From 80d19669ecd34423e85ca04f2210b0e42a47cb16 Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:26:41 -0700 Subject: [PATCH 1/7] net: Refactor XPS for CPUs and Rx queues Refactor XPS code to support Tx queue selection based on CPU(s) map or Rx queue(s) map. Signed-off-by: Amritha Nambiar Signed-off-by: David S. Miller --- include/linux/cpumask.h | 11 +- include/linux/netdevice.h | 98 +++++++++++++++++- net/core/dev.c | 211 +++++++++++++++++++++++++------------- net/core/net-sysfs.c | 4 +- 4 files changed, 244 insertions(+), 80 deletions(-) diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h index bf53d893ad02..57f20a0a7794 100644 --- a/include/linux/cpumask.h +++ b/include/linux/cpumask.h @@ -115,12 +115,17 @@ extern struct cpumask __cpu_active_mask; #define cpu_active(cpu) ((cpu) == 0) #endif +static inline void cpu_max_bits_warn(unsigned int cpu, unsigned int bits) +{ +#ifdef CONFIG_DEBUG_PER_CPU_MAPS + WARN_ON_ONCE(cpu >= bits); +#endif /* CONFIG_DEBUG_PER_CPU_MAPS */ +} + /* verify cpu argument to cpumask_* operators */ static inline unsigned int cpumask_check(unsigned int cpu) { -#ifdef CONFIG_DEBUG_PER_CPU_MAPS - WARN_ON_ONCE(cpu >= nr_cpumask_bits); -#endif /* CONFIG_DEBUG_PER_CPU_MAPS */ + cpu_max_bits_warn(cpu, nr_cpumask_bits); return cpu; } diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h index c6b377a15869..8bf8d6149f79 100644 --- a/include/linux/netdevice.h +++ b/include/linux/netdevice.h @@ -731,10 +731,15 @@ struct xps_map { */ struct xps_dev_maps { struct rcu_head rcu; - struct xps_map __rcu *cpu_map[0]; + struct xps_map __rcu *attr_map[0]; /* Either CPUs map or RXQs map */ }; -#define XPS_DEV_MAPS_SIZE(_tcs) (sizeof(struct xps_dev_maps) + \ + +#define XPS_CPU_DEV_MAPS_SIZE(_tcs) (sizeof(struct xps_dev_maps) + \ (nr_cpu_ids * (_tcs) * sizeof(struct xps_map *))) + +#define XPS_RXQ_DEV_MAPS_SIZE(_tcs, _rxqs) (sizeof(struct xps_dev_maps) +\ + (_rxqs * (_tcs) * sizeof(struct xps_map *))) + #endif /* CONFIG_XPS */ #define TC_MAX_QUEUE 16 @@ -1910,7 +1915,8 @@ struct net_device { int watchdog_timeo; #ifdef CONFIG_XPS - struct xps_dev_maps __rcu *xps_maps; + struct xps_dev_maps __rcu *xps_cpus_map; + struct xps_dev_maps __rcu *xps_rxqs_map; #endif #ifdef CONFIG_NET_CLS_ACT struct mini_Qdisc __rcu *miniq_egress; @@ -3259,6 +3265,92 @@ static inline void netif_wake_subqueue(struct net_device *dev, u16 queue_index) #ifdef CONFIG_XPS int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, u16 index); +int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, + u16 index, bool is_rxqs_map); + +/** + * netif_attr_test_mask - Test a CPU or Rx queue set in a mask + * @j: CPU/Rx queue index + * @mask: bitmask of all cpus/rx queues + * @nr_bits: number of bits in the bitmask + * + * Test if a CPU or Rx queue index is set in a mask of all CPU/Rx queues. + */ +static inline bool netif_attr_test_mask(unsigned long j, + const unsigned long *mask, + unsigned int nr_bits) +{ + cpu_max_bits_warn(j, nr_bits); + return test_bit(j, mask); +} + +/** + * netif_attr_test_online - Test for online CPU/Rx queue + * @j: CPU/Rx queue index + * @online_mask: bitmask for CPUs/Rx queues that are online + * @nr_bits: number of bits in the bitmask + * + * Returns true if a CPU/Rx queue is online. + */ +static inline bool netif_attr_test_online(unsigned long j, + const unsigned long *online_mask, + unsigned int nr_bits) +{ + cpu_max_bits_warn(j, nr_bits); + + if (online_mask) + return test_bit(j, online_mask); + + return (j < nr_bits); +} + +/** + * netif_attrmask_next - get the next CPU/Rx queue in a cpu/Rx queues mask + * @n: CPU/Rx queue index + * @srcp: the cpumask/Rx queue mask pointer + * @nr_bits: number of bits in the bitmask + * + * Returns >= nr_bits if no further CPUs/Rx queues set. + */ +static inline unsigned int netif_attrmask_next(int n, const unsigned long *srcp, + unsigned int nr_bits) +{ + /* -1 is a legal arg here. */ + if (n != -1) + cpu_max_bits_warn(n, nr_bits); + + if (srcp) + return find_next_bit(srcp, nr_bits, n + 1); + + return n + 1; +} + +/** + * netif_attrmask_next_and - get the next CPU/Rx queue in *src1p & *src2p + * @n: CPU/Rx queue index + * @src1p: the first CPUs/Rx queues mask pointer + * @src2p: the second CPUs/Rx queues mask pointer + * @nr_bits: number of bits in the bitmask + * + * Returns >= nr_bits if no further CPUs/Rx queues set in both. + */ +static inline int netif_attrmask_next_and(int n, const unsigned long *src1p, + const unsigned long *src2p, + unsigned int nr_bits) +{ + /* -1 is a legal arg here. */ + if (n != -1) + cpu_max_bits_warn(n, nr_bits); + + if (src1p && src2p) + return find_next_and_bit(src1p, src2p, nr_bits, n + 1); + else if (src1p) + return find_next_bit(src1p, nr_bits, n + 1); + else if (src2p) + return find_next_bit(src2p, nr_bits, n + 1); + + return n + 1; +} #else static inline