ib_verbs.c      |  11 +-
 drivers/infiniband/hw/bnxt_re/ib_verbs.h      |   3 +-
 drivers/infiniband/hw/mlx4/ah.c               |  10 +-
 drivers/infiniband/hw/mlx4/mlx4_ib.h          |   3 +-
 drivers/infiniband/hw/mlx5/ah.c               |  14 +-
 drivers/infiniband/hw/mlx5/gsi.c              |  34 ++-
 drivers/infiniband/hw/mlx5/main.c             |   2 +
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |   4 +-
 drivers/infiniband/hw/mlx5/qp.c               | 122 +++++---
 drivers/infiniband/sw/rxe/rxe_verbs.c         |   8 +-
 drivers/net/bonding/bond_alb.c                |  39 ++-
 drivers/net/bonding/bond_main.c               | 268 +++++++++++++-----
 drivers/net/ethernet/mellanox/mlx5/core/lag.c |  66 +++--
 include/linux/mlx5/driver.h                   |   2 +
 include/linux/mlx5/mlx5_ifc.h                 |   4 +-
 include/linux/mlx5/qp.h                       |   2 +
 include/linux/netdevice.h                     |   6 +
 include/net/bond_alb.h                        |   4 +
 include/net/bonding.h                         |   3 +-
 include/rdma/ib_verbs.h                       |  12 +-
 include/rdma/lag.h                            |  23 ++
 net/core/dev.c                                |  22 ++
 25 files changed, 658 insertions(+), 206 deletions(-)
 create mode 100644 drivers/infiniband/core/lag.c
 create mode 100644 include/rdma/lag.h
Signature · 1 line
2.17.2
#1 lore raw

PATCH v5 1/16 net/core: Introduce netdev_get_xmit_slave

Add new ndo to get the xmit slave of master device. Caller should call dev_put() when it no longer works with slave netdevice.
User can ask to get the xmit slave assume all the slaves can transmit by set all_slaves arg to true.

---
 include/linux/netdevice.h |  6 ++++++
 net/core/dev.c            | 22 ++++++++++++++++++++++
 2 files changed, 28 insertions(+)
include/linux/netdevice.h +6 -0
index 130a668049ab..d1206f08e099 100644
@@ -1389,6 +1389,9 @@ struct net_device_ops {
 						 struct netlink_ext_ack *extack);
 	int			(*ndo_del_slave)(struct net_device *dev,
 						 struct net_device *slave_dev);
+	struct net_device*	(*ndo_get_xmit_slave)(struct net_device *dev,
+						      struct sk_buff *skb,
+						      bool all_slaves);
 	netdev_features_t	(*ndo_fix_features)(struct net_device *dev,
 						    netdev_features_t features);
 	int			(*ndo_set_features)(struct net_device *dev,
@@ -2731,6 +2734,9 @@ void netdev_freemem(struct net_device *dev);
 void synchronize_net(void);
 int init_dummy_netdev(struct net_device *dev);
 
+struct net_device *netdev_get_xmit_slave(struct net_device *dev,
+					 struct sk_buff *skb,
+					 bool all_slaves);
 struct net_device *dev_get_by_index(struct net *net, int ifindex);
 struct net_device *__dev_get_by_index(struct net *net, int ifindex);
 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex);
net/core/dev.c +22 -0
index 9c9e763bfe0e..e6c10980abfd 100644
@@ -7785,6 +7785,28 @@ void netdev_bonding_info_change(struct net_device *dev,
 }
 EXPORT_SYMBOL(netdev_bonding_info_change);
 
+/**
+ * netdev_get_xmit_slave - Get the xmit slave of master device
+ * @skb: The packet
+ * @all_slaves: assume all the slaves are active
+ *
+ * The reference counters are not incremented so the caller must be
+ * careful with locks. The caller must hold RCU lock.
+ * %NULL is returned if no slave is found.
+ */
+
+struct net_device *netdev_get_xmit_slave(struct net_device *dev,
+					 struct sk_buff *skb,
+					 bool all_slaves)
+{
+	const struct net_device_ops *ops = dev->netdev_ops;
+
+	if (!ops->ndo_get_xmit_slave)
+		return NULL;
+	return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
+}
+EXPORT_SYMBOL(netdev_get_xmit_slave);
+
 static void netdev_adjacent_add_links(struct net_device *dev)
 {
 	struct netdev_adjacent *iter;
Signature · 1 line
2.17.2
Reviewed-by on v5 1/16
#2 lore raw

Re: [PATCH V5 mlx5-next 01/16] net/core: Introduce netdev_get_xmit_slave

Quoted 7 lines · Thu, Apr 23, 2020 at 02:55:40PM CEST, [email protected] wrote:

Add new ndo to get the xmit slave of master device. Caller should call dev_put() when it no longer works with slave netdevice.
User can ask to get the xmit slave assume all the slaves can transmit by set all_slaves arg to true.

#3 lore raw

Re: [PATCH V5 mlx5-next 01/16] net/core: Introduce netdev_get_xmit_slave

In reply to Maor Gottlieb
Quoted 3 lines · Maor Gottlieb

Add new ndo to get the xmit slave of master device. Caller should call dev_put() when it no longer works with slave netdevice.

description needs to be updated.

Quoted 14 lines

User can ask to get the xmit slave assume all the slaves can transmit by set all_slaves arg to true.

---
 include/linux/netdevice.h |  6 ++++++
 net/core/dev.c            | 22 ++++++++++++++++++++++
 2 files changed, 28 insertions(+)
include/linux/netdevice.h +0 -0
index 130a668049ab..d1206f08e099 100644

include/linux/netdevice.h struct net_device_ops { · line 1394

Quoted 6 lines
include/linux/netdevice.h +3 -0
@@ -1389,6 +1389,9 @@ struct net_device_ops {
 						 struct netlink_ext_ack *extack);
 	int			(*ndo_del_slave)(struct net_device *dev,
 						 struct net_device *slave_dev);
+	struct net_device*	(*ndo_get_xmit_slave)(struct net_device *dev,
+						      struct sk_buff *skb,
+						      bool all_slaves);

documentation above struct net_device_ops { }; needs to be updated.

Quoted 47 lines
include/linux/netdevice.h +3 -0
 	netdev_features_t	(*ndo_fix_features)(struct net_device *dev,
 						    netdev_features_t features);
 	int			(*ndo_set_features)(struct net_device *dev,
@@ -2731,6 +2734,9 @@ void netdev_freemem(struct net_device *dev);
 void synchronize_net(void);
 int init_dummy_netdev(struct net_device *dev);
 
+struct net_device *netdev_get_xmit_slave(struct net_device *dev,
+					 struct sk_buff *skb,
+					 bool all_slaves);
 struct net_device *dev_get_by_index(struct net *net, int ifindex);
 struct net_device *__dev_get_by_index(struct net *net, int ifindex);
 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex);
net/core/dev.c +22 -0
index 9c9e763bfe0e..e6c10980abfd 100644
@@ -7785,6 +7785,28 @@ void netdev_bonding_info_change(struct net_device *dev,
 }
 EXPORT_SYMBOL(netdev_bonding_info_change);
 
+/**
+ * netdev_get_xmit_slave - Get the xmit slave of master device
+ * @skb: The packet
+ * @all_slaves: assume all the slaves are active
+ *
+ * The reference counters are not incremented so the caller must be
+ * careful with locks. The caller must hold RCU lock.
+ * %NULL is returned if no slave is found.
+ */
+
+struct net_device *netdev_get_xmit_slave(struct net_device *dev,
+					 struct sk_buff *skb,
+					 bool all_slaves)
+{
+	const struct net_device_ops *ops = dev->netdev_ops;
+
+	if (!ops->ndo_get_xmit_slave)
+		return NULL;
+	return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
+}
+EXPORT_SYMBOL(netdev_get_xmit_slave);
+
 static void netdev_adjacent_add_links(struct net_device *dev)
 {
 	struct netdev_adjacent *iter;
#4 lore raw

PATCH v5 2/16 bonding: Export skip slave logic to function

In reply to Maor Gottlieb

As a preparation for following change that add array of all slaves, extract code that skip slave to function.

---
 drivers/net/bonding/bond_main.c | 47 ++++++++++++++++++---------------
 1 file changed, 26 insertions(+), 21 deletions(-)
drivers/net/bonding/bond_main.c +26 -21
index 2e70e43c5df5..f7aded014f08 100644
@@ -4087,6 +4087,29 @@ static void bond_slave_arr_handler(struct work_struct *work)
 	bond_slave_arr_work_rearm(bond, 1);
 }
 
+static void bond_skip_slave(struct bond_up_slave *slaves,
+			    struct slave *skipslave)
+{
+	int idx;
+
+	/* Rare situation where caller has asked to skip a specific
+	 * slave but allocation failed (most likely!). BTW this is
+	 * only possible when the call is initiated from
+	 * __bond_release_one(). In this situation; overwrite the
+	 * skipslave entry in the array with the last entry from the
+	 * array to avoid a situation where the xmit path may choose
+	 * this to-be-skipped slave to send a packet out.
+	 */
+	for (idx = 0; slaves && idx < slaves->count; idx++) {
+		if (skipslave == slaves->arr[idx]) {
+			slaves->arr[idx] =
+				slaves->arr[slaves->count - 1];
+			slaves->count--;
+			break;
+		}
+	}
+}
+
 /* Build the usable slaves array in control path for modes that use xmit-hash
  * to determine the slave interface -
  * (a) BOND_MODE_8023AD
@@ -4156,27 +4179,9 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 	if (old_arr)
 		kfree_rcu(old_arr, rcu);
 out:
-	if (ret != 0 && skipslave) {
-		int idx;
-
-		/* Rare situation where caller has asked to skip a specific
-		 * slave but allocation failed (most likely!). BTW this is
-		 * only possible when the call is initiated from
-		 * __bond_release_one(). In this situation; overwrite the
-		 * skipslave entry in the array with the last entry from the
-		 * array to avoid a situation where the xmit path may choose
-		 * this to-be-skipped slave to send a packet out.
-		 */
-		old_arr = rtnl_dereference(bond->slave_arr);
-		for (idx = 0; old_arr != NULL && idx < old_arr->count; idx++) {
-			if (skipslave == old_arr->arr[idx]) {
-				old_arr->arr[idx] =
-				    old_arr->arr[old_arr->count-1];
-				old_arr->count--;
-				break;
-			}
-		}
-	}
+	if (ret != 0 && skipslave)
+		bond_skip_slave(rtnl_dereference(bond->slave_arr), skipslave);
+
 	return ret;
 }
 
Signature · 1 line
2.17.2
#5 lore raw

PATCH v5 3/16 bonding: Rename slave_arr to usable_slaves

In reply to Maor Gottlieb

Rename slave_arr to usable_slaves, since we will have two arrays, one for the usable slaves and the other to all slaves.

