net: implement mechanism for HW based QOS
Subsystems (3)
RFC PATCH 1/4 net: implement mechanism for HW based QOS
This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged.
Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc.
By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters.
With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with a single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs.
To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc.
This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB.
---
include/linux/netdevice.h | 65 +++++++++++++++++++++++++++++++++++++++++++++ net/core/dev.c | 39 ++++++++++++++++++++++++++- 2 files changed, 103 insertions(+), 1 deletions(-)
include/linux/netdevice.h +65 -0
@@ -646,6 +646,12 @@ struct xps_dev_maps { (nr_cpu_ids * sizeof(struct xps_map *)))#endif /* CONFIG_XPS */ +/* HW offloaded queuing disciplines txq count and offset maps */ +struct netdev_tc_txq { +u16 count; +u16 offset; +}; +/* * This structure defines the management hooks for network devices. * The following hooks can be defined; unless noted otherwise, they are
@@ -1146,6 +1152,10 @@ struct net_device {/* Data Center Bridging netlink ops */ const struct dcbnl_rtnl_ops *dcbnl_ops;#endif +u8 max_tc; +u8 num_tc; +struct netdev_tc_txq *_tc_to_txq; +u8 prio_tc_map[16 ];#if defined(CONFIG_FCOE ) || defined(CONFIG_FCOE_MODULE )/* max exchange id for FCoE LRO by ddp */
@@ -1162,6 +1172,58 @@ struct net_device {#define NETDEV_ALIGN 32static inline +int netdev_get_prio_tc_map (const struct net_device *dev,u32 prio) +{ +return dev->prio_tc_map[prio &15 ]; +} + +static inline +int netdev_set_prio_tc_map (struct net_device *dev,u8 prio,u8 tc) +{ +if (tc >= dev->num_tc) +return -EINVAL ; + + dev->prio_tc_map[prio &15 ] = tc &15 ; +return 0 ; +} + +static inline +int netdev_set_tc_queue (struct net_device *dev,u8 tc,u16 count,u16 offset) +{ +struct netdev_tc_txq *tcp; + +if (tc >= dev->num_tc) +return -EINVAL ; + + tcp = &dev->_tc_to_txq[tc]; + tcp->count = count; + tcp->offset = offset; +return 0 ; +} + +static inline +struct netdev_tc_txq *netdev_get_tc_queue (const struct net_device *dev,u8 tc) +{ +return &dev->_tc_to_txq[tc]; +} + +static inline +int netdev_set_num_tc (struct net_device *dev,u8 num_tc) +{ +if (num_tc > dev->max_tc) +return -EINVAL ; + + dev->num_tc = num_tc; +return 0 ; +} + +static inline +u8 netdev_get_num_tc (const struct net_device *dev) +{ +return dev->num_tc; +} + +static inline struct netdev_queue *netdev_get_tx_queue (const struct net_device *dev,unsigned int index) {
@@ -1386,6 +1448,9 @@ static inline void unregister_netdevice(struct net_device *dev)unregister_netdevice_queue (dev,NULL ); } +extern int netdev_alloc_max_tc (struct net_device *dev,u8 tc); +extern void netdev_free_tc (struct net_device *dev); +extern int netdev_refcnt_read (const struct net_device *dev);extern void free_netdev (struct net_device *dev);extern void synchronize_net (void );
net/core/dev.c +38 -1
@@ -2118,6 +2118,8 @@ static u32 hashrnd __read_mostly;u16 skb_tx_hash (const struct net_device *dev,const struct sk_buff *skb) {u32 hash; +u16 qoffset =0 ; +u16 qcount = dev->real_num_tx_queues;if (skb_rx_queue_recorded (skb)) { hash =skb_get_rx_queue (skb);
