loopback 网卡

分析一下:

[ ] 没有看到 `raise_softirq` 的调用

raise_softirq 和 do_softirq 的调用数量是对等的吗？
物理网卡中，raise_softirq 是发生在的中断处理函数中，但是在 loopback 网卡中，是发生在哪里的？

[ ] `net_tx_action` 会调用到 `napi_poll` 上，既然是没有中断的，为什么还是需要中断聚合的

[ ] CPU A 和 CPU B 使用 softirq 通信，不使用 RPS 的时候，如果 CPU C 使用网络栈的内容，会导致 softirq 迁移吗

[ ] What is the main difference between RSS, RPS and RFS

https://stackoverflow.com/questions/44958511/what-is-the-main-difference-between-rss-rps-and-rfs
调用 raise_softirq 100% 是时钟中断触发的。
__raise_softirq_irqoff 是真正插入 flag 的位置:

从 ip6_finish_output2 的位置开始，关闭和打开 softirq 的:

__raise_softirq_irqoff+1
enqueue_to_backlog+642
netif_rx_internal+58
__netif_rx+20
loopback_xmit+201
dev_hard_start_xmit+217
__dev_queue_xmit+2199
ip6_finish_output2+705
ip6_xmit+1048
inet6_csk_xmit+215
__tcp_transmit_skb+1333
tcp_write_xmit+830
tcp_sendmsg_locked+1283
tcp_sendmsg+40
sock_sendmsg+65
sock_write_iter+151
new_sync_write+371
vfs_write+521
ksys_write+167
do_syscall_64+59
entry_SYSCALL_64_after_hwframe+68

sk_filter_trim_cap+270
tcp_v6_rcv+3100
ip6_protocol_deliver_rcu+205
ip6_input_finish+64
__netif_receive_skb_one_core+99
process_backlog+137
__napi_poll+44
net_rx_action+571
__softirqentry_text_start+238
do_softirq.part.0+152
__local_bh_enable_ip+115
ip6_finish_output2+499
ip6_xmit+1048
inet6_csk_xmit+215
__tcp_transmit_skb+1333
tcp_write_xmit+830
tcp_sendmsg_locked+1283
tcp_sendmsg+40
sock_sendmsg+65
sock_write_iter+151
new_sync_write+371
vfs_write+521
ksys_write+167
do_syscall_64+59
entry_SYSCALL_64_after_hwframe+68

任何一次 softirq

static const struct net_device_ops loopback_ops = {
	.ndo_init        = loopback_dev_init,
	.ndo_start_xmit  = loopback_xmit,
	.ndo_get_stats64 = loopback_get_stats64,
	.ndo_set_mac_address = eth_mac_addr,
};

/* Registered in net/core/dev.c */
struct pernet_operations __net_initdata loopback_net_ops = {
	.init = loopback_net_init,
};

static const struct net_device_ops blackhole_netdev_ops = {
	.ndo_start_xmit = blackhole_netdev_xmit,
};

static const struct ethtool_ops loopback_ethtool_ops = {
	.get_link		= always_on,
	.get_ts_info		= ethtool_op_get_ts_info,
};

loopback_xmit
- __netif_rx
  - enqueue_to_backlog

However this three-way handshake takes some time. And during that time the connection is queued and this is the backlog.

https://stackoverflow.com/questions/36594400/what-is-backlog-in-tcp-connections

XDP

https://www.iovisor.org/technology/xdp

NAPI

似乎这就是 percpu 的 backlog ?

/*
 *	Device drivers call our routines to queue packets here. We empty the
 *	queue in the local softnet handler.
 */

DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
EXPORT_PER_CPU_SYMBOL(softnet_data);

处理 backlog 的地方:

process_backlog
napi_busy_loop

[ ] softirq 的工作流程

[ ] 统计和 loopback 网卡有关的 softirq 的频率，和每次发生的数量

虽然，两个版本的网络栈变化很大，但是从 loopback 的这一个函数而已，其调用过程是非常清晰明了的吧，不存在其他的位置。

iperf

https://load-balancer.inlab.net/manual/performance/measuring-internal-bandwidth-with-iperf/

copy 是不是都是采用 SIMD 的

void irq_exit(void)
{
#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
	local_irq_disable();
#else
	lockdep_assert_irqs_disabled();
#endif
	account_irq_exit_time(current);
	preempt_count_sub(HARDIRQ_OFFSET);
	if (!in_interrupt() && local_softirq_pending())
		invoke_softirq(); ==================================》 __do_softirq

