Changes TUN driver to use napi_gro_receive() upon receiving packets rather than netif_rx_ni(). Adds flag IFF_NAPI that enables these changes and operation is not affected if the flag is disabled. SKBs are constructed upon packet arrival and are queued to be processed later.
The new path was evaluated with a benchmark with the following setup: Open two tap devices and a receiver thread that reads in a loop for each device. Start one sender thread and pin all threads to different CPUs. Send 1M minimum UDP packets to each device and measure sending time for each of the sending methods: napi_gro_receive(): 4.90s netif_rx_ni(): 4.90s netif_receive_skb(): 7.20s Signed-off-by: Petar Penkov <peterpenko...@gmail.com> Cc: Eric Dumazet <eduma...@google.com> Cc: Mahesh Bandewar <mahe...@google.com> Cc: Willem de Bruijn <will...@google.com> Cc: da...@davemloft.net Cc: ppen...@stanford.edu --- drivers/net/tun.c | 133 +++++++++++++++++++++++++++++++++++++++----- include/uapi/linux/if_tun.h | 1 + 2 files changed, 119 insertions(+), 15 deletions(-) diff --git a/drivers/net/tun.c b/drivers/net/tun.c index 3c9985f29950..f16407242b18 100644 --- a/drivers/net/tun.c +++ b/drivers/net/tun.c @@ -121,7 +121,7 @@ do { \ #define TUN_VNET_BE 0x40000000 #define TUN_FEATURES (IFF_NO_PI | IFF_ONE_QUEUE | IFF_VNET_HDR | \ - IFF_MULTI_QUEUE) + IFF_MULTI_QUEUE | IFF_NAPI) #define GOODCOPY_LEN 128 #define FLT_EXACT_COUNT 8 @@ -172,6 +172,7 @@ struct tun_file { u16 queue_index; unsigned int ifindex; }; + struct napi_struct napi; struct list_head next; struct tun_struct *detached; struct skb_array tx_array; @@ -229,6 +230,68 @@ struct tun_struct { struct bpf_prog __rcu *xdp_prog; }; +static int tun_napi_receive(struct napi_struct *napi, int budget) +{ + struct tun_file *tfile = container_of(napi, struct tun_file, napi); + struct sk_buff_head *queue = &tfile->sk.sk_write_queue; + struct sk_buff_head process_queue; + struct sk_buff *skb; + int received = 0; + + __skb_queue_head_init(&process_queue); + + spin_lock(&queue->lock); + skb_queue_splice_tail_init(queue, &process_queue); + spin_unlock(&queue->lock); + + while (received < budget && (skb = __skb_dequeue(&process_queue))) { + napi_gro_receive(napi, skb); + ++received; + } + + if (!skb_queue_empty(&process_queue)) { + spin_lock(&queue->lock); + skb_queue_splice(&process_queue, queue); + spin_unlock(&queue->lock); + } + + return received; +} + +static int tun_napi_poll(struct napi_struct *napi, int budget) +{ + unsigned int received; + + received = tun_napi_receive(napi, budget); + + if (received < budget) + napi_complete_done(napi, received); + + return received; +} + +static void tun_napi_init(struct tun_struct *tun, struct tun_file *tfile, + bool napi_en) +{ + if (napi_en) { + netif_napi_add(tun->dev, &tfile->napi, tun_napi_poll, + NAPI_POLL_WEIGHT); + napi_enable(&tfile->napi); + } +} + +static void tun_napi_disable(struct tun_struct *tun, struct tun_file *tfile) +{ + if (tun->flags & IFF_NAPI) + napi_disable(&tfile->napi); +} + +static void tun_napi_del(struct tun_struct *tun, struct tun_file *tfile) +{ + if (tun->flags & IFF_NAPI) + netif_napi_del(&tfile->napi); +} + #ifdef CONFIG_TUN_VNET_CROSS_LE static inline bool tun_legacy_is_little_endian(struct tun_struct *tun) { @@ -541,6 +604,11 @@ static void __tun_detach(struct tun_file *tfile, bool clean) tun = rtnl_dereference(tfile->tun); + if (tun && clean) { + tun_napi_disable(tun, tfile); + tun_napi_del(tun, tfile); + } + if (tun && !tfile->detached) { u16 index = tfile->queue_index; BUG_ON(index >= tun->numqueues); @@ -598,6 +666,7 @@ static void tun_detach_all(struct net_device *dev) for (i = 0; i < n; i++) { tfile = rtnl_dereference(tun->tfiles[i]); BUG_ON(!tfile); + tun_napi_disable(tun, tfile); tfile->socket.sk->sk_shutdown = RCV_SHUTDOWN; tfile->socket.sk->sk_data_ready(tfile->socket.sk); RCU_INIT_POINTER(tfile->tun, NULL); @@ -613,6 +682,7 @@ static void tun_detach_all(struct net_device *dev) synchronize_net(); for (i = 0; i < n; i++) { tfile = rtnl_dereference(tun->tfiles[i]); + tun_napi_del(tun, tfile); /* Drop read