最近遇到一个问题,当tcp收包的时候,我们的服务器的入向软中断比例很高。
我们知道,napi模式,可以降低收包入向软中断占比,那么,针对napi模式,能不能优化?本文针对2.6.32-358内核进行分析:
static void net_rx_action(struct softirq_action *h) { struct list_head *list = &__get_cpu_var(softnet_data).poll_list; unsigned long time_limit = jiffies + 2; int budget = netdev_budget;----------------这个值可以通过/proc/sys/net/core/netdev_budget修改,默认是300 void *have; int select; struct rps_remote_softirq_cpus *rcpus; local_irq_disable(); while (!list_empty(list)) {-------------------不为空,则一直循环 struct napi_struct *n; int work, weight; /* If softirq window is exhuasted then punt. * Allow this to run for 2 jiffies since which will allow * an average latency of 1.5/HZ. */ if (unlikely(budget <= 0 || time_after(jiffies, time_limit)))----------时间到,或者配额到了,则退出循环。 goto softnet_break; local_irq_enable(); /* Even though interrupts have been re-enabled, this * access is safe because interrupts can only add new * entries to the tail of this list, and only ->poll() * calls can remove this head entry from the list. */ n = list_first_entry(list, struct napi_struct, poll_list); have = netpoll_poll_lock(n); weight = n->weight;--------------napi配置的weight,这个是一次poll的配额,和上面总的配合一起控制收包,这个在netif_napi_add 函数中设置。
/* This NAPI_STATE_SCHED test is for avoiding a race * with netpoll's poll_napi(). Only the entity which * obtains the lock and sees NAPI_STATE_SCHED set will * actually make the ->poll() call. Therefore we avoid * accidently calling ->poll() when NAPI is not scheduled. */ work = 0; if (test_bit(NAPI_STATE_SCHED, &n->state)) { work = n->poll(n, weight);------------回调poll,不同的厂家有不同的实现,比如intel的ixgbe_poll实现 trace_napi_poll(n); } WARN_ON_ONCE(work > weight); budget -= work; local_irq_disable(); /* Drivers must not modify the NAPI state if they * consume the entire weight. In such cases this code * still "owns" the NAPI instance and therefore can * move the instance around on the list at-will. */ if (unlikely(work == weight)) { if (unlikely(napi_disable_pending(n))) { local_irq_enable(); napi_complete(n); local_irq_disable(); } else list_move_tail(&n->poll_list, list); } netpoll_poll_unlock(have); } out: rcpus = &__get_cpu_var(rps_remote_softirq_cpus); select = rcpus->select; rcpus->select ^= 1; local_irq_enable(); net_rps_action(&rcpus->mask[select]); #ifdef CONFIG_NET_DMA /* * There may not be any more sk_buffs coming right now, so push * any pending DMA copies to hardware */ dma_issue_pending_all(); #endif return; softnet_break: __get_cpu_var(netdev_rx_stat).time_squeeze++; __raise_softirq_irqoff(NET_RX_SOFTIRQ); goto out; }
从代码可以看出,限制一次调用net_rx_action的地方,无非是时间,还有netdev_budget,如果把netdev_budget 调大,是不是就可以一次性多收一点包呢,意味着触发软中断的次数就会减少,答案是肯定的。
那么默认值来看,netdev_budget 比napi配置的weight要大。
/* initialize NAPI */
netif_napi_add(adapter->netdev, &q_vector->napi,ixgbe_poll, 64);-------传入64
