mlx5 IB RoCE LAG
Subsystems (2)
cover mlx5 IB RoCE LAG
This feature keeps track, using netdev LAG events, of the bonding and link status of each port's PF netdev.
When both of the card's PF netdevs are enslaved exclusively to the same bond master, LAG state is active.
Bond mode must be one of:
1 (active-backup), 2 (XOR), or 4 (802.3ad).
During LAG, a single IB device is present for both ports. This allows for load balancing and high availability to be managed by the driver, abstracting it away from ULPs. Such an IB device is given a name of "mlx5_bond_X".
LAG mode is determined by the bond driver.
In load balancing (modes 2 and 4), QPs are assigned to ports in a round-robin fashion, on QP transition from RESET->INIT. In high availability (mode 1), all QPs are assigned to the active slave, determined by bond driver.
Please note that if the link state of a port becomes down when mode is load balancing, all QPs will be moved to the other port, and will be moved back once both ports are up again.
This feature itself split into mlx5_core and mlx5_ib parts, while mlx5_core was already submitted and accepted:
1) mlx5_core: Implements RoCE LAG infrastructure and API.
a) Bond device LAG events tracking.
b) Decisions as to whether LAG state is active.
c) Firmware commands to enter and exit LAG state.
d) Manages HA behavior.
2) mlx5_ib: Uses core infrastructure to implement RoCE LAG.
a) Chooses an "mlx5_bond_X" device name when LAG is active.
b) Implements QP load balancing.
c) Merges steering, such that both port's traffic arrives on PF0's root flow table. This is used for processing
both port's usermode Ethernet traffic on PF0.
d) Creates a special flow table that diverts all
non-usermode Ethernet traffic received on port 2
back to PF1's root flow table (since non-usermode Ethernet should not be affected by LAG).
Available in the "topic/roce-lag-mlx5" topic branch of this git repo:git://git.kernel.org/pub/scm/linux/kernel/git/leon/linux-rdma.git
Or for browsing:
https://git.kernel.org/cgit/linux/kernel/git/leon/linux-rdma.git/log/?h=topic/roce-lag-mlx5
Thanks
Aviv Heller (5):
IB/mlx5: Port events in RoCE now rely on netdev events IB/mlx5: Merge vports flow steering during LAG IB/mlx5: Port status track LAG master, when LAG is active IB/mlx5: Set unique device name on LAG
IB/mlx5: LAG QP load balancing
drivers/infiniband/hw/mlx5/main.c | 150 +++++++++++++++++++++++++++++++---- drivers/infiniband/hw/mlx5/mlx5_ib.h | 2 + drivers/infiniband/hw/mlx5/qp.c | 61 ++++++++++++-- 3 files changed, 191 insertions(+), 22 deletions(-)
--
2.7.4
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): 1474193296-28062-1-git-send-email-leon-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org
PATCH 1/5 IB/mlx5: Port events in RoCE now rely on netdev events
From: Aviv Heller <avivh-VPRAkNaXOzVWk0Htik3J/[email protected]>
Since ib_query_port() in RoCE returns the state of its netdev as the port state, it makes sense to propagate the port up/down events to ib_core when the netdev port state changes, instead of relying on traditional core events.
This also keeps both the event and ib_query_port() synchronized.