	tick_irq_exit();
	rcu_irq_exit();
	trace_hardirq_exit(); /* must be last! */
}

问题

如何理解这里的 tcp_delack_timer_handler

#0  loopback_xmit (skb=0xffff888009358700, dev=0xffff8880045d0000) at drivers/net/loopback.c:71
#1  0xffffffff81db57c8 in __netdev_start_xmit (more=false, dev=0xffff8880045d0000, skb=0xffff888009358700, ops=0xffffffff82521c60 <loopback_ops>) at ./include/linux/netdevice.h:4865
#2  netdev_start_xmit (more=false, txq=0xffff88810017c800, dev=0xffff8880045d0000, skb=0xffff888009358700) at ./include/linux/netdevice.h:4879
#3  xmit_one (more=false, txq=0xffff88810017c800, dev=0xffff8880045d0000, skb=0xffff888009358700) at net/core/dev.c:3583
#4  dev_hard_start_xmit (first=first@entry=0xffff888009358700, dev=dev@entry=0xffff8880045d0000, txq=txq@entry=0xffff88810017c800, ret=ret@entry=0xffffc9000013ccec) at net/core/dev.c:3599
#5  0xffffffff81db6422 in __dev_queue_xmit (skb=skb@entry=0xffff888009358700, sb_dev=sb_dev@entry=0x0 <fixed_percpu_data>) at net/core/dev.c:4249
#6  0xffffffff81e6e438 in dev_queue_xmit (skb=0xffff888009358700) at ./include/linux/netdevice.h:3035
#7  neigh_hh_output (skb=<optimized out>, hh=<optimized out>) at ./include/net/neighbour.h:530
#8  neigh_output (skip_cache=<optimized out>, skb=0xffff888009358700, n=0xffff8881424f5200) at ./include/net/neighbour.h:544
#9  ip_finish_output2 (net=<optimized out>, sk=<optimized out>, skb=0xffff888009358700) at net/ipv4/ip_output.c:228
#10 0xffffffff81e709fb in __ip_queue_xmit (sk=0xffff8881035488c0, skb=0xffff888009358700, fl=0xffff888103548c30, tos=<optimized out>) at net/ipv4/ip_output.c:532
#11 0xffffffff81e70c80 in ip_queue_xmit (sk=<optimized out>, skb=<optimized out>, fl=<optimized out>) at net/ipv4/ip_output.c:546
#12 0xffffffff81e93c5f in __tcp_transmit_skb (sk=sk@entry=0xffff8881035488c0, skb=0xffff888009358700, clone_it=clone_it@entry=0, gfp_mask=gfp_mask@entry=0, rcv_nxt=<optimized out>) at net/ipv4/tcp_output.c:1399
#13 0xffffffff81e94af4 in __tcp_send_ack (sk=sk@entry=0xffff8881035488c0, rcv_nxt=<optimized out>) at net/ipv4/tcp_output.c:3983
#14 0xffffffff81e978a7 in __tcp_send_ack (rcv_nxt=<optimized out>, sk=0xffff8881035488c0) at net/ipv4/tcp_output.c:3950
#15 0xffffffff81e98339 in tcp_delack_timer_handler (sk=sk@entry=0xffff8881035488c0) at net/ipv4/tcp_timer.c:316
#16 0xffffffff81e983cc in tcp_delack_timer (t=0xffff888103548d20) at net/ipv4/tcp_timer.c:339
#17 0xffffffff811d5ccf in call_timer_fn (timer=timer@entry=0xffff888103548d20, fn=fn@entry=0xffffffff81e98390 <tcp_delack_timer>, baseclk=baseclk@entry=4296484559) at kernel/time/timer.c:1700
#18 0xffffffff811d5fde in expire_timers (head=0xffffc9000013cf00, base=0xffff88807dd1cdc0) at kernel/time/timer.c:1751
#19 __run_timers (base=base@entry=0xffff88807dd1cdc0) at kernel/time/timer.c:2022
#20 0xffffffff811d60b0 in __run_timers (base=0xffff88807dd1cdc0) at kernel/time/timer.c:2000
#21 run_timer_softirq (h=<optimized out>) at kernel/time/timer.c:2035
#22 0xffffffff8217f704 in __do_softirq () at kernel/softirq.c:571
#23 0xffffffff811308ba in invoke_softirq () at kernel/softirq.c:445
#24 __irq_exit_rcu () at kernel/softirq.c:650
#25 0xffffffff8216c876 in sysvec_apic_timer_interrupt (regs=0xffffc900000e7e38) at arch/x86/kernel/apic/apic.c:1107

接受是如何处理的

只见到了 loopback_xmit 的，对应的 revice

之前说是，在 softirq 的位置会进行接收的。

本站所有文章转发 CSDN 将按侵权追究法律责任，其它情况随意。