The RPS steering logic in netif_rx_internal(), netif_receive_skb_internal() and netif_receive_skb_list_internal() was open-coded three times, each with its own #ifdef CONFIG_RPS block and manual rcu_read_lock()/unlock() pairs. Factor it into two helpers, netif_rps() for the single-skb path and netif_rps_list() for the list path, sharing a common __netif_rps() inner helper that runs the get_rps_cpu()/enqueue_to_backlog() step under RCU. Both wrappers keep the early static_branch_unlikely(&rps_needed) check and only enter an RCU read-side critical section when RPS is actually enabled at runtime, so no extra rcu_read_lock()/unlock() is executed on the hot netif_rx_internal() path when RPS is compiled in but disabled. A new internal NET_RX_UNHANDLED sentinel lets a helper report "RPS did not take this skb" so the caller falls back to the local enqueue / __netif_receive_skb() path; it never escapes to callers. No functional change intended. Signed-off-by: Jemmy Wong --- Changes since v1 [1]: - Move rcu_read_lock()/unlock() inside the static_branch_unlikely(&rps_needed) check, so no RCU read-side section is entered when RPS is compiled in but disabled at runtime. Addresses Paolo's concern that netif_rx_internal() is performance critical and rcu_read_lock() is not a no-op. - Drop the guard(rcu) / scoped_guard(rcu) conversion; the RCU section is now scoped inside netif_rps() / netif_rps_list() themselves. - Split __netif_rps() out of netif_rps() so that netif_rps_list() can reuse the per-skb steering step without re-testing the static branch or dropping and reacquiring the RCU lock per skb. - netif_rps_list() now returns int; when RPS drains the entire list, netif_receive_skb_list_internal() skips the __netif_receive_skb_list() fall-through instead of walking an empty list. The result is decided directly from list_empty(&undo_list): drop the splice on the drained path and only splice back when there is something to fall through. Verified generated code with objdump on aarch64 with CONFIG_RPS=y, CONFIG_XPS=y, CONFIG_SMP=y, CONFIG_PREEMPT=y (so rcu_read_lock() expands to a real preempt_disable(), i.e. not a no-op -- the specific scenario Paolo asked about [2]): scripts/bloat-o-meter net/core/dev.o (before -> after): add/remove: 2/2 grow/shrink: 1/2 up/down: 720/-716 (4) Function old new delta __netif_rps - 660 +660 netif_receive_skb_list_internal 640 700 +60 netif_rx_internal 236 200 -36 netif_receive_skb 204 136 -68 get_rps_cpu 608 - -608 Total: Before=335841, After=335849, chg +0.00% netif_rx_internal fast path (!rps_needed, i.e. static branch not taken -- the path Paolo asked about [2]): - identical instruction sequence to base, no rcu_read_lock/unlock - stack frame 16 bytes smaller (0x50 -> 0x40) - static_branch NOP is preserved netif_receive_skb_internal: inlined in both builds; caller netif_receive_skb shrinks by 68 bytes. netif_receive_skb_list_internal: !rps_needed fast path unchanged; rps_needed=true path grows +60 bytes for LIST_HEAD(undo_list) setup, the explicit list_empty(&undo_list)/list_splice_init() branch and its early NET_RX_SUCCESS return. [1] https://lore.kernel.org/all/20260702152830.39065-1-jemmywong512@gmail.com/ [2] https://lore.kernel.org/all/e53482b8-4e1d-4a04-929c-f5cb443ba59d@redhat.com/ include/linux/netdevice.h | 5 +- net/core/dev.c | 118 +++++++++++++++++++++++--------------- 2 files changed, 74 insertions(+), 49 deletions(-) diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h index 9981d637f8b5..c265b78082e3 100644 --- a/include/linux/netdevice.h +++ b/include/linux/netdevice.h @@ -93,8 +93,9 @@ void netdev_set_default_ethtool_ops(struct net_device *dev, void netdev_sw_irq_coalesce_default_on(struct net_device *dev); /* Backlog congestion levels */ -#define NET_RX_SUCCESS 0 /* keep 'em coming, baby */ -#define NET_RX_DROP 1 /* packet dropped */ +#define NET_RX_UNHANDLED -1 +#define NET_RX_SUCCESS 0 +#define NET_RX_DROP 1 #define MAX_NEST_DEV 8 diff --git a/net/core/dev.c b/net/core/dev.c index 4b3d5cfdf6e0..1a05ac8fdd50 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -5426,6 +5426,65 @@ static int enqueue_to_backlog(struct sk_buff *skb, int cpu, return NET_RX_DROP; } +#ifdef CONFIG_RPS +static inline int __netif_rps(struct sk_buff *skb) +{ + int cpu; + int ret = NET_RX_UNHANDLED; + struct rps_dev_flow voidflow, *rflow = &voidflow; + + cpu = get_rps_cpu(skb->dev, skb, &rflow); + if (cpu >= 0) + ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); + + return ret; +} +#endif + +static inline int netif_rps(struct sk_buff *skb) +{ + int ret = NET_RX_UNHANDLED; + +#ifdef CONFIG_RPS + if (!static_branch_unlikely(&rps_needed)) + return ret; + + rcu_read_lock(); + ret = __netif_rps(skb); + rcu_read_unlock(); +#endif + + return ret; +} + +static inline int netif_rps_list(struct list_head *head) +{ + int ret = NET_RX_UNHANDLED; + +#ifdef CONFIG_RPS + LIST_HEAD(undo_list); + struct sk_buff *skb, *next; + + if (!static_branch_unlikely(&rps_needed)) + return ret; + + rcu_read_lock(); + list_for_each_entry_safe(skb, next, head, list) { + skb_list_del_init(skb); + if (__netif_rps(skb) == NET_RX_UNHANDLED) + list_add_tail(&skb->list, &undo_list); + } + rcu_read_unlock(); + + if (list_empty(&undo_list)) + ret = NET_RX_SUCCESS; + else + list_splice_init(&undo_list, head); +#endif + + return ret; +} + static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) { struct net_device *dev = skb->dev; @@ -5695,34 +5754,18 @@ EXPORT_SYMBOL_GPL(do_xdp_generic); static int netif_rx_internal(struct sk_buff *skb) { + unsigned int qtail; int ret; net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); trace_netif_rx(skb); -#ifdef CONFIG_RPS - if (static_branch_unlikely(&rps_needed)) { - struct rps_dev_flow voidflow, *rflow = &voidflow; - int cpu; - - rcu_read_lock(); - - cpu = get_rps_cpu(skb->dev, skb, &rflow); - if (cpu < 0) - cpu = smp_processor_id(); - - ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); - - rcu_read_unlock(); - } else -#endif - { - unsigned int qtail; + ret = netif_rps(skb); + if (ret != NET_RX_UNHANDLED) + return ret; - ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); - } - return ret; + return enqueue_to_backlog(skb, smp_processor_id(), &qtail); } /** @@ -6389,19 +6432,11 @@ static int netif_receive_skb_internal(struct sk_buff *skb) if (skb_defer_rx_timestamp(skb)) return NET_RX_SUCCESS; - rcu_read_lock(); -#ifdef CONFIG_RPS - if (static_branch_unlikely(&rps_needed)) { - struct rps_dev_flow voidflow, *rflow = &voidflow; - int cpu = get_rps_cpu(skb->dev, skb, &rflow); + ret = netif_rps(skb); + if (ret != NET_RX_UNHANDLED) + return ret; - if (cpu >= 0) { - ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); - rcu_read_unlock(); - return ret; - } - } -#endif + rcu_read_lock(); ret = __netif_receive_skb(skb); rcu_read_unlock(); return ret; @@ -6421,21 +6456,10 @@ void netif_receive_skb_list_internal(struct list_head *head) } list_splice_init(&sublist, head); - rcu_read_lock(); -#ifdef CONFIG_RPS - if (static_branch_unlikely(&rps_needed)) { - list_for_each_entry_safe(skb, next, head, list) { - struct rps_dev_flow voidflow, *rflow = &voidflow; - int cpu = get_rps_cpu(skb->dev, skb, &rflow); + if (netif_rps_list(head) == NET_RX_SUCCESS) + return; - if (cpu >= 0) { - /* Will be handled, remove from list */ - skb_list_del_init(skb); - enqueue_to_backlog(skb, cpu, &rflow->last_qtail); - } - } - } -#endif + rcu_read_lock(); __netif_receive_skb_list(head); rcu_read_unlock(); } -- 2.54.0 (Apple Git-157)