@@ -2126,13 +2128,20 @@ u16 skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb)return hash; } +if (dev->num_tc) { +u8 tc =netdev_get_prio_tc_map (dev, skb->priority); +struct netdev_tc_txq *tcp =netdev_get_tc_queue (dev, tc); + qoffset = tcp->offset; + qcount = tcp->count; + } +if (skb->sk && skb->sk->sk_hash) hash = skb->sk->sk_hash;else hash = (__force u16 ) skb->protocol ^ skb->rxhash; hash =jhash_1word (hash, hashrnd); -return (u16) (((u64 ) hash * dev->real_num_tx_queues) >>32 ); +return (u16) ((((u64 ) hash * qcount)) >>32 ) + qoffset; }EXPORT_SYMBOL (skb_tx_hash);
@@ -5091,6 +5100,33 @@ void netif_stacked_transfer_operstate(const struct net_device *rootdev, }EXPORT_SYMBOL (netif_stacked_transfer_operstate); +int netdev_alloc_max_tc (struct net_device *dev,u8 tcs) +{ +struct netdev_tc_txq *tcp; + +if (tcs >16 ) +return -EINVAL ; + + tcp =kcalloc (tcs,sizeof (*tcp),GFP_KERNEL ); +if (!tcp) +return -ENOMEM ; + + dev->_tc_to_txq = tcp; + dev->max_tc = tcs; +return 0 ; +} +EXPORT_SYMBOL (netdev_alloc_max_tc); + +void netdev_free_tc (struct net_device *dev) +{ + dev->max_tc =0 ; + dev->num_tc =0 ; +memset (dev->prio_tc_map,0 ,sizeof (dev->prio_tc_map)); +kfree (dev->_tc_to_txq); + dev->_tc_to_txq =NULL ; +} +EXPORT_SYMBOL (netdev_free_tc); +#ifdef CONFIG_RPS static int netif_alloc_rx_queues (struct net_device *dev) {
@@ -5699,6 +5735,7 @@ void free_netdev(struct net_device *dev)#ifdef CONFIG_RPS kfree (dev->_rx);#endif +netdev_free_tc (dev);kfree (rcu_dereference_raw (dev->ingress_queue));
RFC PATCH 2/4 net/sched: Allow multiple mq qdisc to be used as non-root
This patch modifies the mq qdisc to allow multiple mq qdiscs to be used. Allowing TX queues to be grouped for management.
This allows a root container qdisc to create multiple traffic classes and use the mq qdisc as a default queueing discipline. It is expected other queueing disciplines can then be grafted to the container as needed.
---
net/sched/sch_mq.c | 73 +++++++++++++++++++++++++++++++++++++++++----------- 1 files changed, 57 insertions(+), 16 deletions(-)
net/sched/sch_mq.c +57 -16
@@ -19,17 +19,42 @@struct mq_sched {struct Qdisc **qdiscs; +u8 num_tc; }; +static void mq_queues (struct net_device *dev,struct Qdisc *sch, +unsigned int *count,unsigned int *offset) +{ +struct mq_sched *priv =qdisc_priv (sch); +if (priv->num_tc) { +struct netdev_tc_txq *tc; +int queue =TC_H_MIN (sch->parent) -1 ; + + tc =netdev_get_tc_queue (dev, queue); +if (count) + *count = tc->count; +if (offset) + *offset = tc->offset; + }else { +if (count) + *count = dev->num_tx_queues; +if (offset) + *offset =0 ; + } +} +static void mq_destroy (struct Qdisc *sch) {struct net_device *dev =qdisc_dev (sch);struct mq_sched *priv =qdisc_priv (sch); -unsigned int ntx; +unsigned int ntx, count;if (!priv->qdiscs)return ; -for (ntx =0 ; ntx < dev->num_tx_queues && priv->qdiscs[ntx]; ntx++) + +mq_queues (dev, sch, &count,NULL ); + +for (ntx =0 ; ntx < count && priv->qdiscs[ntx]; ntx++)qdisc_destroy (priv->qdiscs[ntx]);kfree (priv->qdiscs); }