---
 drivers/net/bonding/bond_alb.c  |  4 ++--
 drivers/net/bonding/bond_main.c | 40 ++++++++++++++++-----------------
 include/net/bonding.h           |  2 +-
 3 files changed, 23 insertions(+), 23 deletions(-)
drivers/net/bonding/bond_alb.c +2 -2
index c81698550e5a..7bb49b049dcc 100644
@@ -1360,7 +1360,7 @@ netdev_tx_t bond_tlb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 				struct bond_up_slave *slaves;
 				unsigned int count;
 
-				slaves = rcu_dereference(bond->slave_arr);
+				slaves = rcu_dereference(bond->usable_slaves);
 				count = slaves ? READ_ONCE(slaves->count) : 0;
 				if (likely(count))
 					tx_slave = slaves->arr[hash_index %
@@ -1494,7 +1494,7 @@ netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 			struct bond_up_slave *slaves;
 			unsigned int count;
 
-			slaves = rcu_dereference(bond->slave_arr);
+			slaves = rcu_dereference(bond->usable_slaves);
 			count = slaves ? READ_ONCE(slaves->count) : 0;
 			if (likely(count))
 				tx_slave = slaves->arr[bond_xmit_hash(bond, skb) %
drivers/net/bonding/bond_main.c +20 -20
index f7aded014f08..2cb41d480ae2 100644
@@ -4120,9 +4120,9 @@ static void bond_skip_slave(struct bond_up_slave *slaves,
  */
 int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 {
+	struct bond_up_slave *usable_slaves, *old_usable_slaves;
 	struct slave *slave;
 	struct list_head *iter;
-	struct bond_up_slave *new_arr, *old_arr;
 	int agg_id = 0;
 	int ret = 0;
 
@@ -4130,11 +4130,10 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 	WARN_ON(lockdep_is_held(&bond->mode_lock));
 #endif
 
-	new_arr = kzalloc(offsetof(struct bond_up_slave, arr[bond->slave_cnt]),
-			  GFP_KERNEL);
-	if (!new_arr) {
+	usable_slaves = kzalloc(struct_size(usable_slaves, arr,
+					    bond->slave_cnt), GFP_KERNEL);
+	if (!usable_slaves) {
 		ret = -ENOMEM;
-		pr_err("Failed to build slave-array.\n");
 		goto out;
 	}
 	if (BOND_MODE(bond) == BOND_MODE_8023AD) {
@@ -4142,14 +4141,14 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 
 		if (bond_3ad_get_active_agg_info(bond, &ad_info)) {
 			pr_debug("bond_3ad_get_active_agg_info failed\n");
-			kfree_rcu(new_arr, rcu);
+			kfree_rcu(usable_slaves, rcu);
 			/* No active aggragator means it's not safe to use
 			 * the previous array.
 			 */
-			old_arr = rtnl_dereference(bond->slave_arr);
-			if (old_arr) {
-				RCU_INIT_POINTER(bond->slave_arr, NULL);
-				kfree_rcu(old_arr, rcu);
+			old_usable_slaves = rtnl_dereference(bond->usable_slaves);
+			if (old_usable_slaves) {
+				RCU_INIT_POINTER(bond->usable_slaves, NULL);
+				kfree_rcu(old_usable_slaves, rcu);
 			}
 			goto out;
 		}
@@ -4169,18 +4168,19 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 			continue;
 
 		slave_dbg(bond->dev, slave->dev, "Adding slave to tx hash array[%d]\n",
-			  new_arr->count);
+			  usable_slaves->count);
 
-		new_arr->arr[new_arr->count++] = slave;
+		usable_slaves->arr[usable_slaves->count++] = slave;
 	}
 
-	old_arr = rtnl_dereference(bond->slave_arr);
-	rcu_assign_pointer(bond->slave_arr, new_arr);
-	if (old_arr)
-		kfree_rcu(old_arr, rcu);
+	old_usable_slaves = rtnl_dereference(bond->usable_slaves);
+	rcu_assign_pointer(bond->usable_slaves, usable_slaves);
+	if (old_usable_slaves)
+		kfree_rcu(old_usable_slaves, rcu);
 out:
 	if (ret != 0 && skipslave)
-		bond_skip_slave(rtnl_dereference(bond->slave_arr), skipslave);
+		bond_skip_slave(rtnl_dereference(bond->usable_slaves),
+				skipslave);
 
 	return ret;
 }
@@ -4197,7 +4197,7 @@ static netdev_tx_t bond_3ad_xor_xmit(struct sk_buff *skb,
 	struct bond_up_slave *slaves;
 	unsigned int count;
 
-	slaves = rcu_dereference(bond->slave_arr);
+	slaves = rcu_dereference(bond->usable_slaves);
 	count = slaves ? READ_ONCE(slaves->count) : 0;
 	if (likely(count)) {
 		slave = slaves->arr[bond_xmit_hash(bond, skb) % count];
@@ -4488,9 +4488,9 @@ static void bond_uninit(struct net_device *bond_dev)
 		__bond_release_one(bond_dev, slave->dev, true, true);
 	netdev_info(bond_dev, "Released all slaves\n");
 
-	arr = rtnl_dereference(bond->slave_arr);
+	arr = rtnl_dereference(bond->usable_slaves);
 	if (arr) {
-		RCU_INIT_POINTER(bond->slave_arr, NULL);
+		RCU_INIT_POINTER(bond->usable_slaves, NULL);
 		kfree_rcu(arr, rcu);
 	}
 
include/net/bonding.h +1 -1
index dc2ce31a1f52..33bdb6d5182d 100644
@@ -200,7 +200,7 @@ struct bonding {
 	struct   slave __rcu *curr_active_slave;
 	struct   slave __rcu *current_arp_slave;
 	struct   slave __rcu *primary_slave;
-	struct   bond_up_slave __rcu *slave_arr; /* Array of usable slaves */
+	struct   bond_up_slave __rcu *usable_slaves; /* Array of usable slaves */
 	bool     force_primary;
 	s32      slave_cnt; /* never change this value outside the attach/detach wrappers */
 	int     (*recv_probe)(const struct sk_buff *, struct bonding *,
Signature · 1 line
2.17.2
#6 lore raw

PATCH v5 4/16 bonding/alb: Add helper functions to get the xmit slave

In reply to Maor Gottlieb

Add two helper functions to get the xmit slave of bond in alb or tlb mode. Extract the logic of find the xmit slave from the xmit flow to function. Xmit flow will xmit through this slave and in the following patches the new .ndo will call to the helper function to return the xmit slave.

---
 drivers/net/bonding/bond_alb.c | 35 +++++++++++++++++++++++++---------
 include/net/bond_alb.h         |  4 ++++
 2 files changed, 30 insertions(+), 9 deletions(-)
drivers/net/bonding/bond_alb.c +26 -9
index 7bb49b049dcc..e863c694c309 100644
@@ -1334,11 +1334,11 @@ static netdev_tx_t bond_do_alb_xmit(struct sk_buff *skb, struct bonding *bond,
 	return NETDEV_TX_OK;
 }
 
-netdev_tx_t bond_tlb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
+struct slave *bond_xmit_tlb_slave_get(struct bonding *bond,
+				      struct sk_buff *skb)
 {
-	struct bonding *bond = netdev_priv(bond_dev);
-	struct ethhdr *eth_data;
 	struct slave *tx_slave = NULL;
+	struct ethhdr *eth_data;
 	u32 hash_index;
 
 	skb_reset_mac_header(skb);
@@ -1369,20 +1369,29 @@ netdev_tx_t bond_tlb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 			break;
 		}
 	}
-	return bond_do_alb_xmit(skb, bond, tx_slave);
+	return tx_slave;
 }
 
-netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
+netdev_tx_t bond_tlb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 {
 	struct bonding *bond = netdev_priv(bond_dev);
-	struct ethhdr *eth_data;
+	struct slave *tx_slave;
+
+	tx_slave = bond_xmit_tlb_slave_get(bond, skb);
+	return bond_do_alb_xmit(skb, bond, tx_slave);
+}
+
+struct slave *bond_xmit_alb_slave_get(struct bonding *bond,
+				      struct sk_buff *skb)
+{
 	struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
-	struct slave *tx_slave = NULL;
 	static const __be32 ip_bcast = htonl(0xffffffff);
-	int hash_size = 0;
+	struct slave *tx_slave = NULL;
+	const u8 *hash_start = NULL;
 	bool do_tx_balance = true;
+	struct ethhdr *eth_data;
 	u32 hash_index = 0;
-	const u8 *hash_start = NULL;
+	int hash_size = 0;
 
 	skb_reset_mac_header(skb);
 	eth_data = eth_hdr(skb);
@@ -1501,7 +1510,15 @@ netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
 						       count];
 		}
 	}
+	return tx_slave;
+}
+
+netdev_tx_t bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev)
+{
+	struct bonding *bond = netdev_priv(bond_dev);
+	struct slave *tx_slave = NULL;
 
+	tx_slave = bond_xmit_alb_slave_get(bond, skb);
 	return bond_do_alb_xmit(skb, bond, tx_slave);
 }
 
include/net/bond_alb.h +4 -0
index b3504fcd773d..f6af76c87a6c 100644
@@ -158,6 +158,10 @@ void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, char
 void bond_alb_handle_active_change(struct bonding *bond, struct slave *new_slave);
 int bond_alb_xmit(struct sk_buff *skb, struct net_device *bond_dev);
 int bond_tlb_xmit(struct sk_buff *skb, struct net_device *bond_dev);
+struct slave *bond_xmit_alb_slave_get(struct bonding *bond,
+				      struct sk_buff *skb);
+struct slave *bond_xmit_tlb_slave_get(struct bonding *bond,
+				      struct sk_buff *skb);
 void bond_alb_monitor(struct work_struct *);
 int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr);
 void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id);
Signature · 1 line
2.17.2
#7 lore raw

PATCH v5 5/16 bonding: Add helper function to get the xmit slave based on hash

In reply to Maor Gottlieb

Both xor and 802.3ad modes use bond_xmit_hash to get the xmit slave. Export the logic to helper function so it could be used in the following patches by the .ndo to get the xmit slave.

