Add support to get xmit slave
Subsystems (6)
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_xmit_slave_get. 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/[email protected]/
Maor Gottlieb (10):
net/core: Introduce master_xmit_slave_get
bonding: Rename slave_arr to usable_slaves
bonding: Add helpers to get xmit slave
bonding: Implement ndo_xmit_slave_get 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 | 139 +++++++++ drivers/infiniband/core/verbs.c | 44 ++- drivers/infiniband/hw/mlx5/ah.c | 4 + 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 | 123 +++++--- drivers/net/bonding/bond_alb.c | 39 ++- drivers/net/bonding/bond_main.c | 272 +++++++++++++----- 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 | 3 + include/net/bond_alb.h | 4 + include/net/bonding.h | 3 +- include/net/lag.h | 32 +++ include/rdma/ib_verbs.h | 2 + include/rdma/lag.h | 22 ++ 20 files changed, 621 insertions(+), 179 deletions(-) create mode 100644 drivers/infiniband/core/lag.c create mode 100644 include/rdma/lag.h
Signature · 1 line
2.17.2
PATCH 1/9 bonding: Rename slave_arr to usable_slaves
This patch renames slave_arr to usable_slaves, since we will have two arrays, one for the usable slaves and the other to all slaves. In addition, exports the bond_skip_slave logic to function.
--- drivers/net/bonding/bond_alb.c | 4 +- drivers/net/bonding/bond_main.c | 85 +++++++++++++++++---------------- include/net/bonding.h | 2 +- 3 files changed, 48 insertions(+), 43 deletions(-)
drivers/net/bonding/bond_alb.c +2 -2
@@ -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 +45 -40
138 lines not on this page · raw message
include/net/bonding.h +1 -1
@@ -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
PATCH 2/9 bonding: Add helpers to get xmit slave
In reply to Maor GottliebThis helpers will be used by both the xmit function and the get xmit slave ndo.
--- drivers/net/bonding/bond_alb.c | 35 ++++++++---- drivers/net/bonding/bond_main.c | 94 +++++++++++++++++++++------------ include/net/bond_alb.h | 4 ++ 3 files changed, 89 insertions(+), 44 deletions(-)
drivers/net/bonding/bond_alb.c +26 -9
@@ -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); }
drivers/net/bonding/bond_main.c +59 -35
173 lines not on this page · raw message
include/net/bond_alb.h +4 -0
@@ -158,6 +158,10 @@ void bond_alb_handle_link_change(struct bonding *bond, struct slave *slave, charvoid 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
PATCH 3/9 bonding: Implement ndo_xmit_slave_get
In reply to Maor GottliebAdd implementation of ndo_xmit_slave_get.
When user sets the LAG_FLAGS_HASH_ALL_SLAVES bit and the xmit slave result is based on the hash, then the slave will be selected from the array of all the slaves.
--- drivers/net/bonding/bond_main.c | 123 +++++++++++++++++++++++++++----- include/net/bonding.h | 1 + 2 files changed, 105 insertions(+), 19 deletions(-)
drivers/net/bonding/bond_main.c +104 -19
@@ -4137,6 +4137,40 @@ 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); +if (usable) +kfree_rcu (usable, rcu); + + all =rtnl_dereference (bond->all_slaves); +rcu_assign_pointer (bond->all_slaves, all_slaves); +if (all) +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
@@ -4147,7 +4181,7 @@ 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 ;
@@ -4159,7 +4193,9 @@ 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; }
@@ -4168,20 +4204,19 @@ 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;
@@ -4191,8 +4226,6 @@ 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);
@@ -4200,14 +4233,17 @@ 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; }
@@ -4313,6 +4349,48 @@ 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, +u16 flags) +{ +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 (flags &LAG_FLAGS_HASH_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);
@@ -4434,6 +4512,7 @@ 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_xmit_get_slave = bond_xmit_get_slave, };static const struct device_type bond_type = {
@@ -4501,9 +4580,9 @@ 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 +4591,16 @@ 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 +1 -0
@@ -201,6 +201,7 @@ struct bonding {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 *all_slaves;/* Array of 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
PATCH 4/9 RDMA/core: Add LAG functionality
In reply to Maor GottliebAdd support to get the RoCE LAG xmit slave by building skb of the RoCE packet and call to master_xmit_slave_get. 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 | 139 +++++++++++++++++++++++++++++++ include/rdma/ib_verbs.h | 2 + include/rdma/lag.h | 22 +++++ 4 files changed, 164 insertions(+), 1 deletion(-) create mode 100644 drivers/infiniband/core/lag.c create mode 100644 include/rdma/lag.h
drivers/infiniband/core/Makefile +1 -1
@@ -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 +139 -0
139 lines not on this page · raw message
include/rdma/ib_verbs.h +2 -0
@@ -894,6 +894,7 @@ struct ib_ah_attr {struct roce_ah_attr {u8 dmac[ETH_ALEN ]; +struct net_device *xmit_slave; };struct opa_ah_attr {
@@ -2709,6 +2710,7 @@ 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 +22 -0
@@ -0,0 +1,22 @@ +/* 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 rdma_ah_attr *ah_attr); +int rdma_lag_get_ah_roce_slave (struct ib_device *device, +struct rdma_ah_attr *ah_attr); + +#endif /* _RDMA_LAG_H_ */
Signature · 1 line
2.17.2
PATCH 5/9 RDMA/core: Get xmit slave for LAG
In reply to Maor GottliebAdd a call to rdma_lag_get_ah_roce_slave when Address handle is created.