@@ -41,21 +66,26 @@ static int mq_init(struct Qdisc *sch, struct nlattr *opt)struct netdev_queue *dev_queue;struct Qdisc *qdisc;unsigned int ntx; +unsigned int count, offset; -if (sch->parent !=TC_H_ROOT ) +if (sch->parent !=TC_H_ROOT && !dev->num_tc)return -EOPNOTSUPP ;if (!netif_is_multiqueue (dev))return -EOPNOTSUPP ; +/* Record num tc's in priv so we can tear down cleanly */ + priv->num_tc = dev->num_tc; +mq_queues (dev, sch, &count, &offset); +/* pre-allocate qdiscs, attachment can't fail */ - priv->qdiscs =kcalloc (dev->num_tx_queues,sizeof (priv->qdiscs[0 ]), + priv->qdiscs =kcalloc (count,sizeof (priv->qdiscs[0 ]),GFP_KERNEL );if (priv->qdiscs ==NULL )return -ENOMEM ; -for (ntx =0 ; ntx < dev->num_tx_queues; ntx++) { - dev_queue =netdev_get_tx_queue (dev, ntx); +for (ntx =0 ; ntx < count; ntx++) { + dev_queue =netdev_get_tx_queue (dev, ntx + offset); qdisc =qdisc_create_dflt (dev_queue, &pfifo_fast_ops,TC_H_MAKE (TC_H_MAJ (sch->handle),TC_H_MIN (ntx +1 )));
@@ -65,7 +95,8 @@ static int mq_init(struct Qdisc *sch, struct nlattr *opt) priv->qdiscs[ntx] = qdisc; } - sch->flags |=TCQ_F_MQROOT ; +if (!priv->num_tc) + sch->flags |=TCQ_F_MQROOT ;return 0 ; err:
@@ -78,9 +109,11 @@ static void mq_attach(struct Qdisc *sch)struct net_device *dev =qdisc_dev (sch);struct mq_sched *priv =qdisc_priv (sch);struct Qdisc *qdisc; -unsigned int ntx; +unsigned int ntx, count; -for (ntx =0 ; ntx < dev->num_tx_queues; ntx++) { +mq_queues (dev, sch, &count,NULL ); + +for (ntx =0 ; ntx < count; ntx++) { qdisc = priv->qdiscs[ntx]; qdisc =dev_graft_qdisc (qdisc->dev_queue, qdisc);if (qdisc)
@@ -94,14 +127,17 @@ static int mq_dump(struct Qdisc *sch, struct sk_buff *skb) {struct net_device *dev =qdisc_dev (sch);struct Qdisc *qdisc; -unsigned int ntx; +unsigned int ntx, count, offset; + +mq_queues (dev, sch, &count, &offset); sch->q.qlen =0 ;memset (&sch->bstats,0 ,sizeof (sch->bstats));memset (&sch->qstats,0 ,sizeof (sch->qstats)); -for (ntx =0 ; ntx < dev->num_tx_queues; ntx++) { - qdisc =netdev_get_tx_queue (dev, ntx)->qdisc_sleeping; +for (ntx =0 ; ntx < count; ntx++) { +int txq = ntx + offset; + qdisc =netdev_get_tx_queue (dev, txq)->qdisc_sleeping;spin_lock_bh (qdisc_lock (qdisc)); sch->q.qlen += qdisc->q.qlen; sch->bstats.bytes += qdisc->bstats.bytes;
@@ -120,10 +156,13 @@ static struct netdev_queue *mq_queue_get(struct Qdisc *sch, unsigned long cl) {struct net_device *dev =qdisc_dev (sch);unsigned long ntx = cl -1 ; +unsigned int count, offset; + +mq_queues (dev, sch, &count, &offset); -if (ntx >= dev->num_tx_queues) +if (ntx >= count)return NULL ; -return netdev_get_tx_queue (dev, ntx); +return netdev_get_tx_queue (dev, offset + ntx); } static struct netdev_queue *mq_select_queue(struct Qdisc *sch,
@@ -203,13 +242,15 @@ static int mq_dump_class_stats(struct Qdisc *sch, unsigned long cl,static void mq_walk (struct Qdisc *sch,struct qdisc_walker *arg) {struct net_device *dev =qdisc_dev (sch); -unsigned int ntx; +unsigned int ntx, count; + +mq_queues (dev, sch, &count,NULL );if (arg->stop)return ; arg->count = arg->skip; -for (ntx = arg->skip; ntx < dev->num_tx_queues; ntx++) { +for (ntx = arg->skip; ntx < count; ntx++) {if (arg->fn (sch, ntx +1 , arg) <0 ) { arg->stop =1 ;break ;
RFC PATCH 3/4 net/sched: implement a root container qdisc sch_mclass
In reply to John FastabendThis implements a mclass 'multi-class' queueing discipline that by default creates multiple mq qdisc's one for each traffic class. Each mq qdisc then owns a range of queues per the netdev_tc_txq mappings.