---
 drivers/net/bonding/bond_main.c | 28 +++++++++++++++++++++-------
 1 file changed, 21 insertions(+), 7 deletions(-)
drivers/net/bonding/bond_main.c +21 -7
index 2cb41d480ae2..8e6305955c75 100644
@@ -4185,6 +4185,23 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 	return ret;
 }
 
+static struct slave *bond_xmit_3ad_xor_slave_get(struct bonding *bond,
+						 struct sk_buff *skb,
+						 struct bond_up_slave *slaves)
+{
+	struct slave *slave;
+	unsigned int count;
+	u32 hash;
+
+	hash = bond_xmit_hash(bond, skb);
+	count = slaves ? READ_ONCE(slaves->count) : 0;
+	if (unlikely(!count))
+		return NULL;
+
+	slave = slaves->arr[hash % count];
+	return slave;
+}
+
 /* Use this Xmit function for 3AD as well as XOR modes. The current
  * usable slave array is formed in the control path. The xmit function
  * just calculates hash and sends the packet out.
@@ -4193,18 +4210,15 @@ static netdev_tx_t bond_3ad_xor_xmit(struct sk_buff *skb,
 				     struct net_device *dev)
 {
 	struct bonding *bond = netdev_priv(dev);
-	struct slave *slave;
 	struct bond_up_slave *slaves;
-	unsigned int count;
+	struct slave *slave;
 
 	slaves = rcu_dereference(bond->usable_slaves);
-	count = slaves ? READ_ONCE(slaves->count) : 0;
-	if (likely(count)) {
-		slave = slaves->arr[bond_xmit_hash(bond, skb) % count];
+	slave = bond_xmit_3ad_xor_slave_get(bond, skb, slaves);
+	if (likely(slave))
 		bond_dev_queue_xmit(bond, skb, slave->dev);
-	} else {
+	else
 		bond_tx_drop(dev, skb);
-	}
 
 	return NETDEV_TX_OK;
 }
Signature · 1 line
2.17.2
#8 lore raw

PATCH v5 6/16 bonding: Add helper function to get the xmit slave in rr mode

In reply to Maor Gottlieb

Add helper function to get the xmit slave when bond is in round robin mode. Change bond_xmit_slave_id to bond_get_slave_by_id, then the logic for find the next slave for transmit could be used both by the xmit flow and the .ndo to get the xmit slave.

---
 drivers/net/bonding/bond_main.c | 56 ++++++++++++++++++---------------
 1 file changed, 30 insertions(+), 26 deletions(-)
drivers/net/bonding/bond_main.c +30 -26
index 8e6305955c75..09c8485e965d 100644
@@ -3923,16 +3923,15 @@ static int bond_set_mac_address(struct net_device *bond_dev, void *addr)
 }
 
 /**
- * bond_xmit_slave_id - transmit skb through slave with slave_id
+ * bond_get_slave_by_id - get xmit slave with slave_id
  * @bond: bonding device that is transmitting
- * @skb: buffer to transmit
  * @slave_id: slave id up to slave_cnt-1 through which to transmit
  *
- * This function tries to transmit through slave with slave_id but in case
+ * This function tries to get slave with slave_id but in case
  * it fails, it tries to find the first available slave for transmission.
- * The skb is consumed in all cases, thus the function is void.
  */
-static void bond_xmit_slave_id(struct bonding *bond, struct sk_buff *skb, int slave_id)
+static struct slave *bond_get_slave_by_id(struct bonding *bond,
+					  int slave_id)
 {
 	struct list_head *iter;
 	struct slave *slave;
@@ -3941,10 +3940,8 @@ static void bond_xmit_slave_id(struct bonding *bond, struct sk_buff *skb, int sl
 	/* Here we start from the slave with slave_id */
 	bond_for_each_slave_rcu(bond, slave, iter) {
 		if (--i < 0) {
-			if (bond_slave_can_tx(slave)) {
-				bond_dev_queue_xmit(bond, skb, slave->dev);
-				return;
-			}
+			if (bond_slave_can_tx(slave))
+				return slave;
 		}
 	}
 
@@ -3953,13 +3950,11 @@ static void bond_xmit_slave_id(struct bonding *bond, struct sk_buff *skb, int sl
 	bond_for_each_slave_rcu(bond, slave, iter) {
 		if (--i < 0)
 			break;
-		if (bond_slave_can_tx(slave)) {
-			bond_dev_queue_xmit(bond, skb, slave->dev);
-			return;
-		}
+		if (bond_slave_can_tx(slave))
+			return slave;
 	}
-	/* no slave that can tx has been found */
-	bond_tx_drop(bond->dev, skb);
+
+	return NULL;
 }
 
 /**
@@ -3995,10 +3990,9 @@ static u32 bond_rr_gen_slave_id(struct bonding *bond)
 	return slave_id;
 }
 
-static netdev_tx_t bond_xmit_roundrobin(struct sk_buff *skb,
-					struct net_device *bond_dev)
+static struct slave *bond_xmit_roundrobin_slave_get(struct bonding *bond,
+						    struct sk_buff *skb)
 {
-	struct bonding *bond = netdev_priv(bond_dev);
 	struct slave *slave;
 	int slave_cnt;
 	u32 slave_id;
@@ -4020,21 +4014,31 @@ static netdev_tx_t bond_xmit_roundrobin(struct sk_buff *skb,
 		if (iph->protocol == IPPROTO_IGMP) {
 			slave = rcu_dereference(bond->curr_active_slave);
 			if (slave)
-				bond_dev_queue_xmit(bond, skb, slave->dev);
-			else
-				bond_xmit_slave_id(bond, skb, 0);
-			return NETDEV_TX_OK;
+				return slave;
+			return bond_get_slave_by_id(bond, 0);
 		}
 	}
 
 non_igmp:
 	slave_cnt = READ_ONCE(bond->slave_cnt);
 	if (likely(slave_cnt)) {
-		slave_id = bond_rr_gen_slave_id(bond);
-		bond_xmit_slave_id(bond, skb, slave_id % slave_cnt);
-	} else {
-		bond_tx_drop(bond_dev, skb);
+		slave_id = bond_rr_gen_slave_id(bond) % slave_cnt;
+		return bond_get_slave_by_id(bond, slave_id);
 	}
+	return NULL;
+}
+
+static netdev_tx_t bond_xmit_roundrobin(struct sk_buff *skb,
+					struct net_device *bond_dev)
+{
+	struct bonding *bond = netdev_priv(bond_dev);
+	struct slave *slave;
+
+	slave = bond_xmit_roundrobin_slave_get(bond, skb);
+	if (slave)
+		bond_dev_queue_xmit(bond, skb, slave->dev);
+	else
+		bond_tx_drop(bond_dev, skb);
 	return NETDEV_TX_OK;
 }
 
Signature · 1 line
2.17.2
#9 lore raw

PATCH v5 7/16 bonding: Add function to get the xmit slave in active-backup mode

In reply to Maor Gottlieb

Add helper function to get the xmit slave in active-backup mode. It's only one line function that return the curr_active_slave, but it will used both in the xmit flow and by the new .ndo to get the xmit slave.

---
 drivers/net/bonding/bond_main.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/net/bonding/bond_main.c +7 -1
index 09c8485e965d..1b0ae750d732 100644
@@ -4042,6 +4042,12 @@ static netdev_tx_t bond_xmit_roundrobin(struct sk_buff *skb,
 	return NETDEV_TX_OK;
 }
 
+static struct slave *bond_xmit_activebackup_slave_get(struct bonding *bond,
+						      struct sk_buff *skb)
+{
+	return rcu_dereference(bond->curr_active_slave);
+}
+
 /* In active-backup mode, we know that bond->curr_active_slave is always valid if
  * the bond has a usable interface.
  */
@@ -4051,7 +4057,7 @@ static netdev_tx_t bond_xmit_activebackup(struct sk_buff *skb,
 	struct bonding *bond = netdev_priv(bond_dev);
 	struct slave *slave;
 
-	slave = rcu_dereference(bond->curr_active_slave);
+	slave = bond_xmit_activebackup_slave_get(bond, skb);
 	if (slave)
 		bond_dev_queue_xmit(bond, skb, slave->dev);
 	else
Signature · 1 line
2.17.2
#10 lore raw

PATCH v5 8/16 bonding: Add array of all slaves

In reply to Maor Gottlieb

Keep all slaves in array so it could be used to get the xmit slave assume all the slaves are active.
The logic to add slave to the array is like the usable slaves, except that we also add slaves that currently can't transmit - not up or active.