void netif_napi_add(struct net_device *dev, struct napi_struct *napi, int (*poll)(struct napi_struct *, int), int weight) { INIT_LIST_HEAD(&napi->poll_list); napi->gro_count = 0; napi->gro_list = NULL; napi->skb = NULL; napi->poll = poll; if (weight > NAPI_POLL_WEIGHT) pr_err_once("netif_napi_add() called with weight %d on device %s ",--------会打印,但是也不会限制 weight, dev->name); napi->weight = weight; list_add(&napi->dev_list, &dev->napi_list); napi->dev = dev; #ifdef CONFIG_NETPOLL spin_lock_init(&napi->poll_lock); napi->poll_owner = -1; #endif set_bit(NAPI_STATE_SCHED, &napi->state);
我们通过将传入的64改大到256,因为在ixgbe_poll 函数中,这个值控制了一次循环收包的个数。
int ixgbe_poll(struct napi_struct *napi, int budget) { struct ixgbe_q_vector *q_vector = container_of(napi, struct ixgbe_q_vector, napi); struct ixgbe_adapter *adapter = q_vector->adapter; struct ixgbe_ring *ring; int per_ring_budget; bool clean_complete = true; #ifdef CONFIG_IXGBE_DCA if (adapter->flags & IXGBE_FLAG_DCA_ENABLED) ixgbe_update_dca(q_vector); #endif ixgbe_for_each_ring(ring, q_vector->tx)----------------------遍历tx队列,跟本文讨论的内容不相关,本文讨论收包 clean_complete &= !!ixgbe_clean_tx_irq(q_vector, ring); if (!ixgbe_qv_lock_napi(q_vector)) return budget; /* attempt to distribute budget to each queue fairly, but don't allow * the budget to go below 1 because we'll exit polling */ if (q_vector->rx.count > 1) per_ring_budget = max(budget/q_vector->rx.count, 1);---通过增大 budget,从64改大到256,增加了一次循环收包的个数 else per_ring_budget = budget; ixgbe_for_each_ring(ring, q_vector->rx)-----------遍历收包队列, clean_complete &= (ixgbe_clean_rx_irq(q_vector, ring, per_ring_budget) < per_ring_budget); ixgbe_qv_unlock_napi(q_vector); /* If all work not completed, return budget and keep polling */ if (!clean_complete) return budget; /* all work done, exit the polling mode */ napi_complete(napi); if (adapter->rx_itr_setting & 1) ixgbe_set_itr(q_vector); if (!test_bit(__IXGBE_DOWN, &adapter->state)) ixgbe_irq_enable_queues(adapter, ((u64)1 << q_vector->v_idx)); return 0; }
通过将 budget 从64改大到256,同样的入向流量,软中断从25%降低到23%,效果很明显。
那么,能否无限制改大呢,显然不行,一则是改大后,各个队列收包就很难均衡,因为每个队列收完对应的报文之后(除非收空了),才能返回,这样,就关中断时间太长了。
可能有人会问,这个改大收包个数,但是处理软中断的总时间应该没变化,为什么会降低呢,取个极限的例子,napi出来就是应对以前单个包就需要一个中断的问题的,所以单次收包
多一些应该是有用的。
除此之外,通过设置网卡的rx-usecs属性,将这个值改大些,也可以降低软中断的占比。
ethtool -c eth0 Coalesce parameters for eth0: Adaptive RX: off TX: off stats-block-usecs: 0 sample-interval: 0 pkt-rate-low: 0 pkt-rate-high: 0 rx-usecs: 1-------------默认是1,改成512
ethtool –C eth0 rx-usecs 512
这个减少了硬中断的触发次数,但是呢,显而易见的是,增加了延迟,如果你的系统是要求实时性极高的,可能要减少该值。
第三个方法就是开启gro了,gro开启之后,收包的时候,如果是同一个流的包,且网卡支持gro属性的话,根据协议的回调会尝试进行合并报文,当前由于目前这个流的宏值
只有最多8个流,超过的会直接送上协议栈,不会合并。当然如果是服务器,发包为主的话,其实gro是有害无益的,因为目前的flow个数限制太多,而且对于tcp来说,ack的报文也不会合并。
所以如果是服务器端,要定制化自己的gro,比如使用hash链来管理flow,使用ack合并来减少软中断消耗。