Using the mclass qdisc the number of tcs currently in use along with the range of queues alloted to each class can be configured. By default skbs are mapped to traffic classes using the skb priority. This mapping is configurable.
To support HW QOS schemes on inflexible HW that require fixed mappings between queues and classes a net device ops ndo_setup_tc is used. The HW setup may be overridden.
Finally, qdiscs graft'd onto the mclass qdisc must be mq like in that they must map queueing disciplines onto the netdev_queue.
---
include/linux/netdevice.h | 3 include/linux/pkt_sched.h | 9 + include/net/sch_generic.h | 1 net/sched/Makefile | 2 net/sched/sch_api.c | 1 net/sched/sch_generic.c | 8 + net/sched/sch_mclass.c | 332 +++++++++++++++++++++++++++++++++++++++++++++ 7 files changed, 354 insertions(+), 2 deletions(-) create mode 100644 net/sched/sch_mclass.c
include/linux/netdevice.h +3 -0
@@ -762,6 +762,8 @@ struct netdev_tc_txq {* int (*ndo_set_vf_port)(struct net_device *dev, int vf, * struct nlattr *port[]); * int (*ndo_get_vf_port)(struct net_device *dev, int vf, struct sk_buff *skb); +* +* int (*ndo_setup_tc)(struct net_device *dev, int tc); */ #define HAVE_NET_DEVICE_OPS struct net_device_ops {
@@ -820,6 +822,7 @@ struct net_device_ops { struct nlattr *port[]);int (*ndo_get_vf_port)(struct net_device *dev,int vf,struct sk_buff *skb); +int (*ndo_setup_tc)(struct net_device *dev,u8 tc);#if defined(CONFIG_FCOE ) || defined(CONFIG_FCOE_MODULE )int (*ndo_fcoe_enable)(struct net_device *dev);int (*ndo_fcoe_disable)(struct net_device *dev);
include/linux/pkt_sched.h +9 -0
@@ -481,4 +481,13 @@ struct tc_drr_stats { __u32 deficit; }; +/* MCLASS */ +struct tc_mclass_qopt { +__u8 num_tc; +__u8 prio_tc_map[16 ]; +__u8 hw; +__u16 count[16 ]; +__u16 offset[16 ]; +}; +#endif
include/net/sch_generic.h +1 -0
@@ -276,6 +276,7 @@ extern struct Qdisc noop_qdisc;extern struct Qdisc_ops noop_qdisc_ops;extern struct Qdisc_ops pfifo_fast_ops;extern struct Qdisc_ops mq_qdisc_ops; +extern struct Qdisc_ops mclass_qdisc_ops;struct Qdisc_class_common {u32 classid;
net/sched/Makefile +1 -1
@@ -2,7 +2,7 @@ # Makefile for the Linux Traffic Control Unit. # -obj-y := sch_generic.o sch_mq.o +obj-y := sch_generic.o sch_mq.o sch_mclass.o obj-$(CONFIG_NET_SCHED) += sch_api.o sch_blackhole.o obj-$(CONFIG_NET_CLS) += cls_api.o
net/sched/sch_api.c +1 -0
@@ -1770,6 +1770,7 @@ static int __init pktsched_init(void)register_qdisc (&bfifo_qdisc_ops);register_qdisc (&pfifo_head_drop_qdisc_ops);register_qdisc (&mq_qdisc_ops); +register_qdisc (&mclass_qdisc_ops);rtnl_register (PF_UNSPEC ,RTM_NEWQDISC , tc_modify_qdisc,NULL );rtnl_register (PF_UNSPEC ,RTM_DELQDISC , tc_get_qdisc,NULL );
net/sched/sch_generic.c +7 -1