---
 drivers/net/bonding/bond_main.c | 78 +++++++++++++++++++++++++--------
 include/net/bonding.h           |  3 +-
 2 files changed, 61 insertions(+), 20 deletions(-)
drivers/net/bonding/bond_main.c +59 -19
index 1b0ae750d732..2de693f0262e 100644
@@ -4120,6 +4120,38 @@ static void bond_skip_slave(struct bond_up_slave *slaves,
 	}
 }
 
+static void bond_set_slave_arr(struct bonding *bond,
+			       struct bond_up_slave *usable_slaves,
+			       struct bond_up_slave *all_slaves)
+{
+	struct bond_up_slave *usable, *all;
+
+	usable = rtnl_dereference(bond->usable_slaves);
+	rcu_assign_pointer(bond->usable_slaves, usable_slaves);
+	kfree_rcu(usable, rcu);
+
+	all = rtnl_dereference(bond->all_slaves);
+	rcu_assign_pointer(bond->all_slaves, all_slaves);
+	kfree_rcu(all, rcu);
+}
+
+static void bond_reset_slave_arr(struct bonding *bond)
+{
+	struct bond_up_slave *usable, *all;
+
+	usable = rtnl_dereference(bond->usable_slaves);
+	if (usable) {
+		RCU_INIT_POINTER(bond->usable_slaves, NULL);
+		kfree_rcu(usable, rcu);
+	}
+
+	all = rtnl_dereference(bond->all_slaves);
+	if (all) {
+		RCU_INIT_POINTER(bond->all_slaves, NULL);
+		kfree_rcu(all, rcu);
+	}
+}
+
 /* Build the usable slaves array in control path for modes that use xmit-hash
  * to determine the slave interface -
  * (a) BOND_MODE_8023AD
@@ -4130,7 +4162,7 @@ static void bond_skip_slave(struct bond_up_slave *slaves,
  */
 int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 {
-	struct bond_up_slave *usable_slaves, *old_usable_slaves;
+	struct bond_up_slave *usable_slaves = NULL, *all_slaves = NULL;
 	struct slave *slave;
 	struct list_head *iter;
 	int agg_id = 0;
@@ -4142,7 +4174,9 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 
 	usable_slaves = kzalloc(struct_size(usable_slaves, arr,
 					    bond->slave_cnt), GFP_KERNEL);
-	if (!usable_slaves) {
+	all_slaves = kzalloc(struct_size(all_slaves, arr,
+					 bond->slave_cnt), GFP_KERNEL);
+	if (!usable_slaves || !all_slaves) {
 		ret = -ENOMEM;
 		goto out;
 	}
@@ -4151,20 +4185,19 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 
 		if (bond_3ad_get_active_agg_info(bond, &ad_info)) {
 			pr_debug("bond_3ad_get_active_agg_info failed\n");
-			kfree_rcu(usable_slaves, rcu);
 			/* No active aggragator means it's not safe to use
 			 * the previous array.
 			 */
-			old_usable_slaves = rtnl_dereference(bond->usable_slaves);
-			if (old_usable_slaves) {
-				RCU_INIT_POINTER(bond->usable_slaves, NULL);
-				kfree_rcu(old_usable_slaves, rcu);
-			}
+			bond_reset_slave_arr(bond);
 			goto out;
 		}
 		agg_id = ad_info.aggregator_id;
 	}
 	bond_for_each_slave(bond, slave, iter) {
+		if (skipslave == slave)
+			continue;
+
+		all_slaves->arr[all_slaves->count++] = slave;
 		if (BOND_MODE(bond) == BOND_MODE_8023AD) {
 			struct aggregator *agg;
 
@@ -4174,8 +4207,6 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 		}
 		if (!bond_slave_can_tx(slave))
 			continue;
-		if (skipslave == slave)
-			continue;
 
 		slave_dbg(bond->dev, slave->dev, "Adding slave to tx hash array[%d]\n",
 			  usable_slaves->count);
@@ -4183,14 +4214,17 @@ int bond_update_slave_arr(struct bonding *bond, struct slave *skipslave)
 		usable_slaves->arr[usable_slaves->count++] = slave;
 	}
 
-	old_usable_slaves = rtnl_dereference(bond->usable_slaves);
-	rcu_assign_pointer(bond->usable_slaves, usable_slaves);
-	if (old_usable_slaves)
-		kfree_rcu(old_usable_slaves, rcu);
+	bond_set_slave_arr(bond, usable_slaves, all_slaves);
+	return ret;
 out:
-	if (ret != 0 && skipslave)
+	if (ret != 0 && skipslave) {
+		bond_skip_slave(rtnl_dereference(bond->all_slaves),
+				skipslave);
 		bond_skip_slave(rtnl_dereference(bond->usable_slaves),
 				skipslave);
+	}
+	kfree_rcu(all_slaves, rcu);
+	kfree_rcu(usable_slaves, rcu);
 
 	return ret;
 }
@@ -4501,9 +4535,9 @@ void bond_setup(struct net_device *bond_dev)
 static void bond_uninit(struct net_device *bond_dev)
 {
 	struct bonding *bond = netdev_priv(bond_dev);
+	struct bond_up_slave *usable, *all;
 	struct list_head *iter;
 	struct slave *slave;
-	struct bond_up_slave *arr;
 
 	bond_netpoll_cleanup(bond_dev);
 
@@ -4512,10 +4546,16 @@ static void bond_uninit(struct net_device *bond_dev)
 		__bond_release_one(bond_dev, slave->dev, true, true);
 	netdev_info(bond_dev, "Released all slaves\n");
 
-	arr = rtnl_dereference(bond->usable_slaves);
-	if (arr) {
+	usable = rtnl_dereference(bond->usable_slaves);
+	if (usable) {
 		RCU_INIT_POINTER(bond->usable_slaves, NULL);
-		kfree_rcu(arr, rcu);
+		kfree_rcu(usable, rcu);
+	}
+
+	all = rtnl_dereference(bond->all_slaves);
+	if (all) {
+		RCU_INIT_POINTER(bond->all_slaves, NULL);
+		kfree_rcu(all, rcu);
 	}
 
 	list_del(&bond->bond_list);
include/net/bonding.h +2 -1
index 33bdb6d5182d..b5e49bedbc9f 100644
@@ -200,7 +200,8 @@ struct bonding {
 	struct   slave __rcu *curr_active_slave;
 	struct   slave __rcu *current_arp_slave;
 	struct   slave __rcu *primary_slave;
-	struct   bond_up_slave __rcu *usable_slaves; /* Array of usable slaves */
+	struct   bond_up_slave __rcu *usable_slaves;
+	struct   bond_up_slave __rcu *all_slaves;
 	bool     force_primary;
 	s32      slave_cnt; /* never change this value outside the attach/detach wrappers */
 	int     (*recv_probe)(const struct sk_buff *, struct bonding *,
Signature · 1 line
2.17.2
#11 lore raw

PATCH v5 9/16 bonding: Implement ndo_get_xmit_slave

In reply to Maor Gottlieb

Add implementation of ndo_get_xmit_slave. Find the slave by using the helper function according to the bond mode. If the caller set all_slaves to true, then it assumes that all slaves are available to transmit.

---
 drivers/net/bonding/bond_main.c | 43 +++++++++++++++++++++++++++++++++
 1 file changed, 43 insertions(+)
drivers/net/bonding/bond_main.c +43 -0
index 2de693f0262e..39b1ad7edbb4 100644
@@ -4347,6 +4347,48 @@ static u16 bond_select_queue(struct net_device *dev, struct sk_buff *skb,
 	return txq;
 }
 
+static struct net_device *bond_xmit_get_slave(struct net_device *master_dev,
+					      struct sk_buff *skb,
+					      bool all_slaves)
+{
+	struct bonding *bond = netdev_priv(master_dev);
+	struct bond_up_slave *slaves;
+	struct slave *slave = NULL;
+
+	switch (BOND_MODE(bond)) {
+	case BOND_MODE_ROUNDROBIN:
+		slave = bond_xmit_roundrobin_slave_get(bond, skb);
+		break;
+	case BOND_MODE_ACTIVEBACKUP:
+		slave = bond_xmit_activebackup_slave_get(bond, skb);
+		break;
+	case BOND_MODE_8023AD:
+	case BOND_MODE_XOR:
+		if (all_slaves)
+			slaves = rcu_dereference(bond->all_slaves);
+		else
+			slaves = rcu_dereference(bond->usable_slaves);
+		slave = bond_xmit_3ad_xor_slave_get(bond, skb, slaves);
+		break;
+	case BOND_MODE_BROADCAST:
+		break;
+	case BOND_MODE_ALB:
+		slave = bond_xmit_alb_slave_get(bond, skb);
+		break;
+	case BOND_MODE_TLB:
+		slave = bond_xmit_tlb_slave_get(bond, skb);
+		break;
+	default:
+		/* Should never happen, mode already checked */
+		WARN_ONCE(true, "Unknown bonding mode");
+		break;
+	}
+
+	if (slave)
+		return slave->dev;
+	return NULL;
+}
+
 static netdev_tx_t __bond_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct bonding *bond = netdev_priv(dev);
@@ -4468,6 +4510,7 @@ static const struct net_device_ops bond_netdev_ops = {
 	.ndo_del_slave		= bond_release,
 	.ndo_fix_features	= bond_fix_features,
 	.ndo_features_check	= passthru_features_check,
+	.ndo_get_xmit_slave	= bond_xmit_get_slave,
 };
 
 static const struct device_type bond_type = {
Signature · 1 line
2.17.2
Reviewed-by on v5 9/16
#12 lore raw

Re: [PATCH V5 mlx5-next 09/16] bonding: Implement ndo_get_xmit_slave

Quoted 6 lines · Thu, Apr 23, 2020 at 02:55:48PM CEST, [email protected] wrote:

Add implementation of ndo_get_xmit_slave. Find the slave by using the helper function according to the bond mode. If the caller set all_slaves to true, then it assumes that all slaves are available to transmit.

I already send review tag. But it got lost apparently:

#13 lore raw

PATCH v5 10/16 RDMA: Group create AH arguments in struct

In reply to Maor Gottlieb

Following patch adds additional argument to the create AH function, so it make sense to group the attribute arguments in struct.