--- drivers/infiniband/hw/mlx5/main.c | 66 ++++++++++++++++++++++++++++++--------- 1 file changed, 51 insertions(+), 15 deletions(-)
drivers/infiniband/hw/mlx5/main.c +51 -15
@@ -106,13 +106,32 @@ static int mlx5_netdev_event(struct notifier_block *this, struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev, roce.nb); -if ((event !=NETDEV_UNREGISTER ) && (event !=NETDEV_REGISTER )) -return NOTIFY_DONE ; +switch (event) { +case NETDEV_REGISTER : +case NETDEV_UNREGISTER : +write_lock (&ibdev->roce.netdev_lock); +if (ndev->dev.parent == &ibdev->mdev->pdev->dev) + ibdev->roce.netdev = (event ==NETDEV_UNREGISTER ) ? +NULL : ndev; +write_unlock (&ibdev->roce.netdev_lock); +break ; -write_lock (&ibdev->roce.netdev_lock); -if (ndev->dev.parent == &ibdev->mdev->pdev->dev) - ibdev->roce.netdev = (event ==NETDEV_UNREGISTER ) ?NULL : ndev; -write_unlock (&ibdev->roce.netdev_lock); +case NETDEV_UP : +case NETDEV_DOWN : +if (ndev == ibdev->roce.netdev && ibdev->ib_active) { +struct ib_event ibev = {0 }; + + ibev.device = &ibdev->ib_dev; + ibev.event = (event ==NETDEV_UP ) ? +IB_EVENT_PORT_ACTIVE :IB_EVENT_PORT_ERR ; + ibev.element.port_num =1 ; +ib_dispatch_event (&ibev); + } +break ; + +default : +break ; + }return NOTIFY_DONE ; }
@@ -2097,14 +2116,19 @@ static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,break ;case MLX5_DEV_EVENT_PORT_UP : - ibev.event =IB_EVENT_PORT_ACTIVE ; - port = (u8 )param; -break ; -case MLX5_DEV_EVENT_PORT_DOWN :case MLX5_DEV_EVENT_PORT_INITIALIZED : - ibev.event =IB_EVENT_PORT_ERR ; port = (u8 )param; + +/* In RoCE, port up/down events are handled in +* mlx5_netdev_event(). +*/ +if (mlx5_ib_port_link_layer (&ibdev->ib_dev, port) == +IB_LINK_LAYER_ETHERNET ) +return ; + + ibev.event = (event ==MLX5_DEV_EVENT_PORT_UP ) ? +IB_EVENT_PORT_ACTIVE :IB_EVENT_PORT_ERR ;break ;case MLX5_DEV_EVENT_LID_CHANGE :
@@ -2509,14 +2533,24 @@ static void get_dev_fw_str(struct ib_device *ibdev, char *str, fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev)); } +static void mlx5_remove_roce_notifier (struct mlx5_ib_dev *dev) +{ +if (dev->roce.nb.notifier_call) { +unregister_netdevice_notifier (&dev->roce.nb); + dev->roce.nb.notifier_call =NULL ; + } +} +static int mlx5_enable_roce (struct mlx5_ib_dev *dev) {int err; dev->roce.nb.notifier_call = mlx5_netdev_event; err =register_netdevice_notifier (&dev->roce.nb); -if (err) +if (err) { + dev->roce.nb.notifier_call =NULL ;return err; + } err =mlx5_nic_vport_enable_roce (dev->mdev);if (err)
@@ -2525,14 +2559,13 @@ static int mlx5_enable_roce(struct mlx5_ib_dev *dev)return 0 ; err_unregister_netdevice_notifier: -unregister_netdevice_notifier (&dev->roce.nb); +mlx5_remove_roce_notifier (dev);return err; }static void mlx5_disable_roce (struct mlx5_ib_dev *dev) {mlx5_nic_vport_disable_roce (dev->mdev); -unregister_netdevice_notifier (&dev->roce.nb); }static void mlx5_ib_dealloc_q_counters (struct mlx5_ib_dev *dev)
@@ -2881,8 +2914,10 @@ err_rsrc:destroy_dev_resources (&dev->devr); err_disable_roce: -if (ll ==IB_LINK_LAYER_ETHERNET ) +if (ll ==IB_LINK_LAYER_ETHERNET ) {mlx5_disable_roce (dev); +mlx5_remove_roce_notifier (dev); + } err_free_port:kfree (dev->port);
@@ -2898,6 +2933,7 @@ static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)struct mlx5_ib_dev *dev = context;enum rdma_link_layer ll =mlx5_ib_port_link_layer (&dev->ib_dev,1 ); +mlx5_remove_roce_notifier (dev);ib_unregister_device (&dev->ib_dev);mlx5_ib_dealloc_q_counters (dev);destroy_umrc_res (dev);
Signature · 6 lines
2.7.4 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
PATCH 2/5 IB/mlx5: Merge vports flow steering during LAG
In reply to a message not in the archiveFrom: Aviv Heller <avivh-VPRAkNaXOzVWk0Htik3J/[email protected]>
This is done in two steps:
1) Issuing CREATE_VPORT_LAG in order to have Ethernet traffic from both ports arriving on PF0 root flowtable, so we will be able to catch all raw-eth traffic on PF0.