@@ -709,7 +709,13 @@ static void attach_default_qdiscs(struct net_device *dev) dev->qdisc = txq->qdisc_sleeping;atomic_inc (&dev->qdisc->refcnt); }else { - qdisc =qdisc_create_dflt (txq, &mq_qdisc_ops,TC_H_ROOT ); +if (dev->num_tc) + qdisc =qdisc_create_dflt (txq, &mclass_qdisc_ops, +TC_H_ROOT ); +else + qdisc =qdisc_create_dflt (txq, &mq_qdisc_ops, +TC_H_ROOT ); +if (qdisc) { qdisc->ops->attach (qdisc); dev->qdisc = qdisc;
net/sched/sch_mclass.c new file +332 -0
332 lines not on this page · raw message
RFC PATCH 4/4 ixgbe: add multiple txqs per tc
In reply to John FastabendThis illustrates the usage model for hardware QOS offloading.
Currently, DCB only enables a single queue per tc. Due to complications with how to map tc filter rules to traffic classes when multiple queues are enabled. And previously there was no mechanism to map flows to multiple queues by priority.
Using the QOS offloading API we allocate multiple queues per tc and configure the stack to hash across these queues. The hardware then offloads the DCB extended transmission selection algorithm. Sockets can set the priority using the SO_PRIORITY socket option and expect ETS to work.
---
drivers/net/ixgbe/ixgbe.h | 2 drivers/net/ixgbe/ixgbe_dcb_nl.c | 4 - drivers/net/ixgbe/ixgbe_main.c | 256 ++++++++++++++++++++++---------------- 3 files changed, 149 insertions(+), 113 deletions(-)
drivers/net/ixgbe/ixgbe.h +1 -1
@@ -243,7 +243,7 @@ enum ixgbe_ring_f_enum { RING_F_ARRAY_SIZE /* must be last in enum set */ }; -#define IXGBE_MAX_DCB_INDICES 8 +#define IXGBE_MAX_DCB_INDICES 64#define IXGBE_MAX_RSS_INDICES 16#define IXGBE_MAX_VMDQ_INDICES 64#define IXGBE_MAX_FDIR_INDICES 64
drivers/net/ixgbe/ixgbe_dcb_nl.c +2 -2
@@ -354,12 +354,12 @@ static u8 ixgbe_dcbnl_set_all(struct net_device *netdev) {struct ixgbe_adapter *adapter =netdev_priv (netdev);int ret; +int tc =max (netdev->num_tc,MAX_TRAFFIC_CLASS );if (!adapter->dcb_set_bitmap)return DCB_NO_HW_CHG ; - ret =ixgbe_copy_dcb_cfg (&adapter->temp_dcb_cfg, &adapter->dcb_cfg, - adapter->ring_feature[RING_F_DCB ].indices); + ret =ixgbe_copy_dcb_cfg (&adapter->temp_dcb_cfg, &adapter->dcb_cfg, tc);if (ret)return DCB_NO_HW_CHG ;
drivers/net/ixgbe/ixgbe_main.c +146 -110
381 lines not on this page · raw message
Quoted 38 lines · John Fastabend
This patch provides a mechanism for lower layer devices to steer traffic using
skb->priorityto tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged.Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific
select_queueor thequeue_mappingaction in the qdisc.By using
select_queuefor this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters.With the mechanism in this patch users can set skb priority using expected methods ie
setsockopt()or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with a single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs.To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc.