---
 drivers/infiniband/core/verbs.c          |  6 +++++-
 drivers/infiniband/hw/bnxt_re/ib_verbs.c | 11 ++++++-----
 drivers/infiniband/hw/bnxt_re/ib_verbs.h |  3 +--
 drivers/infiniband/hw/mlx4/ah.c          | 10 ++++++----
 drivers/infiniband/hw/mlx4/mlx4_ib.h     |  3 +--
 drivers/infiniband/hw/mlx5/ah.c          |  5 +++--
 drivers/infiniband/hw/mlx5/mlx5_ib.h     |  3 +--
 drivers/infiniband/sw/rxe/rxe_verbs.c    |  8 ++++----
 include/rdma/ib_verbs.h                  |  9 +++++++--
 9 files changed, 34 insertions(+), 24 deletions(-)
drivers/infiniband/core/verbs.c +5 -1
index 3bfadd8effcc..1ea69d730492 100644
@@ -502,6 +502,7 @@ static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
 				     u32 flags,
 				     struct ib_udata *udata)
 {
+	struct rdma_ah_init_attr init_attr = {};
 	struct ib_device *device = pd->device;
 	struct ib_ah *ah;
 	int ret;
@@ -521,8 +522,11 @@ static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
 	ah->pd = pd;
 	ah->type = ah_attr->type;
 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
+	init_attr.ah_attr = ah_attr;
+	init_attr.flags = flags;
+	init_attr.udata = udata;
 
-	ret = device->ops.create_ah(ah, ah_attr, flags, udata);
+	ret = device->ops.create_ah(ah, &init_attr);
 	if (ret) {
 		kfree(ah);
 		return ERR_PTR(ret);
drivers/infiniband/hw/bnxt_re/ib_verbs.c +6 -5
index d98348e82422..ebd0da87dbee 100644
@@ -631,11 +631,11 @@ static u8 bnxt_re_stack_to_dev_nw_type(enum rdma_network_type ntype)
 	return nw_type;
 }
 
-int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr,
-		      u32 flags, struct ib_udata *udata)
+int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr)
 {
 	struct ib_pd *ib_pd = ib_ah->pd;
 	struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd);
+	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
 	const struct ib_global_route *grh = rdma_ah_read_grh(ah_attr);
 	struct bnxt_re_dev *rdev = pd->rdev;
 	const struct ib_gid_attr *sgid_attr;
@@ -673,16 +673,17 @@ int bnxt_re_create_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr,
 
 	memcpy(ah->qplib_ah.dmac, ah_attr->roce.dmac, ETH_ALEN);
 	rc = bnxt_qplib_create_ah(&rdev->qplib_res, &ah->qplib_ah,
-				  !(flags & RDMA_CREATE_AH_SLEEPABLE));
+				  !(init_attr->flags &
+				    RDMA_CREATE_AH_SLEEPABLE));
 	if (rc) {
 		ibdev_err(&rdev->ibdev, "Failed to allocate HW AH");
 		return rc;
 	}
 
 	/* Write AVID to shared page. */
-	if (udata) {
+	if (init_attr->udata) {
 		struct bnxt_re_ucontext *uctx = rdma_udata_to_drv_context(
-			udata, struct bnxt_re_ucontext, ib_uctx);
+			init_attr->udata, struct bnxt_re_ucontext, ib_uctx);
 		unsigned long flag;
 		u32 *wrptr;
 
drivers/infiniband/hw/bnxt_re/ib_verbs.h +1 -2
index 18dd46f46cf4..91729e7cd5a6 100644
@@ -170,8 +170,7 @@ enum rdma_link_layer bnxt_re_get_link_layer(struct ib_device *ibdev,
 					    u8 port_num);
 int bnxt_re_alloc_pd(struct ib_pd *pd, struct ib_udata *udata);
 void bnxt_re_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata);
-int bnxt_re_create_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr, u32 flags,
-		      struct ib_udata *udata);
+int bnxt_re_create_ah(struct ib_ah *ah, struct rdma_ah_init_attr *init_attr);
 int bnxt_re_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr);
 int bnxt_re_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr);
 void bnxt_re_destroy_ah(struct ib_ah *ah, u32 flags);
drivers/infiniband/hw/mlx4/ah.c +6 -4
index 02a169f8027b..f98b14c7057d 100644
@@ -141,10 +141,10 @@ static int create_iboe_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr)
 	return 0;
 }
 
-int mlx4_ib_create_ah(struct ib_ah *ib_ah, struct rdma_ah_attr *ah_attr,
-		      u32 flags, struct ib_udata *udata)
-
+int mlx4_ib_create_ah(struct ib_ah *ib_ah, struct rdma_ah_init_attr *init_attr)
 {
+	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
+
 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
 		if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
 			return -EINVAL;
@@ -167,12 +167,14 @@ int mlx4_ib_create_ah_slave(struct ib_ah *ah, struct rdma_ah_attr *ah_attr,
 			    int slave_sgid_index, u8 *s_mac, u16 vlan_tag)
 {
 	struct rdma_ah_attr slave_attr = *ah_attr;
+	struct rdma_ah_init_attr init_attr = {};
 	struct mlx4_ib_ah *mah = to_mah(ah);
 	int ret;
 
 	slave_attr.grh.sgid_attr = NULL;
 	slave_attr.grh.sgid_index = slave_sgid_index;
-	ret = mlx4_ib_create_ah(ah, &slave_attr, 0, NULL);
+	init_attr.ah_attr = &slave_attr;
+	ret = mlx4_ib_create_ah(ah, &init_attr);
 	if (ret)
 		return ret;
 
drivers/infiniband/hw/mlx4/mlx4_ib.h +1 -2
index d188573187fa..ee24993dcfd8 100644
@@ -752,8 +752,7 @@ int mlx4_ib_arm_cq(struct ib_cq *cq, enum ib_cq_notify_flags flags);
 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq);
 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq);
 
-int mlx4_ib_create_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr, u32 flags,
-		      struct ib_udata *udata);
+int mlx4_ib_create_ah(struct ib_ah *ah, struct rdma_ah_init_attr *init_attr);
 int mlx4_ib_create_ah_slave(struct ib_ah *ah, struct rdma_ah_attr *ah_attr,
 			    int slave_sgid_index, u8 *s_mac, u16 vlan_tag);
 int mlx4_ib_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr);
drivers/infiniband/hw/mlx5/ah.c +3 -2
index 80642dd359bc..b3a3df51c44a 100644
@@ -68,10 +68,11 @@ static void create_ib_ah(struct mlx5_ib_dev *dev, struct mlx5_ib_ah *ah,
 	}
 }
 
-int mlx5_ib_create_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr,
-		      u32 flags, struct ib_udata *udata)
+int mlx5_ib_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *init_attr)
 
 {
+	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
+	struct ib_udata *udata = init_attr->udata;
 	struct mlx5_ib_ah *ah = to_mah(ibah);
 	struct mlx5_ib_dev *dev = to_mdev(ibah->device);
 	enum rdma_ah_attr_type ah_type = ah_attr->type;
drivers/infiniband/hw/mlx5/mlx5_ib.h +1 -2
index a4e522385de0..a041a1357a13 100644
@@ -1180,8 +1180,7 @@ void mlx5_ib_db_unmap_user(struct mlx5_ib_ucontext *context, struct mlx5_db *db)
 void __mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq);
 void mlx5_ib_cq_clean(struct mlx5_ib_cq *cq, u32 qpn, struct mlx5_ib_srq *srq);
 void mlx5_ib_free_srq_wqe(struct mlx5_ib_srq *srq, int wqe_index);
-int mlx5_ib_create_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr, u32 flags,
-		      struct ib_udata *udata);
+int mlx5_ib_create_ah(struct ib_ah *ah, struct rdma_ah_init_attr *init_attr);
 int mlx5_ib_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *ah_attr);
 void mlx5_ib_destroy_ah(struct ib_ah *ah, u32 flags);
 int mlx5_ib_create_srq(struct ib_srq *srq, struct ib_srq_init_attr *init_attr,
drivers/infiniband/sw/rxe/rxe_verbs.c +4 -4
index 9dd4bd7aea92..0aa89a070cec 100644
@@ -195,15 +195,15 @@ static void rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 	rxe_drop_ref(pd);
 }
 
-static int rxe_create_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr,
-			 u32 flags, struct ib_udata *udata)
+static int rxe_create_ah(struct ib_ah *ibah,
+			 struct rdma_ah_init_attr *init_attr)
 
 {
 	int err;
 	struct rxe_dev *rxe = to_rdev(ibah->device);
 	struct rxe_ah *ah = to_rah(ibah);
 
-	err = rxe_av_chk_attr(rxe, attr);
+	err = rxe_av_chk_attr(rxe, init_attr->ah_attr);
 	if (err)
 		return err;
 
@@ -211,7 +211,7 @@ static int rxe_create_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr,
 	if (err)
 		return err;
 
-	rxe_init_av(attr, &ah->av);
+	rxe_init_av(init_attr->ah_attr, &ah->av);
 	return 0;
 }
 
include/rdma/ib_verbs.h +7 -2
index bbc5cfb57cd2..c49e4d9ead66 100644
@@ -880,6 +880,12 @@ struct ib_mr_status {
  */
 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult);
 
+struct rdma_ah_init_attr {
+	struct rdma_ah_attr *ah_attr;
+	u32 flags;
+	struct ib_udata *udata;
+};
+
 enum rdma_ah_attr_type {
 	RDMA_AH_ATTR_TYPE_UNDEFINED,
 	RDMA_AH_ATTR_TYPE_IB,
@@ -2403,8 +2409,7 @@ struct ib_device_ops {
 	void (*disassociate_ucontext)(struct ib_ucontext *ibcontext);
 	int (*alloc_pd)(struct ib_pd *pd, struct ib_udata *udata);
 	void (*dealloc_pd)(struct ib_pd *pd, struct ib_udata *udata);
-	int (*create_ah)(struct ib_ah *ah, struct rdma_ah_attr *ah_attr,
-			 u32 flags, struct ib_udata *udata);
+	int (*create_ah)(struct ib_ah *ah, struct rdma_ah_init_attr *attr);
 	int (*modify_ah)(struct ib_ah *ah, struct rdma_ah_attr *ah_attr);
 	int (*query_ah)(struct ib_ah *ah, struct rdma_ah_attr *ah_attr);
 	void (*destroy_ah)(struct ib_ah *ah, u32 flags);
Signature · 1 line
2.17.2
#14 lore raw

Re: [PATCH V5 mlx5-next 10/16] RDMA: Group create AH arguments in struct

include/rdma/ib_verbs.h struct ib_mr_status { · line 886

Quoted 11 lines · Maor Gottlieb
include/rdma/ib_verbs.h +5 -0
index bbc5cfb57cd2..c49e4d9ead66 100644
+++ b/include/rdma/ib_verbs.h
@@ -880,6 +880,12 @@ struct ib_mr_status {
  */
 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult);
 
+struct rdma_ah_init_attr {
+	struct rdma_ah_attr *ah_attr;
+	u32 flags;
+	struct ib_udata *udata;

I prefer to keep the udata as a function argument

Jason

#15 lore raw

PATCH v5 11/16 RDMA/core: Add LAG functionality

In reply to Maor Gottlieb

Add support to get the RoCE LAG xmit slave by building skb of the RoCE packet and call to netdev_get_xmit_slave. If driver wants to get the slave assume all slaves are available, then need to set RDMA_LAG_FLAGS_HASH_ALL_SLAVES in flags.