queue */ tun_queue_purge(tfile); sock_put(&tfile->sk); @@ -631,7 +701,8 @@ static void tun_detach_all(struct net_device *dev) module_put(THIS_MODULE); } -static int tun_attach(struct tun_struct *tun, struct file *file, bool skip_filter) +static int tun_attach(struct tun_struct *tun, struct file *file, + bool skip_filter, bool napi) { struct tun_file *tfile = file->private_data; struct net_device *dev = tun->dev; @@ -677,10 +748,12 @@ static int tun_attach(struct tun_struct *tun, struct file *file, bool skip_filte rcu_assign_pointer(tun->tfiles[tun->numqueues], tfile); tun->numqueues++; - if (tfile->detached) + if (tfile->detached) { tun_enable_queue(tfile); - else + } else { sock_hold(&tfile->sk); + tun_napi_init(tun, tfile, napi); + } tun_set_real_num_queues(tun); @@ -956,13 +1029,28 @@ static void tun_poll_controller(struct net_device *dev) * Tun only receives frames when: * 1) the char device endpoint gets data from user space * 2) the tun socket gets a sendmsg call from user space - * Since both of those are synchronous operations, we are guaranteed - * never to have pending data when we poll for it - * so there is nothing to do here but return. + * If NAPI is not enabled, since both of those are synchronous + * operations, we are guaranteed never to have pending data when we poll + * for it so there is nothing to do here but return. * We need this though so netpoll recognizes us as an interface that * supports polling, which enables bridge devices in virt setups to * still use netconsole + * If NAPI is enabled, however, we need to schedule polling for all + * queues. */ + struct tun_struct *tun = netdev_priv(dev); + + if (tun->flags & IFF_NAPI) { + struct tun_file *tfile; + int i; + + rcu_read_lock(); + for (i = 0; i < tun->numqueues; i++) { + tfile = rcu_dereference(tun->tfiles[i]); + napi_schedule(&tfile->napi); + } + rcu_read_unlock(); + } return; } #endif @@ -1549,11 +1637,25 @@ static ssize_t tun_get_user(struct tun_struct *tun, struct tun_file *tfile, } rxhash = __skb_get_hash_symmetric(skb); -#ifndef CONFIG_4KSTACKS - tun_rx_batched(tun, tfile, skb, more); -#else - netif_rx_ni(skb); -#endif + + if (tun->flags & IFF_NAPI) { + struct sk_buff_head *queue = &tfile->sk.sk_write_queue; + int queue_len; + + spin_lock_bh(&queue->lock); + __skb_queue_tail(queue, skb); + queue_len = skb_queue_len(queue); + spin_unlock(&queue->lock); + + if (!more || queue_len > NAPI_POLL_WEIGHT) + napi_schedule(&tfile->napi); + + local_bh_enable(); + } else if (!IS_ENABLED(CONFIG_4KSTACKS)) { + tun_rx_batched(tun, tfile, skb, more); + } else { + netif_rx_ni(skb); + } stats = get_cpu_ptr(tun->pcpu_stats); u64_stats_update_begin(&stats->syncp); @@ -1980,7 +2082,8 @@ static int tun_set_iff(struct net *net, struct file *file, struct ifreq *ifr) if (err < 0) return err; - err = tun_attach(tun, file, ifr->ifr_flags & IFF_NOFILTER); + err = tun_attach(tun, file, ifr->ifr_flags & IFF_NOFILTER, + ifr->ifr_flags & IFF_NAPI); if (err < 0) return err; @@ -2066,7 +2169,7 @@ static int tun_set_iff(struct net *net, struct file *file, struct ifreq *ifr) NETIF_F_HW_VLAN_STAG_TX); INIT_LIST_HEAD(&tun->disabled); - err = tun_attach(tun, file, false); + err = tun_attach(tun, file, false, ifr->ifr_flags & IFF_NAPI); if (err < 0) goto err_free_flow; @@ -2216,7 +2319,7 @@ static int tun_set_queue(struct file *file, struct ifreq *ifr) ret = security_tun_dev_attach_queue(tun->security); if (ret < 0) goto unlock; - ret = tun_attach(tun, file, false); + ret = tun_attach(tun, file, false, tun->flags & IFF_NAPI); } else if (ifr->ifr_flags & IFF_DETACH_QUEUE) { tun = rtnl_dereference(tfile->tun); if (!tun || !(tun->flags & IFF_MULTI_QUEUE) || tfile->detached) diff --git a/include/uapi/linux/if_tun.h b/include/uapi/linux/if_tun.h index 3cb5e1d85ddd..30b6184884eb 100644 --- a/include/uapi/linux/if_tun.h +++ b/include/uapi/linux/if_tun.h @@ -60,6 +60,7 @@ /* TUNSETIFF ifr flags */ #define IFF_TUN 0x0001 #define IFF_TAP 0x0002 +#define IFF_NAPI 0x0010 #define IFF_NO_PI 0x1000 /* This flag has no real effect */ #define IFF_ONE_QUEUE 0x2000 -- 2.11.0