LINUX protocol stack detailed flow control Tc__linux

The implementation of the streaming TC is mentioned in the packet sending section, and is now further studied.

Packet sending is implemented through Dev_queue_xmit,

if (q->enqueue) {

rc = __DEV_XMIT_SKB (SKB, Q, Dev, txq);

Goto out;

}

If there is a enqueue, then the flow control, or send directly.

Reference function __DEV_XMIT_SKB, we can see the enqueue operation of the SKB, and __qdisc_run start to trigger soft interrupts.

struct Qdisc_ops defines QoS-related operations, most notably Enqueue and dequeue.

When a NIC is registered, call int register_netdevice (struct net_device *dev), which is responsible for initializing the NIC TC

void Dev_init_scheduler (struct Net_device*dev)
{
         dev->qdisc= &noop_qdisc;
         Netdev_for_each_tx_queue (Dev,dev_init_scheduler_queue, &noop_qdisc);
         if (Dev_ingress_queue (dev))
                   dev_init_scheduler_queue (Dev,dev_ingress_queue (Dev), &noop_qdisc);
 
         Setup_timer (&dev->watchdog_timer,dev_watchdog, (unsigned long) dev);

Dev's qdisc was initialized to Noop_qdisc,netdev_for_each_tx_queue responsible for initializing the transport queues in Net_device: Dev->num_tx_queues queue number, dev->_ The TX queue, if there is a ingress queue, is also initialized, where we can see that there is only one queue, called Dev->ingress_queue.

Now look at the Noop_qdisc, initialize the registered QoS, see the code will find that the inside is empty, or simply will call KFREE_SKB will SKB release. This condition will activate the device change after ifconfig up, dev_activate judge, if the discovery is Noop_qdisc, then Attach_default_qdiscs is responsible for assigning the value again.

If the network card is set to a single queue and the queue size is 0, it is noqueue_qdisc;

Single queue, the single queue size is not 0, then pfifo_fast_ops;

Multiple queues, the queue size is 0, then Noqueue_qdisc;

Otherwise, the mq_qdisc_ops is the default policy.

Then, if we write the module to do TC operation, how to do it. Now let's analyze the SCH_RED.C

On the Module_init, Register_qdisc, adding Ops to the qdisc_base, specifically using the TC command at the end of NetLink call Tc_modify_qdisc.

function pointers in the QDISC_OPS structure body are used in the order of invocation is

Ops->init

Ops->attach

Ops->change

Ops->enqueue

Ops->dequeue

Ops->reset

Ops->destroy

The next major analysis

struct Qdisc_ops pfifo_fast_ops __read_mostly = {

. id = "Pfifo_fast",

. priv_size = sizeof (STRUCTPFIFO_FAST_PRIV),

. Enqueue = Pfifo_fast_enqueue,

. dequeue = Pfifo_fast_dequeue,

. Peek = Pfifo_fast_peek,

. init = Pfifo_fast_init,

. reset = Pfifo_fast_reset,

. dump = Pfifo_fast_dump,

. Owner = This_module,

};

This is the default policy under Linux.

Pfifo_fast_init initialization function

Pfifo_fast_enqueue, if the number of frame in the current queue is less than the maximum number in the queue, the band is calculated by Skb->priority, and the queue is finally added to the queue by band the index policy. If the maximum limit is exceeded, the SKB is released.

Pfifo_fast_dequeue are priv->bitmap indexed arrays, and queues are sent out 0 to 21 times, respectively.

Pfifo_fast_peek only extracts data, but not from the queue.

Pfifo_fast_reset finally reset.

Now go back to the Dev_queue_xmit function

1, dev_pick_tx Select Send Queue

2, Q->enqueue pressure queue

3, qdisc_restart,dequeue_skb out the queue

4, Dev_hard_start_xmit send

5, no success is sent, then save SKB to Gso_skb, next time dequeue send this SKB again

There is still one problem, that is the dev_ingress_queue in the ingress flow control is what happened.

The above general introduction of the TC system kernel implementation, TC system is very complex.

http://tldp.org/HOWTO/Traffic-Control-HOWTO/index.html How TC commands are used

Later will continue to further research, the way is to study from the TC command, the kernel implementation, and finally write a module to carry out the flow control.