int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, diff --git a/net/core/dev.c b/net/core/dev.c index dffed642e686..71059558dc39 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -2092,7 +2092,7 @@ static bool remove_xps_queue(struct xps_dev_maps *dev_maps, int pos; if (dev_maps) - map = xmap_dereference(dev_maps->cpu_map[tci]); + map = xmap_dereference(dev_maps->attr_map[tci]); if (!map) return false; @@ -2105,7 +2105,7 @@ static bool remove_xps_queue(struct xps_dev_maps *dev_maps, break; } - RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL); + RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); kfree_rcu(map, rcu); return false; } @@ -2135,31 +2135,58 @@ static bool remove_xps_queue_cpu(struct net_device *dev, return active; } +static void clean_xps_maps(struct net_device *dev, const unsigned long *mask, + struct xps_dev_maps *dev_maps, unsigned int nr_ids, + u16 offset, u16 count, bool is_rxqs_map) +{ + bool active = false; + int i, j; + + for (j = -1; j = netif_attrmask_next(j, mask, nr_ids), + j < nr_ids;) + active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, + count); + if (!active) { + if (is_rxqs_map) { + RCU_INIT_POINTER(dev->xps_rxqs_map, NULL); + } else { + RCU_INIT_POINTER(dev->xps_cpus_map, NULL); + + for (i = offset + (count - 1); count--; i--) + netdev_queue_numa_node_write( + netdev_get_tx_queue(dev, i), + NUMA_NO_NODE); + } + kfree_rcu(dev_maps, rcu); + } +} + static void netif_reset_xps_queues(struct net_device *dev, u16 offset, u16 count) { + const unsigned long *possible_mask = NULL; struct xps_dev_maps *dev_maps; - int cpu, i; - bool active = false; + unsigned int nr_ids; mutex_lock(&xps_map_mutex); - dev_maps = xmap_dereference(dev->xps_maps); + dev_maps = xmap_dereference(dev->xps_rxqs_map); + if (dev_maps) { + nr_ids = dev->num_rx_queues; + clean_xps_maps(dev, possible_mask, dev_maps, nr_ids, offset, + count, true); + + } + + dev_maps = xmap_dereference(dev->xps_cpus_map); if (!dev_maps) goto out_no_maps; - for_each_possible_cpu(cpu) - active |= remove_xps_queue_cpu(dev, dev_maps, cpu, - offset, count); - - if (!active) { - RCU_INIT_POINTER(dev->xps_maps, NULL); - kfree_rcu(dev_maps, rcu); - } - - for (i = offset + (count - 1); count--; i--) - netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i), - NUMA_NO_NODE); + if (num_possible_cpus() > 1) + possible_mask = cpumask_bits(cpu_possible_mask); + nr_ids = nr_cpu_ids; + clean_xps_maps(dev, possible_mask, dev_maps, nr_ids, offset, count, + false); out_no_maps: mutex_unlock(&xps_map_mutex); @@ -2170,8 +2197,8 @@ static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); } -static struct xps_map *expand_xps_map(struct xps_map *map, - int cpu, u16 index) +static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, + u16 index, bool is_rxqs_map) { struct xps_map *new_map; int alloc_len = XPS_MIN_MAP_ALLOC; @@ -2183,7 +2210,7 @@ static struct xps_map *expand_xps_map(struct xps_map *map, return map; } - /* Need to add queue to this CPU's existing map */ + /* Need to add tx-queue to this CPU's/rx-queue's existing map */ if (map) { if (pos < map->alloc_len) return map; @@ -2191,9 +2218,14 @@ static struct xps_map *expand_xps_map(struct xps_map *map, alloc_len = map->alloc_len * 2; } - /* Need to allocate new map to store queue on this CPU's map */ - new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, - cpu_to_node(cpu)); + /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's + * map + */ + if (is_rxqs_map) + new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); + else + new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, + cpu_to_node(attr_index)); if (!new_map) return NULL; @@ -2205,14 +2237,16 @@ static struct xps_map *expand_xps_map(struct xps_map *map, return new_map; } -int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, - u16 index) +int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, + u16 index, bool is_rxqs_map) { + const unsigned long *online_mask = NULL, *possible_mask = NULL; struct xps_dev_maps *dev_maps, *new_dev_maps = NULL; - int i, cpu, tci, numa_node_id = -2; + int i, j, tci, numa_node_id = -2; int maps_sz, num_tc = 1, tc = 0; struct xps_map *map, *new_map; bool active = false; + unsigned int nr_ids; if (dev->num_tc) { num_tc = dev->num_tc; @@ -2221,16 +2255,27 @@ int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, return -EINVAL; } - maps_sz = XPS_DEV_MAPS_SIZE(num_tc); + mutex_lock(&xps_map_mutex); + if (is_rxqs_map) { + maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); + dev_maps = xmap_dereference(dev->xps_rxqs_map); + nr_ids = dev->num_rx_queues; + } else { + maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); + if (num_possible_cpus() > 1) { + online_mask = cpumask_bits(cpu_online_mask); + possible_mask = cpumask_bits(cpu_possible_mask); + } + dev_maps = xmap_dereference(dev->xps_cpus_map); + nr_ids = nr_cpu_ids; + } + if (maps_sz < L1_CACHE_BYTES) maps_sz = L1_CACHE_BYTES; - mutex_lock(&xps_map_mutex); - - dev_maps = xmap_dereference(dev->xps_maps); - /* allocate memory for queue storage */ - for_each_cpu_and(cpu, cpu_online_mask, mask) { + for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), + j < nr_ids;) { if (!new_dev_maps) new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); if (!new_dev_maps) { @@ -2238,73 +2283,81 @@ int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, return -ENOMEM; } - tci = cpu * num_tc + tc; - map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) : + tci = j * num_tc + tc; + map = dev_maps ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; - map = expand_xps_map(map, cpu, index); + map = expand_xps_map(map, j, index, is_rxqs_map); if (!map) goto error; - RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map); + RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); } if (!new_dev_maps) goto out_no_new_maps; - for_each_possible_cpu(cpu) { + for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids), + j < nr_ids;) { /* copy maps belonging to foreign traffic classes */ - for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) { + for (i = tc, tci = j * num_tc; dev_maps && i--; tci++) { /* fill in the new device map from the old device map */ - map = xmap_dereference(dev_maps->cpu_map[tci]); - RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map); + map = xmap_dereference(dev_maps->attr_map[tci]); + RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); } /* We need to explicitly update tci as prevous loop * could break out early if dev_maps is NULL. */ - tci = cpu * num_tc + tc; + tci = j * num_tc + tc; - if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) { - /* add queue to CPU maps */ + if (netif_attr_test_mask(j, mask, nr_ids) && + netif_attr_test_online(j, online_mask, nr_ids)) { + /* add tx-queue to CPU/rx-queue maps */ int pos = 0; - map = xmap_dereference(new_dev_maps->cpu_map[tci]); + map = xmap_dereference(new_dev_maps->attr_map[tci]); while ((pos < map->len) && (map->queues[pos] != index)) pos++; if (pos == map->len) map->queues[map->len++] = index; #ifdef CONFIG_NUMA - if (numa_node_id == -2) - numa_node_id = cpu_to_node(cpu); - else if (numa_node_id != cpu_to_node(cpu)) - numa_node_id = -1; + if (!is_rxqs_map) { + if (numa_node_id == -2) + numa_node_id = cpu_to_node(j); + else if (numa_node_id != cpu_to_node(j)) + numa_node_id = -1; + } #endif } else if (dev_maps) { /* fill in the new device map from the old device map */ - map = xmap_dereference(dev_maps->cpu_map[tci]); - RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map); + map = xmap_dereference(dev_maps->attr_map[tci]); + RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); } /* copy maps belonging to foreign traffic classes */ for (i = num_tc - tc, tci++; dev_maps && --i; tci++) { /* fill in the new device map from the old device map */ - map = xmap_dereference(dev_maps->cpu_map[tci]); - RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map); + map = xmap_dereference(dev_maps->attr_map[tci]); + RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); } } - rcu_assign_pointer(dev->xps_maps, new_dev_maps); + if (is_rxqs_map) + rcu_assign_pointer(dev->xps_rxqs_map, new_dev_maps); + else + rcu_assign_pointer(dev->xps_cpus_map, new_dev_maps); /* Cleanup old maps */ if (!dev_maps) goto out_no_old_maps; - for_each_possible_cpu(cpu) { - for (i = num_tc, tci = cpu * num_tc; i--; tci++) { - new_map = xmap_dereference(new_dev_maps->cpu_map[tci]); - map = xmap_dereference(dev_maps->cpu_map[tci]); + for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids), + j < nr_ids;) { + for (i = num_tc, tci = j * num_tc; i--; tci++) { + new_map = xmap_dereference(new_dev_maps->attr_map[tci]); + map = xmap_dereference(dev_maps->attr_map[tci]); if (map && map != new_map) kfree_rcu(map, rcu); } @@ -2317,19 +2370,23 @@ int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, active = true; out_no_new_maps: - /* update Tx queue numa node */ - netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), - (numa_node_id >= 0) ? numa_node_id : - NUMA_NO_NODE); + if (!is_rxqs_map) { + /* update Tx queue numa node */ + netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), + (numa_node_id >= 0) ? + numa_node_id : NUMA_NO_NODE); + } if (!dev_maps) goto out_no_maps; - /* removes queue from unused CPUs */ - for_each_possible_cpu(cpu) { - for (i = tc, tci = cpu * num_tc; i--; tci++) + /* removes tx-queue from unused CPUs/rx-queues */ + for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids), + j < nr_ids;) { + for (i = tc, tci = j * num_tc; i--; tci++) active |= remove_xps_queue(dev_maps, tci, index); - if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu)) + if (!netif_attr_test_mask(j, mask, nr_ids) || + !netif_attr_test_online(j, online_mask, nr_ids)) active |= remove_xps_queue(dev_maps, tci, index); for (i = num_tc - tc, tci++; --i; tci++) active |= remove_xps_queue(dev_maps, tci, index); @@ -2337,7 +2394,10 @@ int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, /* free map if not active */ if (!active) { - RCU_INIT_POINTER(dev->xps_maps, NULL); + if (is_rxqs_map) + RCU_INIT_POINTER(dev->xps_rxqs_map, NULL); + else + RCU_INIT_POINTER(dev->xps_cpus_map, NULL); kfree_rcu(dev_maps, rcu); } @@ -2347,11 +2407,12 @@ int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, return 0; error: /* remove any maps that we added */ - for_each_possible_cpu(cpu) { - for (i = num_tc, tci = cpu * num_tc; i--; tci++) { - new_map = xmap_dereference(new_dev_maps->cpu_map[tci]); + for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids), + j < nr_ids;) { + for (i = num_tc, tci = j * num_tc; i--; tci++) { + new_map = xmap_dereference(new_dev_maps->attr_map[tci]); map = dev_maps ? - xmap_dereference(dev_maps->cpu_map[tci]) : + xmap_dereference(dev_maps->attr_map[tci]) : NULL; if (new_map && new_map != map) kfree(new_map); @@ -2363,6 +2424,12 @@ int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, kfree(new_dev_maps); return -ENOMEM; } + +int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, + u16 index) +{ + return __netif_set_xps_queue(dev, cpumask_bits(mask), index, false); +} EXPORT_SYMBOL(netif_set_xps_queue); #endif @@ -3384,7 +3451,7 @@ static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb) int queue_index = -1; rcu_read_lock(); - dev_maps = rcu_dereference(dev->xps_maps); + dev_maps = rcu_dereference(dev->xps_cpus_map); if (dev_maps) { unsigned int tci = skb->sender_cpu - 1; @@ -3393,7 +3460,7 @@ static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb) tci += netdev_get_prio_tc_map(dev, skb->priority); } - map = rcu_dereference(dev_maps->cpu_map[tci]); + map = rcu_dereference(dev_maps->attr_map[tci]); if (map) { if (map->len == 1) queue_index = map->queues[0]; diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c index bb7e80f4ced3..b39987c81d53 100644 --- a/net/core/net-sysfs.c +++ b/net/core/net-sysfs.c @@ -1227,13 +1227,13 @@ static ssize_t xps_cpus_show(struct netdev_queue *queue, return -ENOMEM; rcu_read_lock(); - dev_maps = rcu_dereference(dev->xps_maps); + dev_maps = rcu_dereference(dev->xps_cpus_map); if (dev_maps) { for_each_possible_cpu(cpu) { int i, tci = cpu * num_tc + tc; struct xps_map *map; - map = rcu_dereference(dev_maps->cpu_map[tci]); + map = rcu_dereference(dev_maps->attr_map[tci]); if (!map) continue; From 04157469b7b848f4a9978b63b1ea2ce62ad3a0a3 Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:26:46 -0700 Subject: [PATCH 2/7] net: Use static_key for XPS maps Use static_key for XPS maps to reduce the cost of extra map checks, similar to how it is used for RPS and RFS. This includes static_key 'xps_needed' for XPS and another for 'xps_rxqs_needed' for XPS using Rx queues map. Signed-off-by: Amritha Nambiar Signed-off-by: David S. Miller --- net/core/dev.c | 31 +++++++++++++++++++++++++------ 1 file changed, 25 insertions(+), 6 deletions(-) diff --git a/net/core/dev.c b/net/core/dev.c index 71059558dc39..43b5575e40c5 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -2081,6 +2081,10 @@ int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) EXPORT_SYMBOL(netdev_txq_to_tc); #ifdef CONFIG_XPS +struct static_key xps_needed __read_mostly; +EXPORT_SYMBOL(xps_needed); +struct static_key xps_rxqs_needed __read_mostly; +EXPORT_SYMBOL(xps_rxqs_needed); static DEFINE_MUTEX(xps_map_mutex); #define xmap_dereference(P) \ rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) @@ -2168,14 +2172,18 @@ static void netif_reset_xps_queues(struct net_device *dev, u16 offset, struct xps_dev_maps *dev_maps; unsigned int nr_ids; + if (!static_key_false(&xps_needed)) + return; + mutex_lock(&xps_map_mutex); - dev_maps = xmap_dereference(dev->xps_rxqs_map); - if (dev_maps) { - nr_ids = dev->num_rx_queues; - clean_xps_maps(dev, possible_mask, dev_maps, nr_ids, offset, - count, true); - + if (static_key_false(&xps_rxqs_needed)) { + dev_maps = xmap_dereference(dev->xps_rxqs_map); + if (dev_maps) { + nr_ids = dev->num_rx_queues; + clean_xps_maps(dev, possible_mask, dev_maps, nr_ids, + offset, count, true); + } } dev_maps = xmap_dereference(dev->xps_cpus_map); @@ -2189,6 +2197,10 @@ static void netif_reset_xps_queues(struct net_device *dev, u16 offset, false); out_no_maps: + if (static_key_enabled(&xps_rxqs_needed)) + static_key_slow_dec(&xps_rxqs_needed); + + static_key_slow_dec(&xps_needed); mutex_unlock(&xps_map_mutex); } @@ -2297,6 +2309,10 @@ int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, if (!new_dev_maps) goto out_no_new_maps; + static_key_slow_inc(&xps_needed); + if (is_rxqs_map) + static_key_slow_inc(&xps_rxqs_needed); + for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids), j < nr_ids;) { /* copy maps belonging to foreign traffic classes */ @@ -3450,6 +3466,9 @@ static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb) struct xps_map *map; int queue_index = -1; + if (!static_key_false(&xps_needed)) + return -1; + rcu_read_lock(); dev_maps = rcu_dereference(dev->xps_cpus_map); if (dev_maps) { From 755c31cd85aea35cf7a5e7253851b52c08eff6e9 Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:26:51 -0700 Subject: [PATCH 3/7] net: sock: Change tx_queue_mapping in sock_common to unsigned short Change 'skc_tx_queue_mapping' field in sock_common structure from 'int' to 'unsigned short' type with ~0 indicating unset and other positive queue values being set. This will accommodate adding a new 'unsigned short' field in sock_common in the next patch for rx_queue_mapping. Signed-off-by: Amritha Nambiar Signed-off-by: David S. Miller --- include/net/sock.h | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-) diff --git a/include/net/sock.h b/include/net/sock.h index b3b75419eafe..37b09c84504b 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -214,7 +214,7 @@ struct sock_common { struct hlist_node skc_node; struct hlist_nulls_node skc_nulls_node; }; - int skc_tx_queue_mapping; + unsigned short skc_tx_queue_mapping; union { int skc_incoming_cpu; u32 skc_rcv_wnd; @@ -1681,17 +1681,25 @@ static inline int sk_receive_skb(struct sock *sk, struct sk_buff *skb, static inline void sk_tx_queue_set(struct sock *sk, int tx_queue) { + /* sk_tx_queue_mapping accept only upto a 16-bit value */ + if (WARN_ON_ONCE((unsigned short)tx_queue >= USHRT_MAX)) + return; sk->sk_tx_queue_mapping = tx_queue; } +#define NO_QUEUE_MAPPING USHRT_MAX + static inline void sk_tx_queue_clear(struct sock *sk) { - sk->sk_tx_queue_mapping = -1; + sk->sk_tx_queue_mapping = NO_QUEUE_MAPPING; } static inline int sk_tx_queue_get(const struct sock *sk) { - return sk ? sk->sk_tx_queue_mapping : -1; + if (sk && sk->sk_tx_queue_mapping != NO_QUEUE_MAPPING) + return sk->sk_tx_queue_mapping; + + return -1; } static inline void sk_set_socket(struct sock *sk, struct socket *sock) From c6345ce7d361dce1b5d02a2181ccb598c27fd7ae Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:26:57 -0700 Subject: [PATCH 4/7] net: Record receive queue number for a connection This patch adds a new field to sock_common 'skc_rx_queue_mapping' which holds the receive queue number for the connection. The Rx queue is marked in tcp_finish_connect() to allow a client app to do SO_INCOMING_NAPI_ID after a connect() call to get the right queue association for a socket. Rx queue is also marked in tcp_conn_request() to allow syn-ack to go on the right tx-queue associated with the queue on which syn is received. Signed-off-by: Amritha Nambiar Signed-off-by: Sridhar Samudrala Signed-off-by: David S. Miller --- include/net/busy_poll.h | 1 + include/net/sock.h | 28 ++++++++++++++++++++++++++++ net/core/sock.c | 2 ++ net/ipv4/tcp_input.c | 3 +++ 4 files changed, 34 insertions(+) diff --git a/include/net/busy_poll.h b/include/net/busy_poll.h index c5187438af38..9e36fda652b7 100644 --- a/include/net/busy_poll.h +++ b/include/net/busy_poll.h @@ -151,6 +151,7 @@ static inline void sk_mark_napi_id(struct sock *sk, const struct sk_buff *skb) #ifdef CONFIG_NET_RX_BUSY_POLL sk->sk_napi_id = skb->napi_id; #endif + sk_rx_queue_set(sk, skb); } /* variant used for unconnected sockets */ diff --git a/include/net/sock.h b/include/net/sock.h index 37b09c84504b..2b097cc89727 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -139,6 +139,7 @@ typedef __u64 __bitwise __addrpair; * @skc_node: main hash linkage for various protocol lookup tables * @skc_nulls_node: main hash linkage for TCP/UDP/UDP-Lite protocol * @skc_tx_queue_mapping: tx queue number for this connection + * @skc_rx_queue_mapping: rx queue number for this connection * @skc_flags: place holder for sk_flags * %SO_LINGER (l_onoff), %SO_BROADCAST, %SO_KEEPALIVE, * %SO_OOBINLINE settings, %SO_TIMESTAMPING settings @@ -215,6 +216,9 @@ struct sock_common { struct hlist_nulls_node skc_nulls_node; }; unsigned short skc_tx_queue_mapping; +#ifdef CONFIG_XPS + unsigned short skc_rx_queue_mapping; +#endif union { int skc_incoming_cpu; u32 skc_rcv_wnd; @@ -326,6 +330,9 @@ struct sock { #define sk_nulls_node __sk_common.skc_nulls_node #define sk_refcnt __sk_common.skc_refcnt #define sk_tx_queue_mapping __sk_common.skc_tx_queue_mapping +#ifdef CONFIG_XPS +#define sk_rx_queue_mapping __sk_common.skc_rx_queue_mapping +#endif #define sk_dontcopy_begin __sk_common.skc_dontcopy_begin #define sk_dontcopy_end __sk_common.skc_dontcopy_end @@ -1702,6 +1709,27 @@ static inline int sk_tx_queue_get(const struct sock *sk) return -1; } +static inline void sk_rx_queue_set(struct sock *sk, const struct sk_buff *skb) +{ +#ifdef CONFIG_XPS + if (skb_rx_queue_recorded(skb)) { + u16 