Quoted 5 lines
This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB.
Very nice Changelog !
Quoted 25 lines
Signed-off-by John Fastabend <[email protected]>---
include/linux/netdevice.h | 65 +++++++++++++++++++++++++++++++++++++++++++++ net/core/dev.c | 39 ++++++++++++++++++++++++++- 2 files changed, 103 insertions(+), 1 deletions(-)include/linux/netdevice.h +6
-0@@ -646,6 +646,12 @@ struct xps_dev_maps { (nr_cpu_ids * sizeof(struct xps_map *)))#endif /* CONFIG_XPS */ +/* HW offloaded queuing disciplines txq count and offset maps */ +struct netdev_tc_txq { +u16 count; +u16 offset; +}; +/* * This structure defines the management hooks for network devices. * The following hooks can be defined; unless noted otherwise, they are
include/linux/netdevice.h struct net_device { · line 1157
Quoted 6 lines
include/linux/netdevice.h +3
-0@@ -1146,6 +1152,10 @@ struct net_device {/* Data Center Bridging netlink ops */ const struct dcbnl_rtnl_ops *dcbnl_ops;#endif +u8 max_tc; +u8 num_tc; +struct netdev_tc_txq *_tc_to_txq;
Given that this is up to 16*4 bytes (64), shouldnt we embed this in net_device struct to avoid one dereference ?
Quoted 110 lines
include/linux/netdevice.h +56
-0+u8 prio_tc_map[16 ];#if defined(CONFIG_FCOE ) || defined(CONFIG_FCOE_MODULE )/* max exchange id for FCoE LRO by ddp */ @@ -1162,6 +1172,58 @@ struct net_device {#define NETDEV_ALIGN 32static inline +int netdev_get_prio_tc_map (const struct net_device *dev,u32 prio) +{ +return dev->prio_tc_map[prio &15 ]; +} + +static inline +int netdev_set_prio_tc_map (struct net_device *dev,u8 prio,u8 tc) +{ +if (tc >= dev->num_tc) +return -EINVAL ; + + dev->prio_tc_map[prio &15 ] = tc &15 ; +return 0 ; +} + +static inline +int netdev_set_tc_queue (struct net_device *dev,u8 tc,u16 count,u16 offset) +{ +struct netdev_tc_txq *tcp; + +if (tc >= dev->num_tc) +return -EINVAL ; + + tcp = &dev->_tc_to_txq[tc]; + tcp->count = count; + tcp->offset = offset; +return 0 ; +} + +static inline +struct netdev_tc_txq *netdev_get_tc_queue (const struct net_device *dev,u8 tc) +{ +return &dev->_tc_to_txq[tc]; +} + +static inline +int netdev_set_num_tc (struct net_device *dev,u8 num_tc) +{ +if (num_tc > dev->max_tc) +return -EINVAL ; + + dev->num_tc = num_tc; +return 0 ; +} + +static inline +u8 netdev_get_num_tc (const struct net_device *dev) +{ +return dev->num_tc; +} + +static inline struct netdev_queue *netdev_get_tx_queue (const struct net_device *dev,unsigned int index) {@@ -1386,6 +1448,9 @@ static inline void unregister_netdevice(struct net_device *dev)unregister_netdevice_queue (dev,NULL ); } +extern int netdev_alloc_max_tc (struct net_device *dev,u8 tc); +extern void netdev_free_tc (struct net_device *dev); +extern int netdev_refcnt_read (const struct net_device *dev);extern void free_netdev (struct net_device *dev);extern void synchronize_net (void );net/core/dev.c +10