2) Creation of LAG demux flowtable in order to direct all non-raw-eth traffic back to its source port, assuring that normal Ethernet traffic "jumps" to the root flowtable of its RX port (non-LAG behavior).
--- drivers/infiniband/hw/mlx5/main.c | 49 ++++++++++++++++++++++++++++++++++++ drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 + 2 files changed, 50 insertions(+)
drivers/infiniband/hw/mlx5/main.c +49 -0
@@ -2533,6 +2533,47 @@ static void get_dev_fw_str(struct ib_device *ibdev, char *str, fw_rev_min(dev->mdev), fw_rev_sub(dev->mdev)); } +static int mlx5_roce_lag_init (struct mlx5_ib_dev *dev) +{ +struct mlx5_core_dev *mdev = dev->mdev; +struct mlx5_flow_namespace *ns =mlx5_get_flow_namespace (mdev, +MLX5_FLOW_NAMESPACE_LAG ); +struct mlx5_flow_table *ft; +int err; + +if (!ns || !mlx5_lag_is_active (mdev)) +return 0 ; + + err =mlx5_cmd_create_vport_lag (mdev); +if (err) +return err; + + ft =mlx5_create_lag_demux_flow_table (ns,0 ,0 ); +if (IS_ERR (ft)) { + err =PTR_ERR (ft); + goto err_destroy_vport_lag; + } + + dev->flow_db.lag_demux_ft = ft; +return 0 ; + +err_destroy_vport_lag: +mlx5_cmd_destroy_vport_lag (mdev); +return err; +} + +static void mlx5_roce_lag_cleanup (struct mlx5_ib_dev *dev) +{ +struct mlx5_core_dev *mdev = dev->mdev; + +if (dev->flow_db.lag_demux_ft) { +mlx5_destroy_flow_table (dev->flow_db.lag_demux_ft); + dev->flow_db.lag_demux_ft =NULL ; + +mlx5_cmd_destroy_vport_lag (mdev); + } +} +static void mlx5_remove_roce_notifier (struct mlx5_ib_dev *dev) {if (dev->roce.nb.notifier_call) {
@@ -2556,8 +2597,15 @@ static int mlx5_enable_roce(struct mlx5_ib_dev *dev)if (err) goto err_unregister_netdevice_notifier; + err =mlx5_roce_lag_init (dev); +if (err) + goto err_disable_roce; +return 0 ; +err_disable_roce: +mlx5_nic_vport_disable_roce (dev->mdev); + err_unregister_netdevice_notifier:mlx5_remove_roce_notifier (dev);return err;
@@ -2565,6 +2613,7 @@ err_unregister_netdevice_notifier:static void mlx5_disable_roce (struct mlx5_ib_dev *dev) { +mlx5_roce_lag_cleanup (dev);mlx5_nic_vport_disable_roce (dev->mdev); }
drivers/infiniband/hw/mlx5/mlx5_ib.h +1 -0
@@ -156,6 +156,7 @@ struct mlx5_ib_flow_handler {struct mlx5_ib_flow_db {struct mlx5_ib_flow_prio prios[MLX5_IB_NUM_FLOW_FT ]; +struct mlx5_flow_table *lag_demux_ft;/* Protect flow steering bypass flow tables * when add/del flow rules. * only single add/removal of flow steering rule could be done
Signature · 6 lines
2.7.4 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
PATCH 3/5 IB/mlx5: Port status track LAG master, when LAG is active
In reply to a message not in the archiveFrom: Aviv Heller <avivh-VPRAkNaXOzVWk0Htik3J/[email protected]>
When LAG is active, port up/down events should be triggered by tracking the LAG master, and not one of the two slave netdevs.
In the same manner, ib_query_port() should return the details of the LAG master.