---
 drivers/infiniband/core/Makefile |   2 +-
 drivers/infiniband/core/lag.c    | 141 +++++++++++++++++++++++++++++++
 include/rdma/ib_verbs.h          |   1 +
 include/rdma/lag.h               |  23 +++++
 4 files changed, 166 insertions(+), 1 deletion(-)
 create mode 100644 drivers/infiniband/core/lag.c
 create mode 100644 include/rdma/lag.h
drivers/infiniband/core/Makefile +1 -1
index d1b14887960e..870f0fcd54d5 100644
@@ -12,7 +12,7 @@ ib_core-y :=			packer.o ud_header.o verbs.o cq.o rw.o sysfs.o \
 				roce_gid_mgmt.o mr_pool.o addr.o sa_query.o \
 				multicast.o mad.o smi.o agent.o mad_rmpp.o \
 				nldev.o restrack.o counters.o ib_core_uverbs.o \
-				trace.o
+				trace.o lag.o
 
 ib_core-$(CONFIG_SECURITY_INFINIBAND) += security.o
 ib_core-$(CONFIG_CGROUP_RDMA) += cgroup.o
drivers/infiniband/core/lag.c new file +141 -0

141 lines not on this page · raw message

include/rdma/ib_verbs.h +1 -0
index c49e4d9ead66..1d5acf3cc287 100644
@@ -2714,6 +2714,7 @@ struct ib_device {
 	/* Used by iWarp CM */
 	char iw_ifname[IFNAMSIZ];
 	u32 iw_driver_flags;
+	u32 lag_flags;
 };
 
 struct ib_client_nl_info;
include/rdma/lag.h new file +23 -0
new file mode 100644
index 000000000000..8f78e2c0fd7c
@@ -0,0 +1,23 @@
+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
+/*
+ * Copyright (c) 2020 Mellanox Technologies. All rights reserved.
+ */
+
+#ifndef _RDMA_LAG_H_
+#define _RDMA_LAG_H_
+
+#include <net/lag.h>
+
+struct ib_device;
+struct rdma_ah_attr;
+
+enum rdma_lag_flags {
+	RDMA_LAG_FLAGS_HASH_ALL_SLAVES = 1 << 0
+};
+
+void rdma_lag_put_ah_roce_slave(struct net_device *xmit_slave);
+int rdma_lag_get_ah_roce_slave(struct ib_device *device,
+			       struct rdma_ah_attr *ah_attr,
+			       struct net_device **xmit_slave);
+
+#endif /* _RDMA_LAG_H_ */
Signature · 1 line
2.17.2
#16 lore raw

PATCH v5 12/16 RDMA/core: Get xmit slave for LAG

In reply to Maor Gottlieb

Add a call to rdma_lag_get_ah_roce_slave when Address handle is created.
Lower driver can use it to select the QP's affinity port.

---
 drivers/infiniband/core/verbs.c | 53 +++++++++++++++++++++++----------
 include/rdma/ib_verbs.h         |  2 ++
 2 files changed, 39 insertions(+), 16 deletions(-)
drivers/infiniband/core/verbs.c +37 -16
index 1ea69d730492..00062b6c3031 100644
@@ -50,6 +50,7 @@
 #include <rdma/ib_cache.h>
 #include <rdma/ib_addr.h>
 #include <rdma/rw.h>
+#include <rdma/lag.h>
 
 #include "core_priv.h"
 #include <trace/events/rdma_core.h>
@@ -500,7 +501,8 @@ rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
 				     struct rdma_ah_attr *ah_attr,
 				     u32 flags,
-				     struct ib_udata *udata)
+				     struct ib_udata *udata,
+				     struct net_device *xmit_slave)
 {
 	struct rdma_ah_init_attr init_attr = {};
 	struct ib_device *device = pd->device;
@@ -525,6 +527,7 @@ static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
 	init_attr.ah_attr = ah_attr;
 	init_attr.flags = flags;
 	init_attr.udata = udata;
+	init_attr.xmit_slave = xmit_slave;
 
 	ret = device->ops.create_ah(ah, &init_attr);
 	if (ret) {
@@ -551,6 +554,7 @@ struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
 			     u32 flags)
 {
 	const struct ib_gid_attr *old_sgid_attr;
+	struct net_device *slave;
 	struct ib_ah *ah;
 	int ret;
 
@@ -558,8 +562,14 @@ struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
 	if (ret)
 		return ERR_PTR(ret);
 
-	ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
+	ret = rdma_lag_get_ah_roce_slave(pd->device, ah_attr, &slave);
+	if (ret) {
+		rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
+		return ERR_PTR(ret);
+	}
 
+	ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
+	rdma_lag_put_ah_roce_slave(slave);
 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 	return ah;
 }
@@ -598,7 +608,8 @@ struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
 		}
 	}
 
-	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata);
+	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
+			     udata, NULL);
 
 out:
 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
@@ -1642,6 +1653,26 @@ static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
 					  &old_sgid_attr_av);
 		if (ret)
 			return ret;
+
+		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
+		    is_qp_type_connected(qp)) {
+			/*
+			 * If the user provided the qp_attr then we have to
+			 * resolve it. Kerne users have to provide already
+			 * resolved rdma_ah_attr's.
+			 */
+			if (udata) {
+				ret = ib_resolve_eth_dmac(qp->device,
+							  &attr->ah_attr);
+				if (ret)
+					goto out_av;
+			}
+			ret = rdma_lag_get_ah_roce_slave(qp->device,
+							 &attr->ah_attr,
+							 &attr->xmit_slave);
+			if (ret)
+				goto out_av;
+		}
 	}
 	if (attr_mask & IB_QP_ALT_PATH) {
 		/*
@@ -1668,18 +1699,6 @@ static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
 		}
 	}
 
-	/*
-	 * If the user provided the qp_attr then we have to resolve it. Kernel
-	 * users have to provide already resolved rdma_ah_attr's
-	 */
-	if (udata && (attr_mask & IB_QP_AV) &&
-	    attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
-	    is_qp_type_connected(qp)) {
-		ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
-		if (ret)
-			goto out;
-	}
-
 	if (rdma_ib_or_roce(qp->device, port)) {
 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
 			dev_warn(&qp->device->dev,
@@ -1721,8 +1740,10 @@ static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
 	if (attr_mask & IB_QP_ALT_PATH)
 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
 out_av:
-	if (attr_mask & IB_QP_AV)
+	if (attr_mask & IB_QP_AV) {
+		rdma_lag_put_ah_roce_slave(attr->xmit_slave);
 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
+	}
 	return ret;
 }
 
include/rdma/ib_verbs.h +2 -0
index 1d5acf3cc287..22a41493ef4c 100644
@@ -884,6 +884,7 @@ struct rdma_ah_init_attr {
 	struct rdma_ah_attr *ah_attr;
 	u32 flags;
 	struct ib_udata *udata;
+	struct net_device *xmit_slave;
 };
 
 enum rdma_ah_attr_type {
@@ -1273,6 +1274,7 @@ struct ib_qp_attr {
 	u8			alt_port_num;
 	u8			alt_timeout;
 	u32			rate_limit;
+	struct net_device	*xmit_slave;
 };
 
 enum ib_wr_opcode {
Signature · 1 line
2.17.2
#17 lore raw

PATCH v5 13/16 net/mlx5: Change lag mutex lock to spin lock

In reply to Maor Gottlieb

The lag lock could be a mutex, the critical section is short and there is no need that the thread will sleep. Change the lock that protects the LAG structure from mutex to spin lock. It is required for next patch that need to access this structure from context that we can't sleep. In addition there is no need to hold this lock when query the congestion counters.

---
 drivers/net/ethernet/mellanox/mlx5/core/lag.c | 42 +++++++++----------
 1 file changed, 21 insertions(+), 21 deletions(-)
drivers/net/ethernet/mellanox/mlx5/core/lag.c +21 -21
index 93052b07c76c..496a3408d771 100644
@@ -42,7 +42,7 @@
  * Beware of lock dependencies (preferably, no locks should be acquired
  * under it).
  */
-static DEFINE_MUTEX(lag_mutex);
+static DEFINE_SPINLOCK(lag_lock);
 
 static int mlx5_cmd_create_lag(struct mlx5_core_dev *dev, u8 remap_port1,
 			       u8 remap_port2)
@@ -297,9 +297,9 @@ static void mlx5_do_bond(struct mlx5_lag *ldev)
 	if (!dev0 || !dev1)
 		return;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	tracker = ldev->tracker;
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 
 	do_bond = tracker.is_bonded && mlx5_lag_check_prereq(ldev);
 
@@ -481,9 +481,9 @@ static int mlx5_lag_netdev_event(struct notifier_block *this,
 		break;
 	}
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev->tracker = tracker;
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 
 	if (changed)
 		mlx5_queue_bond_work(ldev, 0);
@@ -525,7 +525,7 @@ static void mlx5_lag_dev_add_pf(struct mlx5_lag *ldev,
 	if (fn >= MLX5_MAX_PORTS)
 		return;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev->pf[fn].dev    = dev;
 	ldev->pf[fn].netdev = netdev;
 	ldev->tracker.netdev_state[fn].link_up = 0;
@@ -533,7 +533,7 @@ static void mlx5_lag_dev_add_pf(struct mlx5_lag *ldev,
 
 	dev->priv.lag = ldev;
 