Low driver can use it to select the QP's affinity port.
--- drivers/infiniband/core/verbs.c | 44 ++++++++++++++++++++++----------- 1 file changed, 30 insertions(+), 14 deletions(-)
drivers/infiniband/core/verbs.c +30 -14
@@ -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>
@@ -554,8 +555,14 @@ 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); +if (ret) { +rdma_unfill_sgid_attr (ah_attr, old_sgid_attr); +return ERR_PTR (ret); + } + ah =_rdma_create_ah (pd, ah_attr, flags,NULL ); +rdma_lag_put_ah_roce_slave (ah_attr);rdma_unfill_sgid_attr (ah_attr, old_sgid_attr);return ah; }
@@ -1638,6 +1645,25 @@ 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); +if (ret) + goto out_av; + } }if (attr_mask &IB_QP_ALT_PATH ) {/*
@@ -1664,18 +1690,6 @@ 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,
@@ -1717,8 +1731,10 @@ 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->ah_attr);rdma_unfill_sgid_attr (&attr->ah_attr, old_sgid_attr_av); + }return ret; }
Signature · 1 line
2.17.2
PATCH 6/9 net/mlx5: Change lag mutex lock to spin lock
In reply to Maor GottliebThe 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
@@ -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
PATCH 7/9 net/mlx5: Add support to get lag physical port
In reply to Maor GottliebAdd 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
@@ -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
@@ -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
PATCH 8/9 RDMA/mlx5: Refactor affinity related code
In reply to Maor GottliebMove 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
@@ -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
PATCH 9/9 RDMA/mlx5: Set lag tx affinity according to slave
In reply to Maor GottliebThe 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 | 4 +++ 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 | 53 +++++++++++++++++++--------- include/linux/mlx5/mlx5_ifc.h | 4 ++- include/linux/mlx5/qp.h | 2 ++ 7 files changed, 75 insertions(+), 25 deletions(-)
drivers/infiniband/hw/mlx5/ah.c +4 -0
@@ -51,6 +51,10 @@ 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 (ah_attr->roce.xmit_slave) + ah->xmit_port = +mlx5_lag_get_slave_port (dev->mdev, + ah_attr->roce.xmit_slave); gid_type = ah_attr->grh.sgid_attr->gid_type; memcpy(ah->av.rmac, ah_attr->roce.dmac,
drivers/infiniband/hw/mlx5/gsi.c +27 -7
@@ -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
@@ -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
@@ -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 +36 -17
122 lines not on this page · raw message
include/linux/mlx5/mlx5_ifc.h +3 -1
@@ -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
@@ -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
Quoted 10 lines · Maor Gottlieb
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_xmit_slave_get. Every LAG module can
I do not see ndo_xmit_slave_get introduced ...
Quoted 16 lines
drivers/infiniband/core/Makefile | 2 +- drivers/infiniband/core/lag.c | 139 +++++++++ drivers/infiniband/core/verbs.c | 44 ++- drivers/infiniband/hw/mlx5/ah.c | 4 + 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 | 123 +++++--- drivers/net/bonding/bond_alb.c | 39 ++- drivers/net/bonding/bond_main.c | 272 +++++++++++++----- 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 | 3 +
nor any changes to netdevice.h. Bad spin of the patches?