rx_queue = skb_get_rx_queue(skb); + + if (WARN_ON_ONCE(rx_queue == NO_QUEUE_MAPPING)) + return; + + sk->sk_rx_queue_mapping = rx_queue; + } +#endif +} + +static inline void sk_rx_queue_clear(struct sock *sk) +{ +#ifdef CONFIG_XPS + sk->sk_rx_queue_mapping = NO_QUEUE_MAPPING; +#endif +} + static inline void sk_set_socket(struct sock *sk, struct socket *sock) { sk_tx_queue_clear(sk); diff --git a/net/core/sock.c b/net/core/sock.c index bcc41829a16d..dac6d785186b 100644 --- a/net/core/sock.c +++ b/net/core/sock.c @@ -2818,6 +2818,8 @@ void sock_init_data(struct socket *sock, struct sock *sk) sk->sk_pacing_rate = ~0U; sk->sk_pacing_shift = 10; sk->sk_incoming_cpu = -1; + + sk_rx_queue_clear(sk); /* * Before updating sk_refcnt, we must commit prior changes to memory * (Documentation/RCU/rculist_nulls.txt for details) diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index eecd359595fc..a4731995e899 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -78,6 +78,7 @@ #include #include #include +#include int sysctl_tcp_max_orphans __read_mostly = NR_FILE; @@ -5592,6 +5593,7 @@ void tcp_finish_connect(struct sock *sk, struct sk_buff *skb) if (skb) { icsk->icsk_af_ops->sk_rx_dst_set(sk, skb); security_inet_conn_established(sk, skb); + sk_mark_napi_id(sk, skb); } tcp_init_transfer(sk, BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB); @@ -6420,6 +6422,7 @@ int tcp_conn_request(struct request_sock_ops *rsk_ops, tcp_rsk(req)->snt_isn = isn; tcp_rsk(req)->txhash = net_tx_rndhash(); tcp_openreq_init_rwin(req, sk, dst); + sk_rx_queue_set(req_to_sk(req), skb); if (!want_cookie) { tcp_reqsk_record_syn(sk, req, skb); fastopen_sk = tcp_try_fastopen(sk, skb, req, &foc, dst); From fc9bab24e9c654f62f3d411fc0b041be9e487e9d Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:27:02 -0700 Subject: [PATCH 5/7] net: Enable Tx queue selection based on Rx queues This patch adds support to pick Tx queue based on the Rx queue(s) map configuration set by the admin through the sysfs attribute for each Tx queue. If the user configuration for receive queue(s) map does not apply, then the Tx queue selection falls back to CPU(s) map based selection and finally to hashing. Signed-off-by: Amritha Nambiar Signed-off-by: David S. Miller --- include/net/sock.h | 10 ++++++++ net/core/dev.c | 62 +++++++++++++++++++++++++++++++++------------- 2 files changed, 55 insertions(+), 17 deletions(-) diff --git a/include/net/sock.h b/include/net/sock.h index 2b097cc89727..2ed99bfa4595 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -1730,6 +1730,16 @@ static inline void sk_rx_queue_clear(struct sock *sk) #endif } +#ifdef CONFIG_XPS +static inline int sk_rx_queue_get(const struct sock *sk) +{ + if (sk && sk->sk_rx_queue_mapping != NO_QUEUE_MAPPING) + return sk->sk_rx_queue_mapping; + + return -1; +} +#endif + static inline void sk_set_socket(struct sock *sk, struct socket *sock) { sk_tx_queue_clear(sk); diff --git a/net/core/dev.c b/net/core/dev.c index 43b5575e40c5..08d58e0debe5 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -3459,35 +3459,63 @@ sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) } #endif /* CONFIG_NET_EGRESS */ -static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb) +#ifdef CONFIG_XPS +static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, + struct xps_dev_maps *dev_maps, unsigned int tci) +{ + struct xps_map *map; + int queue_index = -1; + + if (dev->num_tc) { + tci *= dev->num_tc; + tci += netdev_get_prio_tc_map(dev, skb->priority); + } + + map = rcu_dereference(dev_maps->attr_map[tci]); + if (map) { + if (map->len == 1) + queue_index = map->queues[0]; + else + queue_index = map->queues[reciprocal_scale( + skb_get_hash(skb), map->len)]; + if (unlikely(queue_index >= dev->real_num_tx_queues)) + queue_index = -1; + } + return queue_index; +} +#endif + +static int get_xps_queue(struct net_device *dev, struct sk_buff *skb) { #ifdef CONFIG_XPS struct xps_dev_maps *dev_maps; - struct xps_map *map; + struct sock *sk = skb->sk; int queue_index = -1; if (!static_key_false(&xps_needed)) return -1; rcu_read_lock(); - dev_maps = rcu_dereference(dev->xps_cpus_map); + if (!static_key_false(&xps_rxqs_needed)) + goto get_cpus_map; + + dev_maps = rcu_dereference(dev->xps_rxqs_map); if (dev_maps) { - unsigned int tci = skb->sender_cpu - 1; + int tci = sk_rx_queue_get(sk); - if (dev->num_tc) { - tci *= dev->num_tc; - tci += netdev_get_prio_tc_map(dev, skb->priority); - } + if (tci >= 0 && tci < dev->num_rx_queues) + queue_index = __get_xps_queue_idx(dev, skb, dev_maps, + tci); + } - map = rcu_dereference(dev_maps->attr_map[tci]); - if (map) { - if (map->len == 1) - queue_index = map->queues[0]; - else - queue_index = map->queues[reciprocal_scale(skb_get_hash(skb), - map->len)]; - if (unlikely(queue_index >= dev->real_num_tx_queues)) - queue_index = -1; +get_cpus_map: + if (queue_index < 0) { + dev_maps = rcu_dereference(dev->xps_cpus_map); + if (dev_maps) { + unsigned int tci = skb->sender_cpu - 1; + + queue_index = __get_xps_queue_idx(dev, skb, dev_maps, + tci); } } rcu_read_unlock(); From 8af2c06ff4b144064b51b7f688194474123d9c9c Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:27:07 -0700 Subject: [PATCH 6/7] net-sysfs: Add interface for Rx queue(s) map per Tx queue Extend transmit queue sysfs attribute to configure Rx queue(s) map per Tx queue. By default no receive queues are configured for the Tx queue. - /sys/class/net/eth0/queues/tx-*/xps_rxqs Signed-off-by: Amritha Nambiar Signed-off-by: David S. Miller --- net/core/net-sysfs.c | 83 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 83 insertions(+) diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c index b39987c81d53..f25ac5ff48a6 100644 --- a/net/core/net-sysfs.c +++ b/net/core/net-sysfs.c @@ -1283,6 +1283,88 @@ static ssize_t xps_cpus_store(struct netdev_queue *queue, static struct netdev_queue_attribute xps_cpus_attribute __ro_after_init = __ATTR_RW(xps_cpus); + +static ssize_t xps_rxqs_show(struct netdev_queue *queue, char *buf) +{ + struct net_device *dev = queue->dev; + struct xps_dev_maps *dev_maps; + unsigned long *mask, index; + int j, len, num_tc = 1, tc = 0; + + index = get_netdev_queue_index(queue); + + if (dev->num_tc) { + num_tc = dev->num_tc; + tc = netdev_txq_to_tc(dev, index); + if (tc < 0) + return -EINVAL; + } + mask = kcalloc(BITS_TO_LONGS(dev->num_rx_queues), sizeof(long), + GFP_KERNEL); + if (!mask) + return -ENOMEM; + + rcu_read_lock(); + dev_maps = rcu_dereference(dev->xps_rxqs_map); + if (!dev_maps) + goto out_no_maps; + + for (j = -1; j = netif_attrmask_next(j, NULL, dev->num_rx_queues), + j < dev->num_rx_queues;) { + int i, tci = j * num_tc + tc; + struct xps_map *map; + + map = rcu_dereference(dev_maps->attr_map[tci]); + if (!map) + continue; + + for (i = map->len; i--;) { + if (map->queues[i] == index) { + set_bit(j, mask); + break; + } + } + } +out_no_maps: + rcu_read_unlock(); + + len = bitmap_print_to_pagebuf(false, buf, mask, dev->num_rx_queues); + kfree(mask); + + return len < PAGE_SIZE ? len : -EINVAL; +} + +static ssize_t xps_rxqs_store(struct netdev_queue *queue, const char *buf, + size_t len) +{ + struct net_device *dev = queue->dev; + struct net *net = dev_net(dev); + unsigned long *mask, index; + int err; + + if (!ns_capable(net->user_ns, CAP_NET_ADMIN)) + return -EPERM; + + mask = kcalloc(BITS_TO_LONGS(dev->num_rx_queues), sizeof(long), + GFP_KERNEL); + if (!mask) + return -ENOMEM; + + index = get_netdev_queue_index(queue); + + err = bitmap_parse(buf, len, mask, dev->num_rx_queues); + if (err) { + kfree(mask); + return err; + } + + err = __netif_set_xps_queue(dev, mask, index, true); + kfree(mask); + return err ? : len; +} + +static struct netdev_queue_attribute xps_rxqs_attribute __ro_after_init + = __ATTR_RW(xps_rxqs); #endif /* CONFIG_XPS */ static struct attribute *netdev_queue_default_attrs[] __ro_after_init = { @@ -1290,6 +1372,7 @@ static struct attribute *netdev_queue_default_attrs[] __ro_after_init = { &queue_traffic_class.attr, #ifdef CONFIG_XPS &xps_cpus_attribute.attr, + &xps_rxqs_attribute.attr, &queue_tx_maxrate.attr, #endif NULL From a4fd1f4babe3fac5887faa5b1b88bde4cd4d1b0d Mon Sep 17 00:00:00 2001 From: Amritha Nambiar Date: Fri, 29 Jun 2018 21:27:12 -0700 Subject: [PATCH 7/7] Documentation: Add explanation for XPS using Rx-queue(s) map Signed-off-by: Amritha Nambiar Signed-off-by: David S. Miller --- .../ABI/testing/sysfs-class-net-queues | 11 ++++ Documentation/networking/scaling.txt | 61 +++++++++++++++---- 2 files changed, 61 insertions(+), 11 deletions(-) diff --git a/Documentation/ABI/testing/sysfs-class-net-queues b/Documentation/ABI/testing/sysfs-class-net-queues index 0c0df91b1516..978b76358661 100644 --- a/Documentation/ABI/testing/sysfs-class-net-queues +++ b/Documentation/ABI/testing/sysfs-class-net-queues @@ -42,6 +42,17 @@ Description: network device transmit queue. Possible vaules depend on the number of available CPU(s) in the system. +What: /sys/class//queues/tx-/xps_rxqs +Date: June 2018 +KernelVersion: 4.18.0 +Contact: netdev@vger.kernel.org +Description: + Mask of the receive queue(s) currently enabled to participate + into the Transmit Packet Steering packet processing flow for this + network device transmit queue. Possible values depend on the + number of available receive queue(s) in the network device. + Default is disabled. + What: /sys/class//queues/tx-/byte_queue_limits/hold_time Date: November 2011 KernelVersion: 3.3 diff --git a/Documentation/networking/scaling.txt b/Documentation/networking/scaling.txt index f55639d71d35..b7056a8a0540 100644 --- a/Documentation/networking/scaling.txt +++ b/Documentation/networking/scaling.txt @@ -366,8 +366,13 @@ XPS: Transmit Packet Steering