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 }
 
 static void mlx5_lag_dev_remove_pf(struct mlx5_lag *ldev,
@@ -548,11 +548,11 @@ static void mlx5_lag_dev_remove_pf(struct mlx5_lag *ldev,
 	if (i == MLX5_MAX_PORTS)
 		return;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	memset(&ldev->pf[i], 0, sizeof(*ldev->pf));
 
 	dev->priv.lag = NULL;
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 }
 
 /* Must be called with intf_mutex held */
@@ -630,10 +630,10 @@ bool mlx5_lag_is_roce(struct mlx5_core_dev *dev)
 	struct mlx5_lag *ldev;
 	bool res;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev = mlx5_lag_dev_get(dev);
 	res  = ldev && __mlx5_lag_is_roce(ldev);
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 
 	return res;
 }
@@ -644,10 +644,10 @@ bool mlx5_lag_is_active(struct mlx5_core_dev *dev)
 	struct mlx5_lag *ldev;
 	bool res;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev = mlx5_lag_dev_get(dev);
 	res  = ldev && __mlx5_lag_is_active(ldev);
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 
 	return res;
 }
@@ -658,10 +658,10 @@ bool mlx5_lag_is_sriov(struct mlx5_core_dev *dev)
 	struct mlx5_lag *ldev;
 	bool res;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev = mlx5_lag_dev_get(dev);
 	res  = ldev && __mlx5_lag_is_sriov(ldev);
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 
 	return res;
 }
@@ -687,7 +687,7 @@ struct net_device *mlx5_lag_get_roce_netdev(struct mlx5_core_dev *dev)
 	struct net_device *ndev = NULL;
 	struct mlx5_lag *ldev;
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev = mlx5_lag_dev_get(dev);
 
 	if (!(ldev && __mlx5_lag_is_roce(ldev)))
@@ -704,7 +704,7 @@ struct net_device *mlx5_lag_get_roce_netdev(struct mlx5_core_dev *dev)
 		dev_hold(ndev);
 
 unlock:
-	mutex_unlock(&lag_mutex);
+	spin_unlock(&lag_lock);
 
 	return ndev;
 }
@@ -746,7 +746,7 @@ int mlx5_lag_query_cong_counters(struct mlx5_core_dev *dev,
 
 	memset(values, 0, sizeof(*values) * num_counters);
 
-	mutex_lock(&lag_mutex);
+	spin_lock(&lag_lock);
 	ldev = mlx5_lag_dev_get(dev);
 	if (ldev && __mlx5_lag_is_roce(ldev)) {
 		num_ports = MLX5_MAX_PORTS;
@@ -756,18 +756,18 @@ int mlx5_lag_query_cong_counters(struct mlx5_core_dev *dev,
 		num_ports = 1;
 		mdev[MLX5_LAG_P1] = dev;
 	}
+	spin_unlock(&lag_lock);
 
 	for (i = 0; i < num_ports; ++i) {
 		ret = mlx5_cmd_query_cong_counter(mdev[i], false, out, outlen);
 		if (ret)
-			goto unlock;
+			goto free;
 
 		for (j = 0; j < num_counters; ++j)
 			values[j] += be64_to_cpup((__be64 *)(out + offsets[j]));
 	}
 
-unlock:
-	mutex_unlock(&lag_mutex);
+free:
 	kvfree(out);
 	return ret;
 }
Signature · 1 line
2.17.2
#18 lore raw

PATCH v5 14/16 net/mlx5: Add support to get lag physical port

In reply to Maor Gottlieb

Add function to get the device physical port of the lag slave.

---
 drivers/net/ethernet/mellanox/mlx5/core/lag.c | 24 +++++++++++++++++++
 include/linux/mlx5/driver.h                   |  2 ++
 2 files changed, 26 insertions(+)
drivers/net/ethernet/mellanox/mlx5/core/lag.c +24 -0
index 496a3408d771..5461fbe47c0d 100644
@@ -710,6 +710,30 @@ struct net_device *mlx5_lag_get_roce_netdev(struct mlx5_core_dev *dev)
 }
 EXPORT_SYMBOL(mlx5_lag_get_roce_netdev);
 
+u8 mlx5_lag_get_slave_port(struct mlx5_core_dev *dev,
+			   struct net_device *slave)
+{
+	struct mlx5_lag *ldev;
+	u8 port = 0;
+
+	spin_lock(&lag_lock);
+	ldev = mlx5_lag_dev_get(dev);
+	if (!(ldev && __mlx5_lag_is_roce(ldev)))
+		goto unlock;
+
+	if (ldev->pf[MLX5_LAG_P1].netdev == slave)
+		port = MLX5_LAG_P1;
+	else
+		port = MLX5_LAG_P2;
+
+	port = ldev->v2p_map[port];
+
+unlock:
+	spin_unlock(&lag_lock);
+	return port;
+}
+EXPORT_SYMBOL(mlx5_lag_get_slave_port);
+
 bool mlx5_lag_intf_add(struct mlx5_interface *intf, struct mlx5_priv *priv)
 {
 	struct mlx5_core_dev *dev = container_of(priv, struct mlx5_core_dev,
include/linux/mlx5/driver.h +2 -0
index 6f8f79ef829b..7b81b512d116 100644
@@ -1062,6 +1062,8 @@ bool mlx5_lag_is_sriov(struct mlx5_core_dev *dev);
 bool mlx5_lag_is_multipath(struct mlx5_core_dev *dev);
 bool mlx5_lag_is_active(struct mlx5_core_dev *dev);
 struct net_device *mlx5_lag_get_roce_netdev(struct mlx5_core_dev *dev);
+u8 mlx5_lag_get_slave_port(struct mlx5_core_dev *dev,
+			   struct net_device *slave);
 int mlx5_lag_query_cong_counters(struct mlx5_core_dev *dev,
 				 u64 *values,
 				 int num_counters,
Signature · 1 line
2.17.2
#19 lore raw

PATCH v5 15/16 RDMA/mlx5: Refactor affinity related code

In reply to Maor Gottlieb

Move affinity related code in modify qp to function. It's a preparation for next patch the extend the affinity calculation to consider the xmit slave.

---
 drivers/infiniband/hw/mlx5/qp.c | 90 +++++++++++++++++++--------------
 1 file changed, 53 insertions(+), 37 deletions(-)
drivers/infiniband/hw/mlx5/qp.c +53 -37
index 1456db4b6295..a45499809903 100644
@@ -3416,33 +3416,61 @@ static int modify_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp,
 	return 0;
 }
 
-static unsigned int get_tx_affinity(struct mlx5_ib_dev *dev,
-				    struct mlx5_ib_pd *pd,
-				    struct mlx5_ib_qp_base *qp_base,
-				    u8 port_num, struct ib_udata *udata)
+static unsigned int get_tx_affinity_rr(struct mlx5_ib_dev *dev,
+				       struct ib_udata *udata)
 {
 	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
 		udata, struct mlx5_ib_ucontext, ibucontext);
-	unsigned int tx_port_affinity;
+	u8 port_num = mlx5_core_native_port_num(dev->mdev) - 1;
+	atomic_t *tx_port_affinity;
 
-	if (ucontext) {
-		tx_port_affinity = (unsigned int)atomic_add_return(
-					   1, &ucontext->tx_port_affinity) %
-					   MLX5_MAX_PORTS +
-				   1;
+	if (ucontext)
+		tx_port_affinity = &ucontext->tx_port_affinity;
+	else
+		tx_port_affinity = &dev->port[port_num].roce.tx_port_affinity;
+
+	return (unsigned int)atomic_add_return(1, tx_port_affinity) %
+		MLX5_MAX_PORTS + 1;
+}
+
+static bool qp_supports_affinity(struct ib_qp *qp)
+{
+	struct mlx5_ib_qp *mqp = to_mqp(qp);
+
+	if ((qp->qp_type == IB_QPT_RC) ||
+	    (qp->qp_type == IB_QPT_UD &&
+	     !(mqp->flags & MLX5_IB_QP_SQPN_QP1)) ||
+	    (qp->qp_type == IB_QPT_UC) ||
+	    (qp->qp_type == IB_QPT_RAW_PACKET) ||
+	    (qp->qp_type == IB_QPT_XRC_INI) ||
+	    (qp->qp_type == IB_QPT_XRC_TGT))
+		return true;
+	return false;
+}
+
+static unsigned int get_tx_affinity(struct ib_qp *qp, u8 init,
+				    struct ib_udata *udata)
+{
+	struct mlx5_ib_ucontext *ucontext = rdma_udata_to_drv_context(
+		udata, struct mlx5_ib_ucontext, ibucontext);
+	struct mlx5_ib_dev *dev = to_mdev(qp->device);
+	struct mlx5_ib_qp *mqp = to_mqp(qp);
+	struct mlx5_ib_qp_base *qp_base;
+	unsigned int tx_affinity;
+
+	if (!(dev->lag_active && init && qp_supports_affinity(qp)))
+		return 0;
+
+	tx_affinity = get_tx_affinity_rr(dev, udata);
+
+	qp_base = &mqp->trans_qp.base;
+	if (ucontext)
 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x ucontext %p\n",
-				tx_port_affinity, qp_base->mqp.qpn, ucontext);
-	} else {
-		tx_port_affinity =
-			(unsigned int)atomic_add_return(
-				1, &dev->port[port_num].roce.tx_port_affinity) %
-				MLX5_MAX_PORTS +
-			1;
+			    tx_affinity, qp_base->mqp.qpn, ucontext);
+	else
 		mlx5_ib_dbg(dev, "Set tx affinity 0x%x to qpn 0x%x\n",
-				tx_port_affinity, qp_base->mqp.qpn);
-	}
-
-	return tx_port_affinity;
+			    tx_affinity, qp_base->mqp.qpn);
+	return tx_affinity;
 }
 
 static int __mlx5_ib_qp_set_counter(struct ib_qp *qp,
@@ -3554,22 +3582,10 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 		}
 	}
 
-	if ((cur_state == IB_QPS_RESET) && (new_state == IB_QPS_INIT)) {
-		if ((ibqp->qp_type == IB_QPT_RC) ||
-		    (ibqp->qp_type == IB_QPT_UD &&
-		     !(qp->flags & MLX5_IB_QP_SQPN_QP1)) ||
-		    (ibqp->qp_type == IB_QPT_UC) ||
-		    (ibqp->qp_type == IB_QPT_RAW_PACKET) ||
-		    (ibqp->qp_type == IB_QPT_XRC_INI) ||
-		    (ibqp->qp_type == IB_QPT_XRC_TGT)) {
-			if (dev->lag_active) {
-				u8 p = mlx5_core_native_port_num(dev->mdev) - 1;
-				tx_affinity = get_tx_affinity(dev, pd, base, p,
-							      udata);
-				context->flags |= cpu_to_be32(tx_affinity << 24);
-			}
-		}
-	}
+	tx_affinity = get_tx_affinity(ibqp,
+				      cur_state == IB_QPS_RESET &&
+				      new_state == IB_QPS_INIT, udata);
+	context->flags |= cpu_to_be32(tx_affinity << 24);
 
 	if (is_sqp(ibqp->qp_type)) {
 		context->mtu_msgmax = (IB_MTU_256 << 5) | 8;
Signature · 1 line
2.17.2
#20 lore raw

PATCH v5 16/16 RDMA/mlx5: Set lag tx affinity according to slave

In reply to Maor Gottlieb

The patch sets the lag tx affinity of the data QPs and the GSI QPs according to the LAG xmit slave.