--- drivers/infiniband/hw/mlx5/main.c | 31 ++++++++++++++++++++++++++++--- 1 file changed, 28 insertions(+), 3 deletions(-)
drivers/infiniband/hw/mlx5/main.c +28 -3
@@ -117,8 +117,17 @@ static int mlx5_netdev_event(struct notifier_block *this,break ;case NETDEV_UP : -case NETDEV_DOWN : -if (ndev == ibdev->roce.netdev && ibdev->ib_active) { +case NETDEV_DOWN : { +struct net_device *lag_ndev =mlx5_lag_get_roce_netdev (ibdev->mdev); +struct net_device *upper =NULL ; + +if (lag_ndev) { + upper =netdev_master_upper_dev_get (lag_ndev); +dev_put (lag_ndev); + } + +if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev)) + && ibdev->ib_active) {struct ib_event ibev = {0 }; ibev.device = &ibdev->ib_dev;
@@ -128,6 +137,7 @@ static int mlx5_netdev_event(struct notifier_block *this,ib_dispatch_event (&ibev); }break ; + }default :break ;
@@ -142,6 +152,10 @@ static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,struct mlx5_ib_dev *ibdev =to_mdev (device);struct net_device *ndev; + ndev =mlx5_lag_get_roce_netdev (ibdev->mdev); +if (ndev) +return ndev; +/* Ensure ndev does not disappear before we invoke dev_hold() */ read_lock (&ibdev->roce.netdev_lock);
@@ -157,7 +171,7 @@ static int mlx5_query_port_roce(struct ib_device *device, u8 port_num, struct ib_port_attr *props) {struct mlx5_ib_dev *dev =to_mdev (device); -struct net_device *ndev; +struct net_device *ndev, *upper;enum ib_mtu ndev_ib_mtu;u16 qkey_viol_cntr;
@@ -181,6 +195,17 @@ static int mlx5_query_port_roce(struct ib_device *device, u8 port_num,if (!ndev)return 0 ; +if (mlx5_lag_is_active (dev->mdev)) { +rcu_read_lock (); + upper =netdev_master_upper_dev_get_rcu (ndev); +if (upper) { +dev_put (ndev); + ndev = upper; +dev_hold (ndev); + } +rcu_read_unlock (); + } +if (netif_running (ndev) &&netif_carrier_ok (ndev)) { props->state =IB_PORT_ACTIVE ; props->phys_state =5 ;
Signature · 6 lines
2.7.4 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
PATCH 4/5 IB/mlx5: Set unique device name on LAG
In reply to a message not in the archiveFrom: Aviv Heller <avivh-VPRAkNaXOzVWk0Htik3J/[email protected]>
IB bond device name is now 'mlx5_bond_X', instead of 'mlx5_X'.
--- drivers/infiniband/hw/mlx5/main.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
drivers/infiniband/hw/mlx5/main.c +7 -1
@@ -2754,6 +2754,7 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)struct mlx5_ib_dev *dev;enum rdma_link_layer ll;int port_type_cap; +const char *name;int err;int i;
@@ -2786,7 +2787,12 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)MLX5_INIT_DOORBELL_LOCK (&dev->uar_lock); -strlcpy (dev->ib_dev.name,"mlx5_%d" ,IB_DEVICE_NAME_MAX ); +if (!mlx5_lag_is_active (mdev)) + name ="mlx5_%d" ; +else + name ="mlx5_bond_%d" ; + +strlcpy (dev->ib_dev.name, name,IB_DEVICE_NAME_MAX ); dev->ib_dev.owner =THIS_MODULE ; dev->ib_dev.node_type =RDMA_NODE_IB_CA ; dev->ib_dev.local_dma_lkey =0 /* not supported for now */ ;
Signature · 6 lines
2.7.4 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
PATCH 5/5 IB/mlx5: LAG QP load balancing
In reply to a message not in the archiveFrom: Aviv Heller <avivh-VPRAkNaXOzVWk0Htik3J/[email protected]>
When LAG is active, QP tx affinity (the physical port to which a QP is affined, or the TIS in case of raw-eth) is set in a round robin fashion during state transition from RESET to INIT.
--- drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 + drivers/infiniband/hw/mlx5/qp.c | 61 +++++++++++++++++++++++++++++++++--- 2 files changed, 57 insertions(+), 5 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +1 -0
@@ -603,6 +603,7 @@ struct mlx5_roce {rwlock_t netdev_lock;struct net_device *netdev;struct notifier_block nb; +atomic_t next_port; };struct mlx5_ib_dev {
drivers/infiniband/hw/mlx5/qp.c +56 -5
110 lines not on this page · raw message
Signature · 6 lines
2.7.4 -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html