Transmit Packet Steering is a mechanism for intelligently selecting which transmit queue to use when transmitting a packet on a multi-queue -device. To accomplish this, a mapping from CPU to hardware queue(s) is -recorded. The goal of this mapping is usually to assign queues +device. This can be accomplished by recording two kinds of maps, either +a mapping of CPU to hardware queue(s) or a mapping of receive queue(s) +to hardware transmit queue(s). + +1. XPS using CPUs map + +The goal of this mapping is usually to assign queues exclusively to a subset of CPUs, where the transmit completions for these queues are processed on a CPU within this set. This choice provides two benefits. First, contention on the device queue lock is @@ -377,15 +382,40 @@ transmit queue). Secondly, cache miss rate on transmit completion is reduced, in particular for data cache lines that hold the sk_buff structures. -XPS is configured per transmit queue by setting a bitmap of CPUs that -may use that queue to transmit. The reverse mapping, from CPUs to -transmit queues, is computed and maintained for each network device. -When transmitting the first packet in a flow, the function -get_xps_queue() is called to select a queue. This function uses the ID -of the running CPU as a key into the CPU-to-queue lookup table. If the +2. XPS using receive queues map + +This mapping is used to pick transmit queue based on the receive +queue(s) map configuration set by the administrator. A set of receive +queues can be mapped to a set of transmit queues (many:many), although +the common use case is a 1:1 mapping. This will enable sending packets +on the same queue associations for transmit and receive. This is useful for +busy polling multi-threaded workloads where there are challenges in +associating a given CPU to a given application thread. The application +threads are not pinned to CPUs and each thread handles packets +received on a single queue. The receive queue number is cached in the +socket for the connection. In this model, sending the packets on the same +transmit queue corresponding to the associated receive queue has benefits +in keeping the CPU overhead low. Transmit completion work is locked into +the same queue-association that a given application is polling on. This +avoids the overhead of triggering an interrupt on another CPU. When the +application cleans up the packets during the busy poll, transmit completion +may be processed along with it in the same thread context and so result in +reduced latency. + +XPS is configured per transmit queue by setting a bitmap of +CPUs/receive-queues that may use that queue to transmit. The reverse +mapping, from CPUs to transmit queues or from receive-queues to transmit +queues, is computed and maintained for each network device. When +transmitting the first packet in a flow, the function get_xps_queue() is +called to select a queue. This function uses the ID of the receive queue +for the socket connection for a match in the receive queue-to-transmit queue +lookup table. Alternatively, this function can also use the ID of the +running CPU as a key into the CPU-to-queue lookup table. If the ID matches a single queue, that is used for transmission. If multiple queues match, one is selected by using the flow hash to compute an index -into the set. +into the set. When selecting the transmit queue based on receive queue(s) +map, the transmit device is not validated against the receive device as it +requires expensive lookup operation in the datapath. The queue chosen for transmitting a particular flow is saved in the corresponding socket structure for the flow (e.g. a TCP connection). @@ -404,11 +434,15 @@ acknowledged. XPS is only available if the kconfig symbol CONFIG_XPS is enabled (on by default for SMP). The functionality remains disabled until explicitly -configured. To enable XPS, the bitmap of CPUs that may use a transmit -queue is configured using the sysfs file entry: +configured. To enable XPS, the bitmap of CPUs/receive-queues that may +use a transmit queue is configured using the sysfs file entry: +For selection based on CPUs map: /sys/class/net//queues/tx-/xps_cpus +For selection based on receive-queues map: +/sys/class/net//queues/tx-/xps_rxqs + == Suggested Configuration For a network device with a single transmission queue, XPS configuration @@ -421,6 +455,11 @@ best CPUs to share a given queue are probably those that share the cache with the CPU that processes transmit completions for that queue (transmit interrupts). +For transmit queue selection based on receive queue(s), XPS has to be +explicitly configured mapping receive-queue(s) to transmit queue(s). If the +user configuration for receive-queue map does not apply, then the transmit +queue is selected based on the CPUs map. + Per TX Queue rate limitation: =============================