-1@@ -2118,6 +2118,8 @@ static u32 hashrnd __read_mostly;u16 skb_tx_hash (const struct net_device *dev,const struct sk_buff *skb) {u32 hash; +u16 qoffset =0 ; +u16 qcount = dev->real_num_tx_queues;if (skb_rx_queue_recorded (skb)) { hash =skb_get_rx_queue (skb);@@ -2126,13 +2128,20 @@ u16 skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb)return hash; } +if (dev->num_tc) { +u8 tc =netdev_get_prio_tc_map (dev, skb->priority); +struct netdev_tc_txq *tcp =netdev_get_tc_queue (dev, tc); + qoffset = tcp->offset; + qcount = tcp->count; + } +if (skb->sk && skb->sk->sk_hash) hash = skb->sk->sk_hash;else hash = (__force u16 ) skb->protocol ^ skb->rxhash; hash =jhash_1word (hash, hashrnd); -return (u16) (((u64 ) hash * dev->real_num_tx_queues) >>32 ); +return (u16) ((((u64 ) hash * qcount)) >>32 ) + qoffset; }EXPORT_SYMBOL (skb_tx_hash);@@ -5091,6 +5100,33 @@ void netif_stacked_transfer_operstate(const struct net_device *rootdev, }
net/core/dev.c void netif_stacked_transfer_operstate(const struct net_device *rootdev, · line 5110
Quoted 10 lines
net/core/dev.c +8
-0EXPORT_SYMBOL (netif_stacked_transfer_operstate); +int netdev_alloc_max_tc (struct net_device *dev,u8 tcs) +{ +struct netdev_tc_txq *tcp; + +if (tcs >16 ) +return -EINVAL ; + + tcp =kcalloc (tcs,sizeof (*tcp),GFP_KERNEL );
common risk : allocating less than one cache line, and this possibly can have false sharing.
I would just embed the thing.
Quoted 207 lines · Eric Dumazet
Quoted 38 lines · John Fastabend
This patch provides a mechanism for lower layer devices to steer traffic using
skb->priorityto tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged.Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific
select_queueor thequeue_mappingaction in the qdisc.By using
select_queuefor this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters.With the mechanism in this patch users can set skb priority using expected methods ie
setsockopt()or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with a single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs.To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc.
Quoted 5 lines
This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB.
Very nice Changelog !
Quoted 25 lines
Signed-off-by John Fastabend <[email protected]>---
include/linux/netdevice.h | 65 +++++++++++++++++++++++++++++++++++++++++++++ net/core/dev.c | 39 ++++++++++++++++++++++++++- 2 files changed, 103 insertions(+), 1 deletions(-)include/linux/netdevice.h +6
-0@@ -646,6 +646,12 @@ struct xps_dev_maps { (nr_cpu_ids * sizeof(struct xps_map *)))#endif /* CONFIG_XPS */ +/* HW offloaded queuing disciplines txq count and offset maps */ +struct netdev_tc_txq { +u16 count; +u16 offset; +}; +/* * This structure defines the management hooks for network devices. * The following hooks can be defined; unless noted otherwise, they are include/linux/netdevice.h struct net_device { · line 1157
Quoted 6 lines
include/linux/netdevice.h +3
-0@@ -1146,6 +1152,10 @@ struct net_device {/* Data Center Bridging netlink ops */ const struct dcbnl_rtnl_ops *dcbnl_ops;#endif +u8 max_tc; +u8 num_tc; +struct netdev_tc_txq *_tc_to_txq;Given that this is up to 16*4 bytes (64), shouldnt we embed this in
net_devicestruct to avoid one dereference ?Quoted 110 lines
include/linux/netdevice.h +56