For GSI QPs, in case that the link layer is Ethenet (RoCE) we create two GSI QPs, one for each physical port. When the driver selects the GSI QP, it will consider the port affinity result.
For connected QPs, the driver sets the affinity of the xmit slave.

The above, ensure that RC QP and it's corresponding GSI QP will transmit from the same physical port.

---
 drivers/infiniband/hw/mlx5/ah.c      |  9 +++--
 drivers/infiniband/hw/mlx5/gsi.c     | 34 ++++++++++++++----
 drivers/infiniband/hw/mlx5/main.c    |  2 ++
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  1 +
 drivers/infiniband/hw/mlx5/qp.c      | 52 +++++++++++++++++++---------
 include/linux/mlx5/mlx5_ifc.h        |  4 ++-
 include/linux/mlx5/qp.h              |  2 ++
 7 files changed, 77 insertions(+), 27 deletions(-)
drivers/infiniband/hw/mlx5/ah.c +7 -2
index b3a3df51c44a..d73968786ee0 100644
@@ -33,8 +33,9 @@
 #include "mlx5_ib.h"
 
 static void create_ib_ah(struct mlx5_ib_dev *dev, struct mlx5_ib_ah *ah,
-			 struct rdma_ah_attr *ah_attr)
+			 struct rdma_ah_init_attr *init_attr)
 {
+	struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
 	enum ib_gid_type gid_type;
 
 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
@@ -51,6 +52,10 @@ static void create_ib_ah(struct mlx5_ib_dev *dev, struct mlx5_ib_ah *ah,
 	ah->av.stat_rate_sl = (rdma_ah_get_static_rate(ah_attr) << 4);
 
 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
+		if (init_attr->xmit_slave)
+			ah->xmit_port =
+				mlx5_lag_get_slave_port(dev->mdev,
+							init_attr->xmit_slave);
 		gid_type = ah_attr->grh.sgid_attr->gid_type;
 
 		memcpy(ah->av.rmac, ah_attr->roce.dmac,
@@ -98,7 +103,7 @@ int mlx5_ib_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *init_attr)
 			return err;
 	}
 
-	create_ib_ah(dev, ah, ah_attr);
+	create_ib_ah(dev, ah, init_attr);
 	return 0;
 }
 
drivers/infiniband/hw/mlx5/gsi.c +27 -7
index 1ae6fd95acaa..fbae1c094fe2 100644
@@ -119,12 +119,20 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
 	struct mlx5_ib_gsi_qp *gsi;
 	struct ib_qp_init_attr hw_init_attr = *init_attr;
 	const u8 port_num = init_attr->port_num;
-	const int num_pkeys = pd->device->attrs.max_pkeys;
-	const int num_qps = mlx5_ib_deth_sqpn_cap(dev) ? num_pkeys : 0;
+	int num_qps = 0;
 	int ret;
 
 	mlx5_ib_dbg(dev, "creating GSI QP\n");
 
+	if (mlx5_ib_deth_sqpn_cap(dev)) {
+		if (MLX5_CAP_GEN(dev->mdev,
+				 port_type) == MLX5_CAP_PORT_TYPE_IB)
+			num_qps = pd->device->attrs.max_pkeys;
+		else if (dev->lag_active)
+			num_qps = MLX5_MAX_PORTS;
+	}
+
+
 	if (port_num > ARRAY_SIZE(dev->devr.ports) || port_num < 1) {
 		mlx5_ib_warn(dev,
 			     "invalid port number %d during GSI QP creation\n",
@@ -270,7 +278,7 @@ static struct ib_qp *create_gsi_ud_qp(struct mlx5_ib_gsi_qp *gsi)
 }
 
 static int modify_to_rts(struct mlx5_ib_gsi_qp *gsi, struct ib_qp *qp,
-			 u16 qp_index)
+			 u16 pkey_index)
 {
 	struct mlx5_ib_dev *dev = to_mdev(qp->device);
 	struct ib_qp_attr attr;
@@ -279,7 +287,7 @@ static int modify_to_rts(struct mlx5_ib_gsi_qp *gsi, struct ib_qp *qp,
 
 	mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY | IB_QP_PORT;
 	attr.qp_state = IB_QPS_INIT;
-	attr.pkey_index = qp_index;
+	attr.pkey_index = pkey_index;
 	attr.qkey = IB_QP1_QKEY;
 	attr.port_num = gsi->port_num;
 	ret = ib_modify_qp(qp, &attr, mask);
@@ -313,12 +321,17 @@ static void setup_qp(struct mlx5_ib_gsi_qp *gsi, u16 qp_index)
 {
 	struct ib_device *device = gsi->rx_qp->device;
 	struct mlx5_ib_dev *dev = to_mdev(device);
+	int pkey_index = qp_index;
+	struct mlx5_ib_qp *mqp;
 	struct ib_qp *qp;
 	unsigned long flags;
 	u16 pkey;
 	int ret;
 
-	ret = ib_query_pkey(device, gsi->port_num, qp_index, &pkey);
+	if (MLX5_CAP_GEN(dev->mdev,  port_type) != MLX5_CAP_PORT_TYPE_IB)
+		pkey_index = 0;
+
+	ret = ib_query_pkey(device, gsi->port_num, pkey_index, &pkey);
 	if (ret) {
 		mlx5_ib_warn(dev, "unable to read P_Key at port %d, index %d\n",
 			     gsi->port_num, qp_index);
@@ -347,7 +360,10 @@ static void setup_qp(struct mlx5_ib_gsi_qp *gsi, u16 qp_index)
 		return;
 	}
 
-	ret = modify_to_rts(gsi, qp, qp_index);
+	mqp = to_mqp(qp);
+	if (dev->lag_active)
+		mqp->gsi_lag_port = qp_index + 1;
+	ret = modify_to_rts(gsi, qp, pkey_index);
 	if (ret)
 		goto err_destroy_qp;
 
@@ -466,11 +482,15 @@ static int mlx5_ib_gsi_silent_drop(struct mlx5_ib_gsi_qp *gsi,
 static struct ib_qp *get_tx_qp(struct mlx5_ib_gsi_qp *gsi, struct ib_ud_wr *wr)
 {
 	struct mlx5_ib_dev *dev = to_mdev(gsi->rx_qp->device);
+	struct mlx5_ib_ah *ah = to_mah(wr->ah);
 	int qp_index = wr->pkey_index;
 
-	if (!mlx5_ib_deth_sqpn_cap(dev))
+	if (!gsi->num_qps)
 		return gsi->rx_qp;
 
+	if (dev->lag_active && ah->xmit_port)
+		qp_index = ah->xmit_port - 1;
+
 	if (qp_index >= gsi->num_qps)
 		return NULL;
 
drivers/infiniband/hw/mlx5/main.c +2 -0
index 6679756506e6..2db2309dde47 100644
@@ -53,6 +53,7 @@
 #include <linux/list.h>
 #include <rdma/ib_smi.h>
 #include <rdma/ib_umem.h>
+#include <rdma/lag.h>
 #include <linux/in.h>
 #include <linux/etherdevice.h>
 #include "mlx5_ib.h"
@@ -6549,6 +6550,7 @@ static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
 	dev->ib_dev.phys_port_cnt	= dev->num_ports;
 	dev->ib_dev.num_comp_vectors    = mlx5_comp_vectors_count(mdev);
 	dev->ib_dev.dev.parent		= mdev->device;
+	dev->ib_dev.lag_flags		= RDMA_LAG_FLAGS_HASH_ALL_SLAVES;
 
 	mutex_init(&dev->cap_mask_mutex);
 	INIT_LIST_HEAD(&dev->qp_list);
drivers/infiniband/hw/mlx5/mlx5_ib.h +1 -0
index a041a1357a13..bdc36395db4a 100644
@@ -471,6 +471,7 @@ struct mlx5_ib_qp {
 	 * but not take effective
 	 */
 	u32                     counter_pending;
+	u16			gsi_lag_port;
 };
 
 struct mlx5_ib_cq_buf {
drivers/infiniband/hw/mlx5/qp.c +35 -17

121 lines not on this page · raw message

include/linux/mlx5/mlx5_ifc.h +3 -1
index 69b27c7dfc3e..a3b6c92e889e 100644
@@ -1296,7 +1296,9 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8         wol_p[0x1];
 
 	u8         stat_rate_support[0x10];
-	u8         reserved_at_1f0[0xc];
+	u8         reserved_at_1f0[0x8];
+	u8         init2_lag_tx_port_affinity[0x1];
+	u8         reserved_at_1f9[0x3];
 	u8         cqe_version[0x4];
 
 	u8         compact_address_vector[0x1];
include/linux/mlx5/qp.h +2 -0
index ae63b1ae9004..fab88b0c76f9 100644
@@ -66,6 +66,7 @@ enum mlx5_qp_optpar {
 	MLX5_QP_OPTPAR_RETRY_COUNT		= 1 << 12,
 	MLX5_QP_OPTPAR_RNR_RETRY		= 1 << 13,
 	MLX5_QP_OPTPAR_ACK_TIMEOUT		= 1 << 14,
+	MLX5_QP_OPTPAR_LAG_TX_AFF		= 1 << 15,
 	MLX5_QP_OPTPAR_PRI_PORT			= 1 << 16,
 	MLX5_QP_OPTPAR_SRQN			= 1 << 18,
 	MLX5_QP_OPTPAR_CQN_RCV			= 1 << 19,
@@ -315,6 +316,7 @@ struct mlx5_av {
 struct mlx5_ib_ah {
 	struct ib_ah		ibah;
 	struct mlx5_av		av;
+	u8			xmit_port;
 };
 
 static inline struct mlx5_ib_ah *to_mah(struct ib_ah *ibah)
Signature · 1 line
2.17.2