-0+u8 prio_tc_map[16 ];#if defined(CONFIG_FCOE ) || defined(CONFIG_FCOE_MODULE )/* max exchange id for FCoE LRO by ddp */ @@ -1162,6 +1172,58 @@ struct net_device {#define NETDEV_ALIGN 32static inline +int netdev_get_prio_tc_map (const struct net_device *dev,u32 prio) +{ +return dev->prio_tc_map[prio &15 ]; +} + +static inline +int netdev_set_prio_tc_map (struct net_device *dev,u8 prio,u8 tc) +{ +if (tc >= dev->num_tc) +return -EINVAL ; + + dev->prio_tc_map[prio &15 ] = tc &15 ; +return 0 ; +} + +static inline +int netdev_set_tc_queue (struct net_device *dev,u8 tc,u16 count,u16 offset) +{ +struct netdev_tc_txq *tcp; + +if (tc >= dev->num_tc) +return -EINVAL ; + + tcp = &dev->_tc_to_txq[tc]; + tcp->count = count; + tcp->offset = offset; +return 0 ; +} + +static inline +struct netdev_tc_txq *netdev_get_tc_queue (const struct net_device *dev,u8 tc) +{ +return &dev->_tc_to_txq[tc]; +} + +static inline +int netdev_set_num_tc (struct net_device *dev,u8 num_tc) +{ +if (num_tc > dev->max_tc) +return -EINVAL ; + + dev->num_tc = num_tc; +return 0 ; +} + +static inline +u8 netdev_get_num_tc (const struct net_device *dev) +{ +return dev->num_tc; +} + +static inline struct netdev_queue *netdev_get_tx_queue (const struct net_device *dev,unsigned int index) {@@ -1386,6 +1448,9 @@ static inline void unregister_netdevice(struct net_device *dev)unregister_netdevice_queue (dev,NULL ); } +extern int netdev_alloc_max_tc (struct net_device *dev,u8 tc); +extern void netdev_free_tc (struct net_device *dev); +extern int netdev_refcnt_read (const struct net_device *dev);extern void free_netdev (struct net_device *dev);extern void synchronize_net (void );net/core/dev.c +10
-1@@ -2118,6 +2118,8 @@ static u32 hashrnd __read_mostly;u16 skb_tx_hash (const struct net_device *dev,const struct sk_buff *skb) {u32 hash; +u16 qoffset =0 ; +u16 qcount = dev->real_num_tx_queues;if (skb_rx_queue_recorded (skb)) { hash =skb_get_rx_queue (skb);@@ -2126,13 +2128,20 @@ u16 skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb)return hash; } +if (dev->num_tc) { +u8 tc =netdev_get_prio_tc_map (dev, skb->priority); +struct netdev_tc_txq *tcp =netdev_get_tc_queue (dev, tc); + qoffset = tcp->offset; + qcount = tcp->count; + } +if (skb->sk && skb->sk->sk_hash) hash = skb->sk->sk_hash;else hash = (__force u16 ) skb->protocol ^ skb->rxhash; hash =jhash_1word (hash, hashrnd); -return (u16) (((u64 ) hash * dev->real_num_tx_queues) >>32 ); +return (u16) ((((u64 ) hash * qcount)) >>32 ) + qoffset; }EXPORT_SYMBOL (skb_tx_hash);@@ -5091,6 +5100,33 @@ void netif_stacked_transfer_operstate(const struct net_device *rootdev, }net/core/dev.c void netif_stacked_transfer_operstate(const struct net_device *rootdev, · line 5110
Quoted 10 lines
net/core/dev.c +8
-0EXPORT_SYMBOL (netif_stacked_transfer_operstate); +int netdev_alloc_max_tc (struct net_device *dev,u8 tcs) +{ +struct netdev_tc_txq *tcp; + +if (tcs >16 ) +return -EINVAL ; + + tcp =kcalloc (tcs,sizeof (*tcp),GFP_KERNEL );common risk : allocating less than one cache line, and this possibly can have false sharing.
Quoted 2 lines
I would just embed the thing.
Yes, I think you are right plus this simplifies the code a bit. I'll go ahead and do this. Thanks!